Documente Academic
Documente Profesional
Documente Cultură
Aerodynamic Design
& Optimization
Edited & Adapted by:
Ideen Sadrehaghighi, Ph.D.
Baseline Optimized
Optimized
Baseline
ANNAPOLIS, MD
2
Contents
1 Introduction ................................................................................................................................ 22
1.1 Complexity of Flow ............................................................................................................................................. 22
1.2 Computational Cost ............................................................................................................................................ 22
1.3 Aerodynamic Optimization ............................................................................................................................. 24
1.4 Design vs. Off-Design Modeling ..................................................................................................................... 24
2 Design Methodology................................................................................................................. 26
2.1 Direct Design ......................................................................................................................................................... 26
2.1.1 Multi-Objective Optimization ............................................................................................................ 26
2.2 Inverse Design ...................................................................................................................................................... 27
2.2.1 Case Study 1 - Inverse Aerodynamic Design Method for Aircraft Component ............. 27
2.2.1.1 Introduction and Background ............................................................................................... 27
2.2.1.2 Formulation .................................................................................................................................. 28
2.2.1.3 Method of Solution ..................................................................................................................... 28
2.2.1.4 Application and Validation ..................................................................................................... 29
2.2.1.5 Conclusion ..................................................................................................................................... 30
2.2.2 Case Study 2 - A Pseudo-Compressibility Method for Solving Inverse Problems
Based on the 3D Incompressible Euler Equations....................................................................................... 31
2.2.2.1 Introduction & Literature Survey ........................................................................................ 31
2.2.2.2 The Mathematical Model ......................................................................................................... 33
2.2.2.3 Dynamics of Materially Moving Boundaries. .................................................................. 34
2.2.2.4 Numerical Results for 3D Nozzle ......................................................................................... 35
2.2.2.5 Inverse Problem on a Stator Blade ...................................................................................... 36
2.2.2.6 Conclusions ................................................................................................................................... 38
2.2.3 Case Study 3 - Inversely Designed Scramjet Flow-Path.......................................................... 38
2.2.3.1 Introduction .................................................................................................................................. 39
2.2.3.2 Inverse Scramjet 2D Centerline Design Approach........................................................ 40
2.2.3.3 Scramjet Inverse Design Approach ..................................................................................... 41
2.2.3.3.1 Aerodynamics of the 2D ‘Forebody’ Configuration ................................................ 42
2.2.3.3.2 Derivation of the 2D ‘Forebody-Inlet-Isolator’ Configuration ........................... 42
2.2.3.3.3 Design Point at Station A ................................................................................................... 43
2.2.3.3.4 Design Points at Station B ................................................................................................. 43
2.2.3.3.5 Design Points at Station C ................................................................................................. 43
2.2.3.3.6 Design Points at Station D ................................................................................................. 45
2.2.3.4 Computer Aided Design (CAD) Design .............................................................................. 46
2.2.3.4.1 Overview of the 3D Design Process .............................................................................. 46
2.2.3.4.2 The 2D Forebody Construction (Side View).............................................................. 47
2.2.3.4.3 The 2D Inlet Construction (Side View) ........................................................................ 48
2.2.3.4.4 Streamline Preparation of Flow-Field ......................................................................... 48
2.2.3.4.5 Wedge Geometry Extraction From Flow-Field ........................................................ 49
2.2.3.5 The Caret Geometry................................................................................................................... 50
2.2.3.5.1 3D Stream Tube Construction using the Wave Rider Approach ....................... 50
2.2.3.5.2 Transforming Stream Tubes to ‘Star’ Shaped Geometries .................................. 51
2.2.3.5.3 Validation Section................................................................................................................. 51
2.2.3.5.4 2D Simulations....................................................................................................................... 52
2.2.3.5.5 2D Euler Simulations Using AVUS ................................................................................. 52
2.2.3.5.6 2D Isolator Viscous Simulations..................................................................................... 53
3
2.2.3.5.7 3D Simulations....................................................................................................................... 53
2.2.3.5.8 Combustor Diffuser Nozzle Sections ............................................................................ 54
2.2.3.5.9 Scramjet Flow-Path ............................................................................................................. 55
2.2.3.5.10 Conclusion ............................................................................................................................... 56
10.3.4 Case Study 2 – 2D Sensitivity of Aerodynamic Forces in Inviscid Environment ....... 309
10.3.4.1 Kinematics .................................................................................................................................. 310
10.3.4.2 Mathematical Modeling of Aerodynamic Forces ........................................................ 311
10.3.4.3 A Mathematical Model Based on Aerodynamically Steady Motions .................. 312
10.3.4.4 Non-Inertial Form of Conservative Equations ............................................................ 313
10.3.4.5 Euler Equations for Aerodynamically Steady Motions ............................................ 314
10.3.4.6 Flow Equations for Aerodynamically Steady Motions ............................................. 314
10.3.4.7 2D Euler Equations Applied to an Airfoil in Rectilinear Steady Motion ........... 315
10.3.4.8 Flow Sensitivity Equations .................................................................................................. 317
10.3.4.9 Numerical Results: Pitch Rate Sensitivity ..................................................................... 319
10.3.4.10 Validating .................................................................................................................................... 320
10.3.4.11 Conclusions ................................................................................................................................ 321
10.4 Surface Modeling Using Non-Uniform Rational B-Splines (NURBS)........................................... 323
10.4.1 Linear Optimization Loop ................................................................................................................ 325
10.4.2 Case Study 3 - 2D Study of Airfoil Grid Sensitivity via Direct Differentiation
(DD) 326
10.4.2.1 Airfoil Grid, Flow Sensitivity, & Optimization ............................................................ 326
10.4.2.2 Discussions ................................................................................................................................. 327
10.4.3 3D NURBS Surface ............................................................................................................................... 329
10.4.3.1 Advantages of NURBS ............................................................................................................ 330
10.4.3.2 Continuity Constraint ............................................................................................................ 331
10.4.3.3 Mesh Projection........................................................................................................................ 331
10.4.4 Extension of Sensitivity to 3D via Automatic Differentiation (AD) ................................. 332
10.5 Essence of Adjoint Equation ........................................................................................................................ 333
10.5.1 The Adjoint Reverses the Propagation of Information ........................................................ 333
10.5.2 Linear Adjoint Equation .................................................................................................................... 334
10.5.3 Classical Formulation of the Adjoint Variable (AV) Approach to Optimal Design .... 334
10.5.4 Limitations of the Adjoint Approach............................................................................................ 337
10.5.4.1 Constraints ................................................................................................................................. 337
10.5.4.2 Limitations of Gradient-Based Optimization ............................................................... 337
10.5.5 Case Study 3 – Adjoint Aero-Design Optimization for Multistage
Turbomachinery Blades ...................................................................................................................................... 338
10.6 Development of Discrete Adjoint Approach Based on the Lattice Boltzmann Method ...... 339
10.6.1 Introduction and Literature Survey ............................................................................................. 339
10.6.2 Lattice Boltzmann Method ............................................................................................................... 342
10.6.2.1 D2Q9 Model. .............................................................................................................................. 342
10.6.3 General Description of the Discrete Adjoint Method ............................................................ 343
10.6.4 Optimization Problem Statement.................................................................................................. 345
10.6.4.1 Cost Function............................................................................................................................. 345
10.6.4.2 Design Variables....................................................................................................................... 346
10.6.5 Discrete Adjoint Approach Based on the Lattice Boltzmann Method ............................ 346
10.6.5.1 Derivation of the Macroscopic Discrete Adjoint Equation ..................................... 346
10.6.5.2 Derivation of the Microscopic Discrete Adjoint Equation. ..................................... 347
10.6.6 Optimization Algorithm .................................................................................................................... 348
10.6.7 Optimization Process Using the Adjoint Approach Based on the LB Method............. 348
10.6.8 Conclusions and Future Work ........................................................................................................ 349
List of Tables:
13
Table 5.1 Performance Comparison of Initial an Optimize Airfoils (Courtesy of 44) .................... 123
Table 6.1 Analytical optimal Breguet Mach Numbers as a Fraction of Mc for NACA 0012
using M2CL as a ............................................................................................................................................................... 142
Table 6.2 Results for Inviscid Range Optimizations - (Courtesy of [Poole et al.])......................... 143
Table 6.3 Results for Inviscid Range Optimizations with k = 0.04 Induced Drag Factor -
(Courtesy of [Poole et al.]) ........................................................................................................................................ 144
Table 6.4 Results for Drag Minimizations - (Courtesy of [Poole et al.]) ............................................ 146
Table 6.5 Results for Viscous Range Optimizations with k = 0:04 and 0.1 - (Courtesy of
[Poole et al.]) .................................................................................................................................................................. 147
Table 6.6 Optimization Results (CD in Counts) – Courtesy of [Allen et al.]...................................... 167
Table 6.7 Aerodynamic Shape Optimization Problem - (Courtesy of Martins and Hwang) ...... 173
Table 6.8 Mesh Convergence Study for the Baseline CRM Wing - (Courtesy of Martins and
Hwang).............................................................................................................................................................................. 173
Table 6.9 Variations of the Objective Function, Drag (dc=drag count) and Lift Coefficients
– Courtesy of [Brezillon et al.] ................................................................................................................................. 188
Table 7.1 Delta Wing Geometry Specifications ............................................................................................ 210
Table 7.2 Design of Experiments (DoE) result; input and output parameters value ................... 215
Table 7.3 Range of Design Variables ............................................................................................................... 215
Table 7.4 Three Candidates as Final Optimum Design Points ............................................................... 217
Table 7.5 Aerodynamic Coefficients Comparison before and after ..................................................... 217
Table 7.6 Set of operators of the two GAs used – Courtesy of [Vicini & Quagliarella]................. 226
Table 7.7 Design Parameters for the Wing Planform Optimization – Courtesy of [Vicini &
Quagliarella] ................................................................................................................................................................... 230
Table 9.1 Design Space........................................................................................................................................... 277
Table 9.2 Table of CFD Solver Settings ............................................................................................................ 282
Table 9.3 Table Showing Optimized Candidate Point for Routine 1 ................................................... 283
Table 9.4 Table Showing Optimized Candidate Point for Routine 2 ................................................... 284
Table 9.5 Geometrical and Operational Parameters of the VAWT – [Courtesy of Ma et al. )... 288
Table 9.6 Ranges of variables and specific values of variables for NACA0018 airfoil -
[Courtesy of Ma et al.) ................................................................................................................................................. 289
Table 9.7 Sensitivity test of grid size at TSR = 1.0 - [Courtesy of Ma et al.) ..................................... 292
Table 9.8 Settings of multi-island genetic algorithm [Courtesy of Ma et al.)................................... 294
Table 10.1 Computed Pitch-Rate Derivatives............................................................................................... 321
Table 10.2 Aerodynamic Sensitivity Coefficient.......................................................................................... 328
Table 10.3 Design improvement for an Airfoil............................................................................................. 328
Table 11.1 Axial Flow Compressor Design .................................................................................................... 356
Table 11.2 Airfoil Geometry Parameters ........................................................................................................ 366
Table 11.3 Airfoil Design Variables................................................................................................................... 366
Table 11.4 Rotor 37 Optimized Bump Performance Comparison – Courtesy of [John et al.]... 383
Table 12.1 Mathematical Notation for MDO Problem Formulations (Courtesy of Martins &
Lambe) .............................................................................................................................................................................. 398
Table 12.2 Algorithm 1 - Block Gauss–Seidel Multidisciplinary Analysis Algorithm................... 399
List of Figures:
Figure 1.1 Hierarchy of Models for Industrial Applications ...................................................................... 22
Figure 1.2 Cp Contours on High Lift Configuration with 22 M Cells Model ........................................ 23
Figure 2.1 Input, Target, and Airfoil Contours and Surface ....................................................................... 29
Figure 2.2 Input, target, and Computed nacelle contours and surface pressures for
an asymmetric nacelle design problem ........................................................................................................... 30
14
Figure 2.3 Schematic View of the Imposed Boundary Conditions for Inverse Problem on a
3D Nozzle ............................................................................................................................................................................ 35
Figure 2.4 The inverse problem on a three-dimensional nozzle. (a) Initial nozzle
geometry with a wavy upper boundary. (b) Final nozzle geometry as computed by the
inverse method ................................................................................................................................................................. 36
Figure 2.5 Inverse Problem on a Stator Blade ................................................................................................. 37
Figure 2.6 Pod-Mounted Scramjet Concept ...................................................................................................... 40
Figure 2.7 Dual-Mode Scramjet Concept ........................................................................................................... 40
Figure 2.8 Illustration of the Cross Section of the Scramjet ...................................................................... 41
Figure 2.9 Conceptual 2D Centerline Cross-Section of the Forebody-Inlet-Isolator
Scramjet Section with Flow Physics Representation ........................................................................................ 42
Figure 2.10 Nonweiller Caret Wing Wave Rider Configuration ............................................................... 47
Figure 2.11 Preparation for Extracting Information for 2D Base View ................................................ 48
Figure 2.12 Generation of 2D Base View for a Wedge ................................................................................. 49
Figure 2.13 Generation of 2D Caret-Shaped Geometry, Base View ........................................................ 49
Figure 2.14 Wave Rider Derived Stream Tube ............................................................................................... 50
Figure 2.15 AVUS Euler Results Density Contours ........................................................................................ 51
Figure 2.16 Forebody-Inlet-Isolator Validation Study with 2D Slices................................................... 52
Figure 2.17 Centerline 2D Pressure Contours.................................................................................................. 53
Figure 2.18 Mach Contours at the Isolator Exit ............................................................................................... 54
Figure 2.19 A Dual Mode Ramjet-to-Scramjet Concept – Courtesy of Dhanasar M. ....................... 54
Figure 2.20 Illustration of the Transition-Combustor-Nozzle Element................................................ 55
Figure 2.21 2D-3D Geometric Construction from Prescribes Isolator Cross-Section and
Aerodynamic Inputs ....................................................................................................................................................... 55
Figure 2.22 4-Pts Scramjet with Circular Combustor C-Sections ............................................................ 56
Figure 2.23 4-Pts Scram Jet with Square Combustor C-Sections............................................................. 56
Figure 3.1 Basic Three-Factor Designs ............................................................................................................... 60
Figure 3.2 Aerodynamic Characteristics of an Airfoil .................................................................................. 66
Figure 3.3 Rear Fuselage Shape............................................................................................................................. 72
Figure 3.4 Over Wing Mounted Engines Configuration............................................................................... 75
Figure 3.5 Pressure Distribution for wing-pylon-nacelle Configuration; (Initial left),
(refined right) ................................................................................................................................................................... 75
Figure 3.6 Under Wing Mounted Engines Configuration ............................................................................ 76
Figure 3.7 Leading Edge Droop and Vortilons ................................................................................................ 77
Figure 3.8 Rear Fuselage Mounted Engines Configuration........................................................................ 77
Figure 3.9 Simplified Wing Planform of a ......................................................................................................... 78
Figure 3.10 Redesigned Boeing 747 Wing at Mach 0.86 based on Cp Distributions...................... 79
Figure 3.11 Tightly Coupled Two Level Design Process ............................................................................. 80
Figure 3.12 Variable Camber Sealed Flap Example – Courtesy of Sobieczky ..................................... 82
Figure 3.13 Wing Sections for Multi-Component High Lift System, 3D Swept Wing with
Slat 83
Figure 3.14 Wing Parameters and Respective key numbers for Section Distribution, .................. 84
Figure 3.15 Fuselage Parameters and Respective Key Numbers for Cross Section
Definition, ........................................................................................................................................................................... 85
Figure 3.16 Combination of Two Components by a Blended Projection Technique ....................... 85
Figure 3.17 CFD Grid Boundaries Result for Euler Analysis of HSCT Wing-Body in ....................... 87
Figure 4.1 Global Maximum of f (x, y) ................................................................................................................. 89
Figure 4.2 Example of Numerical Optimization ............................................................................................. 90
Figure 4.3 Schematic of a Gradient-Based Optimization with Two Design Variables ................... 91
Figure 4.4 Different Search and Optimization Techniques ........................................................................ 93
15
Figure 6.14 M-CL maps with normalized ML/D Contours of Optimized airfoils (97%
contour highlighted) - (Courtesy of [Poole et al.]) .......................................................................................... 149
Figure 6.15 Volume of solid (VOS) design variables as grey-scale and RSVS profile in red;
1 corresponds to a completely full cell and 0 an empty cell – Courtesy of [Allen et al.] ................. 154
Figure 6.16 Four Levels of Subdivision of a Four Point Control Polygon - Courtesy of
[Allen et al.] ..................................................................................................................................................................... 155
Figure 6.17 Generic Non-Symmetric Airfoil Modes - a Mode 1. b Mode 2. c Mode 3. d Mode
4. e Mode 5....................................................................................................................................................................... 157
Figure 6.18 Surface-Based Control Points and Example Deformation. a Control Points. b
Example Deformation - Courtesy of [Allen et al.] ........................................................................................... 160
Figure 6.19 Surface Mesh and Control Points in 3D – Courtesy of [Allen et al.] ............................ 161
Figure 6.20 Surface and Control Point Modal Deformations. a Mode 1 global. b Mode 3
global. c Mode 5 ............................................................................................................................................................. 162
Figure 6.21 Surface Mesh and off-Surface Control Points – Courtesy of [Allen et al.] ................. 163
Figure 6.22 Domain and Block Boundaries and far-field Mesh – Courtesy of [Allen et al.]....... 167
Figure 6.23 Upper Surface Pressure Coefficient. a Initial Geometry. b Domain Element. c
10 Global Modes ,d 10 Local Modes – Courtesy of [Allen et al.] ................................................................ 168
Figure 6.24 Convergence Histories – Courtesy of [Allen et al.] ............................................................. 169
Figure 6.25 Shape Design Variables are the z-Displacements of 720 FFD Control Points -
(Courtesy of Martins and Hwang) ......................................................................................................................... 171
Figure 6.26 Optimized Wing with Shock-Free with 8.5% Lower Drag – (Courtesy of Lyu
and Martins) ................................................................................................................................................................... 174
Figure 6.27 Insensitivity of Number of Optimization Iterations to Number of Design
Parameters ...................................................................................................................................................................... 176
Figure 6.28 Multipoint Optimization Flight Conditions ........................................................................... 177
Figure 6.29 Multi-Point Optimized - (Courtesy of Lyu and Martins) .................................................. 177
Figure 6.30 Comparison of Baseline, Single, and Multipoint Optimization...................................... 178
Figure 6.31 Flow Field Around the Wing Section of a 3-Element Wing – Courtesy of
[Brezillon et al.] ............................................................................................................................................................. 180
Figure 6.32 Multi-Block Structured Mesh Around the DLR-F11 Model in Full Span Flap
and Slat Configuration – Courtesy of [Brezillon et al.] .................................................................................. 185
Figure 6.33 Objective function according to the wall-clock time - Aerodynamic flows
computed with the structured FLOWer code running sequentially on a NEC-SX8 – Courtesy
of [Brezillon et al.] ........................................................................................................................................................ 187
Figure 6.34 Close View of the Wake Discretization on the Top of the DLR-F11 Wing –
Courtesy of [Brezillon et al.]..................................................................................................................................... 187
Figure 6.35 Objective and Lift Coefficient According to the ................................................................... 188
Figure 6.36 Drag Distribution Along the Spanwise Direction on ......................................................... 189
Figure 6.37 Optimizing Configuration for Slat and Flap – Courtesy of [Brezillon et al.] ............ 190
Figure 7.1 Sketch of Griewank Function on larger scale (left) vs. smaller scale (right) .............. 192
Figure 7.2 Optimization Scheme (GA) ............................................................................................................. 194
Figure 7.3 GA Representation via Control Point of B-Spline for an Airfoil - Courtesy of
[Zhang et al.] ................................................................................................................................................................... 195
Figure 7.4 GA Flowchart - Courtesy of [Zhang et al.] ................................................................................. 196
Figure 7.5 Original NACA0012 and Optimized Airfoil N.S. Solution – Courtesy of [Zhang et
al.] 197
Figure 7.6 Mach Number Distributions on the Wing Surfaces - Courtesy of [Zhang et al.] ....... 197
Figure 7.7 Mach Numbers for the Original and Optimized Airfoil - Courtesy of [Zhang et
al.] 198
17
Figure 7.8 Comparison of the Pressure Distribution between Baseline and Optimized
Design for Maximizing Airfoil Lift – Courtesy of [Ahuja and Hosangadi].............................................. 203
Figure 7.9 Comparison of Cavitation Zones between Baseline (Bottom) and Optimize
Design (Top) For Maximizing Airfoil Lift – Courtesy of [Ahuja and Hosangadi]................................ 203
Figure 7.10 Comparison of Axial Velocity Distribution between Baseline (Bottom) and
Optimize Design (Top) For Maximizing Airfoil Lift – Courtesy of [Ahuja and Hosangadi] ............ 204
Figure 7.11 Solution Schematic and Flowchart ........................................................................................... 206
Figure 7.12 View of two Distinct NACA Supersonic Airfoils................................................................... 208
Figure 7.13 2D View of the Wing in Sweep Angle of 37, 50 and 60 Degrees ................................... 209
Figure 7.14 Delta Wing Geometry Used for Numerical Validation ...................................................... 211
Figure 7.15 Spanwise Cp over the Delta Wing Upper Surface at M = 1.7, AoA = 8 Degree ...... 211
Figure 7.16 Aerodynamic Changes in Sweep Angles of 37 ; 50 and 60 , t/c =10%..................... 212
Figure 7.17 Mach Number at X/L = 0.6 and AoA = 5 deg. ....................................................................... 213
Figure 7.18 Predicted from RSM vs. observed from DOE, Normalized values................................ 216
Figure 7.19 Final and Optimized Wing Geometry ...................................................................................... 217
Figure 7.20 Sensitivity of Objective Functions w.r.t Input Parameters ............................................ 218
Figure 7.21 Hybrid Organic-Optimization Algorithm ............................................................................... 221
Figure 7.22 Flow Chart of Coupled Optimization Framework – Courtesy of [Oh and
Chien] ................................................................................................................................................................................ 222
Figure 7.23 Sketch of the Hybrid Genetic Algorithm – Courtesy of [Vicini & Quagliarella] ...... 225
Figure 7.24 Convergence Histories for the CAST 10 Inverse Design Problem - Courtesy of
[Vicini & Quagliarella] ................................................................................................................................................ 227
Figure 7.25 Scatter of the Results Obtained in 10 Different Runs for the CAST 10 Inverse
Design Problem – Courtesy of [Vicini & Quagliarella]................................................................................... 228
Figure 7.26 Comparison of the Convergence Histories Obtained by Letting the HcO
Operate on All ................................................................................................................................................................ 229
Figure 7.27 Pareto Fronts Obtained for the Wing Optimization – Courtesy of [Vicini &
Quagliarella] ................................................................................................................................................................... 231
Figure 7.28 Comparison Between the Pareto Fronts and the Results Obtained Through
the Gradient Based Method – Courtesy of [Vicini & Quagliarella] ........................................................... 232
Figure 7.29 Surrogate-Based Optimization Framework – Courtesy of [Z.-H. Han et al.]............ 237
Figure 7.30 Wing-Body Transport Aircraft Configuration and FFD box (8 control sections
with 10 FFD Nodes for each Section, Resulting in 40 nodes on upper and lower wing
surfaces, respectively) ................................................................................................................................................ 239
Figure 7.31 Comparison of Surface CP on Fine and Coarse Grids (M = 0.83, Re = 4.34E7, CL
= 0.5) 240
Figure 7.32 CP of the Baseline and SBO Optimized Configurations (M = 0.83, Re = 4.34E7,CL
= 0.5) 241
Figure 7.33 Different Influences on SBO (M = 0.83, Re = 4.34E7, CL = 0.5) ...................................... 242
Figure 7.34 Example of 1D data Interpolation by Kriging, with Confidence Intervals............... 244
Figure 7.35 Prediction Comparison of the Rosenbrock Function Based on Kriging Model
and GEK Model............................................................................................................................................................... 245
Figure 8.1 Artificial Neural Network (ANN) Configuration .................................................................... 247
Figure 8.2 Artificial Neural Networks (ANN) with 1 & 2 Hidden Layers .......................................... 248
Figure 8.3 Sigmoid Function ................................................................................................................................ 248
Figure 8.4 Assessing ANN’s Generalization Ability via Cross-Validation.......................................... 249
Figure 8.5 Operations Done by a Neuron ....................................................................................................... 249
Figure 8.6 Network Diagram for a feed-forward NN with three inputs and one output........... 250
Figure 8.7 Agile AI-Enhanced Design Space Capture and Smart Surfing .......................................... 253
Figure 8.8 Illustration of AI-Enhanced Design Process ............................................................................ 254
18
Figure 8.9 Edge Geometry and definition of flap and slat high-lift rigging ...................................... 254
Figure 8.10 Unsteady Non-Dimensional Pressure Wake Behind the Cylinder after Flow
Initialization without Active Control. The Location of the Velocity Probes is Indicated by the
Black Dots While The Location of the Control Jets is Indicated by the Red Dot ................................. 259
Figure 8.11 Illustration of the Robustness of the Learning Process ................................................... 263
Figure 8.12 Time-Resolved Value of the Drag Coefficient CD in the case without (baseline
curve) and with (controlled curve) Active Flow Control, and Corresponding Normalized
Mass Flow Rate of the Control Jet 1 (Q*1 inset) ............................................................................................... 264
Figure 8.13 Comparison of representative snapshots of the velocity magnitude in the
case without actuation (top), and with active flow control (bottom). The bottom figure
corresponds to the established pseudo-periodic modified regime, which is attained after
the initial transient control. ..................................................................................................................................... 266
Figure 9.1 Schematic of a Wind Turbine Generation System- (Courtesy of Gaurav Kapoor)... 269
Figure 9.2 Schematic of Internal Components of a Modern HAWT- (Courtesy of Gaurav
Kapoor) ............................................................................................................................................................................. 270
Figure 9.3 Profiles of Flat-back and Sharp Trailing Edge Airfoils ........................................................ 271
Figure 9.4 Blade Twist at Span-wise Sections (Airfoils) and Apparent Wind Angles .................. 271
Figure 9.5 Typical Wind Turbine Blade Planform View........................................................................... 272
Figure 9.6 Swirling Flow in the Wind Turbine Wake ................................................................................ 272
Figure 9.7 Typical Wind Turbine Power Output Curve ............................................................................ 273
Figure 9.8 Actuator Disk Concept for Wind Turbine Rotor .................................................................... 274
Figure 9.9 Actuator Disk Concept, Pressure and Velocity Profiles ...................................................... 275
Figure 9.10 AOC 15/50 Blade Geometry ........................................................................................................ 275
Figure 9.11 Turbine Blade Showing Various Radial Stations in ANSYS® DesignModeler ........ 276
Figure 9.12 Response Surface Showing Variation of P3, P9 with respect to P12 (Torque) ...... 278
Figure 9.13 Rotating Body in the Inertial Reference Frame................................................................... 279
Figure 9.14 Grid Independence Study ............................................................................................................. 281
Figure 9.15 Torque (P12) vs Twist_Station3 (P9) for Optimization Routine 2 .............................. 284
Figure 9.16 Schematic diagram of an H-rotor type VAWT: (a) 3D view; (b) top view -
[Courtesy of Ma et al) .................................................................................................................................................. 288
Figure 9.17 Plan Views of the Computational Domain and Boundary Conditions -
[Courtesy of Ma et al.) ................................................................................................................................................. 290
Figure 9.18 Mesh Setup (a) Topology; & BCs. (b) Grids at the Mid-Height Plane - [Courtesy
of Ma et al.) ...................................................................................................................................................................... 290
Figure 9.19 Time-varying power coefficients in a rotational period when adopting
different time steps - [Courtesy of Ma et al.) ..................................................................................................... 292
Figure 9.20 Comparison of average power coefficients between the experiment and
simulation (NACA0018, freestream velocity V∞ = 8 m/s) as well as relative errors of them -
[Courtesy of Ma et al.) ................................................................................................................................................. 293
Figure 9.21 The scheme of the optimization system - [Courtesy of Ma et al.) ................................ 295
Figure 9.22 Profiles of the initial airfoil and optimized airfoil .............................................................. 296
Figure 9.23 Relative thickness distribution of the initial and optimized airfoil profiles -
[Courtesy of Ma et al.) ................................................................................................................................................. 296
Figure 9.24 Comparison of average power coefficients and the growth rate of CP between
the initial and optimized airfoils at different TSRs......................................................................................... 297
Figure 9.25 Instantaneous torque values during one rotational period - [Courtesy of Ma et
al.) 298
Figure 9.26 Instantaneous Contours of Vorticity Magnitude at TSR = 0.9 for the VAWT
with Optimized Airfoils - [Courtesy of Ma et al.) ............................................................................................. 299
Figure 10.1 Methods of Evaluating Sensitivity Derivatives .................................................................... 304
19
Figure 11.17 Total Pressure Loss Coefficient Evolution in time Calculated with URANS
Simulation for both the Baseline and the Optimized Configuration ....................................................... 372
Figure 11.18 Mach Number Contours calculated at Three Different Time Instances with
the HB Method, Based on the OptC1 test case, for both the Baseline (a), (b), (c) and the
Optimized (d), (e), (f) Blade Profile ...................................................................................................................... 372
Figure 11.19 Tip Speed) With Pronounced Negative Camber. From Prince – Courtesy of
[Prince] ............................................................................................................................................................................. 373
Figure 11.20 Schematic Of Shock Structures (A) Datum, (B) S-Shaped Design. Courtesy of
[John et al.] ...................................................................................................................................................................... 375
Figure 11.21 Datum Geometry and Optimized Shock Control Bumps on The Mid- Section
of Nasa Rotor 67- From Mazaheri et al............................................................................................................... 376
Figure 11.22 The R37 CFD Domain Used – Courtesy of [John et al.]................................................... 377
Figure 11.23 Radial Profiles Vs Experimental Data – Courtesy of [ John et al.] ............................. 378
Figure 11.24 (a) 3d Separation (Orange) On The R37 Geometry (Flow Right To Left), (b)
Rel. Mach No. Contour At 60% Span – Courtesy of [John et al.] ................................................................ 379
Figure 11.25 Shock Region Flow Features For RR-Fan At Points A) A, B) B, C) C, D) D, E) E,
F) F. Flow Direction – Courtesy of [John et al.]................................................................................................. 380
Figure 11.26 Example 2d CST Bump (Solid Line) And The Four Polynomials Used To
Construct It (Dashed Lines) – Courtesy of [John et al.] ................................................................................. 381
Figure 11.27 a) Example Individual Bump Geometry............................................................................... 381
Figure 11.28 Spanwise Slice Of The Datum And Optimized R37 Geometries At 60% Span
– Courtesy of [John et al.] .......................................................................................................................................... 382
Figure 11.29 Optimized R37 Bump (Blue) Added To The Datum Blade Geometry (Grey) ....... 382
Figure 11.30 Datum (Left) And Optimized (Right) Rotor 37 Static Pressure Contours.
Flow Direction Right To Left – Courtesy of [John et al.] ............................................................................... 383
Figure 11.31 Datum (Left) And R37 Optimized (Right) ........................................................................... 384
Figure 11.32 Datum (Left) And Optimized (Right) Rotor 37 Separated Flow Contours
(Orange). Flow Direction Right To Left. .............................................................................................................. 384
Figure 11.33 Lift Plots For The Datum and Optimized ............................................................................. 385
Figure 11.34 R37 Optimized Characteristic Vs Datum – Courtesy of [John et al.] ........................ 385
Figure 12.1 a) Single-level optimization method with integrated analyses (first generation
MDO 389
Figure 12.2 Parallel Jacobi that exchanges sub problem solutions at the end of an iteration
(left), 397
Figure 12.3 A block Gauss–Seidel Multidisciplinary Analysis (MDA) Process to Solve a
Three-Discipline Coupled System – (Courtesy of Martins & Lambe ) .................................................... 400
Figure 12.4 A Gradient-Based Optimization (Courtesy of Martins & Lambe) ................................ 401
Figure 12.5 XDSM for Solving the AAO Problem (Courtesy of Martins & Lambe) ........................ 402
Figure 12.6 Diagram for the SAND Architecture (Courtesy of Martins & Lambe) ........................ 404
Figure 12.7 Diagram of the IDF Architecture (Courtesy of Martins & Lambe) ............................... 405
Figure 12.8 Diagram for the MDF Architecture with a Gauss–Seidel Multidisciplinary
Analysis (Courtesy of Martins & Lambe) ........................................................................................................... 406
Figure 12.9 Example of Aero-Structure Coupled Optimization (Courtesy of Martins) ............... 407
Figure 12.10 Classification of the MDO Architectures .............................................................................. 410
Figure 12.11 Diagram for the CSSO Architecture (Courtesy of Martins & Lambe) ....................... 414
Figure 12.12 Diagram for the CO Architecture (Courtesy of Martins & Lambe)............................ 415
Figure 12.13 XDSM for the ECO Architecture (Courtesy of Martins & Lambe) .............................. 417
Figure 12.14 Product, Components and the Supporting Disciplines .................................................. 422
Figure 12.15 Aerodynamic Design Process for Turbomachinery ........................................................ 423
Figure 12.16 Comparison of "Baseline" and "Optimized" Turbine Mean Line Results ............... 425
21
Figure 12.17 The concept of meta-modelling for a response depending on two design
variables – (Courtesy of Ryberg et al.) ................................................................................................................. 427
Figure 12.18 Schematic illustration of Simulation Areas within the Automotive Industry,
example from Saab Automobile (Courtesy of Ryberg et al.) ....................................................................... 428
Figure 12.19 Schematic description of simultaneous optimization of aerodynamic and
structural ......................................................................................................................................................................... 431
22
1 Introduction
1.1 Complexity of Flow
The complexity of fluid flow is well illustrated in Van Dyke’s Album of Fluid Motion. Many critical
phenomena of fluid flow, such as shock waves and turbulence, are essentially nonlinear and the
disparity of scales can be extreme. The flows of interest for industrial applications are almost
invariantly turbulent. The length scale of the smallest persisting eddies in a turbulent flow can be
estimated as of order of 1/Re3/4 in comparison with the macroscopic length scale. In order to resolve
such scales in all three spatial dimensions, a computational grid with the order of Re 9/4 cells would
be required. Considering that Reynolds numbers of interest for airplanes are in the range of 10 to
100 million, while for submarines they are in the range of , the number of cells can easily overwhelm
any foreseeable supercomputer. [Moin and Kim] reported that for an airplane with 50-meter-long
fuselage and wings with a chord length of 5 meters, cruising at 250 m/s at an altitude of 10,000
meters, about 10 quadrillions (1016) grid points are required to simulate the turbulence near the
surface with reasonable details. They
estimate that even with a sustained
performance of 1 Teraflops, it would
take several thousand years to
simulate each second of flight time. RANS (1990s)
Spalart has estimated that if computer
performance continues to increase at
the present rate, the Direct Numerical Euler (1980s)
Simulation (DNS) for an aircraft will be
feasible in 2075.
Consequently mathematical models Non-linear
with varying degrees of simplification Potential (1970s)
have to be introduced in order to make
computational simulation of flow Linear Potential
feasible and produce viable and cost- (1960s)
effective methods. Figure 1.1
indicates a hierarchy of models at
different levels of simplification which
have proved useful in practice. Inviscid Figure 1.1 Hierarchy of Models for Industrial Applications
calculations with boundary layer
corrections can provide quite accurate predictions of lift and drag when the flow remains attached.
The current main CFD tool of the Boeing Commercial Airplane Company is TRANAIR, which uses the
transonic potential flow equation to model the flow. Procedures for solving the full viscous
equations are needed for the simulation of complex separated flows, which may occur at high angles
of attack or with bluff bodies. In current industrial practice these are modeled by the Reynolds
Average Navier Stokes (RANS) equations with various turbulence models1.
1A. Jameson and M. Fatica, “Using Computational Fluid Dynamics for Aerodynamics”, Stanford University,
USA.
23
calculated on a mesh with 160 cells around the section, and 32 cells normal to the section. Using a
new non-linear Symmetric Gauss-Siedel (SGS) algorithm, which has demonstrated “text book”
multigrid convergence (in 5 cycles), two-dimensional calculations of this kind can be completed in
0.5 seconds on a laptop computer (with a 2Ghz processor). A three dimensional simulation of the
transonic flow over a swept wing on a 192 x 32 x 32 mesh (196,608 cells) takes 18 seconds on the
same laptop. Moreover it is possible to carry out an automatic redesign of an airfoil to minimize its
shock drag in 6.25 seconds, and to redesign the wing of a Boeing 747 in 330 seconds2.
Viscous simulations at high Reynolds numbers require vastly greater resources. On the order of 32
mesh intervals are needed to resolve a turbulent boundary layer, in addition to 32 intervals between
the boundary layer and the far field, leading to a total of 64 intervals. In order to prevent
degradations in accuracy and convergence due to excessively large aspect ratios (in excess of 1,000)
in the surface mesh cells, the chord wise resolution must also be increased to 512 intervals.
Translated to three dimensions, this implies the need for meshes with 5-10 million cells (for
example, 512 x 64 x 256 = 8,388,608 cells) for an adequate simulation of the flow past an isolated
wing. When simulations are performed on less fine meshes with, say, 0.5 M to 1 M cells, it is very
hard to avoid mesh dependency in the solutions as well as sensitivity to the turbulence model.
Currently Boeing uses meshes with 15-60 million cells for viscous simulations of commercial
aircraft with their high lift systems deployed3. Figure 1.2 show the Cp contours on High Lift
Configuration (wing) with 22 M Cells (Courtesy of Boeing). Using a multigrid algorithm, 2000 or
more cycles are required to reach a steady state, and it takes 1-3 days to turn around the calculations
on a 200 processor Beowulf cluster.
2 Antony Jameson & Massimiliano Fatica, “Using Computational Fluid Dynamics for Aerodynamics”,
Stanford University.
3 Boing using 22 M Cells on High lift Configuration.
24
4 Glauert, H. (1926), “The Elements of Aero foil and Airscrew Theory”, Cambridge University Press.
5 Prandtl, L. and Tietjens, O.G. (1934), “Applied Hydro and Aerodynamics”, Dover Publications.
6 Sears, W.D. (1947), ”On projectiles of minimum drag”, Q. Appl. Math., 4, 361–366.
7 Jones, R.T. (1981), ”The minimum drag of thin wings in frictionless flow”. J. Aerosol Sci., 18, 75–81.
8 Lighthill, M.J. (1945), ”A new method of two dimensional aerodynamic design”. Rep. Memor. Aero. Res. Coun.
2 Design Methodology
According to different definitions of optimization objectives, the numerical design methods can be
classified into two groups: Direct design vs. Inverse design. The direct design and its application in
optimization, is discussed in the following chapters. Hereby, after brief introduction, we address the
indirect (inverse) design methodology.
10 Zhihui Li, Xinqian Zheng, “Review of design optimization methods for turbomachinery aerodynamics”,
Progress
in Aerospace Sciences, July 2017.
11 Xiaodong Wang, “CFD Simulation of Complex Flows in Turbomachinery and Robust Optimization of Blade
Design”, Submitted to the Department of Mechanical Engineering Doctor of Philosophy at the Vrije Universiteit
Brussel July 2010.
27
2.2.1 Case Study 1 - Inverse Aerodynamic Design Method for Aircraft Component
The motivation for using automated, inverse aerodynamic design methods is to reduce the overall
effort required to develop aircraft geometries possessing favorable aerodynamic performance or
aerodynamic interference characteristics, as investigated by [Malone et al]16. [Garabedian and
McFadden]17 described an iterative aerodynamic design procedure suitable for automated wing
design (referred to here as the GM method). They demonstrated their method by incorporating it
into a three dimensional, full-potential, transonic wing, aerodynamic analysis code. In the GM design
method, an auxiliary partial differential equation governing the spatial location of wing surface
ordinates was solved iteratively in the computational plane, together with the fluid flow equation, to
achieve given target surface-pressure distributions. The technique, was recommended for use over
only a limited portion of the wing geometry. From the present authors' experience, the original GM
method of updating wing ordinates in a normal, or nearly normal, direction to the surface can lead to
irregularities in the final design pressures near the leading edge of the design geometry. This is due
primarily to the non-uniform stretching that can occur near the leading edge where the surface
normal directions are parallel to the longitudinal axis of the component geometry.
12 Junlian Yin K Daneshkah and M Zangeneh, “Parametric design of a Francis turbine runner by means of a three-
dimensional inverse design method”, IOP Conf. Series: Earth and Environmental Science 12, 2010.
13 Lei Tan, Shuliang Cao, Yuming Wang and Baoshan Zhu, “Direct and inverse iterative design method for
inverse design method”, IOP Conf. Series: Earth and Environmental Science, 2010.
15 Mookesh Dhanasar, Frederick Ferguson and Julio Mendez, “Inversely Designed Scramjet Flow-Path”, North
Carolina Agricultural and Technical State University, Greensboro, North Carolina, United States of America,
16 J.B. Malone, J. Vadyak and L.N. Sankar, “Inverse Aerodynamic Design Method for Aircraft Components”, J. of
Aircraft, 1985.
17 Garabedian, P. and McFadden, G., "Design of Supercritical Swept Wings," AIAA Journal, Vol. 20, March 1982.
18 2Malone, J.B., Vadyak, J., and Sankar, L.N., "A Technique for the Inverse Aerodynamic Design of Nacelles and
improve airfoil or wing designs where the control of leading-edge pressures is desirable and to
extend the method to handle different types of configurations, including axisymmetric or asymmetric
body geometries. The surface perturbations generated with the MGM procedure are interpreted as
changes in the coordinate direction perpendicular to the longitudinal axis of the geometry. This
interpretation for the movement of the surface coordinates leads to smoother leading-edge geometry
designs.
2.2.1.2 Formulation
A two-dimensional airfoil design problem is used here to illustrate the salient features of the MGM
procedure. For this two-dimensional application, the original GM auxiliary equation is rewritten as
∂z ∂2 z ∂3 z
β0 + β1 + β2 2 = ∆Q2 = Q2 − q2
∂t ∂x ∂t ∂x ∂t
Eq. 2.1
In Eq. 2.1, the Q's are user specified target pressures expressed as flow velocities, the q computed
flow velocities predicted by the given fluid flow solution procedure, and the 3 user-defined constants
that improve the convergence of the algorithm. The time coordinate in Eq. 2.1 is actually a pseudo
time variable representing different iterations in the solution process. As Q approaches q, the right-
hand side of Eq. 2.1 vanishes and the surface coordinates z(x,t) stop varying with pseudo time.
Partial derivatives with respect to the time coordinate are interpreted as a change in the surface
coordinate Az between any two design iterations. In order to apply Eq. 2.1 correctly to both upper
and lower surfaces, the value of Az must have opposite signs on each surface for equal values of the
quantity, Q2 - q2.
−β1 −2β2
Ai = +
(xi+1 − xi ) (xi+1 − xi )(xi+1 − xi−1 )
Eq. 2.2
A more complete description of the finite-difference expressions used in Eq. 2.2 can be found in19.
Eq. 2.2 is evaluated at each point i, leading to a system of TV equations in TV unknowns (the Azi
values). Note that at each point on the aerodynamic surface, Azi is coupled to values at neighboring
points. The resulting equations form a tridiagonal system that is solved for the values of Azi using the
well-known Thomas algorithm. Special treatment is required at two locations on the design
19 Malone, J.B., Vadyak, J., and Sankar, L.N., "A Technique for the Inverse Aerodynamic Design of Nacelles and Wing
geometry contour. At the leading edge, an ambiguity arises as to the direction to apply the up winding
used to evaluate certain derivatives in Eq. 2.2. To eliminate this problem, the leading-edge point is
constrained to move as the average of both the upper- and lower-surface downstream points. This
also permits the local angle of attack to change during the design process. Also, at the trailing edge,
values of Az are needed to evaluate the second derivative terms. Currently, these values are set to
zero. Thus, during the design process, the trailing-edge thickness remains constant. Actually, the
baseline, or starting geometry used to initiate the design process, serves primarily to fix the trailing-
edge thickness of the final design geometry.
20Malone, J.B. and Sankar, L.N., "Numerical Simulation of Two-Dimensional Unsteady Transonic Flows Using the
Full Potential Equation," AIAA Journal, Vol. 22, Aug. 1984, pp. 1035-1041.
30
2.2.1.5 Conclusion
A well-known design procedure, the [Garabedian-McFadden] method, has been modified to permit
design of airfoil and wing leading-edge regions and to extend the range of configurations that can be
handled by this technique. The modified design procedure has been incorporated by the authors into
several existing aerodynamics programs and sample design problems have been presented for airfoil
and nacelle geometries. Although simple in nature, the changes to the existing design algorithm have,
in all cases examined by the authors, improved the overall versatility of the method. Test problems
that were successful using the original method were not adversely effected by the modified scheme
presented here. However, several classes of problems, an example of which is the case shown here
for the two-dimensional airfoil, were successful only when the modified algorithm was used during
the design process.
21 Vadyak, J. and Atta, E.H., "Approximate Factorization Algorithm for Three-Dimensional Transonic Nacelle/Inlet
Flow field Computations," Journal of Propulsion and Power. Vol. 1, Jan.-Feb. 1985, pp. 58-64.
31
2.2.2 Case Study 2 - A Pseudo-Compressibility Method for Solving Inverse Problems Based on the
3D Incompressible Euler Equations
A numerical technique to solve the 3D inverse problems that arise in aerodynamic design is
presented by [Ferlauto]22. The approach, which is well established for compressible flows, is
extended to the incompressible case via artificial compressibility preconditioning. The modified
system of equations is integrated with a characteristic-based Godunov method. The solution of the
inverse problem is given as the steady state of an ideal transient during which the flow field assesses
itself to the boundary conditions, which are prescribed as design data, by changing the boundary
contour. The main aspects of the Eulerian-Lagrangian numerical procedure are illustrated and the
results are validated by comparisons with theoretical solutions and experimental results.
22 Michele Ferlauto, “A Pseudo-Compressibility Method For Solving Inverse Problems Based On The 3D
Incompressible Euler Equations”, Taylor & Francis in Inverse Problems in Science and Engineering ,2014.
23 O. Pironneau. On optimum design in fluid mechanics. Journal of Fluid Mechanics, 59:117–128, 1972.
24 A. Jameson, L. Martinelli, and N. A. Pierce. Optimum aerodynamic design using the Navier-Stokes equations.
quadratic programming algorithm. AIAA Journal of Propulsion and Power, 17(5):1123–28, 2001.
26 J. M. Lighthill. A new method of two-dimensional aerodynamic design. Aeronautical Research Council Reports
heat transfer, elasticity and materials design. International Journal of Aeronautical & Space Sciences, 2012.
28 A. Iollo, M. Ferlauto, and L. Zannetti. An aerodynamic optimization method based on the inverse problem adjoint
Approaches to the inverse problem solution are based on the potential flow theory and conformal
mapping techniques29-30-31, on stream-function base formulations 32-33, on boundary elements
replacing the body surface34, on the compressible potential flow theory. Several examples of inverse
problem solution methodologies are presented in 35 and references therein. A generalizable way of
solving aerodynamic inverse problems identifies the problem solution as the steady state of an
unsteady evolution during which the flow field tries to accommodate itself to the design data, which
are prescribed as boundary conditions on the unknown surface 36. To do this, the surface is allowed
to move. To give an idea of the physical scenario involved let us consider, for instance, the design of
an airfoil.
First, an arbitrary initial geometry is guessed and the related flow field at the selected flight Mach
number is evaluated. Then, the target pressure distribution along the profile is imposed. In general,
the boundary conditions will not be in agreement with the inner flow field. If we let the profile surface
move while it remains impermeable to the flow, walls start moving to decrease the pressure gap. If
the system reaches a steady state, and this feature relies on the existence of the steady solution for
the Euler equations, then the enquired geometry and the corresponding flow field are found. A
similar approach for the determination of the unknown shape is the elastic membrane motion
concept or the transpiration velocity 37.
One drawback of inverse problems is that they may be ill posed. If certain wall pressure distributions
are required on airfoils, the result is an open or self-intersecting profile. [Lighthill] discovered the
solvability conditions that have to be respected by pressure distributions within an incompressible
potential flow model, whereas the conditions for compressible flows and other issues where
investigated in and references therein38-39. The main advantage of numerical optimization over
inverse problem is that the first allows the maximization or minimization of global quantities, such
as lift or drag, in the presence of constraints, whereas for inverse problems the design is limited to
the selection of the pressure or velocity distribution on the boundary40, which is given on the basis
of designer experience and therefore, somewhat arbitrarily. In addition, a limited control is possible
on the final geometry.
Optimization techniques based on the adjoint method can be adopted to drive inverse problems
towards the maximization or minimization of target functionals. From this point of view, the
numerical solution of an inverse problem becomes attractive, as an alternative route to shape
optimization and to automated design. In the present work a time-dependent technique of solving
Berlin, 1997.
36 L. Zannetti. Time dependent method to solve inverse problems for internal flows. AIAA Journal, 1980.
37 A. Demeulenaere and R. Van den Braembussche. Three-dimensional inverse method for turbomachinery
∂𝐔 ∂𝐅 ∂𝐆 ∂𝐇
+𝚯( + + )=0
∂t ∂x ∂y ∂z
Eq. 2.3
𝜌 𝑐2 0 0 0
𝑢
𝐔={ } , 𝚯 = [0 1 0 0]
𝑣 0 0 1 0
𝑤 0 0 0 1
u v w
p + u2 uv uw
𝐅={ } , 𝐆={ } , 𝐇 = { vw }
uv p + v2
uw vw p + w2
Eq. 2.4
as usual, p is pressure, q = {u, v,w} is the velocity vector. All the variables are normalized to the
reference length lref , density ρref =ρ0, and pressure pref. The unsteady solutions of Eq. 2.3 does not
have a physical meaning, but, from the numerical point of view, the methods of solution for
compressible flow become available. At the steady state, the solution of Eq. 2.4 also satisfies steady
41 L. Zannetti and F. Larocca. Inverse methods for 3D internal flows. AGARD Report 780, 1990.
42 A. J. Chorin. A numerical method for solving incompressible viscous flow problems. Journal of Computational
Physics, 2:12–26, 1967.
43 A. Rizzi and L. E. Eriksson. Computation of inviscid incompressible flow with rotation. Journal of Fluid
Euler equations. The proposed strategy of solving the inverse problem requires the numerical
integration of a 3D flow field on a time-dependent domain. Unsteady flow field are solved, in the
context of artificial compressibility methods, by introducing a dual time stepping technique [26].
Dual time stepping is here unnecessary, since we are interested in over-relaxing the system to the
steady state. Moreover, the time-dependent derivative of the metrics can be neglected without a
significant loss of numerical stability. According to the Gauss formula, an integral form of Eq. 2.3 is
∂
∫ UdΩ + 𝚯 ∫(𝐅i + 𝐆j + 𝐇k). 𝐧 dσ = 0
∂t Ω 𝚺
Eq. 2.5
where ∑ is the boundary of the volume Ω and n the outward normal. Eq. 2.5 is approximated using
a finite volume technique by discretizing the (x, y, z) space with hexahedral cells whose shape
depends on time. The integration in time is carried out according to a Godunov type two-step scheme.
A first order Flux Difference Splitting (FDS) is used at the predictor step: the variables (p, u, v and
w) are assumed as an averaged, constant value inside each cell. The fluxes F, G and H are evaluated
by solving the Riemann problems pertinent to the discontinuities that take place at the cell interfaces.
At the corrector level, the second order of accuracy is achieved by assuming a linear behavior of the
primitive variables inside the cells, according to an Essentially Non-Oscillatory (ENO) procedure [28].
Implicit residual smoothing is applied to increase the numerical stability. The resulting scheme is
second order accurate in space 46.
B(x1 , x2 , x3 , t) = 0
Eq. 2.6
This manifold moves materially within the fluid, that is
dB ∂B ∂B ∂b ∂b
= + 𝐪. ∇B = 0 or = u3 − u1 − u2
dt ∂t ∂t ∂x1 ∂x2
Eq. 2.7
where q = (u1, u2, u3) is the flow velocity expressed in the Γ reference system. In the followings, Eq.
2.6 is coupled to a properly defined Riemann problem. The resulting system is used to compute the
fluxes at the boundary and to evaluate the wall velocity. The latter is then integrated in time to find
the new wall contour. For the inverse problem on a nozzle or diffuser configuration, as shown in
Figure 2.3-a, the pressure distribution is imposed on the unknown boundary 47. Aerodynamic
devices as wings or blades, instead, are composed by two surfaces that represent a closed contour.
46 D. Drikakis and W. Rider. High-Resolution Methods for Incompressible and Low-Speed Flows (Computational
Fluid and Solid Mechanics). Springer Verlag, Berlin, 2005.
47 (b) wave pattern of the Riemann problem at the unknown aerodynamic surface on which the pressure is
prescribed.
35
In this case, the inverse problem formulation must ensure a correct profile closure and non-
overlapping conditions. Among possible approaches, in addition to suitable boundary conditions at
infinity, one may:
1. prescribe the distribution of thickness and pressure jump along the chord and inquire for
the geometry of the camber line;
2. prescribe the distribution of thickness and pressure on one side of the profile and inquire
the geometry of the camber line;
3. prescribe the pressure distribution around the profile and inquire for its geometry.
Case (3) is the most challenging since, when selecting the pressure distribution, additional closure
constraints must be imposed 48. Moreover the convergence rate is low because smaller time-steps
are required and also boundary crossing must be checked at each step. All of the mentioned cases
can be grouped in two approaches in imposing the boundary conditions: the case of a surface on
which a pressure distribution is imposed and the case of two aerodynamic surfaces subjected to a
prescribed pressure jump. In next sections the kinematic constraint is accomplished to the flow field
boundary conditions in order to derive the fluxes and the velocity on the unknown surfaces. For
further info, regarding moving boundary conditions, please refer to [Ferlauto] 49.
Figure 2.3 Schematic View of the Imposed Boundary Conditions for Inverse Problem on a 3D Nozzle
48G. Volpe. Geometric and surface pressure restrictions in airfoil design. AGARD Report 780, 1990.
49 Michele Ferlauto, “A Pseudo-Compressibility Method For Solving Inverse Problems Based On The 3D
Incompressible Euler Equations”, Taylor & Francis in Inverse Problems in Science and Engineering ,2014.
50 Michele Ferlauto, “A Pseudo-Compressibility Method For Solving Inverse Problems Based On The 3D
Incompressible Euler Equations”, Taylor & Francis in Inverse Problems in Science and Engineering ,2014.
36
𝜉 − 𝑖𝑎
W(ξ) = u∞ − iκ log ( ) , Ψ(𝜉) = ℑ(𝑊(𝜉))
𝜁 + 𝑖𝑎
Eq. 2.8
where ξ = z + iy is the position, u∞ is the asymptotic velocity, κ the vortex strength and a is the distance
of the vortex from the x-axis. The two-dimensional nozzle is confined by the streamlines ψ(0) and
ψ(ib), where 0 < b < a. The velocity field can be deduced from Eq. 2.8 as
1 1
u − iw = u∞ − iκ ( − )
𝜁 − 𝑖𝑎 𝜁 + 𝑖𝑎
Eq. 2.9
while the pressure field is obtained from Bernoulli equation, once the total pressure level P0in at inlet
is fixed. The numerical test has been performed on a 90×30×20 grid for the case with b = 0.5 , u∞ =
0, κ = −1. We also set c = 3. With the scope of generating a fully 3D transient flow, we started from
an initial geometry characterized by a wavy upper wall, as shown in Figure 2.4-a. The theoretical
values of total pressure P0in = 1, the flow angles at inlet α = 0 and β = 0, the static pressure distribution
at the outlet pex(y, z), and the target static pressure pup = p(ψ(ib)) along the upper wall are imposed
as boundary conditions. As final solution, the projection of the upper wall geometry on the (x, y)-
plane is expected to match the theoretical curve ψ(ib). Initial and final geometry are shown in Figure
2.4 (a-b).
Figure 2.4 The inverse problem on a three-dimensional nozzle. (a) Initial nozzle geometry with a wavy
upper boundary. (b) Final nozzle geometry as computed by the inverse method
x 𝑥 − 𝑥𝑙
t ( ) = b0 + κca (1 − ξ)√ξ(1 − ξ) , 𝜉=
ca 𝑐𝑎
Eq. 2.10
where xl, xt are the leading and trailing edge axial stations, respectively, and Ca = xt − xl is the axial
chord. The constants are set to b0 = 0.04 and k = 0.12 . Similarly, the blade load is prescribed as
𝑥
∆p ( ) = 𝛾[1 − cos(2πξ)]
𝑐
Eq. 2.11
with γ = 0.675 . The inlet and outlet boundary condition are: P0in = 1, α = 0, β = 0 and pex = 0.99,
respectively. Starting from the initial guess of a blade with planar camber line, the transient dynamics
of the system is integrated numerically until a steady state is reached. At each step the blade
geometry is recovered from the computed velocity of the blade surface. Only the dynamics of the
blade suction-side is computed. Pressure-side surface is deduced from the knowledge of thickness. A
snapshot sequence of the blade moving surface is depicted in Figure 2.5 (a-b-c) 51. The transient
resembles the motion of a flag on the wind, fastened on one side. More precisely, the grid describes a
sliding motion on the tangential direction, since the axial chord of the blade does not change during
the inverse process. At convergence, the blade geometry is found. By switching the code in direct
mode, the flow field is computed with standard boundary conditions, and the blade load is evaluated.
The solutions of the inverse and direct solver computations are in good agreement. The computed
blade load Δp, in direct and inverse modes and the target (Δp)t, defined in Eq. 2.11, are compared in
Figure 2.5-d. A pressure mismatch, with respect to the target blade load distribution of Eq. 2.11,
can be observed close to the blade leading edge. This is a well-known local effect caused by the poor
resolution properties of the H-type grids close to the leading edge. A finer grid gives a more detailed
flow but is meaningless for the test proposed, since the inverse and the direct solvers are acting on
the same grid.
2.2.2.6 Conclusions
A numerical method of solving three-dimensional inverse problems in incompressible flows has been
presented. Details have been given for the derivation of the inverse problem over configurations of
interest in the aerodynamic design as nozzles, diffusers, blades or wings. The numerical method is
based on the pseudo-compressibility technique of preconditioning the unsteady Euler equations.
Time accuracy is unnecessary, so that dual time stepping is avoided. Moreover, the characteristic-
based formulation of the problem simplifies the parallelization of the code. The accuracy of the
numerical procedure has been tested through comparisons with theoretical solutions, numerical
computations performed with direct solvers, and finally, test against experimental data.
Discrepancies at the blade leading edge have been observed, caused by the H-type grid we used.
Further extension of the inverse procedure to the Navier-Stokes equations, in a way similar to 52 is
planned. An intermediate step required is the introduction of efficient overrelaxing techniques, to
increase the convergence ratio. For the inverse design of aerodynamic profiles the surface tracking
algorithm should be modified to deal with C-type grid generation systems that allow for an accurate
evaluation of the viscous stresses.
➢ forebody inlet;
➢ isolator,
➢ combustor,
➢ nozzle.
51 (a-b-c) Evolution of the computational domain and of the blade geometry. Iso-pressure contours are also
shown. (d) Comparison of the target blade load with the corresponding distributions computed by the inverse
procedure and by the direct solver.
52 M. Ferlauto and R. Marsilio. A viscous inverse method for aerodynamic design. Computers & Fluids, 2006.
53 Mookesh Dhanasar, Frederick Ferguson and Julio Mendez, “Inversely Designed Scramjet Flow-Path”, North
Carolina Agricultural and Technical State University, Greensboro, North Carolina, USA and Corrdesa LLC,
Tyrone, Georgia, USA, 2019.
39
Here we focuses on the design of the forebody inlet and the isolator sections of a scramjet engine. In
this framework, key to a functionally efficient scramjet engine lies in the design of its flow-path. This
flow-path design must consider a complex flow-field physics and the interaction of physical surfaces
with this complex flow-field. Many attempts to design efficient scramjet flow-paths have met with
some measured degree of success. This research uses a ‘inverse design’ approach, which is similar to
Darwin’s theory of evolution, where an organism adopts to survive in its environment; the scramjet
flow-path will be carved/extracted from the operational environment. The objective is to naturally
and organically capture, process and direct the flow from the environment; thus preparing it for the
combustion process. This approach uses the ideal 2D oblique shock relations, coupled with Non-
weiler’s caret wave rider theory and streamline marching techniques.
2.2.3.1 Introduction
Driven by the desire to improve air travel and shorten flight time, aircraft engines have evolved from
simple reciprocating internal combustion engines to advance axil flow jet engines. Jet engines fall
into several categories. These include air-breathing, turbine powered, turbojet, turbofan, ramjet
compression and scramjet compression engines. Ramjet and scramjet compression engines are
unique in that they represent the latest development on the evolutionary path of jet engines. The
ramjet, unlike conventional jet engines which uses turbine driven compressors to compress the
incoming air, uses shockwaves to achieve this goal. The compressed air is burnt in the combustor
under sub-sonic conditions. The scramjet is basically an air-breathing jet engine designed to fly at
hypersonic speeds between Mach 4 and 12 or speeds in the range of 1207–2995 m/s (2700–6700
mph).
A scramjet engine captures its airflow from the atmosphere and also compresses it across
shockwaves before the air enters the combustor. Fuel is injected into the combustor where
combustion occurs under supersonic conditions. The hot, high-pressure gas leaving the combustor is
then accelerated to high velocities in the nozzle to produce thrust as it exits the engine. Generally
speaking, the concepts associated with scramjet engines appear at first glance to be very simple. This
however is very misleading as attempts develop a working scramjet engine that has proven to be
quite an engineering challenge.
Several aspects of scramjet engine development are at various stages of development. These include
supersonic fuel-air mixing, aero-thermodynamic heat dissipation from both skin friction and internal
combustion, and other thermal management problems associated with operating an engine at
exceedingly high temperatures for extended periods of time. Combustion chamber components could
experience temperatures on the order of over 3033 K (5000°F). At these temperatures most metals
melt and fluids (air and fuel) ionize, making the physics of their associated behavior unpredictable.
40
This chapter focuses on the design concepts for the forebody, inlet, and isolator sections of an
innovative scramjet engine geometry and some of its flow physics.
conceptual engine design and engines flown to-date all have a common set of components or sub-
sections. Figure 2.6 presents these components/sub-sections for a pod-mounted conceptual
scramjet design. These components/sections are the forebody section, the inlet section, isolator
section, combustor section, and the diffuser-nozzle section. Ideally, the engine concept presented
should be able to function over a wide range of Mach numbers. This gives rise to the idea of a
morphing ramjet/scramjet or dual mode scramjet configurations as presented in Figure 2.754.
Figure 2.7a, presents the dual mode scramjet engine, Figure 2.7b, the pure scramjet mode and
Figure 2.7c, the pure ramjet mode. A typical dual mode scramjet configuration as that presented in
Figure 2.8 was inversely carved out of supersonic and hypersonic flow-fields. The design framework
used in the design of the forebody, inlet and isolator sections forms the core of this study.
All freestream flow-field properties are extracted from the Mach number and altitude55-56. This
information is used in the construction, analysis and definition of the three fundamental aerodynamic
zones, namely; the ‘primary shock’ zone AB, the ‘reflected shock zone’, BC, and the ‘isolator zone’,
CD as presented in Figure 2.9. Also presented in Figure 2.9 is a 2D conceptual representation of
the flow-field physics associated with supersonic flow interaction over a wedge and in a constant
area duct. Details of this flow-field physics and its exploitation in the inverse design approach are
explained in the next section. Also addressed is the derivation of the actual 3-dimensional forebody
54 Dhanasar M. Development of A Benchmark for the Design and Analysis of A Tip-To-Tail Ramjet-Scramjet
Propulsion System [dissertation]. Greensboro: North Carolina Agricultural and Technical State University; 2009.
55 Anderson JD. Fundamentals of Aerodynamics. 3d ed. New York: McGraw-Hill; 2001.
56 Heiser WH, Pratt DT. Hypersonic airbreathing propulsion. AIAA Education Series. Virginia, USA: American
section. This is a two-step process, where in step one, the 2D construction of the ‘forebody’, domain
A-D, is conducted. Step two is where the 3D geometry is obtained.
originating from the leading edge, A, of the wedge. The flow is deflected a second time by a reflected
shock wave, B2C1 emanating from the cowl lip at point B2, of the inlet. The flow enters the isolator
duct and travels once more in a direction that is parallel to the x-axis. To ensure that the flow in the
isolator duct remain supersonic the freestream Mach number must be greater than 3.0 and the shock
wave angle β, greater than 12 and less than 30 degrees. The flow-field behavior within the isolator
duct is of paramount importance. This flow-field may consist of a system of oblique or normal shocks,
as visualized in Figure 2.9. Driving this behavior is the flow-field vicious interactions with the
isolator duct walls. The isolator’s non-dimensional length, L/H, and the pressure differential at the
duct’s entrance and exit also enhance the flow-field’s behavior.
57 Ferguson F, Dhanasar M, Williams R, Blankson I, Kan kam D. Supersonic and hypersonic slender air-breathing
configurations derived from 2D Flow fields. 46th AIAA Aerospace Sciences Meeting and Exhibit; 7–10 January
2008; Reno, NV. DOI: 10.2514/6.2008-163.
58 Curran ET, Murthy SNB. Scramjet Propulsion. Vol. 189. Virginia, USA: AIAA; 2000. ISBN: 1-56347-322-4.
43
experimentally derived in 59-60. It is assumed that in Figure 2.9 the flow travels in the x-direction,
and that the construction of the ‘forebody’ configuration starts at design point, A. The following
account details the logic used to define the locations of design points A, B, C and D:
2
M∞ Sin2 β − 1
θ = atan {2cotβ [ 2 ]}
M∞ (γ + Cos2β) + 2
Eq. 2.12
➢ First, the Mach number, M, behind the primary shock wave, AB2 (see Figure 2.9), is obtained
using Eq. 2.13,
(γ − 1)
1 {1 + [ 2 ] [MSin(β)2 ]}
M=[ ]
Sin(β − θ) [γMSin(β)2 − (γ − 1)/2]
Eq. 2.13
➢ This Mach number, coupled with the free stream parameters are then used with the oblique
shock relations derived in61 for the evaluation of all of flow field properties behind the
primary shock, AB2. The flow-field properties, pressure, P, temperature, T, density ρ, and total
pressure, Pt,2, are evaluated using Eq. 2.14-Eq. 2.17.
59 Billig FS. Research on supersonic combustion. Journal of Propulsion and Power. 1993;9(4):499-514.
60 Waltrup PJ, Billig FS. Prediction of recompression wall pressure distributions in scramjet engines. Journal of
Spacecraft and Rockets. 1973; 10(9):620-622.
61 Anderson JD. Modern Compressible Flow: With Historical Perspective. 3rd ed. New York, USA: McGraw-Hill;
2004.
44
P 2γ(M∞ Sinβ)2 − (γ − 1)
=
P∞ (γ + 1)
Eq. 2.14
ρ (γ + 1)(M∞ Sinβ)2
=
ρ∞ (γ − 1)(M∞ Sinβ)2 + 2
Eq. 2.16
𝛾 1
(γ + 1)(M∞ Sinβ) 2 (γ + 1)
p 𝛾−1 𝛾−1
=[ ] [ ]
p∞ (γ − 1)(M∞ Sinβ)2 + 2 2𝛾(M∞ Sinβ)2 − 𝛾 − 1
Eq. 2.17
➢ B2C1 as seen in Figure 2.9 represent the reflected shock wave. This reflected shock wave is
a the result of a flow-field behind the primary shock wave, AB2, with a supersonic Mach
number, M, once more being deflected by an imaginary wedge, with wedge angle ϴ at design
point B2. This imaginary wedge is oriented in such a manner that it ensures that the deflected
flow travels parallel to the x-axis, Figure 2.9. At this stage updated values for the wedge
angle ϴ and the Mach number, M, are obtained using Eq. 2.12 and Eq. 2.13. A reflection
shock angle is now be defined as ϕ = β1- ϴ. In this expression, β1 is the reflected shock angle.
This reflected shock angle is generated by the interaction of the flow-field with Mach number
M and the imaginary wedge with angle ϴ. Note that β1 is obtained using Eq. 2.12 and
replacing the value of the freestream Mach number, M∞, with that of the supersonic Mach
number, M.
➢ The flow-field properties behind the reflected shock wave B2C1 are now obtained in a similar
manner as described in ‘b’ above. Eq. 2.13 is used to obtain M1, which is the Mach number
behind the reflected shock. In Eq. 2.13 the freestream Mach number, M∞, is replaced with
Mach number M. It is very important to note here that M1 represents the Mach number at the
entrance to the isolator section of the scramjet. Eq. 2.14-Eq. 2.17 are used to derive the
additional flow-field properties of pressure, temperature, density and total temperature, p1,
T1, ρ1 and To, behind the reflected shock. Note that in these equations the value for the
freestream Mach number, M∞, is now replaced with the value of the Mach number, M, from
the flow-field properties behind the primary shock.
➢ Having obtained the parameters ϴ, β and β1 all design points at station C can now be derived.
The y-coordinate and z-coordinate are defined as Cy = 0, and Cz = 0, respectively. The x-
coordinate is obtained with the help of trigonometric relations, and is defined as:
tan(β) − tan(θ)
Cx = [1 + ]B
tan(θ) − tan(β1 − θ) x
Eq. 2.18
45
➢ The coordinates of point C1 are determined as follows: C1x = Cx, C1y = Cxtan(ϴ), and C1z = 0.
➢ Similarly, the coordinates of point C2 are determined from: C2x = Cx, C2y = B2y, and C2z = 0.
➢ First a non-dimensional expression for the ‘normal total’ pressure value, Pn, in, is derived, Eq.
2.19. This expression is a function of isolator entrance conditions, where M1 is treated as the
Min, and the static pressure, P1, as Pin. Note here that the values of M1 and P1 are obtained
from the flow-field properties behind the reflected shock B2C1.
P𝑛,𝑖𝑛 2𝛾𝑀12 − (𝛾 − 1)
=[ ]
Pin (𝛾 + 1)
Eq. 2.19
In determining the isolator length for a design process, the ratio of the entrance to exit pressures,
Pin/Pout, over the range between Pin and Pn,in has to be evaluated. This value is needed to determine
the length of an isolator that can reliably prevent all ‘unstart’ conditions. In this design process, the
ratio, Pout/Pn,in, representing the isolator exit pressure, Pout, to the ‘normal total’ pressure value, Pn,in,
is prescribed. Using this approach, the value for Pin/Pout can be determined by using Eq. 2.20:
2 2 −0.5
γ2 Min [1 + ((γ − 1)/2)Min ] γ−1
Mout ={ 2 − }
(1 − γMin − pout /pin )2 2
Eq. 2.21
Similarly, with the exit Mach number known, the non-dimensional length of the isolator can be
evaluated based on the following experimental relationship developed in 64-65 :
62 Billig FS. Research on supersonic combustion. Journal of Propulsion and Power. 1993;9(4):499-514.
63 Waltrup PJ, Billig FS. Prediction of recompression wall pressure distributions in scramjet engines. Journal of
Spacecraft and Rockets. 1973; 10(9):620-622
64 Billig FS. Research on supersonic combustion. Journal of Propulsion and Power. 1993;9(4):499-514.
65 Waltrup PJ, Billig FS. Prediction of recompression wall pressure distributions in scramjet engines. Journal of
Finally, with the coordinates of all the design points at all stations, A, B, B1, B2, C, C1, C2, D, D1, and D2,
fully defined, the sketch illustrated in Figure 2.9 can be constructed.
66Nonweiler TRF. Aerodynamic problems of manned space vehicles. Journal of the Royal Aeronautic Society.
1959;63:521-528.
47
This approach is further explained the next sub-section and is demonstrated by the construction of a
supersonic 3D wedge followed by a 3D supersonic caret shaped geometry. This caret-shaped
geometry will then be used to generate supersonic star-shaped geometries of interest.
2004.
48
axis, which is compressed by two oblique shock waves; resulting in the flow once again traveling in
a direction parallel to the x-axis. Further examination of Figure 2.14 identifies the primary shock
70Mookesh Dhanasar, Frederick Ferguson and Julio Mendez, “Inversely Designed Scramjet Flow-Path”, North
Carolina Agricultural and Technical State University, Greensboro, North Carolina, USA and Corrdesa LLC,
Tyrone, Georgia, USA, 2019.
51
wave plane as AB3B4, which supports two compression surfaces, ACB3 and ACB4. At this stage the flow
field is no longer parallel to the x-axis. A reflected shock wave is constructed to form the plane, CB3B4.
This specially designed plane, CB3B4, now straightens the flow leaving the shock surface, CB3B4, so
that it once again travels parallel to the x-axis. The reflected flow now forms the stream tube
comprising of the following planar surfaces, CDD3B3, CDD4B4, and B3B4D4B3.
71 Mookesh Dhanasar, Frederick Ferguson and Julio Mendez, “Inversely Designed Scramjet Flow-Path”, North
Carolina Agricultural and Technical State University, Greensboro, North Carolina, USA and Corrdesa LLC,
Tyrone, Georgia, USA, 2019.
72 Mookesh Dhanasar, Frederick Ferguson and Julio Mendez, “Inversely Designed Scramjet Flow-Path”, North
Carolina Agricultural and Technical State University, Greensboro, North Carolina, USA and Corrdesa LLC,
Tyrone, Georgia, USA, 2019.
52
2.2.3.5.4 2D Simulations
Both Euler and viscous studies were conducted on the scramjet forebody, inlet, and isolator sections.
2D Euler flow studies were conducted using the air vehicles unstructured solver (AVUS) 73. AVUS is
a three-dimensional finite volume unstructured-grid Euler/Navier-Stokes flow solver. 2D isolator
viscous simulations were conducted using an in-house computational scheme, the integral
differential scheme (IDS)74. The following contour plots (Figure 2.15-Figure 2.16) represent the
solution of the AVUS software, whose units are in the SI. The IDS is built on the premise of reducing
numerical and modeling errors. As such, the IDS implements the dimensionless form of the Navier
Stokes equations, and therefore it reduces the round-off error.
73 Air Force Research Laboratory, CFD Research Branch, Wright-Patterson AFB, OH, Air Vehicles Unstructured
Solver (AVUS).
74 Ferguson F, Mendez J, Dodoo-Amoo D. Evaluating the hypersonic leading edge phenomena at high Reynolds
and Mach numbers. Recent Trends in Computational Science and Engineering. Rijeka, Croatia: IntechOpen;
2017.
53
flow imitates the conceptual flow-field presented in Figure 2.9. That is, freestream flow is first
processed by the 2D oblique shock, travels parallel to the wedge surface, is processed again by the
reflected shock and travels parallel to the isolator duct walls. Once more we observe the organized
nature of the 2D flow, which is supported by the constant property in the respective zones. The
development of the shock train within the isolator duct is also captured in these figures.
2.2.3.5.7 3D Simulations
3D computational simulations were also conducted on the scramjet forebody, inlet, and isolator
sections. Computational tools used were Fluent and AVUS. In the case of the 3D Euler computational
simulation, a single 3D stream tube, Figure 2.14, was exposed to a Mach 6 flow-field. A similar
process was implemented with AVUS and 2D slices of flow-field data are extracted and presented in
Figure 2.17. On examining these 2D slices of 3D data, it is observed that the stream tube is
processing the flow in an organized consistent manner that is aligned with its design.
75Ferguson F, Mendez J, Dodoo-Amoo D. Evaluating the hypersonic leading edge phenomena at high Reynolds
and Mach numbers. In: Recent Trends in Computational Science and Engineering. Rijeka, Croatia: IntechOpen;
2017.
54
Arguable, this is an Euler analysis, however it is worth pointing out that the stream tube 2-D design
process holds. This is further supported by Figure 2.18 which presents data on the Mach
distribution at the isolator exit.
Figure 2.21 2D-3D Geometric Construction from Prescribes Isolator Cross-Section and Aerodynamic
Inputs
2.2.3.5.10 Conclusion
In this chapter we explored an inverse design approach used in designing scramjet configurations.
The forebody, inlet and isolator sections formed the core focus of the chapter. Ideal oblique 2D shock
relations along with Billig’s isolator relations were first used to generate a centerline 2D geometry.
Streamline marching techniques coupled with Nonweiler’s caret wave rider theory were used to
geometrically construct 3D stream tubes. These 3D stream tubes were later used in the construction
of various star-shaped forebody-inlet-isolator sections. Initial 2D Euler and 2D viscous studies were
performed on the forebody-inlet-isolator sections and the results presented. Initial 3D Euler studies
were also conducted on a single 3D stream tube. All results presented demonstrated the uniform
nature of the flow-field within the stream tube, supporting the inverse design approach. A 3D viscous
analysis of the 3D stream tube is yet to be undertaken.
57
58
3 Airplane Design
3.1 Role of CFD in the Design Process
The actual use of CFD by Aerospace companies is a consequence of the trade-off between perceived
benefits and costs. While the benefits are widely recognized, computational costs cannot be allowed
to swamp the design process76. The need for rapid turnaround, including the setup time, is also
crucial. In current industrial practice, the design process can generally be divided into three phases:
1. Conceptual Design,
2. Preliminary Design,
3. Detailed Design.
76 Antony Jameson and, assisted by, Kui Ou, “Optimization Methods in Computational Fluid Dynamics”,
Aeronautics and Astronautics Department, Stanford University, Stanford, CA, USA.
77 See Previous.
78 Anthony A. Giunta, Steven F. Wojtkiewicz Jr. and Michael S. Eldred, “Overview Of Modern Design Of
79 Gaurav Kapoor, “Exploration Of A Computational Fluid Dynamics Integrated Design Methodology For Potential
Application To A Wind Turbine Blade”, Thesis Submitted to the Department of Aerospace Engineering, College
of Engineering, Embry-Riddle Aeronautical University, Daytona Beach. December 2014.
60
and x2 x3) , as well as the three factor interaction (x1 x2 x3). For the 23-1 fractional factorial indicated
by the solid dots in Figure 3.1 (b), the main effects are aliased (or biased) with the two factor
interactions. [Simpson, et al.]80
80 Timothy W. Simpson, Jesse Peplinski, Patrick N. Koch, and Janet K. Allen, “Meta models For Computer-Based
Engineering Design: Survey And Recommendation”, NASA-NTRS, 2018.
81 Gaurav Kapoor, “Exploration Of A Computational Fluid Dynamics Integrated Design Methodology For Potential
Application To A Wind Turbine Blade”, Thesis Submitted to the Department of Aerospace Engineering, College
of Engineering, Embry-Riddle Aeronautical University, Daytona Beach. December 2014.
61
associated with design of experiments. For further details, reader should consult [Giunta et al.]82.
Overall, Response Surface Methodology (RSM) explores the relationships between several input
variables and one or more response variables, as depicted by [Kapoor]83. The method was introduced
by [Box & Wilson]84. The main idea of response surface methodology is to use a sequence of designed
experiments to converge to an optimal response. Incorporating this routine in the context of design
optimization falls into the category of Surrogate or Response Surface Optimization (RSO). It has
emerged as an effective approach for the design of computationally expensive models such as those
found in aerospace systems, involving aerodynamics, structures, and propulsion.
For a new or a computationally expensive design, optimization based on an inexpensive surrogate,
such as Response Surface Model (also known as surrogate or approximation models). RSO helps in
the determination of an optimum design candidate, and also aids by providing insight into the
workings of the design. A response model not only provides the benefit of low cost for output
evaluations, it also helps revise the problem definition of a design task. Furthermore, it can
conveniently handle the existence of multiple desirable design points and offer quantitative
assessments of trade-offs as well as facilitate global sensitivity evaluations of the design variables.
Thus, the use of Response Surface Models (RSM) in optimization is becoming increasingly popular.
The RSM is not in itself an optimizer, but instead a helper tool for increasing the speed of
optimization. Instead of making direct calls to a computationally expensive numerical analysis code,
such as CFD, an optimization routine takes values from a cheap surrogate model, that is formulated
using a specific set of responses obtained from the numerical code. The popularity of such methods
has probably increased due to the development of approximation methods which are better able to
capture the nature of a multi-modal design space.
The main objective behind creating an RSM is to be able to predict the response of a system for an
operating point without actually performing a simulated analysis at that point. The response of the
system can then be predicted just by inputting the operating point values into the RSM and obtaining
the value of the response. The RSM basically takes the shape of a mathematical equation (𝐱),
essentially a quadratic polynomial, which takes the values of the design variables X as an input, and
returns an approximated value of the system response. Various optimization methodologies can then
be employed to optimize this computationally cheap response model in order to obtain the best
operating point.
82 Anthony A. Giunta, Steven F. Wojtkiewicz Jr. and Michael S. Eldred, “Overview Of Modern Design Of
Experiments Methods For Computational Simulations”, AIAA 2003-0649.
83 Gaurav Kapoor, “Exploration of a Computational Fluid Dynamics Integrated Design Methodology for Potential
Application to a Wind Turbine Blade”, Thesis Submitted to the Department of Aerospace Engineering, College
of Engineering, Embry-Riddle Aeronautical University, Daytona Beach, December 2014.
84 Box, G. E. P., and Wilson, K. B., (1951), “On the Experimental Attainment of Optimum Conditions,” Journal of
➢ Obtaining the specifications of the airplane, selecting the type and determining;
➢ Aerodynamic Considerations;
➢ Wing design;
➢ Optimization of wing loading and thrust loading;
➢ Fuselage design, preliminary design of tail surface and preliminary layout;
➢ Center of Gravity calculation;
➢ the geometric parameters for different surfaces;
➢ Preliminary weight estimation;
➢ Estimates of areas for horizontal and vertical tails;
➢ Engine Selection;
➢ Detail Structural design;
➢ Determination of airplane performance, stability, and structural integrity from flight tests.
The aerodynamic lines of the Boeing 777 were frozen, for example, when initial orders were
accepted, before the initiation of the detailed design of the structure. The starting point is an initial
CAD definition resulting from the conceptual design. The inner loop of aerodynamic analysis is
contained in an outer multi-disciplinary loop, which is in turn contained in a major design cycle
involving wind tunnel testing. In recent Boeing practice, three major design cycles, each requiring
about 4-6 months, have been used to finalize the wing design. Improvements in CFD, might allow
the elimination of a major cycle, would significantly shorten the overall design process and reduce
costs. Moreover, the improvements in the performance of the final design, which might be realized
through the systematic use of CFD, could have a crucial impact.
VL 1 W0 + Wf
Range = log
⏟ ⏟
D SFC ⏟ W0
Aero. Porp. Structure
Eq. 3.3
63
Here V is the speed, L/D is the lift to drag ratio, SFC is the specific fuel consumption of the engines,
W0 is the loading weight (empty weight + payload + fuel resourced), and Wf is the weight of fuel burnt.
Error! Reference source not found. displays the multidisciplinary nature of design. A light structure i
s needed to reduce W0. SFC is the province of the engine manufacturers. The aerodynamic designer
should try to maximize VL/D. This means the cruising speed V should be increased until the onset of
drag rise at a Mach Number M = V/C ∼ 0.85. But the designer must also consider the impact of shape
modifications in structure weight85. An excellent discussion of these and other factors in design of
airplane, is documented by [Tulapurkara]86 and readers encourage to consult the on-line material.
The plan view of supersonic airplanes indicates that the area of cross section of fuselage is decreased
in the region where wing is located. This is called area ruling. A brief note on this topic is presented
below. It was observed that the transonic wave drag of an airplane is reduced when the distribution
of the area of cross section of the airplane, in planes perpendicular to the flow direction, has a
smooth variation. In this context, it may be added that the area of cross section of the fuselage
generally varies smoothly. However, when the wing is encountered there is an abrupt change in the
cross sectional area. This abrupt change is alleviated by reduction in the area of cross section of
fuselage in the region where the wing is located. Such a fuselage shape is called ‘Coke-bottle shape’.
➢ Wing area (S) : This is calculated from the wing loading and gross weight which have been
already decided i.e. S = W/(W/S)
➢ Location of the wing on fuselage : High-, low- or mid-wing
➢ Airfoil : Thickness ratio, camber and shape
85 Antony Jameson, “Airplane Design with Aerodynamic Shape Optimization”, Aeronautics & Astronautics
Department, Stanford University, 2010.
86 E.G. Tulapurkara, ”Airplane design(Aerodynamic)”, Dept. of Aerospace Engineering., Indian Institute of
➢ Sweep : Whether swept forward, swept backward, angle of sweep, cranked wing, variable
sweep.
➢ Aspect ratio : High or low, winglets
➢ Taper ratio : Straight taper or variable taper.
➢ Twist: Amount and distribution
➢ Wing incidence or setting
➢ High lift devices : Type of flaps and slats; values of CLmax, Sflap/S
➢ Ailerons and spoilers : Values of Saileron/S; Sspoiler/S
➢ Leading edge strakes if any;
➢ Dihedral angle.
➢ Other aspects : Variable camber, planform tailoring, area ruling, braced;
➢ Wing, aerodynamic coupling (intentionally adding a coupling lifting surface like canard).
where yc and yt are the ordinates, at location x, of the camber line and the thickness distribution
respectively; tan θ is the slope of the camber line at location x (see also Figure 3.2-C & D). The
leading edge radius is also prescribed for the airfoil. The center of the leading edge radius is located
along the tangent to the mean line at the leading edge. Depending on the thickness distribution, the
trailing edge angle may be zero or have a finite value. In some cases, thickness may be non-zero at
the trailing edge. There are some attempts made by [Xiaoqiang et al. ]87 to decouple the camber from
the thickness so that camber and thickness could be constructed respectively with fewer parameters
for design purposes.
yt =
t
20
0.2969 x − 0.1260x − 0.3516x 2 + 0.2843x 3 − 0.1015x 5
Eq. 3.5
where, t = maximum thickness as fraction of chord. The leading radius is : rt = 1.1019 t2. Figure 3.2-
b shows the shape of NACA 0009 airfoil. It is a symmetrical airfoil by design. The maximum
thickness of all four-digit airfoils occurs at 30% of chord. In the designation of these airfoils, the first
two digits indicate that the camber is zero and the last two digits indicate the thickness ratio as
percentage of chord. The camber line for the four-digit series airfoils consists of two parabolic arcs
tangent at the point of maximum ordinate. The expressions for camber(yc) are :
yc =
m
p 2
2px - x 2 x x ycmax
Eq. 3.6
=
m
(1 - p)2
(1 - 2p) + 2px - x 2 x x ycmax
Where m = maximum ordinate of camber line as fraction of chord and p = chord wise position of
maximum camber as fraction of chord. The camber lines obtained by using different values of m & p
are denoted by two digits, e.g. NACA 64 indicates a mean line of 6% camber with maximum camber
occurring at 40% of the chord. A cambered airfoil of four-digit series is obtained by combining mean
87Lu Xiaoqiang, Huang Jun, Song Leia,, Li Jingb, “An improved geometric parameter airfoil parameterization
method”, Aerospace Science and Technology 78, 241–247, 2018.
66
line and the thickness distribution as described in the previous subsection. For example, NACA 2412
airfoil is obtained by combining NACA 24 mean line and NACA 0012 thickness distribution. This
airfoil has (a) maximum camber of 2% occurring at 40% chord and (b) maximum thickness ratio of
12%.
A&B
C&D
1
y c = k1 x 3 - 3mx 2 + m 2 (3 − m)x
6
0xm
Eq. 3.7
= k1m3 1 - x m x 1
1
6
The value of ‘m’ decides the location of the maximum camber and that of k1 the design lift coefficient.
A combination of m = 0.2025 and k1 = 15.957 gives li C = 0.3 and maximum camber at 15% of chord.
This mean line is designated as NACA 230. The first digit ‘2’ indicates that CL = 0.3 and the
subsequent two digits (30) indicate that the maximum camber occurs at 15% of chord. A typical
five-digit cambered airfoil is NACA 23012. The digits signify : First digit(2) indicates that li C L = 0.3.
Second & third digits (30) indicate that maximum camber occurs at 15% of chord. Last two digits
(12) indicate that the maximum thickness ratio is 12%.
➢ In the approach given by [Lebedinski], the variations, of the following quantities are
obtained when the wing loading is varied.
• (T/W) or (W/P) required for prescribed values of flight speed, absolute ceiling,
(R/C)max and output of a piston engine.
• Weight of the fuel (Wf) required for a given range.
• Distance required for landing.
From these variations, the wing loading which is optimum for each of these items is obtained.
However, the optimum values of W/S in various cases are likely to be different. The final wing
loading is chosen as a compromise.
➢ In the approach followed by [Raymer], (T/W) or (P/W) is chosen from statistical data
correlations and then W/S is obtained from the requirements regarding V max, (Range)max,
maximum based on rate of climb, absolute ceiling, maximum rate of turn, landing distance
and take-off distance.
Finally, W/S is chosen such that the design criteria are satisfied.
The most efficient way of transmitting the load is when the force is transmitted in an axial direction.
In the case of airplane the lift acts vertically upwards and the weights of various components and
the payload act vertically downwards. In this situation, the sizes and weights of structural members
are minimized or the structure is efficient if opposing forces are aligned with each other. This has
led to the flying wing or blended wing-body concept in which the structural weight is minimized as
the lift is produced by the wing and the entire weight of the airplane is also in the wing. However, in
a conventional airplane the payload and systems are in the fuselage. The wing produces the lift and
as a structural member it behaves like a beam. Hence to reduce the structural weight, the fuel tanks,
engines and landing gears are located on the wing, as they act as relieving load. Reduction in number
of cutouts and access holes, consistent with maintenance requirements, also reduces structural
weight.
69
➢ The engines,
➢ Parts of the airframe like control surfaces and high lift devices which significantly change
the airflow direction.
➢ Projections in airflow like landing gear and spoilers.
Considerable research has been carried out to reduce the engine noise. High by-pass ratio engines
with lobed nozzle have significantly lowered the noise level. Noise level inside the cabin has to be
minimal. This is achieved by suitable noise insulation. Further, the clearance between cabin and the
propeller should not be less than the half of the radius of the propeller.
3.2.9.2 Emissions
Combustion of the fuel in an engine produces carbon dioxide, water vapor, various oxides of
nitrogen (NOx), carbon monoxide, unburnt hydrocarbons and Sulphur dioxide (SO2). The
components other than carbon dioxide and water vapor are called pollutants. The thrust setting
changes during the flight and hence the emission levels have to be controlled during landing, take-
off and climb segment up to 3000 ft (1000 m). At high altitudes the NOx components may deplete
ozone layer. Hence, supersonic airplanes may not be allowed to fly above 50000 ft (15 km) altitude.
It may be noted that cruising altitude for Concorde was 18 km. Improvements in engine design have
significantly reduced the level of pollutants. The amount of pollution caused by air transport is
negligible as compared to that caused by road transport, energy generation and industry. However,
the aircraft industry has always been responsive to the ecological concerns and newer technologies
have emerged in the design of engine and airframe.
➢ The maximum rate of turn and the minimum radius of turn in steady level turn depend on
the thrust available, and the permissible load factor. The value of CLmax used here is that
without the flaps. For high speed airplanes the value of CLmax depends also on Mach number;
➢ Take - off run and take - off distance: During take-off an airplane accelerates on the ground.
For an airplane with nose wheel type of landing gear, around a speed of 85% of the take-off
speed, the pilot pulls the stick back. Then, the airplane attains the angle of attack
corresponding to take-off and the airplane leaves the ground. The point at which the main
wheels leave the ground is called the unstick point and the distance from the start of take-off
point to the unstick point is called the ground run. After the unstick, the airplane goes along
a curved path as lift is more than the weight. This phase of take-off is called transition at the
end of which the airplane climbs along a straight line. The take-off phase is said to be over
when the airplane attains screen height which is generally 15 m above the ground. The
horizontal distance from the start of the take off to the where the airplane attains screen
height is called take off distance. The takeoff run and the take-off distance can be estimated
by writing down equations of motion in different phases.
➢ Landing Distance: The landing flight begins when the airplane is at the screen height at a
velocity called the approach speed. During the approach phase the airplane descends along a
flight path of about 3 degrees. Subsequently the flight path becomes horizontal in the phase
called ‘flare’. In this phase the pilot also tries to touch the ground gently. The point where the
main wheels touch the ground is called touch down point. Subsequent to touch down, the
airplane rolls along the ground for about 3 seconds during which the nose wheel touches the
ground. This phase is called free roll. After this phase the brakes are applied and the airplane
comes to halt. In some airplanes, thrust in the reverse direction is produced by changing the
direction of jet exhaust or by reversible pitch propeller. In some airplanes, the drag is
increased by speed brakes, spoilers or parachutes. For airplanes which land on the deck of
the ship, an arresting gear is employed to reduce the landing distance. The horizontal distance
from the start of approach at screen height till the airplane comes to rest is called landing
distance.
1. Operating envelope:
The maximum speed and minimum speed can be calculated from the level flight analysis.
However, the attainment of maximum speed may be limited by other considerations. The
operating envelope for an airplane is the range of flight speeds permissible at different
altitudes. Typical operating envelope for a military airplane is shown in [Tulapurkara]88
where reader are encouraged for detailed view of subject.
2. Energy height technique for climb performance:
The analysis of a steady climb shows that the velocity corresponding to maximum rate of
climb increases with altitude. Consequently, climb with involves acceleration and the rate of
climb will actually be lower than that given by the steady climb analysis. This is because a
part of the engine output would be used to increase the kinetic energy. Secondly, the aim of
the climb is to start from velocity near and at and attain a velocity near at h. To take these
aspects into account, it is more convenient to work in terms of energy height (he) instead of
height(h). The quantity he is defined as :
V2 WV 2
h e = h + Multiplyby W → Wh e = Wh +
2g 2g
Eq. 3.8
The right hand side of the Eq. 3.8 is the sum of the potential energy and the kinetic energy of the
airplane. It is denoted by E. The energy height (he) which is E / W, is also called specific energy. It can
be shown that (dhe/ dt) = (TV – DV)/W and is referred to as specific excess power (Ps). Using energy
height concept the optimum climb path for fastest climb or economical climb can be worked out.
III) Range performance: For commercial airplanes the range performance is of paramount
importance. Hence, range performance with different amounts of payload and fuel on board the
airplane, needs to be worked out. In this context the following three limitations should to be
considered.
a) Maximum payload:
The number of seats and the size of the cargo compartment are limited. Hence maximum payload
capacity is limited.
b) Maximum fuel:
The size of the fuel tanks depends on the space in the wing and the fuselage to store the fuel. Hence,
there is limit on the maximum amount of fuel that can be carried by the airplane.
c) Maximum take-off weight:
The airplane structure is designed for a certain load factor and maximum take-off weight. This value
of weight cannot be exceed the limitations in mind a typical payload vs. range curve.
➢ In the approach given by [Lebedinski], the variations, of the following quantities are obtained
when the wing loading is varied.
• (T/W) or (W/P) required for prescribed values of Vp, Hmax (R/C)max and sto.
• Weight of the fuel (Wf) required for a given range (R).
• Distance required for landing.
• From these variations, the wing loading which is optimum for each of these items is
obtained. However, the optimum values of W/S in various cases are likely to be
different. The final wing loading is chosen as a compromise.
➢ In the approach followed by [Raymer], (T/W) or (P/W) is chosen from statistical data
correlations and then W/S is obtained from the requirements regarding Vmax, Rmax, (R/C)max,
Hmax, max ψ , landing distance and take-off distance.
73
Finally, W/S is chosen such that the design criteria are satisfied. These two approaches are described
in the subsequent sections.
3.2.11.1 Remarks on for choosing Wing Loading and Thrust Loading or Power Loading
It is felt that the approach presented by [Lebedinski] about 50 years ago, is still relevant. The main
features are:
• Derive simplified relations between the chosen performance parameter and the wing loading.
• Obtain the wing loading which satisfies/optimizes the chosen parameter e.g. landing
distance, thrust required for Vp, fuel required for range.
• Examine the influence of allowing small variations in wing loading from the optimum value
and obtain a band of wing loadings. This would give an estimate of the compromise involved
when (W/S) is non-optimum.
• After all important cases are examined, choose the final wing loading as the best compromise.
• With the chosen wing loading, obtain (T/W) or (W/P) which satisfy requirements of V max,
(R/C)max, ceiling (Hmax), take-off field length ( to s ) and maximum turn rate (ψmax). If the
requirements of engine output in these cases are widely different, then examine possible
compromise in specification. After deciding the (T/W) or (W/P) obtain the engine output
required. Choose the number of engine(s) and arrive at the rating per engine. Finally choose
an engine from the engines available from different engine manufacturers.
During the process of optimizing the wing loading, a reasonable assumption is to ignore the changes
in weight of the airplane (W0). However, when W0 is constant but W/S changes, the wing area and in
turn, the drag polar would change. This is taken into account by an alternate representation of the
drag polar.
• It may be added that in actual practice the airplane does not halt on the runway. After
reaching a sufficiently low speed the pilot takes the airplane to the allotted parking place.
• Landing ground run is the distance the airplane covers from the point the wheels first touch
the ground to the point the airplane comes to a stop.
• VA = 1.3(Vs) land, VTD = 1.15(Vs) land (4.1) (Vs)land is the stalling speed in landing
configuration. Exact estimation of landing distance (sland) is difficult as some phases like flare
depend on the piloting technique. based on consideration of landing distance.
W
(W/S) land = p land T0 Eq. 3.9
Wland
74
The weight of the airplane at the time of landing (Wland) is generally lower than WTO. The difference
between the two weights is due to the consumption of fuel and dropping of any disposable weight.
However, to calculate Wland only a part of the fuel weight is subtracted, from the takeoff weight.
Figure 3.5 Pressure Distribution for wing-pylon-nacelle Configuration; (Initial left), (refined right)
According to [EMBRAER], a
civil transport aviation
company out of Brazil, three
major different aerodynamic
configurations were
extensively studied during the
development phase; straight
wing with over wing mounted
engines (Figure 3.4), swept
wing with underwing
mounted engines and swept
wing with rear fuselage
mounted engines. underwing
engine configuration to be Figure 3.4 Over Wing Mounted Engines Configuration
abandoned (Figure 3.6. and
the third major aerodynamic configuration had the engines mounted on pylons on the rear fuselage
(Figure 3.8).
time (full potential 2D airfoil code with coupled boundary layer, 3D inviscid wing full potential code
and a 3D panel method) were not capable of calculating the unfavorable aerodynamic interference
between the jet exhaust and the supersonic flow on the wing upper surface. Although there was prior
qualitative knowledge of the phenomenon and the associated risks, it took a transonic wind tunnel
test.
The third major aerodynamic configuration had the engines mounted on pylons on the rear fuselage
(see Figure 3.8). There was a further increase in fuselage length to accommodate 50 passengers and
the wing was initially the same as that of the underwing configuration. This configuration, with the
changes described below, was the one finally chosen for production. Although good transonic wind
tunnel results had been
obtained for the cruise wing at
the Boeing Transonic Wind
tunnel, low speed wind tunnel
tests at CTA indicated that the
maximum lift coefficient values
would not meet the short take-
off and landing field lengths
required for regional airline
operations. At about the same
time, market surveys indicated
that the potential clients would
not require cruise speeds in
excess of Mach 0.75 to 0.78. This Figure 3.8 Rear Fuselage Mounted Engines Configuration
provided design margins to
allow the leading edge to be
modified with a fixed 'droop' and the wing root flap chord to be extended by 0.15 m. The droop was
designed using the 2D and 3D full potential methods. Additionally, four vortilons were installed on
the lower surface leading edge of the outboard wing panel. During the initial flight test campaign,
some adverse yaw (aileron roll command to the left would produce a slight yawing moment to the
right and vice-versa) was noticed
during climb. The ailerons already
possessed differential gearing (the
aileron whose trailing edge is going
up always deflects more than the
one whose trailing edge is going
down) to counter the theoretically
predicted adverse yaw, but the Figure 3.7 Leading Edge Droop and Vortilons
effect in flight was found to be larger
than expected. Flow visualizations with wool tufts showed that the aileron going down had some
regions of separated flow. This produced additional drag at that wingtip, which in turn produced the
increased adverse yaw. The problem was solved by placing a row of vortex generators in front of the
aileron to 'energize' the boundary layer and delay its separation. (See Figure 3.7).
91Antony Jameson, “Optimum Aerodynamic Design Using CFD and Control Theory”, Department of Mechanical
and Aerospace Engineering, Princeton University, AIAA 95-1729-CP.
78
Then a cost function (I) is selected which might, for example, be the drag coefficient or the lift to drag
ratio, and I is regarded as a function of the parameters αi. The sensitivities I may now be estimated
by making a small variation Sai in each design parameter in turn and recalculating the flow to obtain
the change in I. An alternative approach is to cast the design problem as a search for the shape that
will generate the desired pressure distribution. This approach recognizes that the designer usually
has an idea of the kind of pressure distribution that will lead to the desired performance. Thus, it is
useful to consider the inverse problem of calculating the shape that will lead to a given pressure
distribution. The method has the advantage that only one flow solution is required to obtain the
desired design. Unfortunately, a physically realizable shape may not necessarily exist, unless the
pressure distribution satisfies certain constraints. Thus the problem must be very carefully
formulated. The shape changes in the section needed to improve the transonic wing (shock free)
design are quite small. However, in order to obtain a true optimum design larger scale changes such
as changes in the wing planform (sweepback, span, chord, and taper) should be considered. Because
these directly affect the structure weight, a meaningful result can only be obtained by considering a
cost function that takes account of both the aerodynamic characteristics and the weight. Consider a
cost function (I) is defined as
1
I = α1CD + α 2
2B (p − pd ) 2dS + α3C W Eq. 3.11
where pd is the target pressure and the integral is evaluated over the actual surface area (S).
Antony Jameson, Kasidit Leoviriyakit and Sriram Shankaran, “Multi-point Aero-Structural Optimization of
92
Wings Including Planform Variations”, 45th Aerospace Sciences Meeting and Exhibit, January 8–11, 2007, USA.
79
research during the past decade) and ultimately multidisciplinary system optimization. Figure 3.10
illustrates the result of an automatic redesign of the wing of the Boeing 747, which indicates the
potential for a 5 percent reduction in the total drag of the aircraft by a very small shape modification.
It is also important to recognize that in current practice the setup times and costs of CFD simulations
substantially exceed the solution times and costs. With presently available software the processes of
geometry modeling and grid generation may take weeks or even months. In the preliminary design
of the F22 Lockheed relied largely on wind-tunnel testing because they could build models faster
than they could generate meshes. It is essential to remove this bottleneck if CFD is to be more
effectively used. There have been major efforts in Europe to develop an integrated software
environment for aerodynamic simulations, exemplified by the German “Mega Flow” program.
Figure 3.10 also displays Redesigned Boeing 747 wing at Mach 0.86 with Cp distributions. In the
final-design stage it is necessary to predict the loads throughout the flight envelope. As many as
20000 design points may be considered. In current practice wind-tunnel testing is used to acquire
the loads data, both because the cumulative cost of acquisition via CFD still exceeds the costs of
building and testing properly instrumented models, and because a lack of confidence in the reliability
of CFD simulations of extreme flight conditions93.
Figure 3.10 Redesigned Boeing 747 Wing at Mach 0.86 based on Cp Distributions
93Antony Jameson and, assisted by, Kui Ou, “Optimization Methods in Computational Fluid Dynamics”,
Aeronautics and Astronautics Department, Stanford University, Stanford, CA, USA.
80
94M. B. Giles, “Some thoughts on exploiting CFD for turbomachinery design”, Oxford University Computing
Laboratory, 1998.
81
In the future, there may be a shift to a more tightly-coupled two-level design system, as illustrated in
Figure 3.11 (right). The overall system design will begin, as now, with a preliminary design based
on past empiricism. This will provide the starting point for the detailed component design. The main
reason a tightly coupled design system is not used today is time. The design time for an engine or
aircraft project is strictly limited.
95 Helmut Sobieczky, “Geometry Generator for CFD and Applied Aerodynamics”, Geometry Generator for CFD and
Applied Aerodynamics.
82
distributions for specified operation conditions and numerical results are found for airfoil and wing
shapes. The status of these methods is reviewed in the next book chapter. Given the designer’s
experience in aerodynamics for selecting suitable pressure distributions, choice of a few basic
functions and parameters may provide a dense set of data just like geometry coordinates are
prescribed, the amount of needed parameters for typical attractive pressure distributions about the
same as for the direct airfoil modelling.
96 Flap deflection as a function of angle of attack variations, for constant lift. Airfoil in transonic flow, M ∞ = 0.75,
Re = 4 x 106, CL = 0.7, (MSES analysis).
83
every phase of start and landing configurations. Figure 3.13 illustrates a multicomponent high lift
system in 2D and 3D.
Figure 3.13 Wing Sections for Multi-Component High Lift System, 3D Swept Wing with Slat
and Flaps - Courtesy of Sobieczky
Figure 3.14 Wing Parameters and Respective key numbers for Section Distribution,
Planform, an/Dihedral, Twist, Thickness Distribution and Airfoil Blending - Courtesy of Sobieczky
3.6.5 Axially Defined Components (Fuselage Bodies, Nacelles, Propulsion and Tunnel Geometries)
This group of shapes is basically aligned with the main flow direction, the usual development is
directed toward creating volume for payload, propulsion or, in internal aerodynamics (and
hydrodynamics) the development of channel and pipe geometries. The parameters of cross sections
are quite different to those of airfoils; the quality of their change along a main axis with constraints
for given areas within the usually symmetrical contour is the design challenge. Fuselages are
therefore described by another set of “keys” which is defined along the axis. This axis may be a curve
in 3D space, with available gradients providing cross section planes normal to the axis. For simple
straight axes in the cartesian x-direction key 40 defines axial stations just like key 20 defines
spanwise x-stations. With the simplest cross section consisting of super elliptic quarters allowing a
choice of the half axes or crown lines and body planform, plus the exponents ( with the value of 2. for
ellipses), 8 parameters (key 41 - 48) are given (Figure 3.15).
Basic bodies are described easily this way, with either explicitly calculating the horizontal coordinate
85
y(x, z) for given vertical coordinate z, or the vertical upper and lower coordinate z(x, y) for given
points y within the planform, at each cross section station x = const. More complex bodies are defined
by optional other shape definition subprograms with additional keys (49 - 59) needed for geometric
details. These may be of various kind but of paramount interest is the aerodynamically optimized
shape definition of wing-body junctures. In the following a simple projection technique is applied
requiring only a suitable wing root geometry to be shifted toward the body, but more complex
junctures require also body surface details to suitably meet the wing geometry.
Figure 3.15 Fuselage Parameters and Respective Key Numbers for Cross Section Definition,
Planform and Crown Lines, Super Elliptic Exponents - Courtesy of Sobieczky -
components F1 and F2 with the condition that for the first component one coordinate (here the
spanwise y) needs to be defined by an explicit function y = F1(x , z), while the other component F2
may be given as a dataset for a number of surface points. Using a blending function for a portion of
the spanwise coordinate, all surface points of F2 within this spanwise interval may be moved toward
the surface F1 depending on the local value of the blend key function. Figure 3.16 shows that this
way the wing root (F2) emanates from the body (F1), wing root fillet geometry can be designed as
part of the wing prior to this wrapping process. Several refinements to this simple projection
technique have been implemented to the program.
97 Cosentino, G. B., Holst, T. L. Numerical Optimization Design of Advanced Transonic Wing Configurations.
Journal of Aircraft, Vol. 23, pp. 192-199, 1986
98 Zhu, Y., Sobieczky, H. Numerical Optimization Method for Transonic Wing Design. Proc. 6th Asian Congress of
rotor section with a sealed slat periodically drooped nose have been carried out and the results of
unsteady Navier Stokes analysis suggest a concept for dynamic stall control 99.
3.6.11 Applications
Case studies for new generation supersonic transport aircraft have been carried out through the past
years in research institutions and in the aircraft industry. Our present tool to shape such
configurations needs to be tested by trying to model the basic features of various investigated
geometries. Knowing that the fine-tuning of aerodynamic performance must be done by careful
selection of wing sections, wing twist distribution and the use of sealed slats and flaps, with initial
exercises we try to geometrically model some of the published configurations, generate CFD grids
around them and use optimization strategies to determine the sensitivity of suitable geometry
parameters. This is still a difficult task but tackling its solution greatly contributes to building up the
knowledge base of high speed design.
3.6.12 Case Study - Generic High Speed Civil Transport (HSCT) Configuration
Figure 3.17 illustrate data visualization of a generated configuration derived from a Boeing HSCT
design case for Mach 2.4 100. The configuration consists of 10 components, engine pylons are not yet
included. Wing and horizontal and vertical tail components are spanwise defined, fuselage and
engines are axially defined components. The wing has a subsonic leading edge in the inner portion
and a supersonic leading edge on the outer portion. For this study a minimum of support airfoils
(Figure 3.13) is used to get a reasonable pressure distribution a rounded leading edge section in
most of the inner wing and a wedge-sharp section in the outer wing portion define the basic shape of
the wing. Wing root fillet blending, the smooth transition between rounded and sharp leading edge
and the tip geometry are effectively shaped by the previously illustrated wing keys, the fuselage here
is a simple slender body requiring just the baseline body tool with elliptical cross sections.
Preprocessing input data for CFD requires providing a grid surrounding the configuration. For
application of either structured or unstructured grids additional geometric shapes need to be
provided. In the case of the generic HSCT with given supersonic flight Mach number the far-field
boundary is chosen to engulf the expected bow shock wave (Figure 3.17-a) and a cross sectional
grid for both wing and body is generated, either as simple algebraic trajectories or using elliptic
Figure 3.17 CFD Grid Boundaries Result for Euler Analysis of HSCT Wing-Body in
Supersonic flow M∞ = 2.4 - Courtesy of Sobieczky
99Geissler, W., Sobieczky, H. Unsteady Flow Control on Rotor Airfoils. AIAA 95-1890, 1995
100 Kulfan, R. High Speed CivilTransport Opportunities, Challenges and Technology Needs. Lecture at Taiwan IAA
34th National Conference, 1992.
88
equations. Short runs using an the inviscid flow Euler option of a flow solver101 were carried out on
a coarse (33 x 81 x 330) grid, here only to get an idea about the needed wing section and twist
modifications for acceptable pressure distributions. Visualization of the results in various grid
surfaces is needed, like pressure distributions in cross section planes as shown in Figure 3.17-b.
Kroll, N., Radespiel, R. An Improved Flux Vector Split Discretization Scheme for Viscous Flows. DLR-FB 93-53,
101
1993.
89
4 Optimization Problem
4.1 Mathematical Optimization
In mathematics and computer science, an optimization problem is the problem of finding the best
solution from all feasible solutions. In the simplest terms, an optimization problem consists of
maximizing or minimizing a real function by systematically choosing input values from within an
allowed set and computing the value of the
function. Figure 4.1 shows a graph of a
paraboloid given by z = f(x, y) = − (x² + y²) + 4.
The global maximum at (x, y, z) = (0, 0, 4) is
indicated by a blue dot. Optimization is the
process of obtaining the most suitable solution
to a given problem, while for a specific problem
only a single solution may exist, and for other
problems there may exist multiple potential
solutions [Skinner and Zare-Behtash]102. Thus,
optimization is the process of finding the `best'
solution, where `best' implies that the solution
is not the exact solution but is sufficiently
superior. Optimization tools should be used
for supporting decisions rather than for Figure 4.1 Global Maximum of f (x, y)
making decisions, i.e. should not substitute
decision-making process103. Another example
of numerical optimization is defined as
102 S. N. Skinner and H. Zare-Behtash, ”State-of-the-Art in Aerodynamic Shape Optimization Methods”, Article in
Applied Soft Computing , September 2017, DOI: 10.1016/j.asoc.2017.09.030.
103 103 Dragan Savic,” Single-objective vs. Multi-objective Optimization for Integrated Decision Support”, Centre
for Water Systems, Department of Engineering School of Engineering and Computer Science, University, UK,
104 Joaquim R. R. A. Martins, “Multidisciplinary Design Optimization”, 7th International Fab Lab Forum and
Symposium on Digital Fabrication Lima, Peru, August 18, 2011, (Remote presentation).
105 L. J. Kedward , A. D. J. Payot y, T. C. S. Rendall, C. B. Allen, “Efficient Multi-Resolution Approaches for
106J. Sobieszczanski-Sobieski and R.T. Haftka. “Multidisciplinary Aerospace Design Optimization: Survey of
Recent Developments”. In 34th Aerospace Sciences Meeting and Exhibit, AIAA 96-0711, volume 70, Reno,
Navada, 1996.
91
variables. Constrained optimization problems can be furthered classified according to the nature of
the constraints (e.g., linear, nonlinear, convex) and the smoothness of the functions (e.g.,
differentiable or no differentiable). Gradient-based algorithms typically use an iterative two-step
method to reach the optimum as described by [Venter]107 (2010). The first step is to use gradient
information to find the search direction and the second step is to move in that direction until no
further progress can be made or until a new constraint is reached. The second step is known as the
line search and provides the optimum step size. The two-step process is repeated until the optimum
is found, see Figure 4.3 108. Depending on the scenario, different search directions are required.
107 Venter, G. “Review of optimization techniques. In R. Blockley, and W. Shyy (Eds.), Encyclopedia of
aerospace engineering (Vol. 8: System Engineering). Chichester, West Sussex, UK: John Wiley and Sons.
108 Schematic picture of a gradient-based optimization algorithm for the case with two design variables. The
response values are indicated by the iso-curves and the star represents the optimum solution for a)
unconstrained optimization and b) constrained optimization. The unfeasible region violating the constraints is
marked by the shaded areas.
92
or choosing a portfolio might involve maximizing the expected return while minimizing the risk. In
practice, problems with multiple objectives often are reformulated as single objective problems by
either forming a weighted combination of the different objectives or by replacing some of the
objectives by constraints.
Dragan Savic,” Single-objective vs. Multi-objective Optimization for Integrated Decision Support”, Centre for
109
Water Systems, Department of Engineering School of Engineering and Computer Science, University of Exeter,
United Kingdom,
93
Local Gradient
Direct
Numerical Stochastic
Methods
Methods and
Otimization
Randum Search
DoE
Complex/Simplex
Hybrid Methods
Single-objective optimization identifies a single optimal alternative which can be used within
the multi-objective framework. This does not involve accumulating different objectives into a single
objective function, but entails setting all except one of them as constraints in the optimization
process. However, most design and planning problems are characterized by a large and often infinite
number of alternatives. Thus, multi-objective methodologies are more likely to identify a wider range
of these alternatives since they do not need to specify for which level of one objective a single optimal
solution is obtained for another. Figure 4.4 illustrates different optimization techniques and
searches. Be advised that in Direct search methods perform hill climbing in the function space by
moving in a direction related to the local gradient. Where else in indirect methods, the solution is
sought by solving a set of equations resulting from setting the gradient of the objective function to
zero110. A more orthodox figure could be devised as appears in [Andersson]111, where optimization
methods could be divided into derivative and non-derivative methods.
110 Dragan Savic, ”Single-objective vs. Multi-objective Optimization for Integrated Decision Support”, Centre for
Water Systems University of Exeter, UK,
111 Johan Andersson, “A survey of multi objective optimization in engineering design”, Technical Report: LiTH-
IKP-R-1097.
94
dominated sorting GA (NSGA), and positioned Pareto GA. Note that all these techniques are
essentially non-exclusive in nature. Simulated annealing (SA) performs reasonably well in solving
single-objective optimization problems. However, its application for solving multi-objective
problems has been limited, mainly because it finds a single solution in a single run instead of a set of
solutions. This appears to be a critical bottleneck in multi-objective optimization. However, SA has
been found to have some favorable characteristics for multimodal search. The advantage of SA stems
from its good selection technique112.
112 Bandyopadhyay, S., Saha, S, “Some Single- and Multi-objective Optimization Techniques”, Chapter 2”, ISBN:
973-3-642-35450-5, 2013.
113 Wikipedia.
95
Hill climbing finds optimal solutions for convex problems – for other problems it will find only local
optima (solutions that cannot be improved upon by any neighboring configurations), which are not
necessarily the best possible solution (the global optimum) out of all possible solutions (the search
space). Examples of algorithms that solve convex problems by hill-climbing include the simplex
algorithm for linear programming and binary search.[1]:253 To attempt to avoid getting stuck in local
optima, one could use restarts (i.e. repeated local search), or more complex schemes based on
iterations (like iterated local search), or on memory (like reactive search optimization and tabu
search), or on memory-less stochastic modifications (like simulated annealing).
The relative simplicity of the algorithm makes it a popular first choice amongst optimizing
algorithms. It is used widely in artificial intelligence, for reaching a goal state from a starting node.
Different choices for next nodes and starting nodes are used in related algorithms. Although more
advanced algorithms such as simulated annealing or tabu search may give better results, in some
situations hill climbing works just as well. Hill climbing can often produce a better result than other
algorithms when the amount of time available to perform a search is limited, such as with real-time
systems, so long as a small number of increments typically converges on a good solution (the optimal
solution or a close approximation). At the other extreme, bubble sort can be viewed as a hill climbing
algorithm (every adjacent element exchange decreases the number of disordered element pairs), yet
this approach is far from efficient for even modest N, as the number of exchanges required grows
quadratically114.
1. Local maximum : It is a state which is better than its neighboring state however there exists
a state which is better than it(global maximum). This state is better because here value of
objective function is higher than its neighbors.
2. Global maximum : It is the best possible state in the state space diagram. This because at this
state, objective function has highest value.
3. Plateua/flat local maximum : It is a flat region of state space where neighboring states have
the same value.
114 Wikipedia.
115 From GeeksforGeeks.
96
4. Ridge : It is region which is higher than its neighbours but itself has a slope. It is a special kind
of local maximum.
5. Current state : The region of state space diagram where we are currently present during the
search.
6. Shoulder : It is a plateau that has an uphill edge.
➢ Local maximum : At a local maximum all neighboring states have a values which is worse
than the current state. Since hill climbing uses greedy approach, it will not move to the worse
state and terminate itself. The process will end even though a better solution may exist.
➢ To overcome local maximum problem : Utilize backtracking technique. Maintain a list of
visited states. If the search reaches an undesirable state, it can backtrack to the previous
configuration and explore a new path.
➢ Plateau : On plateau all neighbors have same value . Hence, it is not possible to select the best
direction.
➢ To overcome plateaus : Make a big jump. Randomly select a state far away from current state.
Chances are that we will land at a non-plateau region
➢ Ridge : Any point on a ridge can look like peak because movement in all possible directions
is downward. Hence the algorithm stops when it reaches this state.
➢ To overcome Ridge : In this kind of obstacle, use two or more rules before testing. It implies
moving in several directions at once116.
In continuous adjoint method, the nonlinear flow equations in a partial differential equation form
are linearized with respect to a design variable. Then an adjoint system will be derived from the
linearized flow equations, followed by discretization. In the discrete adjoint method, the flow
equations in a partial differential equation form are discretized first, followed by the linearization
and adjoint formulation. The discrete adjoint method can produce the exact gradient of an objective
function with respect to the variables in a discretized flow system, which will ensure that a design
process converges quickly and fully. However it is very difficult to develop discrete adjoint codes,
particularly by hand and with high-order upwind schemes and sophisticated turbulence models
117 Cramer, E. J., Dennis Jr., J. E., Frank, P. D., Lewis, R. M., and Shubin, G. R.. “Problem formulation for
multidisciplinary optimization”. SIAM Journal on Optimization, 4(4), 754-776, 1994.
118 Kodiyalam, S., and Sobieszczanski-Sobieski, J. “Multidisciplinary design optimization – some formal methods,
framework requirements, and application to vehicle design”. International Journal of Vehicle Design, 2001.
119 Sobieszczanski-Sobieski, J. ”Optimization by decomposition: a step from hierarchic to non-hierarchic systems”.
2nd NASA/Air Force Symposium on Recent Advances in Multidisciplinary Analysis and Optimization. Hampton,
Virginia, USA, 1998.
98
I = I(U, α)
Eq. 4.3
Then, the relationship between the flow variable and the design variable is determined through the
flow equation,
R(U, α) = 0
Eq. 4.4
The gradient of the objective function relative to the design variable is
dI ∂I ∂I ∂U
= +
dα ∂α ∂U ∂α
Eq. 4.5
where ∂I/∂α and ∂I/∂U are to be calculated. However, the calculation of the flow variable sensitivity,
∂U/∂α, involves solving the flow equations,
dR ∂R ∂U
=
dα ∂U ∂α
Eq. 4.6
This linearized equation also depends on the design variable, which means that each design variable
requires one solution of the flow equation. The key is to find a way to decouple the influence of the
design variables on the objective function by means of the flow sensitivity, i.e. eliminating the explicit
120 D. X. Wang, L. He, “Adjoint Aerodynamic Design Optimization for Blades in Multistage Turbomachines—Part
I: Methodology and Verification”, Journal of Turbomachinery · April 2010.
99
dependency of the objective function sensitivity on the flow variable sensitivity, ∂U/∂α. To achieve
this goal, the right side of the flow equation is multiplied by the adjoint factor, λ. (For a complete
derivation of AV method see section 4.7).
121 Zhihui Li, Xinqian Zheng, “Review of design optimization methods for turbomachinery aerodynamics”,
Progress in Aerospace Sciences, July 2017.
122 W.C. Carpenter, “Effect of Design Selection on Response Surface Performance”, (NASA - CR 4520), 1993.
123 R. Unal, R.A. Lepsch, M.L. McMillin, “Response surface model building and multidisciplinary optimization using
D-optimal designs”, 7th AIAA/USAF/NASA/ISSMO Sym. on Multidisciplinary Analysis & Optimization, 1998.
124 T.J. Mitchell, “An algorithm for the construction of D-optimal” experimental designs”, Technimetrics, 1974.
100
ym (𝐱) = yt (𝐱) + ε
Eq. 4.7
where ym is the measured response, yt is the true
response, and ε is a random error term. In many cases,
the ε values are assumed to be independent and
identically distributed. For ease of explanation in this
text, consider ε to be a normal (i.e., Gaussian) random
variable with a mean value of zero and a variance of
unity. Because of the random error term, ym is non-
repeatable even when exactly the same values of x are
used in two measurements. One of the goals of a typical
design of experiments study is to estimate and predict
the trends in the response data. While there are many
forms for the approximation functions, a generic
approximation model is used here with the form:
Figure 4.7 An illustration of the effect of
random errors in producing an estimated
ŷ(𝐱) = f(𝐱𝑆 , 𝐲𝒎 (𝐱𝑆 )) linear model (dashed) that has a different
Eq. 4.8 slope than the true linear model (solid).
where the sample points are denoted as xs, the measured
response data are represented as , f is a user-selected function, and is the approximate response value
computed at an arbitrary design point, x. Typical functions for f include low-order polynomials,
splines, kriging, neural networks, and radial basis functions. The key feature of Eq. 4.7 is the random
error term and the accompanying assumption that ε is always present due to sources such as
measurement error, inherent fluctuations in the response quantity (e.g., turbulence), or other
sources. Another critical assumption is that the experimenter has knowledge of the general trends of
the true response, yt. Based on these assumptions, the goal of most classical DoE methods is to place
a fixed number of samples in the de-sign space so as to minimize the influence of the random error
term in subsequent computations, e.g., where the approximation model given in Eq. 4.8 is computed.
In classical DOE, the goal of minimizing the effects of random error has the effects of placing the
sample sites near or on the boundaries of the design space. Myers and Montgomery1 provide a
detailed description of why this occurs, but the general idea can be observed in Figure 4.7 and
Figure 4.8. (see further info please refer to [Giunta et al.]126.
125Myers, R. H., and Montgomery, D. C., “Response Surface Methodology: Process and Product Optimization
Using Designed Experiments”, John Wiley & Sons, Inc., New York, NY, 1995.
126 Anthony A. Giunta, Steven F. Wojtkiewicz Jr, and Michael S. Eldred, “Overview Of Modern Design Of
In Figure 4.7, there are two sample points x1 and x2. The measured response value for each point is
indicated by the diamond symbol. The true trend is shown by the solid line, and the estimated trend,
ŷ(x), is shown by the dotted line. In this illustration, the random errors, ε1 and ε2, cause the estimated
trend to be a poor approximation of the true trend.
Figure 4.8 illustrates the effect of moving sample sites
x1 and x2 to the lower and upper bounds of x. In this case,
ε1 and ε2, remain the same, but the resulting estimated
trend is a better approximation to the true trend. From
a statistical perspective, it is easier to discriminate the
response trend component from the random error
component when the sample sites are spaced as far
apart as possible. This principle is followed in classical
DoE methods that place samples on the boundaries
and/or vertices of the de-sign space, but place very few
samples in the interior of the design space. Another
feature of classical DoE that differs from modern DoE is
the use of replicated sampling. Since the measured
response, ym, contains a random error term, repeated
measurements taken for identical design variable
Figure 4.8 By moving the samples to the
values will result in slightly different ym values. Classical boundaries of the design space, the effect
DoE methods typically employ replicated sampling to of the random error terms is reduced.
permit a lack-of-fit statistical analysis and to measure Now, the estimated linear model (dashed)
the magnitude of the error term in Eq. 4.7. is a better approximation to the true
linear model (solid).
4.4.1.2 Classical DoE Example
Figure 4.9 illustrates a commonly used classical DOE technique, central composite design (CCD), for
the design space [0,1]2. In CCD, the number of samples grows with the dimension of the design space,
n, according to the formula 2n + 2n + 1. The 2n samples
correspond to the “corner” points at the vertices of
[0,1]2, while the 2n samples correspond to the points
that lie outside of [0,1]2 (where the distance of these
points from the center of the design space changes
with respect to n. If this CCD were to be used on a
laboratory experiment, replicated samples would be
taken, at least, at the center of the design space, and
at all sample sites if economically possible. Figure
4.9 clearly illustrates some of the drawbacks to
classical DOE. That is, at best, the number of samples
in CCD scales as 2n; a rate that can be unacceptable if
n is large and/or if experiments are expensive.
Furthermore, CCD and other classical DoE methods
tend to place samples on or near the boundary of the
design space, leaving the interior of the design space Figure 4.9 A central composite design from
largely unexplored. For example, in the two- classical DOE for n=2. Note that the number
dimensional CCD shown in Figure 4.9, eight of the of sample sites (stars) scales as the number
nine samples are on or outside the boundary of the of vertices, i.e., as 2n
design space, and only one sample, the center point,
lies in the interior of the design space. Similar trends are exhibited by Box-Behnken Design (BBD)
and many other classical DoE methods.
102
127 Myers, R. H., and Montgomery, D. C., “Response Surface Methodology: Process and Product Optimization
Using Designed Experiments”, John Wiley & Sons, Inc., New York, NY, 1995.
128 Metropolis, N., and Ulam, S., “The Monte Carlo Method,” Journal of the American Statistical Association, Vol.
by enforcing simple geometric properties (e.g., to produce MC samples in a circular region, inscribe
the circle in a square, sample over the square and discard samples that fall outside of the circle), or
by applying some type of transformation that maps the boundary of the rectangular design space to
the boundary of the convex design space. For nonconvex design spaces, MC sampling can be more
difficult to employ, especially in high-dimensional design spaces, depending on how the nonconvex
region is defined. A nontrivial aspect of MC sampling is the selection of a reliable algorithm to
generate random numbers. This topic is covered in numerous texts on numerical methods and
statistics and is not addressed here. While MC sampling is simple to implement, a set of MC samples
will often leave large regions of the design space un-explored. This occurrence stems from the
random and independent nature of the sample sites produced by a random number generator.
Several modern DOE methods have been developed to address this deficiency of basic MC sampling.
Some of these are variants are described below.
129 Anthony A. Giunta, Steven F. Wojtkiewicz Jr, and Michael S. Eldred, “Overview Of Modern Design Of
Experiments Methods For Computational Simulations”, AIAA 2003-0649.
130 M.D. McKay, R.J. Beckman, W.J. Conover, “A comparison of three methods for selecting values of input variables
with the range of each input design variable divided into n intervals with each observation on the
input variable made in each interval using random sampling. Thus, there are n observations for each
of the d input variables. One of the observations on variable X1 is randomly selected (each
observation can be selected equally), matched with a randomly selected observation on X2, and so on
through Xd to build a design vector, X1. One of the remaining observations on X1 is then matched at
random with one of the remaining observation on X2 and so on to get X2. The procedure is followed
for X3, X4, …, Xn, resulting in n LHD
sampling points.
In real designs, some combinations of
the variables are not feasible or can
even crash the CFD code. However,
LHD allows the flexibility to adjust a
variable some without undermining
the fundamental properties of the
LHD sample. Various optimal LHDs
have been proposed based on
minimax, minimum mean square
error, maximum entropy or
orthogonal algorithms. The LHD Figure 4.12 Comparison of the Full Factorial and Latin
sample size can be controlled by the Hypercube Data Points (Courtesy of [Li & Zheng])
designer based on their time, budget
or other limitations. There is no comprehensive theory about the number of design points required
to construct the response surface models with LHD. Figure 4.12 compares the Full Factorial method
using 33 variable combinations with the LHD sampling method using only 3 samples. LHD methods
have been frequently coupled with surrogate models for optimization of turbomachinery designs131-
132.
131 W. Yi, H. Huang, W. Han, “Design optimization of transonic compressor rotor using CFD and genetic
algorithm”,
ASME Turbo Expo, (ASME Paper GT 2006–90155), 2006.
132 D. Pasquale, P. Giacomo, R. Stefano, “Optimization of turbomachinery flow surfaces applying a CFD-based
the searching efficiency. Kriging and radial basis function are preferred to conventional quadratic
models for robust design optimization because of a superior ability to simulate non-linear responses.
Kriging is known as Gaussian process modelling consisting of a regression model and a stochastic
process [Sacks et al..]135.
∆E
(− )
P(∆E) = e T
Eq. 4.9
P is the Boltzmann probability of accepting the new design and T is the current “temperature” of the
system. A random number in the range {0..1} is generated, and if it is less than P the new design
replaces the current design. The temperature is simply a control parameter in the same units as the
objective function. As the annealing processes proceeds the temperature is decreased according to a
cooling scheme. The temperature controls the probability that a worse design is accepted. This allows
the algorithm to avoid local optima in the beginning of the search when the temperature is high. At
the end of the search, when the temperature is low the probability of accepting worse designs is very
low. Thus, the search converges to an optimal solution. An advantage of SA is that it can handle mixed
discrete and continues problems. The parameters settings for a SA algorithm determines how new
solutions should be generate, the initial temperature and what the cooling scheme should be.
[Andersson]138.
135 Sacks, Jerome, William J Welch, Toby J Mitchell, and Henry P Wynn. "Design and analysis of computer
experiments." Statistical science:409-423, 1989.
136 See above.
137 Johan Andersson, “A survey of multi objective optimization in engineering design”, Technical Report: LiTH-
IKP-R-1097.
138 Johan Andersson, “A survey of multi objective optimization in engineering design”, Technical Report: LiTH-
IKP-R-1097.
106
139 D.E. Goldberg, “Genetic and evolutionary algorithms come of age”, Comm. ACM 37 (1994) 113–120.
140 F. Zhang, S. Chen and M. Khalid, “Optimizations Of Airfoil And Wing Using Genetic Algorithm”, ICAS 2002.
107
Figure 4.13 The Progress of the Complex Method for a Two Dimensional Example, with the Optimum
Located in the Middle of the Circles. (Courtesy of J. Andersson).
141 Box M. J., “A new method of constraint optimization and a comparison with other methods,”, 1965.
142 Guin J. A., “Modification of the Complex method of constraint optimization,” Computer Journal, vol. 10, 1968.
143 Spendley W., Hext G. R., and Himsworth F. R., “Sequential application of Simplex designs in optimization and
IKP-R-1097.
146 Haque M. I., “Optimal frame design with discrete members using the complex method,” Computers &
implementation they can handle mixed continues and discrete problems. However, they usually
shows quit slow convergence. Here just one among many methods is described. It is an iterative
procedure where the next point xt+1 is calculated as
𝐱 t+1 = 𝐱 t + α. 𝐒
Eq. 4.10
where α is a step size parameter and S is a random generated unit vector in witch to search. If xt+1
has a better function value than xt , xt+1 replaces xt . Otherwise a new search direction S is generated
until a better solution is found. If no better solution can be found the step size is reduced and new
search directions is generated. The search is considered to be converged when no better solutions
can be found and the step size is reduced beyond a prescribed value.
1. Pipelining Hybrids
2. Asynchronous Hybrids
3. Hierarchical Hybrids
4. Additional Operators
147Yen J., Liao J., Lee B., and Randolph D., “A hybrid approach to modeling metabolic systems using a genetic
algorithm and simplex method,” IEEE Transaction on systems, man and cybernetics, vol. 28, pp. 173-191, 1998.
109
method. For instance, [Yen et al.] has introduced a simplex method as a new way of generating
children in a genetic algorithm. At each iteration a certain percentage of the population is generated
by employing a simplex method on group of promising individuals. The rest of the population is
generated using the usual genetic operators. This has improved both the convergence speed as well
as the ability to find the absolute optima. There are many additional hybrids of SA and GA algorithms
for instance where among other things the replacement is done according to a simulated annealing
scheme. The thermodynamic genetic algorithm is also an example of such a hybrid. [Gunel and
Yazgan]148 have presented a method where they combine a random search strategy with the complex
method. The basic idea of this method is to modify the reflection strategy of the complex method. If
the new reflected point is still the worst point it is not moved against the centroid as in the normal
Complex, but a random strategy is employed in order to generate the new point. Additional
information is available in [Andersson]149.
148 Gunel T. and Yazgan B., “New global hybrid optimization algorithm,” presented at 2nd
Biennial European Joint
Conference on Engineering Systems Design and Analysis, London, England, 1994.
149 Johan Andersson, “A survey of multi objective optimization in engineering design”, Technical Report: LiTH-
IKP-R-1097.
150 Borup L. and Parkinson A., “Comparison of four non-derivative optimization methods on two problem
containing heuristic and analytic knowledge,” presented at ASME Advances in Design Automation, Scottsdale,
Arizona, 1992.
110
151 J. Shinkyu, C. Kazuhisa, O. Shigeru, “Data mining for aerodynamic design space”, J. Aerospace. Computation
Inf. Comm. 452–469, 2005.
111
Few researchers have investigated the effect of different shape parameterization methods on
optimization process. [Sripawadkul]154 studied and compared five airfoil parameterization methods,
Ferguson’s curves, Hicks-Henne bump functions, B-Spline, PARSEC and Class/Shape function
transformation method (CST), in terms of parsimony, completeness, orthogonality, flawlessness and
intuitiveness. Five parameterization methods were scored to assist to select the proper method
respect to specific issue. [Song and Keane]155 investigated effect of two parameterization methods,
orthogonal basis function and B-Spline, on inverse fitting the different airfoils. The results showed
the B-spline could provide higher accuracy than orthogonal basis function using high number of
design variables. [Castonguay]156 studied the effect of four parameterization methods, mesh points,
B-Splines Hicks-Henne bump function and PARSEC, on inverse design and drag minimization in 2D
airfoil. The results demonstrated the mesh points method provides the highest level of accuracy
comparing to other methods, and PARSEC may be unable to provide high flexibility since it failed in
inverse design case. [Mousavi]157 performed the 2D airfoil inverse design, 2D drag minimization and
152 Samareh, J.A., “Survey of shape parameterization techniques for high-fidelity multidisciplinary shape
optimization”. AIAA Journal, 2001.
153 See the previous.
154 Sripawadkul, V., M. Padulo and M.Guenov, “A Comparison of Airfoil Shape Parameterization Techniques for
AIAA-2007-59. 2007.
157 Mousavi, A., Castonguay P., and S. K. Nadarajah, “Survey of Shape Parameterization Techniques and Its Effect
3D wing drag minimization using mesh points, B-Spline and CST methods. It showed the mesh points
method provided the best results in all test cases. The B-Spline and CST methods were able to provide
the reasonable accuracy with low number of design variables. The CST was able to eliminate the
shock wave using very low number of variables in drag minimization case.
[He et al.]158 studied an aerodynamic shape optimization problem that demonstrates the importance
of robustness. The problem was to minimize the drag at a transonic condition while constraining lift
starting from a circular shape. This required CFD simulations of many shapes a human designer
would not consider, such as the shape shown in the second frame of Figure 5.1. Although RANS
cannot accurately predict the aerodynamic performance of such designs due to the massive
separation, the gradients provided the correct trends, and the optimization eventually converged to
an optimal supercritical airfoil for which RANS is valid.
Figure 5.1 The optimizer started with the initial shape of a circle (1st frame) and converged to a super-
critical airfoil (3rd frame). In this process, the optimizer had to go through infeasible intermediate shapes,
such as the one shown here (2nd frame)
➢ The level of information fidelity required from the flow solver, depending on problem ;
➢ Scope of parametrized design space;
➢ Types of design variables, e.g. discrete and/or continuous;
➢ Single or Multi-Objective optimization;
➢ Constraints handling;
➢ Properties of the design space, e.g. number of local optima, discontinuities.
158 He, X., Li, J., Mader, C. A., Yildirim, A., and Martins, J. R. R. A., \Robust aerodynamic shape optimization from
a circle to an airfoil," Aerospace Science and Technology, Vol. 87, 2019.
113
Multiple Objectives:
Minimize F(x) = [F1(x) , F2(x) , . . . , Fk (x)]T
subject to gj (F(x)) ≤ 0 , j = 1, 2, . . . , m , and hL (F(x)) = 0, L = 1, 2, . . . , e ,
F(x) ∈ Ek are also called objectives, criteria, payoff functions, cost functions, or
value functions, where k is the number of objective functions, m is the number of
inequality constraints, and e is the number of equality constraints. x ∈ En is a vector
of design variables (also called decision variables), where n is the number
independent variables.
It is important to note that no optimization procedure guarantees the global optima of the objective
function f(x) will be found the process may only converge towards a locally optimal solution.
Typically in this situation there are three possibilities:
Multi-Level Optimization:
Minimize (x , y) , F(x , y)
subject to: y ∊ S(x); G(x, y) ≤ 0
where S(x) denotes the set of solutions of the lower level problem as:
Minimize (y) : f(x , y) subject to g(x, y) ≤ 0
114
programming problem159:
[Fathi & Shadaram]160 introduced a of Multi-Level, Multi-Objective, as well as, Multi-Point
aerodynamic optimization of the axial compressor blade. Generally, they versioned an approach to
the problem to build an objective function which is the summation of penalty terms, to limit the
violations of the constraints. To reduce the computational effort, optimization procedure is working
on two levels. Fast but approximate prediction methods has been used to find a near-optimum
geometry at the firs-level, which is then further verified and refined by a more accurate but expensive
Navier–Stokes solver. Genetic algorithm and gradient-based optimization were used to optimize the
parameters of first-level and second-level, respectively.
β1 + β2 + ⋯ + β𝑛 = 1 , g = β1 g1 + β2 g 2 + ⋯ + β𝑛 g 𝑛
Eq. 5.1
159 Jane J. Ye, “Necessary optimality conditions for multi-objective bi-level programs”, 2010.
160 A. Fathi · A. Shadaram, “Multi-Level Multi-Objective Multi-Point Optimization System for Axial Flow
Compressor 2D Blade Design”, Arab J Scientific Engineering, 2013.
161 Jameson, A., Leoviriyakit, K., and Shankaran, S., "Multi-point Aero-Structural Optimization of Wings Including
Planform Variations", 45th Aerospace Sciences Meeting and Exhibit, AIAA-2007-764, Reno, NV, 8–11 Jan 2007.
115
The gradient to determine an assigned a weight overall loss or a gain for the design is created by
summing all the gradients times each respective weight (see Eq. 5.1). What this allows for is if a
change drastically improves takeoff performance but results in a slight hit on cruising performance,
the cruising hit can override the takeoff gain due to weighting. Setting the simulation up in this
manner can significantly improve the designs produced by the software. This version of the modeler,
however, adds yet another complexity to the initial conditions, and a slight error on the designer’s
behalf can have a significantly larger effect on the resulting design. The calculation efficiency
improvement takes advantage of the multiple variables. (See Figure 5.2). The problem observed is
that changes that boosted one point of interest directly conflicted with the other, and the resulting
compromise severely hampers the improvement gained. Current research involves a better way to
resolve the differences and achieve an improvement similar to the single-point optimizations.
Figure 5.2 Multi-Point Design Process as Envisioned by Jameson – (Courtesy of Jameson et al.)
162B., Chiguluri, “Multi-Point optimization of Airfoils”, Undergraduate Research Thesis, Georgia Institute of
Technology, 2011.
116
based on the phase of the flight. Slat is used to delay stall such that an increment in the angle of attack
doesn’t cause adverse effect on the lift. The flap is used to increase the camber of the airfoil so that
additional lift is obtained. Figure 5.3 summarizes the typical configuration of the wing at different
phases of flight level flight, take-off, and landing.
In the cruise phase i.e. when the slat and
flap are retracted, the multi-element airfoil
can be simplified (by ignoring the small
gaps between the surfaces) to a single
element airfoil. The simplified single
element airfoil’s aerodynamic properties
are often used to design an optimum wing
cross section. Often, landing could be
ignored because it is easily achieved
through aileron deployment. Therefore,
the two phases of flight that govern the
airfoil design are the steady flight and take-
off conditions. In most studies, the
optimization process is applied to the
cruise level condition while ignoring the
take-off conditions. The results often result
in inefficient take-off conditions which
result in excess fuel procurement. Figure 5.3 Wing configurations at different flight
Therefore, it is important to design the phases (Courtesy of Chiguluri)
airfoil for both the steady-flight and take-
off conditions163.
Hampton, Va.
117
laminated plate structures, to high-fidelity tools, such as nonlinear CFD and detailed CSM. For a
multidisciplinary problem, the application must also use a consistent parameterization across all
disciplines. An MDO application requires a common geometry data set that can be manipulated and
shared among various disciplines. In addition, an accurate sensitivity derivative analysis is required
for gradient-based optimization. The sensitivity derivatives are defined as the partial derivatives of
the geometry model or grid-point coordinates with respect to a design variable. The sensitivity
derivatives of a response, F, with respect to the design variable vector D, can be written as:
𝐅 Field Grid
∂𝐅 ∂𝐅 ∂𝐑 𝐅 ∂𝐑 𝐒 ∂𝐑 𝐆 𝐒 Surface Grid
=[ ][ ][ ][ ] where {
∂𝐃 ⏟∂𝐑 𝐅 ⏟∂𝐑 𝐒 ⏟ ∂𝐑 𝐆 ⏟∂𝐃 𝐆 Geometery
I II III IV 𝐃 Design Variables
Eq. 5.2
The 1st term on the right-hand side of Eq. 5.2 represents the sensitivity derivatives of the response
with respect to the field grid point coordinates. The 2nd term on the right-hand side is vector of the
field grid-point sensitivity derivatives with respect to the surface grid points. The sensitivity
derivative vector must be provided by the field grid generator, but few grid generation tools have the
capability to provide the analytical grid-point sensitivity derivatives? The third term on the right-
hand side of Eq. 5.2 denotes the surface grid sensitivity derivatives with respect to the shape design
variables, which must be provided by the surface grid generation tools. The fourth term on the right-
hand side of Eq. 5.2 signifies the geometry sensitivity derivatives with respect to the design variable
vectors; this must be provided by the geometry construction tools166. An important ingredient of
shape optimization is the availability of a model parameterized with respect to the airplane shape
parameters such as planform, twist, shear, camber, and thickness. The parameterization techniques,
according to [Samareh], are divided into the following categories:
➢ Basis vector,
➢ Domain Element,
➢ Partial Differential Equation (PDE),
➢ Discrete,
➢ Polynomial,
➢ Spline Representation,
➢ CAD-Based,
➢ Analytical,
➢ Free Form Deformation (FFD),
➢ Modified FFD.
Among those, we attend to Discrete, Analytical, PDE, CST, Spline Representation and Free From
Deformation (FFD).
leads to high cost and a difficult optimization problem to solve. The natural design approach is a
variation of the discrete approach that uses a set of fictitious loads as design variables. These
fictitious loads are applied to the boundary points, and the resulting displacements, or natural shape
functions, are added to the baseline grid to obtain a new shape. Consequently, the relationship
between changes in design variables and grid-point locations is established through a finite element
analysis. [Zhang and Belegundu]167 provided a systematic approach for generating the sensitivity
derivatives and several criteria to determine their effectiveness. The typical drawback of the natural
design variable method is the indirect relationship between design variables and grid-point
locations.
For an MDO application, grid requirements are different for each discipline. So, each discipline has a
different grid and a different parameterized model. Consequently, using the discrete
parameterization approach for an MDO application will result in an inconsistent parameterization168.
The most attractive feature of the discrete approach is the ability to use an existing grid for
optimization. The model complexity has little or no bearing on the parameterization process. It is
possible to have a strong local control on shape changes by restricting the changes to a small area.
When the shape design variables are the grid-point coordinates, the grid sensitivity derivative
analysis is trivial to calculate; the third and fourth terms in Error! Reference source not found. can be c
ombined to form an identity matrix.
167 Zhang, S. and Belegundu, A. D., "A Systematic Approach for Generating Velocity Fields in Shape Optimization,"
Structural Optimization, Vol. 5, No. 1-2, 1993, pp. 84-94.
168 Jamshid A. Samareh, “A Survey of Shape Parameterization Techniques”, NASA Langley Research Center, VA.
169 Hicks, R. M. and Henne, P. A., "Wing Design by Numerical Optimization," Journal of Aircraft, 1978.
170 Elliott, 3. and Peralre, J., "Practical Three-Dimensional Aerodynamic Design and Optimization Using
Unstructured Meshes," AIAA Journal, Vol. 35, No. 9, 1997, pp. 1479-1486.
171 Hager, J. O., Eyi, S., and Lee, K. D., "A Multi-Point Optimization for Transonic Airfoil Design," AIAA 92-4681.
172 Bloor, M. I. G. and Wilson, M. J., "Efficient Parameterization of Genetic Aircraft Geometry," Journal of Aircraft,
AIAA 12th Computational Fluid Dynamics Conference, AIAA, Jun. 1995, pp. 452-462, also AIAA-95-1687
119
that must model the internal structural elements such as spars, ribs, and fuel tanks. As a result, this
method is suitable for problems involving a single discipline with relatively simple external geometry
changes.
174 M.J. Martın, E. Andres, M Widhalm, P. Bitrian, and C. Lozano,“Non-uniform rational B-splines-based
aerodynamic shape design optimization with the DLR TAU code”, Proceedings of the Institution of Mechanical
Engineers, Part G: Journal of Aerospace Engineering, 2011.
175 S. Peigin and B. Epstein. “Multi-constrained Aerodynamic Design of Business Jet By CFD Driven Optimization
Kriging and Space Mapping”. 57th AIAA/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials
Conference, AIAA 2016-0418, number January, pages 1-10, San Antonio, TX, USA, 2016.
177 P.D. Vecchia and F. Nicolosi. “Aerodynamic Guidelines in The Design and Optimization of New Regional
fuselage/wing juncture [Vecchia and Nicolosi]179. Geometry definition through the use of partial
differential equations (PDEs) are not as commonly used as well-established spline-based methods
but are just as versatile for geometry surface definition. [Athanasopoulos et al.]180 show that for
equivalently complex surface construction PDEs require fewer design variables, resulting in a more
compact design space. Due to the small set of design parameters required by the PDE method the
computational cost associated with the optimization of a given aerodynamic surfaces can be reduced.
In a PDE-based method the parameters are boundary values to the PDE, hence the relationship
between the value of the design parameter and the geometry can be unclear making method-official
surface deformations tedious. This is likely why the aerodynamic definition of a body in an
optimization scheme does not used PDE representation even though it may initially seem a more
appropriate method. Comparatively, spline-based methods are conceptually simpler and will provide
a more direct relationship between design parameters and the resulting geometry and thus allow
better control over the range of geometries that can be generated.
If optimization establishes performance metrics from CFD, the simplest methods for body surface
definitions are reformative ones. In reformative methods the mesh points on the surface of the body
are directly treated as design variables,181 and their position can be perturbed by the optimizer in
order to generate new shapes. These approaches have the significant advantage that any geometry
the mesh generation algorithm is capable of can be evaluated, however it is likely to require many
hundreds of design variables; deformations are therefore usually limited to single-degree-of-
freedom deformations.
Figure 5.5 Free-Form Deformation (FFD) Parametrizing Wing with 720 Control Points - (Courtesy of
Kenway and Martins)
179 P.D. Vecchia and F. Nicolosi. “Aerodynamic Guidelines in The Design and Optimization of New Regional
Turboprop Aircraft”, Aerospace Science and Technology, Oct 2014.
180 M. Athanasopoulos, H. Ugail, and G.G. Castro. “Parametric Design of Aircraft Geometry Using Partial
[Kenway and Martins]182. This approach embeds the solid geometry within a FFD hull volume
(volumes are typically referends of Bezier splines, Non-Uniform Rational B-splines of NURBS), which
are parametrized by a series of control points as shown in Figure 5.5. These control points deform
the volume which translate to geometric changes of the solid geometry rather than redefining the
whole geometry itself which can give a relatively more efficient set of design variables. A key
assertion of the FFD approach, when applied within a CFD environment, is that a geometry has
constant topology through-out the optimization process; this is typical of high-fidelity optimizations
where the initial geometry considered is sufficiently close to the optimal solution. Figure 5.5 shows
the FFD hull volume enclosing a wing with 720 geometric control points used by [Lyu et al.]183 which
control shape deformation in the vertical (z) axis. The initial random wing deformation and
associated optimized wing cross-sections at select locations are also shown. A similar method is
based on Radial Basis Function (RBF) interpolation which defines data sets of design variables and
their global relationships. [Fincham & Friswell]184 and [Poole et al.]185 use radial basis functions to
optimize morphing aero-foils and report that they provide a means to deform both aerodynamic and
structural meshes and interpolate performance metrics between two non-coincident meshes.
Volumetric-based body representation have been used for optimization but rarely in the field of
aerodynamics, a recent review of the applicability of volumetric parametrization for aerodynamic
optimization is given by [Hall et al.]186.
182 G. Kenway, G. Kennedy, and J.R.R.A. Martins. “A CAD-Free Approach to High-Fidelity Aero-structural
Optimization”. 13th AIAA/ISSMO Multidisciplinary Analysis Optimization Conference, AIAA 2010-9231.
183 Z. Lyu, G. Kenway, and J.R.R.A. Martins, “Aerodynamic Shape Optimization Investigations of the Common
AIAA 2013.
186 J. Hall, D.J. Poole, T.C.S. Rendall, and C.B. Allen. “Volumetric Shape Parameterization for Combined
Component Shapes," 11th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, 2006.
189 Marco Ceze, Marcelo Hayashiy and Ernani Volpe, “A Study of the CST Parameterization Characteristics”, AIAA
2009-3767.
190 Arash Mousavi, Patrice Castonguay and Siva K. Nadarajah, “Survey Of Shape Parameterization Techniques
ς U (ψ) = C N1
N2 (ψ).SU (ψ) + ψ.ΔςU
x z
where ψ = and ς = Eq. 5.3
N2 (ψ).S L (ψ) + ψ.ΔςL
ς L (ψ) = C N1 c c
The last terms define the upper and lower trailing edge thicknesses. Equation uses the general class
function to define the basic profile and the shape function to create the specific shape within that
geometry class. The general class function is defined as:
N2 (ψ) = ψ .(−ψ)
C N1 N1 N2
Eq. 5.4
For a general NACA type symmetric airfoil with a round nose and pointed aft end, N1 is 0.5 and N2 is
0 in the class function. This classifies the final shape as being within the "airfoil" geometry class,
which forms the basis of CST airfoil representation. This means that all other airfoils represented by
the CST method are derived from the class function airfoil. Further details can be found in [Lane and
Marshall]191, or [Ceze et al.]192. The 2D process for airfoils is easily extended to wings as a simple
extrusion of parameterized airfoils. This greatly increases the number of design variables for an
optimization scheme. However, it is no less powerful. By controlling the distribution of airfoils, any
smooth wing can be represented. Also, such characteristics as sweep, taper, geometric twist, and
aerodynamic twist can be included. The definition of a 3D surface follows a similar structure to that
of a 2D surface. Again, for complete description of method, readers are encouraged to consult193. For
further and complete information, readers encourage to consult [Su et al.]194.
191 Kevin A. Lane and David D. Marshall, “A Surface Parameterization Method for Airfoil Optimization and High
Lift 2D Geometries Utilizing the CST Methodology”, AIAA 2009-1461.
192 Marco Ceze, Marcelo Hayashiy and Ernani Volpe, “A Study of the CST Parameterization Characteristics”, AIAA
2009-3767.
193 See 40.
194 Hua Su, Chunlin Gong, and Liangxian Gu, “Three-Dimensional CST Parameterization Method Applied in
Aircraft Aero elastic Analysis”, Hindawi, International Journal of Aerospace Engineering Volume 2017, Article
ID 1874729, 15 pages, https://doi.org/10.1155/2017/1874729.
195 See 40.
123
where many other solution methods are not valid. This complicates the process greatly. Since CFD is
to be used by an optimizer, both the meshing and solution processes must be automated. Therefore,
the meshing process must be robust enough to handle any airfoil selected by the optimizer. However,
the meshing process will still be sensitive to the given airfoil geometry. If the airfoil selected by the
optimizer is too unlike the airfoil used to develop the meshing automation, the meshing process is
prone to errors. Therefore, constraints are used to force the optimizer to select airfoils that somewhat
resemble the initial airfoil. Constraints implemented to ensure an airfoil successfully passes the
meshing stage include limits on maximum thickness and minimum thickness. An additional
constraint was placed so that the upper surface does not cross the lower surface. The optimizer is
currently being tested for a cruise condition of Mach 0.8 at an altitude of 35,000 feet. To ensure that
a constant CL is maintained, the objective function estimates the current airfoil's lift curve by fitting
a line to CL values taken from CFD solutions at different angles of attack. This is used to obtain the
angle of attack that should produce the desired Cl. This angle of attack is used for the final CFD
solution of the objective function from which L/D is taken and read by the optimizer. The initial airfoil
selected for the optimization scheme was the RAE 2822 transonic airfoil previously used in the class
function coefficient optimization study. A CL of 0.322
was selected to correspond to the CL at cruise of the Initial Optimized
airfoil used by a next generation commercial airliner CL 0.329 0.318
currently being studied at Cal Poly. Table 5.1 displays
CD 0.0273 0.0140
a comparison between the performance of the initial
and optimized airfoils. The CL values differ somewhat L/D 12.05 22.78
and displays some error in the selection of angle of AoA 0.96 1.5
attack. However, the CD of the optimized airfoil is
dramatically lower than that of the initial airfoil. The CD Table 5.1 Performance Comparison of
value drops from 273 drag counts to 40, which is a Initial an Optimize Airfoils (Courtesy of 44)
reduction of 33 drag counts or about 49%. This also
causes the L/D to increase from 2.05 to 22.78, which is an increase of about 89%. Figure 5.6 displays
contours of Mach number over the initial and optimized airfoils. The initial airfoil is displayed on the
left while the optimized airfoil is shown on the right. The maximum Mach number in the flow over
the initial airfoil is much higher than that of the optimized airfoil. This is because the upper surface
of the optimized airfoil is much flatter than that of the initial airfoil, which causes the flow to
accelerate less over the upper surface of the optimized airfoil. This is the cause of the dramatic drag
reduction. The lowest point of the lower surface of the optimized airfoil is forward from that of the
initial airfoil. This allows for lower speed flow and therefore higher pressure.
Figure 5.6 Contours of the Initial Airfoil (Left) an Optimize Airfoil (Right)44
124
➢ strategies that preserve only feasible solutions with no constraint violations: infeasible
solutions are deleted;
➢ strategies that allow feasible and infeasible solutions to co-exist in a population, however
penalty functions penalize the infeasible solutions (constraint based reasoning);
➢ strategies that create feasible solutions only;
➢ strategies that artificially modify solutions to boundary constraints if boundaries are
exceeded; and
➢ strategies that repair/modify infeasible solutions.
Most commonly optimizations apply weighted penalties to the objective function if the constraint(s)
are violated. The reason for this is that penalty functions are often deemed to ease the optimization
process, and bring the advantage of transforming constrained problems into unconstrained one by
directly enforcing the penalties directly to the objective function. With this method Pareto-optimal
solutions with good diversity and
reliable convergence for many
algorithms can be obtained easily
when the number of constraints are
small; fewer than 20 constraints. It
becomes more difficult to reach
Pareto-optimal solutions efficiently
as the number of constraints
increase, and the number of
analyses of objectives and
constraints quickly becomes
prohibitively expensive for many
applications.
This is because the selection
pressure decreases due to the
reduced region in which feasible
solutions exist. [Kato et al.]198
suggest that in certain
circumstances Pareto-optimal
solutions may exist in-between Figure 5.7 Concept of using Parallel Evaluation Strategy of
regions of solution feasibility and Feasible and Infeasible Solutions to Guide Optimization
Direction in a GA
196 S. Koziel and Z. Michalewicz. “Evolutionary Algorithms, Homomorphous Mappings, and Constrained
Parameter Optimization”. Evolutionary Computation, 7(1):19{44, 1999.
197 J. Siens and M.S. Innocente. “Particle Swarm Optimisation: Fundamental Study and its Application to
Optimisation and to Jetty Scheduling Problems”. Trends in Engineering Computational Technology, pages 103-
126, 2008.
198 T. Kato, K. Shimoyama, and S. Obayashi. “Evolutionary Algorithm with Parallel Evaluation Strategy of Feasible
and Infeasible Solutions Considering Total Constraint Violation”. IEEE, 1(978):986-993, 2015.
125
infeasibility. This is illustrated in Figure 5.7, where it is seen that feasible and infeasible solutions
could be evaluated in parallel to guide the optimization search direction towards feasible design
spaces. This is intuitively true for single discipline aerodynamic optimization problems where often
small modifications to design variables can largely impact the performance rendering designs
infeasible. Algorithm understanding of infeasible solutions can help in the betterment of feasible
solutions though algorithm learning/training and constraint based reasoning. [Robinson et al.]199,
comparing the performance of alternative trust-region constraint handling methods, showed that
reapplying knowledge of constraint information to a variable complexity wing design optimization
problem reduced high-fidelity function calls by 58% and additionally compare the performance to
alternative constraint managed techniques.
Elsewhere, [Gemma and Mastroddi]200 demonstrated that for the multi-disciplinary, multi-objective
aircraft optimizations the objective space of feasible and infeasible design candidates are likely to
share no such definitive boundary. With the adoption of utter constraints, structural constraints, and
mission constraints solutions defined as infeasible under certain conditions would otherwise be
accepted, hence forming complex Pareto fronts. Interdisciplinary considerations such as this help to
develop and balance conflicting constraints. For example, structural properties which may be
considered feasible, but are perhaps heavier than necessary will inflict aero-elastic instabilities at
lower frequencies. In the aerospace industry alone there are several devoted open-source
aerodynamic optimization algorithms with built-in constraint handling capability. Some studies
have also adopted MATLAB's optimization tool-box for successful optimization constraint
management.
199 T.D. Robinson, K.E. Willcox, M.S. Eldred, and R. Haimes. “Multi-fidelity Optimization for Variable-Complexity
Design”. 11th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, pages 1-18, Portsmouth,
VA, 2006. AIAA 2006-7114.
200 S. Gemma and F. Mastroddi. “Multi-Disciplinary and Multi-Objective Optimization of an Unconventional
Aircraft Concept”. 16th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, AIAA 2015-2327,
pages 1-20, Dallas, TX, 2015.
126
201 G. Kenway and J.R.R.A. Martins. “Aerodynamic Shape Optimization of the CRM Configuration Including Buffet-
Onset Conditions”, 54th AIAA Aerospace Sciences Meeting, AIAA 2016-1294, pages 1-23, San Diego, California.
202 M. Secanell and A. Suleman. “Numerical Evaluation of Optimization Algorithms for Low-Reynolds-Number
demands when handing many hundreds of design variables; this makes them well suited for
optimizing shapes based on irregular geometric parametrizations. optimizing shapes based on
deformities geometric parametrizations.
204 O. Pironneau. “On Optimum Design In Fluid Mechanics”. Journal of Fluid Mechanics, 64(1): 97-110, 1974.
205 C.A. Mader, J.R.R.A. Martins, J.J. Alonso, and E.V. Der Weide. “An Approach for the Rapid Development of
Discrete Adjoint Solvers”. AIAA Journal, 46(4):863-873, 2008.
206 Z. Lyu, G. Kenway, C. Paige, and J.R.R.A. Martins. “Automatic Differentiation Adjoint of the Reynolds-Averaged
Navier-Stokes Equations with a Turbulence Model”, 21st AIAA Computational Fluid Dynamics Conference, AIAA
2013-2581, pages 1-24, San Diego, California, 2013.
128
higher-order effects must be included for accurate induced drag prediction and hence for meaningful
optimizations. This work was followed by [Gagnon and Hicken]207 for the aerodynamic optimization
of un-conventional aircraft configurations. These global geometric variables are generally not
considered in high-fidelity simulation. The cost of allowing such geometric variation away from the
baseline under high-fidelity optimization limited how many variables could be considered in any one
optimization process. The authors observed limited optimization in some wing configurations
because of this. More details regarding the evaluation of sensitivity analysis is available in section 6.
207 H. Gagnon and D.W. Zingg. “High-Fidelity Aerodynamic Shape Optimization of Unconventional Aircraft
through Axial Deformation”. 52nd Aerospace Sciences Meeting, AIAA 2014-0908, pages 1-18, Maryland, 2014.
208 J.R.R.A. Martins, J.J. Alonso, and J.J. Reuther. “A Coupled-Adjoint Sensitivity Analysis Method for High-fidelity
Morphing Trailing Edge Wing”. 15th AIAA/ISSMO Multi-disciplinary Analysis and optimization Conference,
AIAA paper 2014-3275, pages 1-13, 2014.
210 Z. Zhang, P.C. Chen, Z. Zhou, S. Yang, Z. Wang, and Q. Wang. “Adjoint Based Structure and Shape Optimization
with Flutter Constraints”. 57th AIAA/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials
Conference, AIAA 2016-1176, pages 1-23, San Diego, California, 2016.
211 B. Grossman, R.T. Haftka, P.J. Kao, D.M. Polen, and M. Rais-Rohani. “Integrated Aerodynamic-Structural Design
the total number of function calls and sensitivity calculations, the wing performance gain was limited.
When performing sequential optimization the optimizer does not have sufficient information
necessary for aero elastic tailoring. This limitation of sequential optimization is further explained by
[Chittick and Martins]212. A significant drawback of all gradient-based algorithms is the requirement
for continuity and low-modality throughout the design space otherwise the algorithm may become
sub-optimally trapped. The challenge is that an aerodynamic shape analysis throughout a
geometrically varying search space will encounter both non-continuous topological and local flow
changes, each providing local optima. Gradient-dependent algorithms' robustness significantly
decreases in the presence of discontinuity and lack of convergence, usually related to turbulence
modelling, making the objective function noisy. [Kenway] encountered such a problem with
aerodynamic shape optimization with a separation-based constraint formulation to mitigate buffet-
onset behavior at a series of operating conditions. The discontinuity from the function arose from
monitoring the wing local surface for separated flow; this resulting in locally negative skin friction
coefficients. To address this issue blending functions were to be implemented to smooth the
discontinuity, smearing the separation sensor value around the separated flow region.
212 I.R. Chittick and J.R.R.A. Martins. “An Asymmetric Sub-optimization Approach to Aero-structural
Optimization”. Optimization and Engineering, 10(1):133-152, 2009.
213 G. Kenway and J.R.R.A. Martins. “Aerodynamic Shape Optimization of the CRM Configuration Including Buffet-
Onset Conditions”, 54th AIAA Aerospace Sciences Meeting, AIAA 2016-1294, pages 1-23, San Diego, California.
130
214 A. Beechook1, J. Wang, “Aerodynamic Analysis of Variable Cant Angle Winglets for Improved Aircraft
Performance”, Proceedings of the 19th International Conference on Automation & Computing, Brunel
University, London, UK, 13-14 September 2013.
215 P. Bourdin_, A. Gatto_, and M.I. Friswell, “The Application of Variable Cant Angle Winglets for Morphing
favorable coupling in pitch and roll with regard to turning maneuvers. (See Figure 6.3).
6.4.1 Comparison of Cant Angle Winglet for Simulation vs. Wind Tunnel
The comparison of lift-to-drag ratio values from wind tunnel test and CFD simulations is presented
in Figure 6.4. The L/D values obtained from wind tunnel experiments were very low compared to
those obtained from simulations. This is due to the high CD values from the wind tunnel results. From
the simulations results, winglet with cant angle 45° has the highest L/D compared to all other
configurations. From the wind tunnel results, the L/D for each winglet configuration varies for
different angle of attack, e.g. the winglet at cant angle 45° has the highest L/D at 12° angle of attack.
Figure 6.4 Lift-to-Drag Ratio, L/D (Wind Tunnel and CFD Comparison)
132
The CFD simulations and wind tunnel test results showed that the different winglet configurations
have different aerodynamic characteristics when the angle of attack is varied. At low angles of attack,
ideally at cruise angle of attack, winglets at cant angle 45° and 60° showed improved the aerodynamic
performance in terms of lift and drag coefficients. The winglets at cant 45° and 60° did not provide
optimum performance at high angles of attack, for example, at higher angle of attack, the winglets at
cant 45° produced more lift compared to other winglet configurations. Hence, varying the winglets’
cant angle at different flight phases can improve the aircraft efficiency and optimize performance.
Therefore, it can be concluded that the investigated concept of variable cant angle winglets appears
to be a promising alternative for traditional fixed winglets. However, this study involved only the
flow study and there are many other important factors to consider while designing new devices for
aircraft, e.g. structural weight and cost. Similar results obtained by [Bourdin at al.]216 for moments
attainable by folding up or down the right winglet.
If designed appropriately, winglets, and indeed all non-planar wingtip variants, can be made to show
aerodynamic advantages when compared to conventional designs but often fail as they usually lead
to structurally heavier wings with detrimental increases in parasitic drag. C-wings have been
considered a compromise between a box wing and a winglet; theoretically providing a reduction in
the induced drag that approaches that of the closed box wing arrangement whilst additionally
reducing the viscous drag penalty incurred by large wetted areas. The C-wing and box wing have also
been recognized to have the potential to replace the conventional horizontal stabilizer to provide
pitch control. However, owing to large and heavy wingtip extensions, the C-wing. is inherently
sensitive to structural and aero elastic issues; they are not closed systems like box wing
arrangements which are, by comparison, much stiffer as the upper wing is fixed. Despite the aero
elastic concerns C-wings are seemingly prone to, conceptually the auxiliary lifting surface at the
main-wing wingtip could be used to introduce substantial damping to modes of vibration; the
characteristics of such a design is not obvious.
The optimal loading condition indicated in Figure 6.5 indicate that the circulation of the main-wing
is carried onto the side wing, acting much like a winglet, thus loaded inward towards the fuselage218.
The circulation is then further extended onto the top-wing producing a net down-loaded surface for
minimum induced drag at a fixed total lift and wingspan. The goal of minimizing the induced drag of
the system requires the gradients of circulation, where possible, to be minimized219. Conventional
planar wings shed strong vortices at the wingtips and the circulation tends to zero. Hence,
distributing the vorticity in the wake over an effectively longer wingspan would reduce the wake
sheet intensity (weaker vortices shed), in addition to moving the wingtip vortices closer together
than that for a conventional wing, accelerating the breakdown of the wake system. The down-loading
of the top-wing surface will naturally have an effect on the structural weight, performance and
control, and may provide a means of stability that is less effected by the main-wing down-wash such
as conventional horizontal stabilizers.
6.6 Case Study 1 - Comparison of Point Design and Range-Based Objectives for
Transonic Airfoil Optimization
The most common airfoil optimization problem considered is lift-constrained drag minimization at
a fixed design point, however, shock-free solutions can result which can lead to poor off-design
performance. As such, [Poole et al.]220 presents a study into the construction of the airfoil
optimization problem and its elected of the performance over a range of operating conditions. Single-
and Multi-point optimizations of airfoils in transonic flow are considered and an improved range-
based optimization problem subject to a constraint on fixed non-dimensional wing loading with a
varying design point is formulated. This problem is more representative of the aircraft design
problem though similar in cost to single-point drag minimization. An analytical treatment using an
approximation of wave drag is also presented which demonstrates that the optimum Mach number
for a fixed shape is supercritical if the required loading is above a critical threshold. Optimizations
are presented that show that to dene an effective objective function, 3D effects modelled via an
induced drag term must be introduced. The general trend is to produce solutions with higher Mach
numbers and lower lift coefficients, and that shocked solutions perform better when considering the
performance in range over the operating space.
218 Demasi, L., Dipace, A., Monegato, G., Torino, P., Cavallaro, R., “Invariant formulation for the minimum induced
drag conditions of nonplanar wing systems”, AIAA J. 52 (10), 2223–2240, 2014.
219 Prandtl, L., “Induced drag of multiplanes”. Technical Report TN 182, NACA, Reproduction of Der Induzierte
Wilderstands Von Mehreckern. Technische Bar, vol. 3, no. 7, pp. 309–315, 1924.
220 D.J. Poole, C.B. Allen, T.C.S. Rendallz, “Comparison of Point Design and Range-Based Objectives for Transonic
section. In the first two of a trilogy of papers, [Morawetz]221-222 proved that shock-free solutions in
inviscid transonic ow around airfoils were isolated. Due to this result, it was considered difficult, if
not impossible, to obtain a shock-free airfoil design. However, the hodograph method223 allowed
shock-free designs to be achieved224-225. [Harbeck & Jameson]226 later quantified the front in the M-
CL space between where shock-free solutions were and were not able 3 of 42 American Institute of
Aeronautics and Astronautics to be obtained. Nowadays, shock-free designs for transonic flows
around airfoils are commonly obtained; for example, of the benchmark cases from the AIAA
Aerodynamic Design and Optimization Discussion Group (ADODG) a shock-free results are readily
available for case 2 (transonic, viscous, drag minimization of RAE2822)227-228-229, and a low drag
solution has recently been published for case 1 (inviscid, non-lifting, drag minimization of
NACA0012) by the authors230.
The goal of point design, that can lead to shock-free solutions, is the substantial performance
improvement at the design point. However, the off-design performance is often severely
compromised, and may be much worse than the initial solution. It was proved in the final paper of
the trilogy by [Morawetz]231 that shock-free airfoils in transonic ow would have a shock if the
freestream Mach number was perturbed. The flow structure for these types of airfoils tends to be a
single shock for an increase in the freestream Mach number and a double shock for a decrease in the
Mach number. Hence, using a single-point design for airfoil optimization can be problematic. This
issue was also considered when designing the NASA supercritical airfoils; [Harris]232 stated
permitting a weak shock rather than trying to design for a shock-free design point also reduces the
off-design penalties usually associated with point design airfoils".
A common way of dealing with the off-design problems is to use multi-point design, where the
objective is a combination of the objective at different design points. The idea is to reach a
compromised solution which, while not being optimal at a number of discrete design points, is a
trade-off of the performance at those design points, which leads to lower off-design penalties. Many
221 Morawetz, C. S., “On the non-existence of continuous transonic flows past profiles I", Communications on Pure
and Applied Mathematics, Vol. 9, No. 1, 1956, pp. 45-68.
222 Morawetz, C. S., “On the non-existence of continuous transonic flows past profiles II," Communications on Pure
Aerospace Sciences Meeting and Exhibit, Reno, Nevada, 2005, AIAA Paper 2005-1041.
227 Carrier, G., Destarac, D., Dumont, A., Meheut, M., Salah El Din, I., Peter, J., Ben Khelil, S., Brezillon, J., and
Pestana, M., “Gradient-Based Aerodynamic Optimization with the elsA Software," 52nd AIAA Aerospace Sciences
Meeting, National Harbor, Maryland, 2014, AIAA Paper 2014-0568.
228 Bisson, F., Nadarajah, S. K., and Shi-Dong, D., “Adjoint-Based Aerodynamic Optimization Framework," 52nd
AIAA Aerospace Sciences Meeting, National Harbor, Maryland, 2014, AIAA Paper 2014-0412.
229 Poole, D. J., Allen, C. B., and Rendall, T. C. S., “Control Point-Based Aerodynamic Shape Optimization Applied to
AIAA ADODG Test Cases," 53rd AIAA Aerospace Sciences Meeting, Kissimmee, Florida,
2015, AIAA Paper 2015-1947.
230 Masters, D. A., Taylor, N. J., Rendall, T. C. S., and Allen, C. B., “Multilevel Subdivision Parameterization Scheme
for Aerodynamic Shape Optimization," AIAA Journal, Vol. 55, No. 10, 2017, pp. 3288-3303.
231 Morawetz, C. S., “On the non-existence of continuous transonic flows past profiles III," Communications on
Pure and Applied Mathematics, Vol. 11, No. 1, 1958, pp. 129-144.
232 Harris, C. D., “NASA Supercritical Airfoils: A Matrix of Family-Related Airfoils," Tech. rep., Langley Research
examples of multi-point design can be found in the literature, see, for example233,234,235, or the AIAA
ADODG case 4 results236,237. Two common issues tend to arise when applying multipoint
optimization. The first is that to apply it successfully requires careful selection of both the design
points (these are often known a priori but can be determined using gradient span analysis), and the
weightings between the objectives at those design points. This issue can be eased by using automated
weight selections, an integral approach or a probabilistic approach. The second common issue is the
cost surrounding multi-point optimization.
By the nature of the problem, multipoint design requires a flow solution at each design point per
objective function evaluation. This makes performing high-fidelity, multi-point optimization on fine
numerical grids prohibitively expensive for more than a handful of design points. Further issues with
multi-point optimization were highlighted in a comprehensive study by [Drela]238. Drela considered
single , two and four-point optimizations and hypothesized that to avoid point design at each of the
considered design points, the number of chosen operating points should be of the order of the
number of design variables (this was later validated from a mathematical argument). The problem
tends to go back to the proven theory of [Morawetz], where, unless there is a shocked solution, a
shock will result for a deviation in
freestream Mach number, even if multiple
points are considered. Hence, the problem
of posing a suitable transonic airfoil
optimization problem is still an open one.
An alternative approach to the construction
of the aerodynamic optimization problem,
including the choice of design point, design
variables, objective function and
constraints is considered here alongside the
conventional single- and multi-point drag
minimization problem. Maximization of the
Breguet range parameter, ML/D, is
considered, subject to constant non-
dimensional wing loading. This design
problem is not often studied in Figure 6.6 Range Variation with Mach Number for
aerodynamic optimization, however, Boeing 747 (where l is the Non-Dimensional Wing
examples of it can be found in historical Loading) - (Courtesy of [Poole et al.])
aircraft design. For example, Figure 6.6
233 Lyu, Z., Kenway, G. K. W., and Martins, J. R. R. A., “Aerodynamic Shape Optimization Investigations of the
Common Research Model Wing Benchmark," AIAA Journal, Vol. 53, No. 4, 2015, pp. 968-985.
234 Epstein, B., Jameson, A., Peigin, S., Roman, D., Harrison, N., and Vassberg, J., “Comparative Study of Three-
Dimensional Wing Drag Minimization by Different Optimization Techniques," Journal of Aircraft, Vol. 46, No. 2,
2009, pp. 526-541.
235 Buckley, H. P., Zhou, B. Y., and Zingg, D. W., “Airfoil Optimization Using Practical Aerodynamic Design
of Benchmark Problems Using Jetstream," 53rd AIAA Aerospace Sciences Meeting, Orlando, Florida, 2015, AIAA
Paper 2015-0262.
237 Meheut, M., Destarac, D., Carrier, G., Anderson, G., Nadarajah, S., Poole, D., Vassberg, J., and Zingg, D.,
“Gradient-Based Single and Multi-points Aerodynamic Optimizations with the elsA Software," 53rd AIAA
Aerospace Sciences Meeting, Kissimmee, Florida, AIAA Paper 2015-0263, 2015.
238 Drela, M., “Pros and Cons of Airfoil Optimization," in Caughey, D. and Hafez, M., editions .,Frontiers of
(which was constructed using the data in the book of [Mair and Birdsall]239, which itself is a processed
form of the data from other researches, shows the range parameter variation with Mach number for
different non-dimensional wing loadings of early variants of the Boeing 747.
Further to considering this design problem, the design point is also considered here as a design
variable, with Mach number and lift coefficient allowed to vary, however to fully model the trade-offs
of speed, lift and drag with range, an induced drag penalty is also introduced. Hence the cost of a
range-based optimization problem is close to equivalent to a single-point design problem and
therefore much cheaper than a multi-point problem. The extra cost only comes from using one extra
design variable in the optimization (Mach number). The overall goal is to consider an aerodynamic
optimization problem which has an optimal solution that is shocked, allowing better overall
performance across the design space.
239 Mair, W. A. and Birdsall, D. L., “Aircraft Performance”, Cambridge University Press, 1992.
137
⏟
Minimize J(α) Subject to g(α) ≤ 0 , and h(α) = 0
α∈ℝn
Eq. 6.2
The design variables used are from a singular value decomposition (SVD) approach240, which
decomposes a training library of airfoils into constituent modes and this has the advantage of
producing an optimal reduced set of shape modes according to the optimality theory of SVD241. To
determine design variables using an SVD approach, a training library of airfoils is collated. From this,
shape variations between all of the airfoils in the library are calculated to produce a variations matrix,
ΔX. Performing an SVD then decomposes the variations matrix into a matrix of mode shapes, U, a
matrix of singular values, Σ and a weighting matrix, V, where the decomposition is given by:
∆𝐗 = 𝐔𝚺𝐕 𝐓
Eq. 6.3
The mode shapes, which are the columns of U, are then design parameters. Since the SVD
process ranks mode shapes by their singular values, the U matrix can be truncated to the first
n columns (where n is the required number of design variables in the optimization) to
produce U’. A new airfoil shape, Xnew, is then given by a linear combination of n modes
superimposed onto an initial airfoil shape, Xold:
n
➢ Global Optimizer which is the global optimization algorithm used is an agent-based method,
where a population of agents are used to traverse the design space in search of a solution.
The location of an agent within the search space of n design variables is α = [α1 , α2 , , , αn]T
and in an agent-based optimization algorithm moves to a new location at the next iteration
240 Poole, D. J., Allen, C. B., and Rendall, T. C. S., “Metric-Based Mathematical Derivation of Efficient Airfoil Design
Variables," AIAA Journal, Vol. 53, No. 5, 2015, pp. 1349-1361.
241 Eckart, C. and Young, G., “The Approximation of One Matrix by Another of Lower Rank," Psychometrika, Vol.
orthogonal modal design variables with a constrained global optimizer," Computers & Fluids, Vol. 143, 2017.
138
244 Zhou, J. L., Tits, A. L., and Lawrence, C. T., “Users Guide for FSQP Version 3.7 : A Fortran Code for Solving
Optimization Programs, Possibly Minimax, with General Inequality Constraints and Linear Equality Constraints,
Generating Feasible Iterates, "Tech. rep., Institute for Systems Research, University of Maryland, 1997.
245 Mayne, D. Q. and Polack, E., “A super linearly convergent algorithm for constrained optimization problems,"
uL W1 L 1 W1
r= log ( ) or r = u log ( )
cD W2 ⏟D ⏟
SFC ⏟ W2
Aerodynamic Porplusion Structure
Eq. 6.7
where u is the aircraft velocity, c is the specific fuel consumption (SFC), L and D, are the lift and drag,
and W1 and W2 are the initial and final cruise weights respectively. Under the assumption of constant
speed of sound through cruise (so u/M) and constant SFC, the range factor can be extracted, R =
ML/D, which can be used as the objective function in optimization. The equivalent expression using
non-dimensional force coefficients is R = MCL/CD. In this scenario, the aerodynamic optimization
problem is enriched with the operating point (characterized by M and CL), which is allowed to vary.
A similar optimization problem has also be considered by [Buckley and Zingg]250, albeit for low-speed
UAV design. The work here expands into inviscid and viscous transonic airfoil design, and puts this
type of optimization problem in the context of suitability for design. It is also worth noting that a
simplified engine model can also be introduced to change the assumption of constant SFC. For
example, [Jameson et al.]251 assumed that c ∝ √M, hence this leads to a range factor of √ML=D. Using
this simplified model would likely alter the optimum design point. However, in this work the MCL=CD
factor is considered since it is the trade-offs in the aerodynamic performance that are studied.
Before attempting geometric optimization it is useful to consider the optimization of the operating
point in isolation. A common optimization approach is to constrain CL, however, this is not suitable
in a process where Mach number varies, as equilibrium flight at a fixed weight demands a fixed
dimensional lift. A more appropriate constraint is therefore M2CL, which is a non-dimensional
measure of wing loading (W/S =CL1/2γM2P). This means that there is only a single free parameter,
Mach number, with the airfoil trimmed to achieve the required lift coefficient for that Mach number.
The optimization problem is therefore written as:
CL
⏟
Maximize =M subjected to M2 CL = 𝑙
M
CD
Eq. 6.8
where the parameter, l, is some non-dimensional wing loading that must be maintained to yield the
same physical lift. In the purely unconstrained problem, where range is maximized with no lift
constraint, assuming two-dimensional inviscid flow, the solution is known to be the critical Mach
number (the Mach number at which flow over the body first becomes sonic). However, by realistic
selection of the lift constraint, a transonic, and shocked solution can be forced. While it is useful to
consider the isolated effect of changing Mach number, the usual aerodynamic optimization process
involves modifying some shape to improve the objective. Often this is subject to an internal volume
248 Morawetz, C. S., “On the non-existence of continuous transonic flows past profiles III," Communications on
Pure and Applied Mathematics, Vol. 11, No. 1, 1958, pp. 129-144.
249 Drela, M., “Pros and Cons of Airfoil Optimization," in Caughey, D. and Hafez, M., eds., “Frontiers of
Transonic Wings," Journal of Aircraft, Vol. 47, No. pp. 505-514, 2010.
140
(V) requirement to represent the need to house structure or fuel. Hence, if any general shape changes
are included, defined here by a vector, Δx, the full optimization problem is now described as:
CL
⏟
Maximize =M subjected to M 2 CL = 𝑙 and V ≥ Vinitial
ΔX , M
CD
Eq. 6.9
In the context of optimization, the cost of solving problem of Eq. 6.9 is similar to that of a single-
point drag minimization. The cost associated with adding a further design variable in both optimizers
is small. Hence, the multi-point optimization is performed using N design points, so the single-point
and range-based optimizations offers a cost reduction of O(N) of the cost of the multi-point. However,
while the cost is lower for range-based optimization, it is worth noting that the primary disadvantage
of considering this is the context of aircraft optimization is that improvement at a specified design
condition is not the goal (as it would be in point optimization). For the work considered in this paper
this is not a problem since the objective here is to consider the effect of a range-based problem on
the resulting optimal geometries. An interesting aspect of this problem is that in the circumstance of
fixing the design point, the problem reduces to that of a single-point drag minimization. Hence, it
would be theoretically possible, though prohibitively expensive, to obtain similar results if the M-CL
space was densely sampled and a single-point drag minimization performed at each of those sample
points. [Poole et al.]252.
252D.J. Poole, C.B. Allen, T.C.S. Rendallz, “Comparison of Point Design and Range-Based Objectives for Transonic
Airfoil Optimization”, AIAA Journal · June 2018.
141
C𝐷𝑤 = k(M − M𝑐 )4
Eq. 6.10
Although k can be found through treatment of incompressible data, it is more straightforward and
accurate to use two-dimensional transonic CFD. Lock's work was originally aimed at deriving a
measure of compressible airfoil performance superior to only considering critical Mach number, for
which Lock proposed the use of k, but the Mach-drag scaling remains useful in the transonic regime,
where analytical results, however approximate, are relatively rare. The objective here is to apply this
in the context of a constrained ML=D operating point. Since Lock's result uses Mc as a reference Mach
number, the influence of lift on wave drag is included in addition to the effect of Mach umber, because
any increase in lift corresponds to a lower value of Mc, thus raising drag (the critical Mach number
drops as the minimum Cp becomes more negative, which is equivalent to higher Cl at incompressible
speeds). The physical trade-off for the operating point is very important. At low Mach, lift coefficient
must be high. This drives a low critical Mach number and consequently a higher wave drag. At high
Mach, the wave drag naturally increases due to the increased offset from Mc. It follows that in
between these extremes there lies an optimum where neither the lift coefficient nor Mach number
are too high, and it is this optimum that shall be explored with a basic analytical treatment. It is shown
that transonic results arise naturally if M2CL is large for the problem given by Eq. 6.8 (where the only
design variable is Mach number). After some differentiation, outlined in [Poole et al.]253, the optimum
Mach number, Mopt, for maximizing range subject to a constraint on fixed loading as:
Mc
Mopt =
dM
5 − 4 ( c)
dM
Eq. 6.11
where Mc is critical Mach number. To solve this, an outer bisection loop on Mopt is used, with Mc found
through an inner bisection loop. dMc/dM is calculated thereafter using the calculated Mc.
6.6.5.2 Results
First, it is interesting to explore the shape of the governing polynomial with the caveat that two-
dimensional inviscid flow is considered (meaning CD0 = 0) and any roots to the left of Mc are non-
253D.J. Poole, C.B. Allen, T.C.S. Rendallz, “Comparison of Point Design and Range-Based Objectives for Transonic
Airfoil Optimization”, AIAA Journal · June 2018.
142
physical. Figure 6.8 shows the polynomial for NACA 0012 alongside the Mc values; it is important
to note that the Mc value plotted corresponds to the Mc value for the rightmost root, therefore, the
Mc does not necessarily correspond to a root of the plotted function (this is intrinsic because Mc is a
function of M due to the M2CL constraint). Indeed, this is the important point
to note from the function shape, because for l > 0.15 the root is to the right of l Mopt/Mc
Mc, i.e. it is transonic. The limit case is also shown, illustrating that, as 0.20 1.071
expected, in this scenario for l = 0.15, the optimum Mach number sits just on 0.25 1.149
top of the root of the function, and for any higher value of l the root shifts to 0.35 1.230
the right of Mc and becomes transonic. For any lower value there will be no 0.45 1.288
root to the right of Mc. Table 6.1 shows the Mopt=Mc values for varying l for 0.55 1.340
the NACA0012, indicating that the value of 1.23 for l = 0.35 compares 0.60 1.366
reasonably to the CFD value of 1.33 (see below), at least to a margin Table 6.1
consistent with the assumptions underpinning Lock's relationship. Further Analytical optimal
increases in l drive a trend towards a higher Mc, consistent with the root in Breguet Mach
Figure 6.8 being driven to the right as the function curves increasingly below Numbers as a
Fraction of Mc for
the axis. Lock's result is limited in accuracy, but it gives a clear indicator of a
NACA 0012 using
transonic optimal point in this case (i.e. where Mopt/Mc ≥ 1). A final M2CL as a
interesting point is what value of l necessitates a transonic optimum. This is Constraint
found from enforcing Mopt = Mc, which may be done with a bisection loop
around the analytical root solver. This reveals that for the NACA 0012 case l = 0.15 is the approximate
limiting value, while the equivalent value for RAE 2822 is l = 0.26. Below these watershed points it is
possible to find a subcritical optimal point that satisfies the constraint, whilst above this only a
supercritical optimal condition is possible.
Figure 6.9 Mach Sweeps at Fixed l showing ML/D - (Courtesy of [Poole et al.])
143
the sweep resolution (and since Mc = 0.49 for this trim point, M/Mc= 1.33, compared to an analytical
prediction of 1:23), while for 2822 at l = 0.45, Mopt = 0.7 (and Mc = 0.55 so for this trim point, M/Mc =
1.27). The limited accuracy of the Lock result means that the absolute value of the optimal Mach
number differs between the algebraic and the CFD (0.75 versus 0.65 for NACA0012, and 0.8 versus
0.7 for RAE2822) by some margin, but as a fraction of Mc the agreement is surprising. Also, Lock notes
an offset in drag rise as a function of Mach of 0.1, which is observable compared to CFD and
attributable to the compressible flow simplifications in his analysis, and a similar shift subtracting an
increment of 0.1 in optimal Mach number would bring the absolute results much closer to CFD.
CL
⏟
Maximize M subjected to M 2 CL = 𝑙 and V ≥ Vinitial
α ,M
CD + kCL2
Eq. 6.13
objective function, this change will at some point balance out to result in a shocked optimum. Range
optimizations are now presented for the NACA0012 at three different values of non-dimensional
wing loading: l = 0.35, 0.4, 0.45. An initial set of optimizations are presented for no induced drag
penalty (k = 0) to provide a datum. For these optimizations, the shape and Mach number are allowed
to change with a constraint placed on the internal volume. The optimization results are presented in
Table 6.2.
The three optimizations have produced shock-free solutions [see Poole et al.]254. Furthermore, as was
shown in the analytical treatment, higher values of l result in lower optimum range factors, which
was also found in the
optimization results.
However, in the analytical
treatment it was also shown
that the optimum is
supercritical assuming that
the lift value chosen is
sufficiently high, but this does
not necessarily lead to a
shocked solution and, in fact,
all of the optimizations
presented are supercritical
(as shown in the surface Mach
plot), though shock-free. If
freestream Mach number was
to be increased further, to
where a shock forms, then Table 6.3 Results for Inviscid Range Optimizations with k = 0.04
wave drag increases Induced Drag Factor - (Courtesy of [Poole et al.])
approximately with the
fourth power of Mach number, according to Lock255.
The subsequent increase in objective function due to the increase in Mach number is off-set by the
reduction in objective function due to the increase in wave drag. Hence, the optimizer has increased
lift coefficient to maximize the objective function, at the expense of higher Mach numbers leading to
a shock-free solution. Optimizations for a value of induced drag representative of a wing with aspect
ratio 7.95 - which is reasonably typical for a conventional jet-powered airliner{(k = 0.04) are now
presented. Table 6.3 shows the results for these while Figure 6.10 shows the optimized pressure
distributions for these optimizations. The induced drag penalty has acted to lower the optimal lift
coefficient and to increase the optimal Mach against not having an representation of induced drag. It
is clear that for an induced drag penalty factor that is a representative value, shocked optima result
for all four of the l values considered. The general trend in the optimum solution with an increasing
value of l is that the optimum Mach number reduces. The reduction in the optimal range that would
result from a lower Mach number is compensated for by a greater increase in lift coefficient,
indicating that any further increase in the Mach number would substantially increase the wave drag.
The final surface shapes of each of these optimizations are shown in Figure 6.10 demonstrating that
not only are shocked solutions forced, but that the global form of the surfaces are reasonably
independent of the value of l chosen for a fixed k. It is also interesting to note that the resulting shapes
are supercritical in nature, displaying the at upper surface pressure distribution with delayed shock,
254 D.J. Poole, C.B. Allen, T.C.S. Rendallz, “Comparison of Point Design and Range-Based Objectives for Transonic
Airfoil Optimization”, AIAA Journal · June 2018.
255 Lock, C. N. H., “The Ideal Drag Due to a Shock Wave Parts I and II," Tech. rep., Aeronautical Research Council,
and the trailing edge cusp. This result is particularly promising considering the airfoils were all
initialized as a NACA0012.
Figure 6.10 Surface Shapes and CP for Inviscid Range Optimizations of NACA0012 at Different Lift Values
with Induced Drag Penalty - (Courtesy of [Poole et al.])
⏟
Minimize CD subjected to CL ≥ CL initial and V ≥ Vinitial
α
Eq. 6.14
Second, a multi-point optimization is also presented. The conventional weighted-sum approach,
where the overall objective function is a weighted sum of the objective from the individual design
points, is considered. A lift constraint on each of the design points is used as well as the volume
constraint on the overall shape. For N design points, where each has an objective weighting, λi, the
multi-point optimization problem is therefore given as:
256Chernukhin, O. and Zingg, D. W., “Multimodality and Global Optimization in Aerodynamic Design," AIAA
Journal, Vol. 51, No. 6, 2013, pp. 1342-1354.
146
⏟
Minimize ∑ 𝜆𝑖 𝐶𝐷𝑖 subjected to 𝐶𝐿𝑖 ≥ 𝐶𝐿𝑖 𝑖𝑛𝑡𝑖𝑎𝑙 1 ≤ i ≤ N
α 𝑖=1
and V ≥ Vinitial
Eq. 6.15
where the single-point optimization is performed on case 2 and the multi-point optimization is
performed for equal weightings on each case, hence the objective is (1/2)CD1+(1/2)CD2. As before,
12 modal design parameters are used for the surface deformations. The results of the viscous
optimizations are given in Table 6.4. The final drag value of the single-point case at design condition
2 (which is the optimized condition) is slightly lower than the drag value of the multi-point case at
the same condition and this is to be expected since the trade-off with the second design condition in
the multi-point case restricts, somewhat, the result compared to considering one design point in
isolation. The surface shapes of the two optimizations (given in Figure 6.12-(c) as well as the
resulting pressure coefficients (given in Figure 6.12 (a-b) differ considerably. Both solutions reduce
the lift over the fore body of the airfoil to minimize the leading edge acceleration and therefore avoid
a shock forming while the lift is recovered closer to the trailing edge. However, the suction peak for
condition 2 is pronounced in the single-point case, whereas this is more suppressed in the multi-
point. The multi-point case has resulted in a slightly more cambered airfoil. Clearly, the addition of
condition 1 to the optimization has meant a slight compromise has been required in the multi-point
optimization, which is unsurprising. For further information, please refer to [Poole et al.]258.
257 Drela, M., “Pros and Cons of Airfoil Optimization," in Caughey, D. and Hafez, M., eds., “Frontiers of
Computational Fluid Dynamics," World Scientific, pp. 363-381, 1998.
258 D.J. Poole, C.B. Allen, T.C.S. Rendallz, “Comparison of Point Design and Range-Based Objectives for Transonic
viscous range optimization is also performed with the induced drag factor considered in the inviscid
cases, though two values are considered here: k = 0.04 and k = 0.1. The starting airfoil is the RAE2822,
which is in flow at Re = 6.5
X 106 (as per the drag
K = 0.04
minimizations).
The final optimization
results are given in Table
6.5. The first clear trend is
that for increasing non-
dimensional wing loading,
the optimum range
reduces for both induced
drag penalties
considered. Furthermore,
the increase in induced
drag penalty means that
higher lift coefficients are
penalized so the resulting
optimum design point is
at a higher Mach number
and lower lift coefficient. K = 0.1
These were both seen in
the inviscid results, so
these trends are expected.
The surface pressure
distributions of the range
optimizations are given in
Figure 6.11 (a-b) and
the surface shapes in
Figure 6.11 (c-d). An
interesting result is that
despite the fact that an
induced drag penalty has
been added, the k = 0.04
value is not enough to
force a shocked solution.
This is further shown in
Table 6.5 Results for Viscous Range Optimizations with k = 0:04 and
Figure 6.11(e-f), which
0.1 - (Courtesy of [Poole et al.])
gives the field pressure
contours of the
optimizations at the two different induced drag values at l = 0.35, which clearly show that the lower
value is not sufficient to penalize a shock-free solution. On the other hand, moving to a higher induced
drag factor is enough to create an optimum at a high enough Mach and CL combination such that a
shock-free optimum is not possible.
148
(c) Surface Shapes for k = 0.04 (d) Surface Shapes for k = 0.1
Figure 6.12 Surface Shapes for Drag Minimizations
Figure 6.11 Surface Shapes & Cp for Viscous Range Optimizations - (Courtesy of [Poole et al.])
149
Figure 6.14 M-CL maps with normalized ML/D Contours of Optimized airfoils (97% contour highlighted)
- (Courtesy of [Poole et al.])
150
by which gives a measure of robustness. (please refer to [Poole et al.]259 for full description). This
area has been calculated and the results are given for all of the viscous results. Clearly, moving to the
range optimization has resulted in an overall larger area in the M-CL space that contains high range
performance, hence the range optimizations are more robust to changes in the operating conditions.
It is also observed that the optimum range often does not occur at (or even close to) the final
optimized design point. Since the optimum point for the range optimizations must lie on the non-
dimensional wing loading constraint, the optimization is somewhat restricted. Since there is a better
range that exists at a different value of l, the question is raised of is there an overall optimum value
of l.
259 D.J. Poole, C.B. Allen, T.C.S. Rendallz, “Comparison of Point Design and Range-Based Objectives for Transonic
Airfoil Optimization”, AIAA Journal · June 2018.
260 Christian B. Allen, Daniel J. Poole, Thomas C. S. Rendall, “Wing aerodynamic optimization using efficient
parameters have previously been tested on geometric shape recovery problems and aerodynamic
shape optimization in two dimensions, and shown to be efficient at covering a large portion of the
design space; the work is extended here to consider their use in three dimensions. Wing shape
optimization in transonic flow is performed using an upwind flow-solver and parallel gradient-based
optimizer, and a small number of global deformation modes are compared to a section-based local
application of these modes and to a previously-used section-based domain element approach to
deformations. An effective geometric deformation localization method is also presented, to ensure
global modes can be reconstructed exactly by superposition of local modes. The modal approach is
shown to be particularly efficient, with improved convergence over the domain element method, and
only 10 modal design variables result in a 28% drag reduction.
261 Hicks RM, Henne PA, “Wing design by numerical optimization”. J Aircr 15(7):407–412, 1978.
262 Qin N, Vavalle A, Le Moigne A, Laban M, Hackett K, Weinerfelt P, “Aerodynamic considerations of blended wing
47(2):638–646, 2010.
264 Lyu Z, Kenway GKW, Martins JRR, “Aerodynamic shape optimization investigations of the common research
element parameterization and mesh deformation”. Inter J Numerical Meth Fluids 58(8):827–860, 2008.
267 Morris AM, Allen CB, Rendall TCS, “Domain-element method for aerodynamic shape optimization applied to a
functions for shape parameterization and mesh deformation”. Optimum Eng. 14:97–118.
152
The quality of the optimization result obtained is driven, primarily, by the quality and type of
numerical optimization algorithm used in the ASO framework, and the two primary types of
optimization algorithms are local methods and global methods. The local methods are usually built
around the gradient-based approach, which uses the local gradient of the design space as a basis
around which to construct a search direction. The optimization algorithm therefore traces a
movement path through the design space until the gradient values become very small where the
result has converged. These approaches are the most common methods used in the ASO framework,
driven primarily by the low cost associated with them compared to global methods, and an efficient
gradient-based optimizer is used here.
The aerodynamic model defines the parameter space of the problem, but the problem design space,
which the optimization algorithm interrogates, is constructed to fully interrogate the true design
space (which contains every possible design) is driven by the ability of the degrees of freedom of the
parameterization scheme to represent any shape within the design space, and so this is a critical
aspect of any optimization scheme. The level of flexibility generally increases with the number of
design variables, but the use of a low number of design variables is advantageous, since good
convergence of optimization algorithms tends to correlate with small numbers of design variables,
and so there is a definite requirement for an efficient parameterization scheme. The work presented
at this juncture is considers aerodynamic shape optimization using a novel method of deriving design
variables. The design variables used here are derived by a mathematical technique that is based on
singular value decomposition (SVD), that extracts an orthogonal set of geometric ‘modes’. The
method itself has been presented recently by the authors [Poole et al.]269, and has been shown to
outperform other commonly-used parameterization schemes (Masters et al.]270 when considering
geometric inverse design in two dimensions, often requiring fewer than a dozen variables to
represent a large design space [Poole et al.]271.
Methods are classified as either constructive, reformative or unified. In-depth reviews have been
269 Poole DJ, Allen CB, Rendall TCS, “Metric-based mathematical derivation of efficient airfoil design variables”.
AIAA J 53(5):1349–1361, 2015.
270 Masters DA, Taylor NJ, Rendall TCS, Allen CB, Poole DJ, “Geometric comparison of airfoil shape
design variables with a constrained global optimizer”. Computer Fluids 143:1–15, 2017.
272 L. J. Kedward , A. D. J. Payot y, T. C. S. Rendall, C. B. Allen, “Efficient Multi-Resolution Approaches for
presented by [Samareh]273, [Castonguay and Nadarajah]274, [Mousavi et al. ]275 and [Masters et al.
]276. Constructive methods consider the definition of the surface and the deformation of the surface
separately. Examples of these methods are CST [Kulfan]277, PARSEC [Sobieczky]278, PDEs [Bloor and
Wilson]279 and splines [Braibant and Fleury]280. Other approaches that combine various
parameterizations in a hybrid approach can also be found. Because of the constructive nature of these
approaches, perturbation of the base geometry through the optimization process requires that the
new surface be reconstructed, which subsequently requires automatic mesh generation tools for
production of a new surface and volume mesh. This extra difficulty can make it advantageous to
consider approaches that manipulate an existing mesh.
An alternative to constructive methods are deformities methods which unify the geometry creation
and perturbation. This tends to make them simpler to integrate with mesh deformation tools and
allows the use of previously generated meshes; a considerably cheaper alternative to regeneration,
although the mesh deformation scheme is a separate algorithm. Analytic [Hicks and Henne]281 and
discrete [Jameson]282 methods are examples of deformities approaches. A further refinement of
unifying geometry creation and perturbation is the integration with a mesh deformation algorithm.
Methods of this type typically have some interpolation that describes a link between the surface and
volume, often via a set of control points that are independent of both, such that deformation of the
control points results in deformation of the surface and CFD mesh. These approaches are commonly
used in ASO, and the methods included in this unified category are free-form deformation
[Samareh]283, domain elements (Morris et al.]284 and direct manipulation [Yamazaki et al.]285.
Surface parameterizations developed around the FFD and domain element approach are very
popular in wing optimization as this type of approach allows the design space to be reduced from
thousands of design parameters to hundreds of design parameters. Such techniques have been
developed by [Zingg] and colleagues [Hicken and Zingg]286 ; [Leung and Zingg]287, and have shown
273 Samareh JA, “Survey of shape parameterization techniques for high fidelity multidisciplinary shape
optimization”, AIAA J 39(5):877–884, 2001.
274 Castonguay P, Nadarajah SK, “Effect of shape parameterization on aerodynamic shape optimization”, 45th
AIAA Aerospace Sciences Meeting and Exhibit, Reno, Nevada, AIAA Paper 2007–59.
275 Mousavi A, Castonguay P, Nadarajah SK, “Survey of shape parameterization techniques and its effect on three
dimensional aerodynamic shape optimization”.18th AIAA CFD dynamics conference, Florida, 2007.
276 Masters DA, Taylor NJ, Rendall TCS, Allen CB, Poole DJ, “Geometric comparison of aerofoil shape
NW, Keane AJ, Holden CME (2010) Geometric filtration using proper orthogonal decomposition for
aerodynamic design optimization. AIAA J 48(5):916–928, 1998.
279 Bloor MIG, Wilson MJ, “Generating parameterizations of wing geometries using partial differential equations”.
44(3):247–267, 1984.
281 Hicks RM, Henne PA, “Wing design by numerical optimization”. J Aircraft 15(7):407–412, 1978.
282 Jameson A (1988) ,“Aerodynamic design via control theory”. J Scientific Computation 3(3):233–260.
283 Samareh JA, “Novel multidisciplinary shape parameterization approach”. J Aircraft- 38(6):1015–1024, 2001.
284 Morris AM, Allen CB, Rendall TCS, “CFD-based optimization of airfoils using radial basis functions for domain
element parameterization and mesh deformation”, Int J Numer Meth Fluids 58(8):827–860, 2008.
285 Yamazaki W, Mouton S, Carrier G, “Geometry parameterization and computational mesh deformation by
that these types of methods can be flexible enough to allow the molding of a sphere into an aircraft
like shape under certain optimization conditions [Gagnon and Zingg]288. Further work has been
performed by Martins and others [Mader and Martins]289; [Lyu and Martins]290 who showed results
for blended-wing-body optimizations, and [Yamazaki et al.]291 who further reduced the number of
design variables by considering the direct manipulation method for wing optimization. A novel
method, recently developed by the authors, is to extract airfoil design variables using a mathematical
approach. The approach utilizes singular value decomposition in a manner that analyses an initial
library of airfoils and decomposes that library into a reduced set of optimum variables that are
geometrically orthogonal to each another. The method is based on perturbations, so is independent
of the initial geometry and can fit into any of the three categories outlined above; the deformities
formulation is used in this work, to allow a unified application of the design variables to control both
the surface and volume mesh within the ASO framework. Previous work has considered the method’s
ability to represent a wide-range of airfoil shapes [Poole et al.]; [Masters et al.], and their effectiveness
in airfoil optimization [Poole et al.]; [Masters et al.], wherein the efficiency of this modal approach
was clearly demonstrated. Hence, the aim of the work presented here is to develop an effective
method to apply these novel mathematically-extracted design variables, which have been extracted
as two-dimensional quantities, in three dimensions, and determine their effectiveness when applied
to aerodynamic optimization, in particular drag minimization of wings in transonic flow.
Figure 6.15 Volume of solid (VOS) design variables as grey-scale and RSVS profile in red; 1
corresponds to a completely full cell and 0 an empty cell – Courtesy of [Allen et al.]
288 Gagnon H, Zingg DW, “Two-level free-form deformation for high-fidelity aerodynamic shape optimization”.
12th AIAA aviation technology, integration and operations (ATIO) conference and 14th AIAA/ISSMO
multidisciplinary analysis optimization conference, Indianapolis, Indiana, 2012.
289 Mader CA, Martins JRRA, “Stability-constrained aerodynamic shape optimization of flying wings”. J Aircraft
50(5):1431–1449, 2013.
290 Lyu Z, Kenway GKW, Martins JRRA , “Aerodynamic shape optimization investigations of the common research
the fraction of each cell within a geometry built from this information. This process is shown for a
simple grid in Figure 6.15. This parameterization procedure provides intuitive handling of topology
change without maintaining explicit control of it. It is important that topology is not controlled
explicitly as this would lead to a severely discontinuous design space which would not be usable with
many of the traditional local and global optimizers used for aerodynamic optimization. Further
details can be obtained from [Kedward et al.]292.
Figure 6.16 Four Levels of Subdivision of a Four Point Control Polygon - Courtesy of [Allen et al.]
This subdivision matrix encompasses two operations: a uniform topological refinement of the mesh
(splitting) and a smoothing of the result (averaging), demonstrated in Figure 6.16. Both operations
are local and can hence be performed very efficiently. [Kedward et al.]293.
‖𝐌 − 𝐌 k ‖F ≤ ‖𝐌 − 𝐦‖F
Eq. 6.17
where m is any matrix of rank k and ‖ . ‖F is the Frobenius norm. Hence, the error in the low rank
approximation (found from SVD) will always be at least as good as the error between any other rank
k matrix and the full rank matrix. The SVD thus produces an optimal low order projection of the
higher dimensional space into the lower dimensional one, which is significant for optimization
parameters. The SVD method first requires a training library of Na airfoils to be collated from which
the airfoil deformation modes are extracted. Each airfoil surface is parameterized by N surface points,
where the i-th surface point has a position in the space (xi , zi). To ensure consistency of the surface
description of the training data all airfoils are parameterized with the same parametric distribution.
The x distribution is often defined as the controlling parameterization, with zi = f (xi), but this is not
the most flexible approach; instead all airfoil surfaces are parameterized in terms of peripheral
distance s ∈ [0 , 1] and then exactly the same si distribution is defined for all airfoils. Following the
surface point distribution, each airfoil has a rigid body translation, scaling and then rotation applied
to it to map the geometry into a consistent form where each section has unit chord and z(0) = z(1) =
0: A matrix is built from which SVD is performed, by evaluating the vector difference of the i-th
surface point between all airfoils, producing Ndef = Na(Na -1)/2 airfoil deformations. The x and z
deformations are stacked into a single vector of length 2N, for each airfoil deformation, so a matrix is
built of the airfoil deformations which has 2N rows and Ndef columns:
∆x1,1 ⋯ ∆x1,Ndef
⋮ ⋱ ⋮
∆xN,1 ⋯ ∆xN,Ndef
𝐌=
∆z1,1 … ∆z1,Ndef
⋮ ⋱ ⋮
[ ∆zn,1 … ∆zn,Ndef ]
Eq. 6.18
Performing a SVD decomposes the matrix into three constituent matrices:
𝐌 = 𝐔𝚺𝐕 𝐓
Eq. 6.19
where U is a matrix of vectors, each of length 2N. The structure is analogous to the decomposed
matrix, so the columns of this matrix are the airfoil mode shapes. Σ is a diagonal matrix of the singular
values, arranged in descending order. These can be considered the ‘relative energy’ of the modes, and
represent the ‘importance’ of the mode shapes in the original library. The total number of possible
mode shapes is governed by the number of singular values, which is the minimum of the number of
columns and rows of the decomposed matrix. A truncation of the U matrix, based on a certain total
energy required, then gives the number of design variables used in the optimization. The training
library is based on deformations, and this is an important choice such that design variables that result
from the decomposition are also deformations, ensuring they are independent of the topology of the
airfoils that are used. This allows direct insertion into an aerodynamic shape optimization
framework where deformation of the surface and mesh is important. If the constructive formulation
is used, however, then the columns of the training matrix, M, are absolute positions of the airfoil
surface points as opposed to deformations between surface points.
In this work, a generic, non-symmetric training library is considered based on the optimization being
performed. The library contains 100 different airfoils, extracted from a larger library by quantifying
their performances in the transonic regime using the Korn technology factor [Poole et al.]294. The first
six modes of the library are shown in Figure 6.17; all modes are scaled up for illustration purposes,
and have been added to a NACA0012 section. Also the relative ‘energy’ of the first 20 modes, i.e. the
singular values of each mode normalized by the sum of all singular values. Once the design variables
Figure 6.17 Generic Non-Symmetric Airfoil Modes - a Mode 1. b Mode 2. c Mode 3. d Mode 4. e Mode 5.
f Mode 6 – Courtesy of [Allen et al.]
294Poole DJ, Allen CB, Rendall TCS, “Metric-based mathematical derivation of efficient airfoil design variables”.
AIAA J 53(5):1349–1361, 2015.
158
have been extracted and the total number of modes has been truncated, a new airfoil can be formed
by a weighted combination of m modal parameters, as shown in Eq. 6.20. The weighting vector, β,
represents the magnitude of the modal deformations which are then the design variable values that
the optimizer works with. The truncation of the total number of modes, which is often very large,
down to a number which is useful for the optimization can either be user-specified or based on the
requirement for a total amount of energy to be preserved, e.g., if 99.0% of the energy of the original
library is required to be preserved then the first, say, six modes may cumulatively have 99.1% of the
energy so six modes would be used. In this work, a number of modes is specified and those modes
with the highest amount of energy are taken.
m
new old
𝐗 =𝐗 + ∑ βn 𝐔 n
n=1
Eq. 6.20
The modes extracted here are two-dimensional deformations, based on a large database of
aerodynamic surfaces, and so are effective in two dimensions, and this has been proven previously
[Poole et al.]. Hence, it would make sense to consider a similar approach in three dimensions.
However, this would require a database of wings, something that would not be easy to create, and
with variable parameters such as taper and sweep, and surface discontinuities such as crank
locations, would also require a complex parametric transformation to a normalized space.
Furthermore, this would still not contain global variables, and so it is more sensible and flexible to
consider a more local sectional approach. This is the approach considered here.
For surface and volume mesh deformation, it is sensible to use Polynomials are not included here,
due to their growing radial influence, and so (superscript c represents a control point):
∆x1c α1x
∆𝐗 𝐜 = [ ⋮ ] , 𝛂n = [ ⋮ ]
c
∆xnc αxnc
Eq. 6.23
(analogous definitions hold for y and z coordinates) and the control point dependence
matrix, C, takes the form
φ11 … φ1nc
𝐂=[ ⋮ ⋱ ⋮ ], where φij = φ(‖𝐱 ic − 𝐱 jc ‖)
φnc 1 … φncnc
Eq. 6.24
For surface and volume mesh deformation, it is sensible to use decaying basis functions, to give the
interpolation a local character and ensure deformation is contained in a region near the moving body,
and [Wendland’s] C2 function is used here. It is also sensible to omit polynomial terms, since these
will transfer deformation throughout the entire mesh. Hence, in the case considered here the global
influence on any point in the aerodynamic mesh (denoted by superscript a) from the control points
is determined by Eq. 6.21, which is applied as
n𝑐
performance, using the Korn technology factor, and a form of library filtering applied to down-select
the library; see [Poole et al.]. A set of control points is used to control the aerodynamic surface; these
points are independent of a base geometry, and the surface deformations are defined in terms of
perturbations so the modes can be applied to any geometry. A ‘shrink-wrapping’ method is used to
map them onto the geometry being considered. Here, 24 control points are used; more than this are
not necessary unless small wavelength changes are required. The modal deformations can be defined
using a larger number of points than the N value in Eq. 6.18 and projected onto these 24 points, but
here N = 24 is used in the SVD extraction process. Figure 6.18 shows the surface control points and
an example deformation of the fourth mode for a NACA0012 mesh.
Figure 6.18 Surface-Based Control Points and Example Deformation. a Control Points. b Example
Deformation - Courtesy of [Allen et al.]
161
Figure 6.19 Surface Mesh and Control Points in 3D – Courtesy of [Allen et al.]
162
where Rj = R ( α0 +αtwist +αpitch) and xrj is the local rotation center. Hence, in this case there are m
design parameters. For local deformations, i.e. one design variable for each mode at each of the ns
sections, this can be formulated as:
𝐧𝒔 𝐦
where φ (j , s) is a basis function. In this case there are m _ ns design parameters. Hence, this basis
function can be used to determine how the deformation of each of the ns sections affects the other
sections, i.e. controls the zone of influence. This can be left to the global interpolation, but is defined
here to allow control of the decay. A basis function can be defined such that if the effect of the
sectional deformation decays to zero at the neighboring sections each side, a global modal
deformation can be recovered exactly. In this case βn would have a single value for all sections, and
so it can be shown that if
ns
∑ φ (j, s) = 1
s=1
Eq. 6.29
Figure 6.20 Surface and Control Point Modal Deformations. a Mode 1 global. b Mode 3 global. c Mode 5
global. d Mode 1 local. e Mode 3 local. f Mode 5 local – Courtesy of [Allen et al.]
163
at any span wise point, Eq. 6.26 and Eq. 6.27 are equivalent. Hence, a trigonometric function of (j,
s) is used. Figure 6.20 shows the control locations and resulting surface mesh for deformations
using the first, third, and fifth modes; the upper row shows a global modal deformation, and the lower
row shows local modal deformations of the fifth control point section. The modal deformation
magnitude is exaggerated to 10% local chord for illustration purposes. This improved localization
process is also adopted to improve the application of off-surface domain element perturbations, used
previously by the authors. Figure 6.21 shows two views of the surface and domain element points,
again using ns = 10 and nint = 4. An exaggerated movement of all the points on the fifth section is
shown; magnitude 20% local chord. (see [Allen et al.]295).
Figure 6.21 Surface Mesh and off-Surface Control Points – Courtesy of [Allen et al.]
Within the context of aerodynamic shape optimization, the presence of a multi-modal search space
is highly dependent on the extent of the surface representation and the fidelity of the flow analysis
tool. The issue of degree of multi-modality in aerodynamic optimization problems is an unanswered
question, with work presented showing that multimodality exists in a number of cases, but unimodal
295Christian B. Allen, Daniel J. Poole, Thomas C. S. Rendall, “Wing aerodynamic optimization using efficient
mathematically-extracted modal design variables”, Optimum Engineering, 2018.
164
cases also exist (Namgoong et al]296, [Khurana et al.]297; [Buckley et al.]298; [Chernukhin and Zingg].
[Chernukhin & Zingg] have considered this issue by testing a number of different optimization
problems and have shown that for a b-spline parameterization of the surface, viscous, compressible
drag minimization of the RAE2822 airfoil has one global optimum. They also showed multiple local
optima for other three-dimensional problems. For maximum flexibility and efficiency, a gradient-
based method is used here, with a second-order finite-difference approach for gradient evaluation.
This approach allows a ‘wrap-around’ approach, i.e. any flow-solver can be implemented within the
framework.
1 T
Minimize: ∆𝛃 𝐇∆𝛃sqp + ∇(𝛃)T ∆𝛃 sqp
2 sqp
Subject to: ∇g i (𝛃)T ∆𝛃sqp + g i (𝛃) ≤ 0 i = 1, , , , , G
Eq. 6.31
where J is the objective function and gi is the ith inequality constraint of a total of G inequality
constraints. FSQP augments the SQP descent direction by a feasible step direction, Δβf , that is a
fraction of either ⊽J(β) or ⊽gi(β), depending on the constraint value. The overall step direction is
then a blend of the SQP and feasible step directions
296 Namgoong H, Crossley W, Lyrintzis AS, “Global optimization issues for transonic airfoil design”, 9th
AIAA/ISSMO symposium on multidisciplinary analysis and optimization, Atlanta, AIAA Paper 2002–5641.
297 Khurana MS, Winarto H, Sinha AK, “Airfoil optimization by swarm algorithm with mutation and artificial
neural networks.” ,47th AIAA aerospace sciences meeting including the new horizons forum and aerospace
Exposition, Orlando, Florida, AIAA Paper 2009–1278.
298 Buckley HP, Zhou BY, Zingg DW, “Airfoil optimization using practical aerodynamic design requirements”. J
correction direction is found such that both descent and feasibility are ensured by solving a further
quadratic programmed while avoiding the need for further constraint and function evaluations. The
exact implementation of the rules that govern the computation of the step size are given in Zhou et
al. (1997). The Hessian, or an approximation to the Hessian, at every major iteration is required,
which in turn requires sensitivity of the objective function and constraints with respect to the design
variables. The Hessian is approximated by the Broyden–Fletcher–Goldfarb–Shanno (BFGS) update
scheme where the Hessian approximation is initialized as the identity matrix. The gradients are
obtained by a second Aerodynamic optimization using efficient modal variables order central-
difference scheme, so the number of objective function evaluations is proportional to the number of
design variables. Once the search arc has been determined, the no monotone line search proceeds
(Zhou and Tits 1993). For this, the conventional backtracking line search requirement of requiring a
suitable reduction in the objective function from iteration t to t + 1 is relaxed, such that a reduction
in the objective function to iteration t + 1 is required against the maximum objective function from
the last four iterations. This has been shown to be highly effective for unconstrained optimization
when compared to a conventional backtracking search (Grippo et al. 1986), and when implemented
for FSQP shows similar results for constrained optimization (Zhou and Tits 1993).
The algorithm iterates until the Kuhn–Tucker conditions are satisfied, which then represent a
converged solution using a constrained gradient-based optimizer. For computational efficiency, the
sensitivity evaluation has been parallelized based on the number of design variables such that the
evaluation of the sensitivity of the objective function and constraints with respect to the design
variables is split between the number of CPUs available (Morris et al. 2008, 2009). This is necessary
as within the ASO environment, an objective function evaluation represents a CFD solution, so this
formulation allows parallel evaluation of the required sensitivities; second-order finite-differences
are used for the sensitivities. It is well known that the accuracy of the gradient evaluation is a critical
issue, and the authors have performed several studies on perturbation size for finite-differences; see
for example Morris et al. (2008). A relative perturbation of 10-4 is adopted here, i.e. a deformation
magnitude of 0.01% of local chord. Constraint and step-size evaluations and optimizer updates occur
in the master process, and each CPU controls the geometry (and CFD volume mesh) perturbations
corresponding to the different design variables, and calls the flow solver. Flow-solver results are then
returned to the master for optimizer updates.
301 Allen CB, “Multigrid convergence of inviscid fixed- and rotary-wing flows”. International J Numerical Meth
Fluids 39(2):121–140, 2002.
302 Allwright S, “Multi-discipline optimization in preliminary design of commercial transport aircraft”.
wing optimization for drag minimization subject to strict constraints is investigated, and so the
problem definition is:
1. Twist case to address the induced drag effect, a simple case was first run using a global linear
twist variable, plus a global pitch variable to allow lift balancing. This results in two variables.
2. Individual point deformation case Conventional off-surface domain element, with individual
deformations of each point, normal to the local chord, in each of the 10 slices, plus a global
linear twist variable and a global pitch variable to allow lift balancing. This results in 10 x16
+ 2 = 162 variables.
3. Global mode case Global modal deformations of all 10 sectional slices using 6, 8, and 10
modes, a global linear twist variable, plus a global pitch variable to allow lift balancing. This
results in 6 + 2; 8 +2 or 10+ 2 = 8; 10 or 12 variables. A global mode is a single deformation
of all control points, with the modes scaled and rotated according to the local geometry.
4. Local mode case Local modal deformation using 6, 8, and 10 modes at each of the 10 sectional
locations, a global linear twist variable, plus a global pitch variable to allow lift balancing. This
results in 10 x 6 + 2; 10 x 8 + 2 or 10 x 10 + 2 = 62 ; 82 or 102 variables. Again at each section,
the local modes are scaled and rotated according to the local geometry. Global modes are not
included, since these can be recovered exactly from a combination of the local modes.
303 Allen CB, “Towards automatic structured multi-block mesh generation using improved transfinite
interpolation”. International J Numerical Meth Eng. 74(5):697–733, 2008.
167
Figure 6.22 Domain and Block Boundaries and far-field Mesh – Courtesy of [Allen et al.]
6.7.10.2 Results
Table 6.6 presents results for the four sets of variables. The twist variables are clearly effective at
reducing the induced drag, and the finer surface deformations then reduce the pressure and wave
drag. Figure 6.23 shows the upper surface pressure contours for the baseline case, and
optimizations using 16-point domain element deformations at each section, 10 global modes, and 10
local modes. Sectional pressure coefficient variations are also presented in [Allen et al.]304. The
convergence histories in terms of iterations and function evaluations are shown in Figure 6.24 in
terms of the objective function convergence. Also shown in Table 6.6 is the total CPU time, i.e. the
304Christian B. Allen, Daniel J. Poole, Thomas C. S. Rendall, “Wing aerodynamic optimization using efficient
mathematically-extracted modal design variables”, Optimum Engineering, 2018.
168
number of objective evaluations (flow solutions) multiplied by run-time per solution. All cases were
run on the University of Bristol HPC cluster, comprising Intel Sandy Bridge 2.6 GHz cores. Also shown
in the table are optimization run times and the number of cores used for each. Note that the costs
presented do not include the cost of the twist-only case run first. The optimizer adopts a second-
order central finite-difference gradient evaluation, and so each iteration requires two flow solutions
per variable, and a further one or two solutions for the step size evaluation, and the step size
evaluation is always performed in serial on the master node. All cases were run with one core per
design variable and one for the master process. Hence, the total cost scales with the number of design
variables, but the parallel framework means the optimization run time scaling can be reduced to the
number of iterations. It is clear that the global modes are particularly efficient, requiring significantly
fewer evaluations than the off-surface domain element, for similar drag reduction. However, neither
of these approaches has eliminated the wave drag entirely, whereas the local modes have achieved
Figure 6.23 Upper Surface Pressure Coefficient. a Initial Geometry. b Domain Element. c 10 Global
Modes ,d 10 Local Modes – Courtesy of [Allen et al.]
169
this for significantly lower cost than the domain element approach. an improved distribution.
6.7.11 Conclusions
Aerodynamic shape optimization has been considered, using mathematically derived design
variables. Orthogonal design variables have been extracted by a singular value decomposition
approach where a training library of airfoils is analyzed and decomposed to obtain an efficient and
reduced set of design variables. They are geometric ‘modes’ of the original library, representing
typical airfoil design parameters. In the aerodynamic shape optimization framework a surface and
mesh deformation algorithm is required, and a control point approach has been adopted. This adopts
a small number of control points which are linked to the numerical mesh points by a global volume
interpolation using radial basis functions to allow large, smooth deformations of the mesh. The
performance of the mathematical design variables has been demonstrated in three dimensions, with
results of optimization of the MDO wing in transonic flow. The modal deformations have been applied
as both local and global variables. An important aspect of effective geometric application of these
two-dimensional variables is localization of the deformation field. A basis function deformation
control approach has been developed and presented, allowing improved local control of
deformations, and ensuring exact recovery of global modes from local modes. It has been
demonstrated that the modal approach gives better results than the domain element approach, for
significantly fewer design variables and, furthermore, using global modes, an impressive result is
achieved with only O(10) variables.
305Martins, J. R. R. A. and Hwang, J. T., “Review and Unification of Methods for Computing Derivatives of
Multidisciplinary Computational Models," AIAA Journal, Vol. 51, No. 11, pp. 2582-2599, 2013.
170
operating cost and the emission of green-house gases. Despite considerable research on
aerodynamic shape optimization, there is no standard benchmark problem allowing researchers to
compare results. This was also address by [Mavriplis]306 which he complained about the lack of
certification by analysis. Aerodynamic shape optimization can be dated back to the 16th century,
when Sir Isaac Newton307 used calculus of variations to minimize the fluid drag of a body of revolution
with respect to the body's shape. Although there were many significant developments in
optimization theory after that, it was only in the 1960s that both the theory and the computer
hardware became advanced enough to make numerical optimization a viable tool for everyday
applications.
The application of gradient-based optimization to aerodynamic shape optimization was pioneered in
the 1970s. The aerodynamic analysis at the time was a full-potential small perturbation inviscid
model, and the gradients were computed using finite differences. [Hicks et al.]308 first tackled airfoil
design optimization problems. [Hicks and Henne]309 then used a three-dimensional solver to
optimize a wing with respect to 11design variables representing both airfoil shape and the twist
distribution.
Because small local changes in wing shape have a large effect on performance, wing design
optimization is especially effective for large numbers of shape variables. As the number of design
variables increases, the cost of computing gradients with finite differences becomes prohibitive. The
development of the adjoin method addressed this issue, enabling the computation of gradients at a
cost independent of the number of design variables. For a review of methods for computing
aerodynamic shape derivatives, including the adjoin method, see [Peter and Dwight]310. For a
generalization of the adjoin method and its connection to other methods for computing derivatives,
see [Martins and Hwang]311.
6.8.1 Methodology
This section describes the numerical tools and methods that we used for the shape optimization
studies. These tools are components of the framework for multidisciplinary design optimization
(MDO) of aircraft configurations with high fidelity312. It can perform the simultaneous optimization
of aerodynamic shape and structural sizing variables considering aero-elastic directions313.
However, here we use only the components which are relevant for aerodynamic shape optimization,
namely, the geometric parametrization, mesh perturbation, CFD solver, and optimization algorithm.
306 Dimitri Mavriplis, Department of Mechanical Engineering, University of Wyoming and the Vision CFD2030
Team, “Exascale Opportunities for Aerospace Engineering”, AIAA 2007-4048.
307 Newton, S. I., Philosophic Naturalis Principia Mathematica, Londini, jussi Societatus Regiaeac typis Josephi
Multidisciplinary Computational Models," AIAA Journal, Vol. 51, No. 11, pp. 2582-2599, 2013.
312 Kenway, G. K. W., Kennedy, G. J., and Martins, J. R. R. A., “Scalable Parallel Approach for High-Fidelity Steady-
State Aero-elastic Analysis and Adjoint Derivative Computations," AIAA Journal, Vol. 52, No. 5, pp. 935-951, 2014.
313 Kenway, G. K. W. and Martins, J. R. R. A., “Multi-point High-fidelity Aero-structural Optimization of a Transport
Aircraft Configuration," Journal of Aircraft, Vol. 51, No. 1, pp. 144-160, 2014.
171
Figure 6.25 Shape Design Variables are the z-Displacements of 720 FFD Control Points - (Courtesy of
Martins and Hwang)
The FFD volume parametrizes the geometry changes rather than the geometry itself, resulting in a
more efficient and compact set of geometry design variables, thus making it easier to manipulate
complex geometries. Any geometry may be embedded inside the volume by performing a Newton
search to map the parameter space to the physical space. All the geometric changes are performed
on the outer boundary of the FFD volume. Any modification of this outer boundary indirectly
modifies the embedded objects. The key assumption of the FFD approach is that the geometry has
constant topology throughout the optimization process, which is usually the case in wing design. In
addition, since FFD volumes are B-spline volumes, the derivatives of any point inside the volume can
be easily computed. Figure 6.25 demonstrations the FFD volume and the geometric control points
(red dots) used in the aerodynamic shape optimization. The shape design variables are the
displacement of all FFD control points in the vertical (z) direction.
314 Carrier, G., Destarac, D., Dumont, A., Meheut, M., Din, I. S. E., Peter, J., Khelil, S. B., Brezillon, J., and Pestana,
M., “Gradient-Based Aerodynamic Optimization with the elsA Software,” 52nd Aerospace Sciences Meeting, 2014.
315 Kenway, G. K., Kennedy, G. J., and Martins, J. R. R. A., “A CAD-free Approach to High-Fidelity Aero-structural
Optimization,” Proceedings of the 13th AIAA/ISSMO Multidisciplinary Analysis Optimization Conference, Fort
Worth, TX, 2010.
316 De Boor, C., A Practical Guide to Splines, Springer, New York, 2001.
172
algebraic and linear elasticity methods, developed by [Kenway et al.]317. The idea behind the hybrid
scheme is to apply a linear-elasticity-based perturbation scheme to a coarse approximation of the
mesh to account for large, low-frequency perturbations, and to use the algebraic warping approach
to attenuate small, high-frequency perturbations.
317 Kenway, G. K., Kennedy, G. J., and Martins, J. R. R. A., “A CAD-free Approach to High-Fidelity Aero-structural
Optimization," Proceedings of the 13th AIAA/ISSMO Multi-disciplinary Analysis Optimization Conference”, 2010.
318 Saad, Y. and Schultz, M. H., “GMRES: A Generalized Minimal Residual Algorithm for Solving Non-symmetric
Linear Systems," SIAM Journal on Scientific and Statistical Computing, Vol. 7, No. 3, 1986, pp. 856-869.
319 Gill, P. E., Murray, W., and Saunders, M. A., “SNOPT: An SQP Algorithm for Large-Scale Constrained
Optimization," SIAM Journal on Optimization, Vol. 12, No. 4, 2002, pp. 979-1006.
173
Table 6.8 Mesh Convergence Study for the Baseline CRM Wing - (Courtesy of Martins and Hwang)
Table 6.7 Aerodynamic Shape Optimization Problem - (Courtesy of Martins and Hwang)
320Zhoujie Lyu and J. R. R. A. Martins. “Aerodynamic Shape Optimization Investigations of the Common
Research Model Wing Benchmark”. AIAA Journal, 2014.
174
Figure 6.26 Optimized Wing with Shock-Free with 8.5% Lower Drag – (Courtesy of Lyu and Martins)
175
to the theoretical minimum for a planar wake. This is achieved by fine-tuning the twist distribution
and airfoil shapes. The baseline wing has a near-linear twist distribution. The optimized design has
more twist at the root and tip, and less twist near mid-wing. The overall twist angle changed only
slightly: from 8.06 degree to 7.43 degree. The optimized thickness distribution is significantly
different from that of the baseline, since the thicknesses are allowed to decrease to 25% of the
original thickness, and there is a strong incentive to reduce the airfoil thicknesses in order to reduce
wave drag. The volume is constrained to be greater than or equal to the baseline volume, so the
optimizer drastically decreases the thickness of the gained value drag trade off more promising. To
ensure that the result of our single-point optimization has sufficient accuracy, we conducted a grid
convergence study of the optimized design.
successively finer grids is greatly reduced. Thanks to the optimization with the coarser grids, only 18
iterations are needed on the L0 grid to converge the optimization. However, the L0 grid requires the
largest computational effort, due to the high cost of the flow and adjoin solutions on this fine grid.
Given that the cost per optimization iteration in the L0 grid is 770 process-hr (compared to 2.9
process-hr for the L2 grid) it is not feasible to perform an optimization using only the L0 grid.
Assuming that the same number of iterations used for the L2 grid (638) would be needed for the L0
grid, the computational cost would be 23 times higher than that of the multilevel approach, which
would correspond to 16 days using 1248 processors.
performing a multi-point optimization. The optimization is performed on the L2 grid. We choose five
equally weighted flight conditions with different combinations of lift coefficient and the Mach
number. The flight conditions are the nominal cruise, 10% of cruise CL, and 0.01 of cruise M, as shown
in Figure 6.28. More sophisticated ways of choosing multipoint flight conditions and their
associated weights can be used. The objective function is the average drag coefficient for the have
flight conditions, and the moment constraint is enforced only for the nominal flight condition.
A comparison of the single-point and multi-point optimized designs is shown in Figure 6.29. Unlike
the shock-free design obtained with single-point optimization, the multipoint optimization settled on
an optimal compromise between the flight conditions, resulting in a weak shock at all conditions. The
leading edge is less sharp than that of the single-point optimized wing. Additional fight conditions,
such as a low-speed flight condition, would be needed to further improve the leading edge. The
overall pressure distribution of the multipoint design is similar to that of the single-point design. The
twist and lift distributions are nearly identical. Most of the differences are in the chord wise Cp
distributions in the outer wing section. The drag coefficient under the nominal condition is
approximately two counts higher. However, the performance under the off design conditions is
considerably improved.
nominal Mach of 0.85, which corresponds to a dip in a drag divergence plot. For the multipoint
optimization, the optimized flight conditions are distributed in the Mach-CL space, resulting in an
attended ML=D variation near the maximum, which means that we have more uniform performance
for a range of flight conditions. In aircraft design, the 99% value of the maximum ML = D contour is
often used to examine the robustness of the design. The point with the highest Mach number on that
contour line corresponds to the Long Range Cruise (LRC) point, which is the point at which the
aircraft can fly at a higher speed by incurring a 1% increase in fuel burn. In this case, we see that the
99% value of the maximum ML = D contour of the multipoint design is larger than that of the single-
point optimum, indicating a more robust design.
321 J. Brezillon, R.P. Dwight, J. Wild, “Numerical Aerodynamic Optimization of 3D High-Lift Configurations”,
German Aerospace Center, Institute of Aerodynamics and Flow Technology, Braunschweig, Germany, 26th
International Congress Of The Aeronautical Sciences, ICASE 2008.
322 Thiede P. EUROLIFT - Advanced high lift aerodynamics for transport aircraft. AIR&SPACE EUROPE, 2001.
323 Rudnik R, and Geyr H. The European high lift project EUROLIFT II – Objectives, Approach, and Structure. 25th
As a result of these efforts, we present the numerical optimization activity carried out at DLR for the
optimization of 3D multi-element aircrafts. Since nowadays the time required by the design process
is a key element to capture new market, a challenging time constraint for the 3D optimization has
been set to 2 weeks. The purpose here is to assess the tools and methods needed for the optimization
of a 3D high-lift model within a given (short) wall clock time. After exposing the CFD challenges to
simulate an airplane in high-lift configuration, the relevant key technologies to perform the
optimization are presented and then applied on 2 different high-lift configurations. For further
information, please consult the [Brezillon et al]326.
Figure 6.31 Flow Field Around the Wing Section of a 3-Element Wing – Courtesy of [Brezillon et al.]
German Aerospace Center, Institute of Aerodynamics and Flow Technology, Braunschweig, Germany, 26 th
International Congress Of The Aeronautical Sciences, ICASE 2008.
181
In order to capture all of these critical flow features a grid of high quality is required. This implies the
use of thin cells not only normal to the wall for the accurate discretization of the boundary layers but
also behind the trailing edge of the upstream element and up to the downstream element in order to
accurately resolve the wakes until its confluence with the boundary layer of the following elements.
A thick layer of thin cells has then to be generated orthogonal to the wall in order to capture the
mixing shear layers and the ability of the boundary layer to sustain separation. Furthermore, the
high-lift configuration has complicated areas like slots, slat and flap coves, which have to be meshed
as well. The selection of the meshing procedure for high-lift design is then a compromise between
the flexibility to handle complex configurations, the number of cells for fast computations and the
capability to control the space discretization in particular areas.
A comparison between block-structured and hybrid unstructured approaches for high-lift
configurations has already been investigated 327. On a given simplified high-lift configuration
and with the same mesh size, the block structured mesh allows a better stream wise resolution
on the leading edges and normal resolution. Additionally a higher cell stretching in spanwise
direction is feasible than for hybrid unstructured mesh. However, the hybrid approach clearly
simplifies the meshing procedure and permits to consider more complex configurations. A
simplification in the structured grid generation process can be achieved with the so-called
overset technique, also commonly called chimera technique. This approach has been
successfully used for the flow analysis on 3D high-lift configuration.
Maintaining a constant mesh quality in high-lift design is a second great challenge as the change of
geometry can be rather large. The most relevant design parameters are indeed the relative position
of the slat and the flap with respect to the main wing, the so called setting. The shape may also be
changed, but since the outer shape of the aircraft is designed for cruise condition, only the shape that
is hidden when the devices are retracted is allowed to be modified. In the scope of this paper only the
settings of the elements are considered as design parameters. The optimization of the slot imposes a
modification of the mesh surrounding the trailing edge of the preceding elements, which is a critical
area in aerodynamics. A change of the mesh quality in this region can artificially change the flow
physics and may slow down the design process or even lead to a non-physical optimum. For all these
reasons, the selection of the meshing procedure is driven by its capability to preserve a constant
mesh quality during the optimization process. Various mesh deformation procedures exist and have
successfully been used for shape optimization problems328 and even for 2D high-lift optimization329.
However, this technique applied on large geometric changes can degrade the quality of the resulting
mesh due to highly skewed cells. The overset approach seems to be a better strategy since the original
mesh is not modified and only cells masked and overlapped are adjusted during the setting
optimization. This approach has been used in 2D high-lift optimization but the procedure to mask
the cells influences the flow around the multielement wing and can lead to a wrong design. A last
approach is to generate new meshes during the optimization process. If the replay mode of the mesh
generator permits to control the mesh properties in almost the complete domain, this procedure will
guarantee a sufficiently constant mesh quality during the design. However, the major drawback here
is its fastidious setting to get a robust mesh generation process for a large range of setting changes.
327 Melber-Wilkending, S, Rudnik R, Ronzheimer A, Schwarz T. Verification of MEGAFLOW-software for high lift
applications. MEGAFLOW – Numerical Flow Simulation for Aircraft Design. Notes on Numerical Fluid Mechanics
and Multidisciplinary Design (NNFM), Vol. 89, pp. 163-178, ISBN 3-540-24383-6, 2005.
328 Mavriplis DJ. Multigrid solution of the discrete adjoint for optimization problems on unstructured meshes. AIAA
method. 40th AIAA Aerospace Sciences Meeting and Exhibit, AIAA Paper 2002-844, Reno NV, January 2002.
182
1 Differential Evolution (DE) that belongs to the group of evolutionary algorithms 330;
2 Gradient free subspace simplex method (SubPlex) 331;
3 Gradient based sequential quadratic programming algorithm NLPQLP 332, available within
mode FRONTIER 333.
These strategies were already evaluated for the design of the shape and position of a 2D flap in high-
lift configuration. It has been observed that the evolutionary algorithm is the most robust to failure
and uncertainty in the numerical chain, is easily parallelizable and permits to reach the best optimum
on multimodal design space. However, this strategy requires at least one order of magnitude more
time to converge than gradient based strategy. On the other hand, the gradient-based strategy is
more tedious to use in high-lift configuration since it requires a highly accurate solution with as less
uncertainty on the goal function as possible. At this time, free approach obtained the best
compromise between efficiency and robustness. Currently, these strategies are tested on the DLR-
F11 configuration in order to check their capabilities to efficiently complete the high-lift
optimizations problem within 2 weeks. See [Brezillon et al]334.
330 Storn R, Price K. Differential Evolution - a simple and efficient adaptive scheme for global optimization over
continuous spaces. Technical Report, International Computer Science Institute, TR-95-012, Berkley, 1995.
331 Rowan T. Functional stability analysis of numerical algorithms, Thesis, Department of Computer Sciences,
German Aerospace Center, Institute of Aerodynamics and Flow Technology, Braunschweig, Germany, 26 th
International Congress Of The Aeronautical Sciences, ICASE 2008.
335 Kallinderis Y. Hybrid grids and their applications. Handbook of Grid Generation, edited by J.F. Thompson, B.K.
Soni, and N. Weatherill, CRC Press, Boca Raton, FL, 1999, pp. 25.1-25.18.
183
latter run in batch modus during the optimization process. This approach has been applied for the
optimization of the DLR-F11 configuration, a 3D high-lift configuration featuring full span flap and
slat. However, the increasing complexity of the configuration to be designed limits the application of
purely structured meshes mainly because of the intricate mesh topology needed.
As a second alternative, a mixed mesh approach has been tested. This relatively new development in
mesh generation techniques combines the advantages of structured meshing for the resolution of
boundary layers and wakes with the advantages of unstructured mesh generation for its flexibility
and cell growing in all 3 dimensions. Pure block structured meshes are employed where the geometry
is not too complicated and more topologically difficult areas are meshed with prisms like in the
classical hybrid unstructured grid approach. A further step towards quality of these elements has
recently been achieved by developing an automatic generator for smooth prismatic layers applying
parabolized Laplacian equations336. The rest of the flow field is meshed using tetrahedral elements.
This procedure has been implemented in MegaCads by incorporating the 3D triangulation software
NETGEN of [Schöberl]337 as black box for the generation of the unstructured mesh part. The
structured hexahedrons are connected to the unstructured part by collapsing the outer hexahedral
layer into prisms and tetrahedrons in order to achieve a smooth transition of the control volumes for
the reduction of numerical errors and instabilities. It has already been shown that applying this mesh
type can greatly speed up RANS flow computations without losing solution accuracy, mainly due the
reduced number of points338. This second approach has been selected for the setting optimization of
the FNG configuration that is a 3D high-lift configuration featuring a slat and a flap that does not
extend the whole wing span.
tools for aerodynamic applications in DLR, aerospace industry and universities344-345. Both codes
336 Wild J, Niederdrenk P, Gerhold, T. Marching generation of smooth structured and hybrid meshes based on
metric identity. 14th International Meshing Roundtable, edited by B. Hanks, Springer, Berlin, pp. 109-127, 2005.
337 Schöberl J. NETGEN - An advancing front 2D/3D mesh generator based on abstract rules. Computing and
ERCOFTAC Design Optimization: Methods & Applications, Paper ERCODO2004_221, Athens (Greece), 2004.
339 Raddatz J, Fassbender JK. Block structured Navier-Stokes solver FLOWer. MEGAFLOW – Numerical Flow
Simulation for Aircraft Design. Notes on Numerical Fluid Mechanics and Multidisciplinary Design (NNFM), Vol.
89, pp. 27-44, ISBN 3-540-24383-6, 2005.
340 Kroll N, Radepiel R, Rossow CC. Accurate and efficient flow solvers for 3D applications on structured meshes.
Design. Notes on Numerical Fluid Mechanics and Multidisciplinary Design (NNFM), Vol. 89, pp. 81-92, ISBN 3-
540-24383-6, 2005.
343 Schwamborn D, Gerhold T, Heinrich R. The DLR TAU-Code: Recent Applications in Research and Industry.
ECCOMAS CDF 2006, In proceedings of “European Conference on Computational Fluid Dynamics“, Delft The
Netherland, 2006.
344 Kroll N, Rossow CC, Schwamborn D. The MEGAFLOW-project - numerical flow simulation for aircraft. A.
DiBucchianico, R.M.M. Mattheij, M.A. Peletier (Edts.), Progress in Industrial Mathematics at ECMI 2004,
Springer, New York, S. 3 – 33, 2005.
345 Kroll N, Fassbender JK. MEGAFLOW – Numerical flow simulation for aircraft design. MEGAFLOW - Numerical
Flow Simulation for Aircraft Design. Notes on Numerical Fluid Mechanics and Multidisciplinary Design (NNFM),
Vol. 89, ISBN 3-540-24383-6, 2005.
184
346 Brezillon J, Dwight R. Discrete adjoint of the Navier-Stokes equations for aerodynamic shape optimization.
Evolutionary and Deterministic Methods for Design, Optimization, and Control with Applications to Industrial and
Societal Problems EUROGEN 2005, edited by R. Schilling, W. Haase, J. Pé riaux, H. Baier and G. Bugeda, FLM, ISBN:
3-00-017534-2, Munich, 2005.
347 Brezillon J, Brodersen O., Dwight R, Ronzheimer A, Wild J. Development and application of a flexible and
efficient environment for aerodynamic shape optimization. ONERA-DLR Aerospace Symposium (ODAS),
Toulouse, 2006.
348 Saad Y, Schultz MH. A generalized minimum residual algorithm for solving non-symmetric linear systems. SIAM
constant the number of grid points while varying the design variables. This process can be done
during the CFD run and once the aerodynamic and adjoint states are available, the gradient is
extremely fast to compute. See [Brezillon et al]349.
Figure 6.32 Multi-Block Structured Mesh Around the DLR-F11 Model in Full Span Flap and Slat
Configuration – Courtesy of [Brezillon et al.]
German Aerospace Center, Institute of Aerodynamics and Flow Technology, Braunschweig, Germany, 26th
International Congress Of The Aeronautical Sciences, ICASE 2008.
350 Rudnik R, Thiede P. European research on high lift commercial aircraft configurations in the EUROLIFT
too heavy a mechanism. The relation between the horizontal displacement and the penalty is set
according to industrial specifications.
Figure 6.34 Close View of the Wake Discretization on the Top of the DLR-F11 Wing – Courtesy of
[Brezillon et al.]
6.9.5.2 Synthesis
All optimizations presented above were obtained within 14 days but with different solvers or levels
of convergence. For the sake of comparison, the optimums previously obtained are recomputed with
the same flow solver and the same level of convergence and the deviations to the baseline
configuration are summarized in Table 6.9. One can sees that all optimization strategies give almost
the same improvement with a slight advantage of the NLPQLP strategy, independently to the way of
computing the gradients.
Figure 6.37 (a-b)
present a comparison of
the slat and flap positions
at a given spanwise
position and confirm that
all optimums are located at
almost the same position
and that both optimums
given by the NLPQLP Table 6.9 Variations of the Objective Function, Drag (dc=drag count)
strategy are identical. It and Lift Coefficients – Courtesy of [Brezillon et al.]
can be guessed that the
design space is rather shallow close to the optimum and the optimum can be reached only with a fully
converged optimizations process, as for the NLPQLP cases. The performance improvement is made
189
6.9.6 Optimization of Figure 6.36 Drag Distribution Along the Spanwise Direction on
the FNG the Baseline and Optimized configurations – Courtesy of [Brezillon et
Configuration al.]
The configuration used for the second design case, the so-called FNG wing351, is similar to the first
configuration a wing-body configuration with deployed high-lift devices. The difference to the
previous DLR-F11 is that for the FNG the high-lift devices do not extend over the full wing span but
are limited to 95% relative span. Figure 6.37 (a) Slat setting for the baseline and optimized
configurations at the middle of the outer wing. Figure 6.37 (b) Flap setting for the baseline and
optimized configurations at the middle of the outer wing. By this the topology of the geometry is no
longer suited for block structured mesh generation like the DLR-F11 configuration. The design again
is made for a take-off configuration at flight conditions (M = 0.18, Re = 18 x 106) and a constant lift
coefficient of CL =1.25. In order to achieve the latter the calculations are all made in target lift mode,
where the angle of attack is adapted during the flow calculation. The objective of the optimization is
the minimization of drag coefficient. In order to eliminate even smallest deviations of the lift
coefficient and its influence on the induced drag, the value used was the actual drag coefficient minus
the ideal elliptic induced drag, commonly known as profile drag
(CL )2
CD 𝑝 = CD −
πΛ
Eq. 6.34
351 Dargel, G., Hansen, H., Wild, J., Streit, T., Rosemann, H. Aerodynamische flügelauslegung mit
multifunktionalen steuerflächen (aerodynamic design of wings with multi-functional control devices). DGLR
Jahrbuch 2002, DGLR-2002-096, Vol I, DGLR, (2002).
190
at this stage homogenously along the device span. Since the geometric restrictions at the device ends
and their interference with the remaining clean wing tip do not allow for an easy application of block
structured meshes, here the mixed mesh type was used.
6.9.6.2 Results
For this design only the SubPlex method was applied. The gradient method based on the adjoint could
not be applied here due to the variation of the number of grid points in the unstructured mesh part.
The subspace simplex method was run for 5 cycles resulting in overall 150 flow calculations. For the
flow calculations 96 processors have been used, resulting in an overall turn-around time of roughly
375 hours or 15 days. The optimum solution has already been found after the third loop, the rest of
the optimization secures that there is no better optimum within the range of the parametric changes.
Neglecting these iterations would decrease the turn-around time to approximately 11 days. The
objective was reduced by 0.83%, which corresponds to approximately 4 dc. The drag reduction
corresponds to a slight gap increase over the whole wing span. The lift coefficient was kept constant
within a margin of 0.03%. The relatively small amount of improvement may be explained by the fact
that the wing section of the FNG high-lift wing has already been designed applying numerical
optimization in 2D and the spanwise variation of the geometry was not included in full extend into
the presented exercise.
(a) Slat Setting for the Baseline (b) Flap Setting for the Baseline
and Optimised Configurations and Optimised Configurations
Figure 6.37 Optimizing Configuration for Slat and Flap – Courtesy of [Brezillon et al.]
6.9.6.3 Conclusion
The methods and tools developed and successfully applied to perform the settings optimizations of
3D high-lift configurations within a limited time frame were presented. The major key technology to
optimize such complex configurations is the meshing procedure and two ways followed at DLR were
discussed in detail. The development of efficient and robust numerical methods and the increase of
the computer performance permit fast computations of complex flows and makes feasible numerical
optimizations within limited turn-around time. It was observed that the evolutionary algorithm is
very attractive thanks to his inherent possibility to compute a population in parallel. For strategies
involving only sequence of simulations, the reduction of the turnaround time can be achieved by
conducting CFD in parallel. The gradient-based strategy tested here has permitted to fully converge
the optimizations problem within a very short time frame and provides here the best solution. The
inclusion of the adjoint approach gives new opportunities to treat more complex optimizations
problems in limited turnaround time and will be further investigated at DLR.
191
192
Figure 7.1 Sketch of Griewank Function on larger scale (left) vs. smaller scale (right)
capable of constructing insightful design trade-off relationships, referred to as Pareto fronts, between
objectives. GAs are well-suited to complex optimization tasks as they can easily use both discrete
and continuous variables, and can easily handle non-linear, non-convex, and non-continuous
objective functions. [Chiba et al.] suggest that GAs have four distinct advantages which encourage
their use in aerodynamic/aero-structural optimizations:
➢ GAs have the ability to find multiple optimal solutions and design trade-offs;
➢ GAs process information in parallel, optimizing from multiple points within the design space;
➢ High-fidelity CFD codes can be adapted to GAs without any modifications;
➢ GAs are insensitive to numerical noise that may be present in the computation.
The main drawbacks associated with these algorithms are high computational cost, poor constraints
handling, and the requirement for problem specific tuning and limitations in how many variables are
feasible to handle. Studies have shown that GAs are very fast at identifying regions of optimality
within a design space but demonstrate slow convergence as they moves nearer optimal solutions.
Some studies have tried to build on the classical GA to enhance its applications to aerodynamic
optimization. GA optimization is, in its most basic form, an iterative process which can be
summarized as:
Details for these and other methods can be found in [Skinner & Zare-Behtash]352, as well as in
Wikipedia.
7.1.1 Framework for the Shape Optimization of Aerodynamic using Genetic Algorithms (GA)
To demonstrate the Genetic Algorithm GA, [López et al.]353 developed a framework for the shape
optimization of aerodynamics profiles using computational fluid dynamics (CFD) and genetic
algorithms. A genetic algorithm code and a commercial CFD code were integrated to develop a CFD
shape optimization tool, as illustrated in Figure 7.2. The results obtained demonstrated the
effectiveness of the developed tool. The shape optimization of airfoils was studied using different
strategies to demonstrate the capacity of this tool with different GA parameter combinations. Details
of procedures can be found in354. Optimization was performed using a simple GA that follows the
352 S. N. Skinner and H. Zare-Behtash, ”State-of-the-Art in Aerodynamic Shape Optimisation Methods”, Article in
Applied Soft Computing , September 2017, DOI: 10.1016/j.asoc.2017.09.030.
353 D. López, C. Angulo, I. Fernández de Bustos, and V. García, “Framework for the Shape Optimization of
Aerodynamic Profiles Using Genetic Algorithms”, Hindawi Publishing Corporation, Mathematical Problems in
Engineering, Volume 2013, Article ID 275091, 11 pages http://dx.doi.org/10.1155/2013/275091.
354 See above.
194
7.1.2 Case Study 1 - Optimizations of Airfoil & Wing using Genetic Algorithm (GA)
The study by [Zhang et al.]355 demonstrates how the Genetic Algorithm (GA), coupled with the CFD
solvers for 2D and 3D problems, can successfully be applied to the airfoil and wing aerodynamic drag
minimization. The geometry of airfoil and wing sections is represented by a B-spline curve. The
actual values of the coordinates of the control nodes for the B-spline curve are designated as the
design variables. The NACA0012 airfoil is optimized for a free stream Mach number M∞=0.30, at
which the drag coefficient is reduced by 46.1% for a constant lift coefficient of CL = 0.55. For the
corresponding straight wing, the Mach number is set at M∞=0.80. The drag coefficient is reduced by
13.5% under the constant lift coefficient of CL = 0.30.
355F. Zhang, S. Chen and M. Khalid, “Optimizations Of Airfoil And Wing Using Genetic Algorithm”, ICAS2002.
356Blaize M, Knight D and Rasheed K. “Automated Optimal-Design of 2-Dimensional Supersonic Missile Inlets”.
Journal of Propulsion and Power, Vol. 14, Iss. 6, 1998.
195
357David E. Goldberg. “Genetic Algorithms in Search, Optimization, and Machine Learning”. Massachusetts,
Addison-Wesley, 1989.
196
Figure 7.4 shows the flowchart describing the GA application to aerodynamic optimization for an
airfoil (or a wing). The CFD solver calculates the objective function (Cl/CD) and sends it to GA, which
uses it as a fitness value. For the 3D wing configuration, a straight wing with constant chord and
thickness and an aspect ratio of 1.5 is chosen. The GA operations are applied to each section, defined
in the figure, in the same way as that for an airfoil. The procedure only modifies the section shapes.
Numerical experiments showed that the CPU time spent for the GA operations is negligible when
compared to the CPU time of flow solution. For a complete information, please refer to [Zhang et
al.]358.
358 F. Zhang, S. Chen and M. Khalid, “Optimizations Of Airfoil And Wing Using Genetic Algorithm”, ICAS2002.
198
should be mentioned here that the maximum fitness corresponds to the best member in each
generation and the averaged fitness of the entire members in the generation. The trend of the fitness
in this figure strongly shows that the optimum was approaching from one generation to another,
demonstrating the reliability of the Genetic Algorithm. Figure 7.5 displays the original NACA0012
airfoil in comparison with the optimized airfoil.
The required free stream angle of attack for the original airfoil was computed to have a value of α =
5.49 degrees to create the required lift coefficient, while it was 1.91 degrees to keep the same lift
coefficient for the optimized airfoil. Compared with the original airfoil, the radius of leading edge of
the optimized airfoil is greatly reduced. Both smaller angle of attack and leading edge radius result
in the decrease of the pressure peak value on the suction surface. At the rearward part of the airfoil,
the curvature on the both lower and upper surfaces is increased, which creates the bigger pressure
difference between the two surfaces. This would then compensate the lift lost at the forward part of
the airfoil in order to keep the lift coefficient constant. The Mach number contours for both original
and the resulting optimized airfoils are displayed in Figure 7.7. The maximum local Mach number
is reduced from 0.561 for the original airfoil to 0.432 for the optimized airfoil.
Figure 7.7 Mach Numbers for the Original and Optimized Airfoil - Courtesy of [Zhang et al.]
As is well known, one of the main contributions to aerodynamic drag in transonic flow is from shock
199
waves. Therefore, in the present study, the reduction of the shock wave in both strength and size
contribute to the drag coefficient decrease from 0.03686 for the original wing to 0.03190 for the
optimized wing, which represents 13.5% reduction under the fixed lift coefficient CL = 0.30. It is not
surprising that the drag reduction is smaller for a wing than that for an airfoil. For a wing, each wing
section must satisfy the geometry constraints, which is equal to imposing much more constraints
than those for an airfoil. The more constraints imposed, the less benefits obtained. Close to the
trailing edge, the increased curvature on both lower and upper surfaces of the optimized wing creates
a greater loading. This would then compensate for the lift lost at the forward part of the wing sections
due to the reduced expansion on the upper surface in order to keep the lift coefficient constant. It
should be noted that the pressure distribution on each wing section is subject to both stream wise
and span wise (3D effects aspects of the problem are comprehensively accounted for in the
computations. It was noted that, for the original wing, the free stream angle of attack was set at α =
4.31 degrees to provide a lift coefficient of CL = 0.30. For the optimized wing, the angle of attack was
reduced to α = 3.40 degrees to provide the same lift.
7.1.2.4 Conclusions
The Genetic Algorithm has been successfully applied to the airfoil and wing aerodynamic drag
minimization. The results demonstrated the powerful reliability and robustness of the genetic
algorithms. Numerical experiment showed that the CPU time spent for GA operations is negligible
when compared to the CPU time of flow solution during the optimization. The genetic algorithm is
independent of the CFD solvers used. This means that the ability of the optimization, to deal
accurately and efficiently with subsonic, transonic or supersonic flows, or potential, Euler or Navier-
Stokes solutions, strongly depends on the CFD solver. The optimized airfoil has smaller leading edge
radius and angle of attack compared with the original one, leading to the smaller suction peak value
on the upper surface. The optimized 3D wing has weaker shock waves on the upper surface, leading
to the smaller drag coefficient. Both the optimized airfoil and wing have a greater aft loading to
compensate for the lift lost at the forward part in order to keep the lift coefficient constant.)
7.1.3 Case Study 2 - Cavitation Airfoil Optimization of Multi-Phase Flow-Fields using the
Evolutionary Algorithms (GA)
Shape Optimization of airfoils operating in multi-phase flow regimes is discussed by [Ahuja and
Hosangadi]359 in this section. Such airfoil design efforts in cavitation flow regimes are important
because of their application to turbo pumps operating with cryogenic fluids, where cavitation
breakdown significantly affects performance characteristics and in marine propeller systems where
cavitation erosion in of primary concern.
359Vineet Ahuja and Ashvin Hosangadi, “Design Optimization of Complex Flow fields Using Evolutionary
Algorithms and Hybrid Unstructured CFD”, Article · June 2005.
200
the selection, crossover and mutation operators on the design space in manner analogous to
evolution in nature. Many variants of these operators exist and in our particular implementation we
utilize tournament selection with elitism and uniform crossover. In the Simple Genetic Algorithm
(SGA) (see Goldberg360) mutation played a dominant role in avoiding premature convergence in
multi-modal design landscapes. However, placing undue reliance on the mutation is inefficient
because mutation usually results in designs with lower fitness. Preservation of diversity in the design
population is ensured by niching or fitness sharing by sub-populations that define a certain niche in
a multi-modal design space. In particular, following the lead of [Carroll], we have used Goldberg’s
multi-dimensional phenotypic sharing scheme with a triangular sharing function. For multi-objective
optimization problems we follow the ranking scheme of [Fonseca and Fleming]361-362 in obtaining the
Pareto optimal set of solutions.
The Navier-Stokes simulations are performed with the multi-element unstructured CRUNCH CFD®
code discussed in a following section. The CRUNCH CFD® code used for fitness evaluations, can
simulate multi-phase flow fields with cavitation. Pure incompressible flow fields devolve naturally as
a subset of the generalized multi-phase system and will be utilized in this paper. Since genetic
algorithms effectively search the entire design space, the geometry can undergo radical changes,
thereby requiring a new grid to be generated. For each design, a new grid is generated through a
semi-automated grid generation procedure utilizing the scripting language in GRIDGEN ©. The
scripting language efficiently reconstructs the grid from the baseline geometry to new designs and is
discussed in detail in the following section.
∂𝐐 ∂𝐄 ∂𝐅 ∂𝐆
+ + + = 𝐒 + 𝐃V
∂t ∂x ∂y ∂z
Eq. 7.2
Here Q is the vector of dependent variables, E, F and G are the flux vectors, S the source terms and DV
represents the viscous fluxes. The viscous fluxes are given by the standard full compressible form of
Navier Stokes equations [Hosangadi]365. The vectors Q, E and S are given below and a detailed
discussion on the details of the cavitation source terms is provided in our previously published
works.
360 Goldberg, D.E., “Genetic Algorithms in Search, Optimization, and Machine Learning,” Addison-Wesley,
Reading, MA, 1989.
361 Fonseca, C.M., and Fleming, P. J., “An Overview of Evolutionary Algorithms in Multi objective Optimization,”
Mixtures," AIAA-99-3330, 14th AIAA CFD Conf., Norfolk, VA, June 28-July 1, 1999.
364 Ahuja, V., Hosangadi, A. and Arunajatesan, S., "Simulations of Cavitating Flows Using Hybrid Unstructured
Dimensional Combusting Flows," Journal of Propulsion and Power, Vol. 12, No. 3, pp. 494-503, May-June 1996.
201
ρm ρm u 0
ρm u ρm u2 + P 0
ρm v ρm uv 0
ρm w ρm uw 0
𝐐= ρ ϕ , 𝐄= ρ ϕ u , 𝐒= S
g g g g g
ρm hm ρm hm u Sh
ρm k ρm ku Sk
[ ρm ε ] [ ρm εu ] [ Sε ]
Eq. 7.3
Here, ρm and are the mixture density and enthalpy respectively, and φg the volume fraction or
porosity of the vapor phase. The mixture energy equation has been formulated with the assumption
that the contribution of the pressure work on the mixture energy is negligible which is a reasonable
assumption for this flow regime. The mixture density and gas porosity are related by the following
relations locally in a given cell volume:
ρm = ρg ϕg + ρL ϕl , ϕg + ϕL = 1
Eq. 7.4
where ρg , ρL are the physical material densities of the gas and liquid phase respectively and in general
are functions of both the local temperature and pressure. The equation system as formulated in Eq.
7.2 is very stiff since the variations in density are much smaller than the corresponding changes in
pressure. Therefore to devise an efficient numerical procedure we wish to transform Eq. 7.2 to a
pressure based form where pressure rather than density is the variable solved for. An acoustically
accurate two-phase form of Eq. 7.2 is first derived, followed by a second step of time-scaling or
preconditioning to obtain a well-conditioned system. We begin by defining the acoustic form of
density differential for the individual gas and liquid phase as follows:
1 1
dρg = dP , dρL = dP
cg2 cL2
Eq. 7.5
Here cg is the isothermal speed of sound (∂P/∂ρg )T in the pure gas phase, and cL is the corresponding
isothermal speed of sound in the liquid phase, which is a finite-value. We note that in Eq. 7.5 the
variation of the density with temperature has been neglected in the differential form. This
assumption was motivated by the fact that the temperature changes are primarily due to the source
term and not by the pressure work on the fluid i.e. the energy equation is a scalar equation. This
simplifies the matrix algebra for the upwind flux formulation significantly, at the potential expense
of numerical stability in a time-marching procedure. However, more importantly, there is no impact
on the accuracy since the fluid properties themselves are taken directly from the thermodynamic
data bank for each fluid. Following the discussion above, the differential form of the mixture density
ρm using Eq. 7.5 is written as,
1 1 𝜙𝑔 𝜙𝐿
dρm = (ρ𝑔 − ρ𝐿 )dϕ𝑔 + dP , = +
𝑐𝜙2 𝑐𝜙2 𝑐𝑔2 𝑐𝐿2
Eq. 7.6
Here, cφ is a variable defined for convenience and is not the acoustic speed, cm, in the mixture, which
will be defined later. Using Eq. 7.6, Eq. 7.2 may be rewritten as:
202
∂𝐐𝐕 ∂𝐄 ∂𝐅 ∂𝐆
Γ + + + = 𝐒 + 𝐃𝐕 , 𝐐𝐕 = [ρ , u , v , w , ϕg , k , ε ] T
∂t ∂x ∂y ∂z
Eq. 7.7
The numerical characteristics of the Eq. 7.7 are studied by obtaining the eigenvalues of the matrix,
Γ−1(∂E/Qv) . The eigenvalues of the system are derived to be:
𝚲 = [u + c𝑚 , u − c𝑚 , u , u , u , u , u ]
Eq. 7.8
where cm turns out to be the well-known, harmonic expression for the speed of sound in a two-phase
mixture and is given as:
1 𝜙𝑔 𝜙𝐿
= ρ𝑚 [ + ,]
2
𝑐𝑚 𝜌𝑐𝑔2 𝜌𝑐𝐿2
Eq. 7.9
The behavior of the two-phase speed of sound is interestingly a function of the gas porosity; at either
limit the pure single-phase acoustic speed is recovered. However, away from the single-phase limits,
the acoustic speed rapidly drops below either limit value and remains at the low-level in most of the
mixture regime. As a consequence, the local Mach number in the interface region can be large even
in low speed flows. To obtain an efficient time-marching numerical scheme, preconditioning is now
applied to the system in Eq. 7.7, in order to rescale the eigenvalues of the system so that the acoustic
speeds are of the same order of magnitude as the local convective velocities.
∆𝑄𝑉 𝑉
Γ𝑃 + ∫ 𝐹(𝑄𝑉 , 𝑛)𝑑𝑠 = ∫ 𝑆 𝑑𝑉 + ∫ 𝐷(𝑄𝑉 , 𝑛)𝑑𝑠
∆𝑡
𝜕𝑄𝑖 𝑄𝑖 𝜕𝑄𝑖
Eq. 7.10
The inviscid flux procedure involves looping over the edge list and computing the flux at the dual face
area bisecting each edge. The Riemann problem at the face is then solved using higher order
reconstructed values of primitive variables at the dual face. Presently a second-order linear
reconstruction procedure366 is employed to obtain a higher-order scheme. For flows with strong
gradients, the reconstructed variables are limited to obtain a stable TVD scheme. While the inviscid
flux as outlined above is grid-transparent, the viscous flux calculation is computed on an element
basis since the diffusion of scalars is sensitive to the skewness of the element type. The numerical
366 Barth, T. J., "A 3-D Upwind Euler Solver for Unstructured Meshes,” Paper No. AIAA-91-1548.
203
framework is operational on distributed memory systems for parallel computations. Details of the
parallel architecture and viscous flux formulation are described in [Barth]367.
7.1.3.4 Optimization
In our optimization work, the profile
of the airfoil was defined as a cubic Optimized
Bezier curve and the coordinates of
the control points were specified as
design variables. The shape of the
airfoil was assumed to be symmetric
and the length of the airfoil was kept
constant. The airfoil was operating in
water at a 4 degree incidence angle
with a freestream cavitation number
of 0.6, and a freestream velocity of
12.2 m/s. The baseline design Base Line
represented a NACA 0012 airfoil
profile and the fitness function was
defined by lift experienced by the
airfoil in the cavitation flow regime.
Fitness evaluations were carried out
with CFD simulations utilizing a
multi-phase version of CRUNCH CFD®
with the multi-phase equation Figure 7.8 Comparison of the Pressure Distribution
system. Cavitation in this system is between Baseline and Optimized Design for Maximizing
triggered by finite rate source terms Airfoil Lift – Courtesy of [Ahuja and Hosangadi]
and the liquid-vapor interface is
captured as part of the solution
procedure for the mixture Navier-
Stokes equation system. This case is
particularly interesting because the
lift characteristics of the airfoil
strongly correlate with the extent of
cavitation on the suction side of the
airfoil surface. Therefore, it becomes
imperative to use a high-fidelity
solution procedure that has been
extensively validated for cryogenic
and cavitation flow regime 17,18 such
as the one utilized for fitness
evaluations in this paper. Fitness
calculations for multi-phase flows are
expensive because of more rigorous
numerical stability considerations
when compared to conventional
single phase systems. In our case, Figure 7.9 Comparison of Cavitation Zones between
fitness calculations for a particular Baseline (Bottom) and Optimize Design (Top) For
design were initiated from the Maximizing Airfoil Lift – Courtesy of [Ahuja and Hosangadi]
367 Barth, T. J., "A 3-D Upwind Euler Solver for Unstructured Meshes,” Paper No. AIAA-91-1548.
204
369 Anderson JD, “A history of aerodynamics and its impact on flying machines”. Cambridge University Press,
Cambridge, 1997.
370 Secanell M, Suleman A, Gamboa P (2006) Design of a morphing airfoil using aerodynamic shape optimization.
AIAA J 44:1550–1562.
371 Lock RC, Bridgewater J (1967) Theory of aerodynamic design for swept-winged aircraft at transonic and
376 Li P, Sobieczky H, Seebass R (1995) A design method for supersonic transport wings. AIAA Meet Pap.
377 Elmer K, Welge H, Salamone J, Cowart R (2013) SCAMP: supersonic passenger transport transonic
acceleration flight profiles with considerations of focused sonic boom. AIAA Meet Pap.
378 Mavris DN, Hayden WT (1996) Formulation of an IPPD methodology for the design of a supersonic business
business jet (SSBJ) based on the Learjet airframe. AIAA Meet Pap.
207
conceptual design of commercial supersonic aircraft with the highest Mach number of 2 and a cruise
flight range of 6000 km. Aerodynamic characteristics of this vehicle were modeled and incorporated
into an aircraft sizing code through the use of Response Surface Methodology, Design of Experiments
techniques and mathematical modeling.
A detailed aerodynamic design of a complete aircraft configuration was presented by [Sturdza]381.
Euler equations along with multipoint optimization techniques were employed in the design. In
addition, a full-scale Reynolds number laminar flow experiment was presented. The experiment was
based on a low-sweep wing with sharp nosed sections of the supersonic Aerion Business Jet. The
concept of a natural laminar flow was the central feature of the investigation. [Lee et al.]382 carried
out a multidisciplinary design with multiple objectives for a supersonic fighter wing with sweep
angle of 30 and 40 degrees. They utilized Response Surface Methodology and Genetic Algorithm.
For investigating aerodynamic efficiency, nine wing and airfoil parameters were chosen and four
structural variables were added to determine the wing skin thickness. They applied three-
dimensional Euler equations to calculate the transonic and supersonic aerodynamic properties of the
fighter wing. Accomplishing the higher lift-to-drag ratio at the cruise flight speed, improving the
maneuverability at the transonic flight condition and boosting the maximum speed of the aircraft
were the main objectives of their project.
[Yoshida]383 at the Japan Aerospace Exploration Agency (JAXA) has developed an advanced drag
reduction technique for next generation supersonic transport (SST). The method is based on
pressure drag reduction concept. It is mainly built on supersonic linear theory and a friction drag
reduction technique. To obtain the most optimized geometry that maximizes the drag reduction
effects, he applied an extensive numerical and experimental study on arrow planform. This type of
warped wing alongside an area-ruled body configuration is favorable for reducing the wing–body
interference. In [Reuther’s]384 work, to optimize a supersonic aircraft that cruises at Mach 2.2, a fluid
dynamics algorithm using Euler equations is modeled. Design variables are chosen as Hicks-Henne
shape functions due to its consistency and simplicity of use. Eighteen design variables are defined for
wing cross-sections.
According to their final results, drag coefficient is reduced from 0.0088 to 0.00812. The study of
[Kiyici and Ardag]385 focuses on the aerodynamic design and shape optimization of a supersonic
business jet through designing an algorithm called BF-1. By obtaining the baseline geometry, CFD
analysis is performed and Euler equations are used to solve a compressible flow. They have
developed a numerical methodology at Mach number 1.5 to optimize the standard geometry for
getting minimum CD and maximum CL values for the wing with NACA 65-210 airfoil. According to
their results for the final optimized geometry, drag coefficient is decreased by 14% and lift-to-drag
ratio increased by 10%.
381 Sturdza P (2007) Extensive supersonic natural laminar flow on the Aerion business jet. AIAA Meeting.
382 Kim Y, Jeon Y, Lee D, Multi-objective and multidisciplinary design optimization of supersonic fighter wing. J
Aircraft, 2006.
383 Yuhara T, Rinoie K, Makino Y (2014) Conceptual design study on LH2 fueled supersonic transport considering
configurations via an adjoint formulation on distributed memory parallel computers. J Computer Fluid.
385 Kiyici F, Aradag S (2015) Design and optimization of a supersonic business jet. AIAA Aviation Forum.
386 Parashar S, Bloebaum C (2006) Multi-objective genetic algorithm concurrent subspace optimization
387 Kenway GKW, Martins JRRA (2015) Multipoint aerodynamic shape optimization investigations of the common
research model wing. AIAA J 10(2514/1):J054154.
388 Lyu Z, Xu Z, Martins J R R A (2014) Benchmarking optimization algorithms for wing aerodynamic design
optimization. 8th International conference on computational fluid dynamics (ICCFD8). Chengdu, Sichuan.
389 Kim Y, Jeon Y, Lee D (2006) Multi-objective and multidisciplinary design optimization of supersonic fighter
wing. J Aircraft.
390 Holland J (1992) Adaption in natural and artificial systems. MIT Press, Cambridge.
391 Goldberg DE (1989) Genetic algorithms in search, optimization and machine learning. Wesley Longman
Publishing, Boston 26. Sforza P (2014) Commercial airplane design principles, 1 st ed. Butterworth-Heinemann,
Oxford, pp 119–212.
209
produced than that on a conventional airfoil. Furthermore, this group of airfoils with their exclusive
design are used for maintaining laminar flow on the wing. Laminar airfoils have the potential to delay
flow separation in specific cases.
For the 3D wing design and analysis in this study, effect of sweep angle is combined with effect of
aspect ratio and taper ratio of this specific wing. For investigating the particular effects of sweep
angle, three analogous geometries with sweep angle of 37, 50 and 60 degrees with two distinctive
thickness to chord ratio are generated. Free stream Mach number is constant during this study and
it equals with 1.4 and the angles of attack are 0, 5 and 10 degrees. Extracted aerodynamic coefficients
in this work are based on wing’s root length as well as surface area (when viewed from top), which
is calculated and equal to 1.01 square meters. In this phase of study, aspect ratio of the wing is
constant and equal to 3.3. Figure 7.13 depicts the 2D view of geometries in three distinct sweep
angles.
Figure 7.13 2D View of the Wing in Sweep Angle of 37, 50 and 60 Degrees
∂ ∂
(ρui ) + (ρui uj )
∂t ∂xj
∂p ∂ ∂ui ∂uj 2 ∂ui ∂
=− + [μ ( + − δij )] + ̅̅̅̅̅
(−ρu ′ u′ )
i j
∂xj ∂xj ∂xj ∂xi 3 ∂xi ∂xj
Eq. 7.11
where ü, ú, p, ρ, μ, respectively, denote time-averaged fluid velocity, the fluctuating velocity, fluid
pressure, fluid density and fluid dynamic viscosity. Reynolds stresses, qu’iu’j, must be modeled to
close the equation. The Spalart–Allmaras turbulence model393 is used in this work for modeling
turbulent flow. It is a 1- equation model that solves a transport equation for the kinematic eddy
turbulent viscosity. This model was designed exclusively for aerospace applications and has shown
392 Wu Z, Zhang Q, Bao T, Li L, Deng J, Hu Z (2016) Experimental and numerical study on ethanol and dimethyl
ether lifted flames in a hot vitiated co-flow. J Fuel 184:620–628.
393 Spalart P, Allmaras S (1992) A one-equation turbulence model for aerodynamic flows. AIAA Meet Pap.
210
to provide decent results for boundary layers exposed to adverse pressure gradients394-395.
394 Dehghan Manshadi M, Jamalinasab M (2016) Optimizing a two element wing model with morphing flap by
means of the Response Surface Method. Iran J Sci Technol Trans Mech Eng.
395 Catalano P, Amato M (2003) An evaluation of RANS turbulence modelling for aerodynamic applications. J
concept at supersonic conditions by means of the response surface method”, Journal of the Brazilian Society of
Mechanical Sciences and Engineering (2018) 40:254.
397 See Previous.
211
7.1.4.7 Validation
To ensure the integrity of present numerical method, computational results obtained by [Oyama et
al.]398 are compared with the numerical simulation data. The delta wing examined here has leading
edge sweep angle of 65 degree. The leading edge is sharp and lee-surface is flat. Geometric
specification of the delta wing is presented in Table 7.1 and different views of its geometry are
shown in Figure 7.14. Flow
field around the wing geometry
is supersonic and Mach number
is 1.7. RANS equations along
with Spalart–Allmaras
turbulence model are utilized
for solving the governing
equations in domain of study.
Spanwise distributions of
pressure coefficient on the
upper surface at 95% chord
position of the delta wing are
compared in Figure 7.15.
According to the extracted
results, qualitative objectives of Figure 7.14 Delta Wing Geometry Used for Numerical Validation
the present computational
approach are satisfactory.
Figure 7.15 Spanwise Cp over the Delta Wing Upper Surface at M = 1.7, AoA = 8 Degree
_
398 OyamaA, Imai G, Ogawa A, Fujii K (2008) Aerodynamic characteristics of a delta wing at high angles of attack.
AIAA Meet Pap.
212
(c) CL/CD vs CL
prevent the extension of this pressure gradient, which in turn increases the overall value of L/D. As
sweep wings have the potential of achieving comparatively low drag at supersonic lifting conditions,
for almost all high-speed aircraft, a rational increase in sweep angle has a positive outcome. However,
a decrease in stall angle of attack and tip stall of a swept wing are some of the major constraints. Thus,
acquiring an optimal wing design is extremely challenging, especially in supersonic condition, and a
trade-off between several essential parameters has to be well thought-out.
For investigating the effect of t/c on the aerodynamic coefficients, thickness of the airfoil is reduced
from 10 to 6%. Figure 7.16-d show drag coefficient values for two distinct thickness to chord ratios
at Mach number 1.4. It can be seen that decreasing the thickness of the wing in any given lift
coefficient has favorable effects on CD and a rise in CL/CD ratio. Thus, to obtain a higher maximum lift-
to-drag ratio, the airfoil ought to be thinner, on the premise of good stall characteristics. It is clear
that the eventual goal and mission of each aircraft are major constraints for excessive reduction of
t/c.
Reducing thickness of the wing, not only has structural and operational limitations, but also
decreases the internal space of wing which is suitable for fuel tanks and other equipment. Pressure
coefficient (Cp) charts is display the effects of sweep angle and thickness ratio of the airfoil on
pressure distribution over the wing399. Reduction in t/c with a systematic increase of sweep angle,
reduces the free stream velocity over wing surface, and by averting the undesirable effects of adverse
pressure gradient range on the wing causes an improvement in lift force. Formation of shock waves
in flow around wing does not have a positive outcome. This is because much of the flow energy is
wasted by drag and friction400.
Mach number contours in the direction of wing root chord with t/c equal to 10 % for sweep angles
of (37, 60), and 5 deg, angle of attack is, respectively, demonstrated in Figure 7.17 (a-b). Velocity
distribution is evidently changed for both upper and lower side of the wing for all cases. As can be
seen, applying sweep angle to the wing results in delaying of curved and oblique shocks. A high Mach
399 Mojtaba Dehghan Manshadi, Soheil Aghajanian, “Computational aerodynamic optimization of wing-design
concept at supersonic conditions by means of the response surface method”, Journal of the Brazilian Society of
Mechanical Sciences and Engineering , 2018.
400 Roelof V, Farokhi S (2015) Introduction to transonic aerodynamics. Springer, New York.
214
number causes a strong acceleration of flow around the wing and generation of a large supersonic
region with a shock wave. This can have a significant influence on aircraft and often leads to
undesirable effects such as loss of performance or structural damage.
Thus, sweep angle by diminishing strength of shock waves, which in turn leads to a decrease in wave
drag of supersonic flow, increases the aerodynamic efficiency of the wing. At cruise condition, the
flow around wing of a highspeed aircraft is classified by the presence of shock waves on the suction
side of the profile. The result is the existence of shocks and the shear layers resulting from the
interaction with the boundary layers developing over a surface401.
y = β0 + β1 x1 +. . . . . . . +βk xk + ε
Eq. 7.12
In the case of any form of curvature in the system, higher-order polynomials have to be used:
k k
401 Sartor F, Timme S (2015) Reynolds averaged Navier–Stokes simulations of shock buffet on half wing body
configuration. AIAA SciTech Forum.
402 Monotgormery D (2008) Design and analysis of experimental. Wiley, Amsterdam.
403 Myers RH, Montgomery DC, Anderson-Cook CM (2009) Response surface methodology: process and product
Press, Biometrika.
215
where xi and xj are the design variables, β are the tuning variables and e represents the noise or error
observed in the response y. In this paper, to achieve a high quality and comprehensive model, by
applying DOE module, based on CCD and RSM, some of the important wing shape parameters are
investigated. As shown in Table 7.3, four design parameters including leading edge sweep angle,
thickness to chord ratio of the airfoil, angle of attack through cruise flight and aspect ratio of the wing
are designated. Similar procedures and process for stipulating the design variables have been
performed by [Yoshida] and
[Chiba et al.]405 for multi- Aspect Ratio AoA t/ c Leading Edge Sweep Angle
objective aerodynamic 2.7 0 0.06 37 37 deg.
optimization of a supersonic
3,3 3 0.1 65 deg.
wing shape. In selection of
variables range and lower and
higher bounds of design space, Table 7.3 Range of Design Variables
some predefined conditions as
well as geometric limitations have been taken to account. The goal of this project was to analyze and
to achieve a supersonic wing design in a prearranged criterion; like, sweep angle between 37–60
Table 7.2 Design of Experiments (DoE) result; input and output parameters value
405 ChibaK, Makino Y, Takatoya T (2007) Multidisciplinary design exploration of wing shape for Silent supersonic
technology demonstrator. AIAA Meet Pap.
216
degrees, maximum aspect ratio of 3.3. As shown in various charts of the previous sections, sweep
angle has a great effect in controlling aerodynamic parameters of this type of wing.
That’s the reason of 5 degree increase in its higher bounds for a chance of achieving an even better
aerodynamic efficiency. Table 7.3 displays the defined range of design variables in this work. ANSYS
software’s CCD algorithms generate 25 design points for 4 input parameters406. RSM has the potential
to quickly provide the approximated values of the output parameters throughout the design space
without having to perform a complete solution. However, in this study, for increasing the accuracy of
response surfaces, 25 CCD generated design points were numerically solved by the FLUENT 16.0
solver and outputs were manually inserted.
Each configuration corresponds to a different geometry for which all the output parameters were
numerically evaluated by the solver. Clearly, the accuracy of the response surface outcome is a
function of the data fitting of the response surfaces to the DOE design points. In addition, verification
points are applied to determine the accuracy of fit criterion which compares the response surface
outputs with the DoE results
used to create them. For
calculating the error after
creating the response
surface, verification points
are placed in locations that
maximize the distance from
current DOE points. These
points are never used for
training the model. If the
model can fit well with these
data without using them for
training, optimization Figure 7.18 Predicted from RSM vs. observed from DOE, Normalized
procedure continues further values
which is a sign of good
predictability. Hereby, by specifying the refinements and increasing the number of verification points
of the solver, relatively accurate surface model is achieved. Table 7.2 shows the DoE result of this
study. Figure 7.18 displays the scatter chart that values predicted from the response surface versus
the values observed from the design points. It specifies that response surface correctly fits the points
of the DoE.
7.1.5 Optimization
Once the RSM data are obtained, it is possible to run a multi-objective genetic algorithm (MOGA)
and also carry out a sensitivity analysis, which is provided in the following sections. In multi-objective
optimization, also known as Pareto optimization, several objective functions exist, and the goal is
optimizing those functions by getting their minimum or maximum values simultaneously. Most of
these objectives are in contrast to each other, so that improving one of them worsens the other
responses that are not appropriate. Evolutionary algorithms are popular approaches for generating
Pareto optimal solutions to a multi-objective optimization problem. One of the most prominent and
widely used evolutionary algorithms is the improved version of Non-dominated Sorting Genetic
Algorithm, called NSGA-II, which is suggested by [Deb et al.]407. NSGA-II is easy to use, almost
parameter-less and is fast due to its non-dominated sorting with the crowding distance as a diversity
metric.
406ANSYS© Academic Research, Release 16.0, Help System, Workbench Guide, ANSYS, Inc
407Deb K, Pratap A, Agarwal S, Meyarivan T (2002) A fast and elitist multi objective genetic algorithm: NSGA-II.
IEEE Trans Evol Comput 6:182–197.
217
NSGA-II is implemented in Design-Xplorer module of ANSYS® commercial CFD solver, version 16.0.
Lift and drag coefficients are considered as objective functions of this study. Final goal of optimization
process is minimizing the drag and maximizing the corresponded lift to achieve the finest lift-to-drag
ratio. By applying GA, as shown in Table 7.4, three candidates are derived, as final optimum design
points. The first candidate was selected as the Table 7.3 Defined range for design variables Aspect
ratio AoA , t/c Leading edge sweep angle optimum wing geometry. From a design perspective, a
single point must be chosen
from amongst these Pareto Design Parameters First Second Third
solutions as the optimum or Nominee Nominee Nominee
most acceptable solution. Sweep Angle (deg.) 64.78 63.01 60.09
In this study, MOGA method Aspect Ratio 3.277 3.113 3.248
generated 10,000 samples t = c (%) 6.25 6.43 6.64
initially and performed 5000 AoA (deg.) 2.60 2.47 2.36
samples per iteration. Solution CL 0.1745 0.1681 .0.1664
converged after 23,842 CD 0.0237 0.0234 0.0251
evaluations and 410 iterations, CL/CD 7.36 7 7.18 6.63
which generated 3 final
candidates. These generated Table 7.4 Three Candidates as Final Optimum Design Points
points belong to the best Pareto
fronts. In the trade-off chart,
samples are ranked by non-dominated Pareto frontier. In the preliminary phase of this study and
before conducting the design of experiments, 18 different wing geometries are numerically analyzed.
Throughout this level, maximum CL/CD achieved at t/c 6%, CL of 0.2183 in a corresponding angle of
attack of 5 deg. Table 7.5 tabulates CL, CD, and CL/CD for both optimized geometry and the finest
design case selected among
the preliminary geometries. Cases CL CD CL/CD
Aerodynamic efficiency is Best Geometry Among Initial 0.2183 0.0342 6.38
increased by 15% and CD is Cases
reduced by approximately Optimized Geometry 0.1745 0.0237 7.36
30%, as a result of
optimization study of this Table 7.5 Aerodynamic Coefficients Comparison before and after
supersonic wing geometry. Optimization
In addition, to ensure the
reliability of
the
optimization
process, the
optimized
geometry is
evaluated with
the CFD solver.
By generating
precise mesh
and grid
system, flow
field around
the optimized
wing is
analyzed and Figure 7.19 Final and Optimized Wing Geometry
through
218
numerical simulations final results acquired. Lift and drag coefficient values, as objective functions,
evaluated by numerical calculations are 0.1716 and 0.0228, respectively. This yields to the significant
design element of any aerial vehicle, aerodynamic efficiency of 7.52. The aerodynamic cruise
efficiency at a particular Mach number is significant for all aircraft, as it has a direct impact on the
achievable range. It can be seen that there is negligible variance between calculated values of the
numerical solution and predicted values by the optimization algorithm. This indicates the high
capability of the whole solution procedure in simulating flow field and solving the related governing
equations for offering an optimized wing geometry for supersonic flights. Optimized wing geometry
is depicted in Figure 7.19.
represents the beginning of it, which is the lower bound of the input parameter as specified in Table
7.3. According to this non-dimensionalized chart, lift has higher sensitivity to changes in sweep angle
than drag coefficient. As sweep angle increases from 37 to 50 degrees, drag coefficient decreases by
an amount of 0.4892 along the Y-Axis, while CL decreases by a quantity of 0.8187 along the X-Axis.
This implies that sweep angle has pronounced impact and influences on variations in both lift and
drag coefficients.
Figure 7.20-b indicates the effects and sensitivity of changing just aspect ratio of the optimized wing
on drag and lift coefficients. Steep slope indicates the sensitivity of design parameters and outputs
on each other. Both of the objective functions have high sensitivity to fluctuations in this input
variable. Thus, by increasing aspect ratio in the specified range according to Table 7.3, CL and CD
take ascending trend. This trend is consistent for any distinct aspect ratio. However, as aspect ratio
increases from lower bound of 2.7 up to the optimized point of 3.277, CL shows marginally more
sensitivities to this factor than drag. Figure 7.20-c shows how much two objective functions are
inclined to changes in thickness to chord ratio. It can be seen that any increase in t/c causes a
continuous negative trend on CL. By increasing t/c more than a certain limit, drag coefficient rises
remarkably, which indicates a higher sensitivity of drag on wing’s overall thickness.
7.1.5.2 Conclusion
The successful application of multi-objective optimization design for supersonic wing configurations
has been presented. The multi-objective design and aerodynamic shape optimization of a wing with
capability of flying efficiently and economically at supersonic regime has been explored using
numerical simulations, design of experiment techniques and Genetic Algorithm. The objective
functions of the optimization process are lift and drag coefficients of the wing. In this study, the effects
of different sets of aerodynamic design variables are explored in the geometry design and
optimization procedure. The compressible flow field around the wing geometries is solved using the
CFD tools and numerical calculations. Values of aerodynamic parameters are validated by the results
published in reliable studies. The effects of the design parameters on lift, drag and pressure
coefficients are studied by various charts and figures. Furthermore, the particular design variables,
which have severe effects on the objective functions, are found. Wing sweep angle alongside aspect
ratio have strong effects on the objective functions. It can be seen that, achieving a final geometry in
multi-objective optimization, for a supersonic wing shape design is extremely challenging and trade-
offs among several variables have to be pondered. In a fixed state, with the optimized supersonic
wing geometry compared to the best geometry among initial cases, maximum lift-to-drag ratio
increased from 6.38 to 7.36. For further information, please see the work by [Manshadi &
Aghajanian]408.
➢ Pre-hybridization, where, for example, the population of the (GA) is pre-optimized using the
Gradient-Based Method (GBM);
➢ Organic-hybridization, in which the (GBM) is used as an operator within the (GA)s for
improving each population member in each generation;
➢ Post-hybridization, in which the (GA)s final population is used to provide an initial design for
the (GBM).
It is should be noted that these classifications are not limited to the hybridization of GAs and GBMs,
however, the most frequent hybrid algorithm found in aerospace application have been designed in
attempt to combine the best characteristics of GAs and GBM. [Gage et al.]409 present the post-
hybridization of a classical GA with sequential quadratic programming for the topological design of
wings and trusses. This work is notable for hybridization as it is one of the first hybrid methods (HM)
employed in aerodynamic design optimization. They demonstrated that post-hybridization is
effective for final refinement of the GA's candidate solutions. By switching to a GBM once the GAs
population is sufficiently mature, computational demands can be reduced and superior solutions can
be found relative to allowing the GA to continue. In more recent work, [Kim et al.]410 also use post-
optimization, to improve the aerodynamic and acoustic performance of a axial-flow fan, by combining
the multi-objective real-encoded NSGA-II from which the Pareto-optimal solutions are further
optimized using sequential quadratic programming. The specific difficulty in this method of
hybridization is the transition from a multi-objective problem to a single-objective problem. There
are typically two ways to transition:
➢ Combine all of the objectives into a composite objective using a weighted-sum approach for
example;
➢ Sequential optimization, optimizing one objective at a time while treating all other objectives
as equality constraints.
[Kim et al.]411 adopted the latter method, pointing out that it did not preserve Pareto optimality and
created a set of optimal solutions for each objective with many duplications forming. Similar pre-
optimization strategies have been employed by [Xing & Damodaran]412 which combine GA stochastic
searching and GBMs analytical optimization procedures in the optimization of nozzle shapes.
Compiled optimization results from [Chernukhin & Zingg]413 for a series of optimization problems
was found to significantly outperform other algorithms for highly-multi-modal problems; namely
the Griewank function which offers hundreds of locally optimal solutions.
The classical GA has been extended for the purpose of more efficient aerodynamic shape design
optimization by [Catalano et al.]414 to include two new operations based on gradient search in a
hybrid organic-optimization algorithm. Figure 7.21 shows how the classical genetic algorithm has
been modified to include such operators. The activation of each gradient operator is controlled
probabilistically, inspired by classical crossover and mutation operation use. The first gradient-based
operator, which acts on the whole population, has two further probabilistic controls determining
409 P. Gage, I. Kroo, and P. Sobieski. “Variable-Complexity Genetic Algorithm for Topological Design”. AIAA J, 1995.
410 J. Kim, B. Ovgor, K. Cha, J. Kim, S. Lee, and K. Kim. “Optimization of the Aerodynamic and Aero acoustic
Performance of an Axial-Flow Fan”. AIAA Journal, 52(9):2032-2044, 2014.
411 See previous.
412 X. Xing and M. Damodaran. “Design of Three-Dimensional Nozzle Shape using NURBS, CFD and Hybrid
Optimization Strategies”, 10th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, 2004.
413 O. Chernukhin and D.W. Zingg. “Multimodality and Global Optimization in Aerodynamic Design”. AIAA J, 2013.
414 L.A. Catalano, D. Quagliarella, and P.L. Vitagliano. “Aerodynamic Shape Design Using Hybrid Evolutionary
Computing and Multigrid-Aided Finite-Difference Evaluation of Flow Sensitivities”. International Journal for
Computer Aided Engineering and Software, 32(2):178-210, 2014.
221
behavior with each candidate solution. The first determines the maximum number of iterations and
the second determines the sensitivity analysis method the gradient optimizer uses. The second
gradient operator is relatively simpler, applying only one optimization iteration to the current best
solution according to the steepest descent rule with a random step size. Finally there is an exclusive
mechanism which preserve the test solutions from each generation and reintroduces the best known
solution from all generations into the current population.
For the single-objective optimization of an aero-foil with 24 variables, [Catalano et al.]415 find that the
complex HM developed is comparable to the standard GBM in terms of the number of function
evaluations only achieving a 0.4% better objective. Under different settings (population, number of
iterations, sensitivity analysis, etc.), they find that the hybrid algorithm slightly outperforms the
gradient method, reducing the objective function further by 2.2%, but requires many more function
evaluations. The classical GA is outperformed by all other algorithms. Therefore, the hybridized GA-
GBM showed accelerated performance in aerodynamic optimization relative to a standard GA and
capable of matching, and in one case able to outperform, the GBM. The GBM (with a suitable
sensitivity analysis) is shown overall more effective, the hybrid algorithm performed well however
added unnecessary complexity to the optimization framework. [Jong-Taek Oh & Nguyen Ba Chien]416
purpose an optimization procedure coupling (CFD) and genetic algorithms (GAs) as depicted in
Figure 7.22. Current the procedure is deployed relying on the open-source software OpenFOAM in
415 L.A. Catalano, D. Quagliarella, and P.L. Vitagliano. “Aerodynamic Shape Design Using Hybrid Evolutionary
Computing and Multigrid-Aided Finite-Difference Evaluation of Flow Sensitivities”. International Journal for
Computer Aided Engineering and Software, 2014.
416 Jong-Taek Oh and Nguyen Ba Chien, “Optimization Design by Coupling CFD and Genetic Algorithm”, 2018.
222
CFD block, the DAKOTA in optimization block, and the interface script to connect two blocks.
Figure 7.22 Flow Chart of Coupled Optimization Framework – Courtesy of [Oh and Chien]
7.2.1 Case Study - Airfoil and Wing Design Through Hybrid Optimization Strategies
Real world design problems need robust and effective system-level optimization tools, as they are
ruled by several criteria, most often in multidisciplinary environments. In this work by [Vicini &
Quagliarella]417 hybrid optimization algorithm has been obtained by adding a gradient based
technique among the set of operators of a multi-objective genetic algorithm. This way it has been
possible to increase the computational efficiency of the genetic algorithm, while preserving its
favorable features of robustness, problem independence and multi-objective optimization
capabilities. The results here illustrated regard aerodynamic shape design problems, including both
airfoil and wing design.
417 A. Vicini, D. Quagliarella, “Airfoil and Wing Design Through Hybrid Optimization Strategies”, AIAA-98-2729.
418 Optimum Design Methods for Aerodynamics, AGARD-R-803, Nov. 1994.
223
• Robustness, intended as the capability of avoiding local optima, vs. the need of human
interaction and expertise;
• Capability of multiple objective optimization vs. single-objective one;
• Computational efficiency vs. the need of large computational resources.
From this point of view, the choice of one particular optimization technique implies the renunciation
of some possible advantages in favor of some others. On the other hand, due to the fact that
aerodynamic shape design represents only a part of the overall design of a flying vehicle, and that the
need for an effective multidisciplinary approach to the design task is arising, it is important for an
optimization tool to combine as much as possible all the favorable characteristics stated above while
avoiding the shortcomings. In this sense, a hybrid approach to optimization, in which techniques of
different nature are used at the same time, may result extremely beneficial.
Hybrid optimization may in fact exploit the most favorable features of the methods which are
combined while masking the corresponding shortcomings. In this work a hybrid optimization tool,
developed by incorporating a gradient based optimization routine among the operators of a multi
objective genetic algorithm, is applied to aerodynamic shape design problems. Genetic algorithms
belong to the class of evolutionary optimization procedures, which finds its philosophical basis in
Darwin's theory of survival of the fitness test419. In an attempt to mimic the process of biological
evolution, a set of design alternatives, representing a population in this metaphorical transposition,
is let evolve through successive generations so as to promote the individuals which better adapt
themselves to the environment, i.e. those which better meet the design requirements. Each element
is characterized by the value of its fitness, which is the measure of how fit it is for the given
environment; in other words, how good the corresponding solution is for the problem at hand. The
process of evolution is realized in the reproduction phase using a selection criteria driven by the
value of the fitness of the individuals, so that bias is allocated to the best fit members of the
population.
The individuals selected for reproduction are recombined using genetic operators (crossover,
mutation), so that a combination of their most desirable characteristics may be obtained in the off-
springs, and hence elements characterized by higher fitness are produced in subsequent generations.
These search methods rely only on the evaluation of the fitness of the elements and do not require
the computation of gradients; therefore, they are less susceptible to pitfalls of convergence to a local
optimum, and can successfully deal with disjoint or non-convex design spaces. Moreover, they are
capable of facing the problem of multiple objectives optimization in a straightforward fashion, using
the notion of domination among solutions. These characteristics make GAs very attractive
optimization tools, and explain the considerable growth of interest which has been devoted to them
in recent years for applications of engineering interest420.
The major weakness of GAs lies in their relatively poor computational efficiency, as they generally
require a very high number of evaluations of the objective function. For this reason, the use of GAs
may become unpractical when this evaluation is expensive, as happens for aerodynamic optimization
applications where the solution of complex partial differential equation systems is necessary.
Coupling a genetic algorithm with a different optimization technique can be an effective way to
overcome its lack of efficiency while preserving its favorable features. Many different strategies to
hybridize the GA can be realized; the simplest one is that of using the best solution found by the GA
as starting point for a subsequent optimization with the other method adopted421. However, a closer
419 Holland, J. H., “Adaptation in Natural and Artificial Systems”, The University of Michigan Press, 1975.
420 Ahmed, Q., Krishnakumar, K., and Neidhoefer, J., “Applications of Evolutionary Algorithms to Aerospace
Problems - A Survey", in Computational Methods in Applied Sciences '96, John Wiley & Sons Ltd., England, 1996,
421 Poloni, C., “Hybrid GA for Multi Objective Aerodynamic Shape Optimization", Winter G. et al., editors, Genetic
Algorithms in Engineering and Computer Science, John Wiley & Sons Ltd., England, Dec. 1995, pp. 397-416.
224
interaction between the different algorithms, rather than the two-stage optimization described, may
more favorably combine the best features of both methods, and provide results better than those
obtainable using either of the two techniques. Some representative results obtained with the hybrid
genetic algorithm (HGA) will be demonstrated for aerodynamic shape design problems, including
both airfoil and wing optimization test cases.
However, the rate of mutation needed for these two tasks may be different; in particular, while
mutation is very good for maintaining population diversity, its refining capabilities may not be
optimal for every class of problems. There is in fact a broad class of problems, namely the ones where
the fitness function is differentiable, for which gradient based techniques are much more efficient to
locally improve a given solution. This suggests the introduction of a gradient based routine among
the set of operators of the GA; mutation is then prevalently left with the role of keeping the diversity
among population elements at an optimum level. The genetic algorithm developed adopts a bit string
codi_cation of the design variables; anyway, this does not prevent the use of operators requiring real
number list encoding, such as extended intermediate crossover and word level mutation422.
In these cases the binary string is decoded into a real number list, the operator is applied and the set
of modified variables is encoded back into a bit string. This scheme allows the use of a free mix of
different type of operators; among these, as said before, a routine performing a gradient based
optimization (with a conjugate gradients technique) has been included, and called “hill climbing
operator" (HcO). The HcO operates as follows: through the application of the selection, crossover
and mutation operators, an intermediate generation is created from the current one. Afterwards, if
the hybrid option is activated, some individuals may be selected and fed into the HcO to be improved,
and then introduced into the new generation, as sketched in Figure 7.23. Regarding the choice of
the elements to be fed into the HcO, in the case of single objective optimization three different
strategies are possible:
Of course, these strategies determine different levels of selection pressure, decreasing from strategy
422Vicini, A., and Quagliarella, D., “Inverse and Direct Airfoil Design Using a Multi objective Genetic Algorithm",
AIAA Journal, Vol. 35, No. 9, Sep. 1997, pp. 1499-1505.
225
#1 to strategy #3; the relative performance will therefore depend on the optimization problem. The
above described scheme can be naturally extended for multi-objective optimization. In this case the
elements are not ranked on the basis of a scalar fitness function, but are just divided into two classes:
the dominated and the not dominated ones423. The set of not-dominated individuals (Pareto front),
updated after each new generation, is composed by all potential solutions of the problem, satisfying
the design criteria at different levels of compromise. When multi-objective problems are formulated,
strategy 1) becomes the (random) selection of a number of elements determined by an assigned
probability from the current set of Pareto optimal solutions, while strategies 2) and 3) remain the
same. Of course, the HcO is by its nature capable of dealing only with scalar objective functions; thus,
when multi-objective problems are faced, the objective function fed into the HcO is obtained through
a weighted linear combination of the n problem objectives, i.e. as obj = α obj1 + (1−α) obj2 in the case
of n = 2.
Hill Climbing
Figure 7.23 Sketch of the Hybrid Genetic Algorithm – Courtesy of [Vicini & Quagliarella]
The weighting factor α can be chosen at random or assigned explicitly to favor one of the objectives.
It must be remarked, on the basis of what previously stated, that the aid provided by the HcO simply
consists in its capability to improve to some extent the selected individuals; in other words, the role
of the HcO is that of introducing improvements which will then be processed and exploited by the
GA, which remains the driving engine of the procedure. From this point of view, the use of the HcO
has to be limited to the minimum necessary to provide the desired effect, which is that of improving
the convergence characteristics of the procedure without causing premature convergence to local
minima. On the basis of these considerations, the use of the HcO has been subject to the following
rules:
• Hybrid mechanism sketched in Figure 7.23 is not used at each generation, but only after
each assigned block of K generations;
• It is not necessary { and it would be detrimental for the computational performances } for the
HcO to carry out each time a converged optimization.
Therefore, only one or two gradient iterations are generally prescribed. Furthermore, as the HcO has
basically to behave as an improved mutation operator, the beneficial effects of hybridization can be
obtained also by making it operate only on a subset of the active design variables, thus reducing the
number of objective function evaluations that the HcO needs to carry out each time; the success in
423Goldberg, D. E., “Genetic Algorithms in Search, Optimization and Machine Learning”, Addison-Wesley,
Reading, Massachusetts, Jan. 1989.
226
this case will depend on the degree of cross correlation among the design variables. The total number
of evaluations of the objective function needed, Ne, can be estimated as follows:
φ
Ne = Npop N gen [1 + N (η Nvar + 4)]
K it
Eq. 7.14
where Npop and Ngen are the population size and the total number of generations, φ is the HcO
probability, K is the frequency in terms of number of generations for the activation of the HcO, Nit is
the number of gradient iterations, Nvar is the number of design variables, and η = [0; 1] the factor
determining the size of the subset of design variables that is passed to the HcO. In the applications
that will be illustrated, the design variables that need to be frozen are chosen at random each time
the HcO is used; in this way, a different subset of design variables is passed each time to the HcO.
Table 7.6 Set of operators of the two GAs used – Courtesy of [Vicini & Quagliarella]
424 Poloni, C., “Hybrid GA for Multi Objective Aerodynamic Shape Optimization", Winter G. et al., editors, Genetic
Algorithms in Engineering and Computer Science, John Wiley & Sons Ltd., England, Dec. 1995, pp. 397-416.
425 Vicini, A., and Quagliarella, D., “Inverse and Direct Airfoil Design Using a Multi objective Genetic Algorithm",
and a 50 individuals population evolved for 100 generations; the hybrid strategies have been
activated so as to select on average only one individual every other generation, and carry out 2
gradient iterations (φ = 0:02, K = 2, Nit = 2, η = 1). Hence, to consider the same total number of
objective function evaluations, the hybrid strategies must be judged approximately at generation 70.
Two different GAs, characterized by the set of operators described in Table 7.6, have been used with
and without hybridization. Figure 7.24 illustrates the convergence histories, each one averaged
over 10 successive trials characterized by different starting populations; the convergence history
obtained by the application of the gradient based method by itself is also shown in the same figure;
besides, it must be noted that a restart procedure had to be used in this case to take the solution out
of a local minimum where it got stuck after a few iterations.
Figure 7.24 Convergence Histories for the CAST 10 Inverse Design Problem - Courtesy of [Vicini &
Quagliarella]
As can be seen, for a given GA, hybridization is always beneficial, meaning that a better result can be
found with the same amount of computations, or that the same result can be obtained with a
substantial reduction of computation needed (ranging in this case from 30 to 75%). In particular,
strategy #1, when the HcO is applied only to the best fit individuals, appear as the less effective,
probably due to an excessive selection pressure. At the same time, the behavior of the gradient based
method is considerably improved from the point of view of the robustness.
Another important characteristic that needs to be considered is the statistical dispersion of the
results obtained starting from different initial populations; in fact, if it is correct to judge the
convergence characteristics of a given GA by averaging the results of a number of runs, from an
application-oriented point of view it is more important for the algorithm to guarantee satisfactory
convergence performances even on a single run basis. Figure 7.25 shows all the values of the
objective function obtained at the end of each of the 10 different runs, for each one of the algorithms
used; it can be observed how the scatter of the results provided by both basic GAs is much higher
than that obtained using the corresponding hybrid algorithms.
In particular, the best behavior from this point of view is obtained when the elements to be fed into
the HcO are chosen at random, so that the level of selection pressure is not increased too much. The
same runs have then been repeated using η = 0.5; in this way, the HcO acts on a subset of 9 design
variables out of 18, that are chosen at random each time. The convergences obtained are illustrated
in Figure 7.26, limitedly to the hybrid strategies #2 and #3, i.e. those giving the best performances.
228
In this case the actual number of fitness evaluations has been used for the x axis, in order to better
compare the results. We see how freezing some of the design variables has a positive effect when GA
#1 is used, whereas for GA #2 convergence is slowed down to some extent. Considering that the
design variables for this problem are strongly cross-correlated, as it is not possible to move one
control point of the B-splines independently from the others, this result shows that this approach can
generally be used with success. An example of multi-objective optimization is then presented,
consisting in reducing the wave drag of the airfoil while keeping the corresponding pitching moment
under control, for a fixed lift coefficient and maximum thickness.
Figure 7.25 Scatter of the Results Obtained in 10 Different Runs for the CAST 10 Inverse Design
Problem – Courtesy of [Vicini & Quagliarella]
The airfoil chosen as initial geometry is the RAE 2822,10 at a design point M = 0.78, CL = 0.75. The
constraint on lift coefficient is satisfied by letting the flow solver find the angle of attack that produces
the specified lift, while the thickness of the airfoil is scaled to the desired value after each geometry
modification; in this way, every solution is a feasible one. The two objective functions have been
evaluated as obj1 = Cdw= C2L , obj2 = C2m. A population of 100 individuals was let evolve for 100
generations; selection was carried out by means of a 3 steps random walk, with one-point crossover
(Pc = 1) and bit mutation (Pm = 0.02). Differently from the inverse design previously described, the
geometry of the airfoil has been represented as a linear combination of the initial one, yo, and a
number of modification functions, yi :
y = y0 + ∑ x i yi
i=1
Eq. 7.16
The coefficients xi of this combination are the design variables; the functions yi have been obtained
as the difference between the initial geometry and the geometries of other airfoils chosen from a
database, so that a particular aerodynamic effect can be associated to each design variable. The
allowable range assigned was xi ∈ [−0.2; 1.2], i = 1, N, and 12 design variables were used. The same
run was then repeated using the HcO, on average, on one element per generation, chosen at random
from the current Pareto front, and carrying out only one gradient iteration. The run in this case was
stopped at generation #86, in order to establish the comparison for the same total number of
229
evaluations of the objective functions. It can be seen how the result provided by the hybrid algorithm
is a Pareto front characterized by solutions of higher quality, and more uniformly distributed; only
12 of the solutions found by the GA are not dominated by those obtained with the HGA. (see [Vicini
& Quagliarella]427.
Figure 7.26 Comparison of the Convergence Histories Obtained by Letting the HcO Operate on All
Design Variables or on a 50% Subset – Courtesy of [Vicini & Quagliarella]
427A. Vicini, D. Quagliarella, “Airfoil and Wing Design Through Hybrid Optimization Strategies”, AIAA-98-2729.
Doorly, D. J., Peir_o, J., and Oesterle, J.- P., “Optimization of Aerodynamic and Coupled Aerodynamic-Structural
428
Design Using Parallel Genetic Algorithms", 6th AIAA/NASA/USAF Multidisciplinary Analysis & Optimization
Symposium, AIAA Paper 96-4027, Bellevue, Seattle, WA, U.S.A., Sep. 1996.
230
2822 airfoil; for simplicity, the wing planform is maintained trapezoidal, so that all geometric
characteristics vary linearly from the root section to the tip. A total of 5 design variables have been
used: 4 of these act directly on the wing planform, namely the taper ratio θ, the sweep angle at 25%
of the chord Λ, the aspect ratio AR and the twist angle θ; moreover, the thickness at the wing root has
also been included among the design parameters, while the thickness at the wing tip has been fixed
at t/c t = 10%. The wing surface is kept constant, so that the average wing loading is not changed
during optimization.
Table 7.7 Design Parameters for the Wing Planform Optimization – Courtesy of [Vicini & Quagliarella]
In Table 7.7 the initial values of the design parameters are reported together with the prescribed
allowable ranges. The wing twist is distributed symmetrically between the root and the tip; in other
words, a twist angle θ corresponds to an increase of local incidence of θ/2 at the tip, and a decrease
of θ/2 at the root. The wing weight is computed using the algebraic equation; this equation combines
analytical and empirical (statistical) methods, and shows design sensitivity and prediction accuracy
that make it possible to use it with success for preliminary design. As can be seen from , two separate
runs have been carried out exploring separately the positive or negative sweep design spaces; in fact,
the choice of a positive or negative swept wing is based on considerations of different nature,
including stability and handling characteristics.
The selection has been carried out through a 3 steps random walk, with one-point crossover (Pc = 1)
and bit mutation (Pm = 0.1); a population of 64 individuals was let evolve for 50 generations. The HcO
has been used on one element for each generation, selected from the current Pareto front, with Nit =
1. The Pareto fronts obtained are illustrated in Figure 7.27 (where Wo is a reference weight),
together with the planform of the wings corresponding to the extremities and to the center of the
fronts. It can be seen that, for a given value of aerodynamic drag, the negative swept wings are heavier
than the corresponding positive swept ones; therefore, almost all the solutions with a negative sweep
angle would be dominated if the two Pareto fronts were merged. The fronts are populated by 116
and 100 individuals, in the case of positive and negative swept wings, respectively. The use of the
HcO doesn't prevent the development of the complete Pareto front; on the contrary, the solutions are
uniformly distributed along the fronts without the need of niching techniques. In order to evaluate
how well these solutions are representative of the real Pareto fronts, the same test case has been
solved using the gradient based method by itself, with the problem formulated through the weighted
linear combination approach (i.e. the same used by the HcO): obj = α obj1 + (1− α) obj2 ; 5 different
231
Figure 7.27 Pareto Fronts Obtained for the Wing Optimization – Courtesy of [Vicini & Quagliarella]
The solutions thus obtained are compared in Figure 7.28 with the Pareto fronts provided by the
HGA. As can be observed, these solutions lie at most on the Pareto Fronts, and in some cases they fall
in the dominated solutions region. It is also interesting to observe that in neither cases the gradient
method is capable of finding the solution of minimum drag, i.e. that corresponding to α = 1. The values
of the design parameters of the solutions belonging to the Pareto fronts are shown as a function of
aerodynamic drag. As can be expected, the sweep angle varies in an almost linear fashion from the
maximum allowable values (positive or negative) to zero. Similarly, the aspect ratio is at its maximum
at the low-drag end of the front, and rapidly diminishes to the minimum as drag increases. (see [Vicini
& Quagliarella]429.
It can be seen how changing the aspect ratio from 8 to 6 implies an increase of aerodynamic drag of
about 80%; this increase is composed for 60% by induced drag, and by wave drag for the remaining
40%. The taper ratio remains approximately at the minimum allowable value, λ = 0.1, for most part
of the front, assuming higher values only for the solutions corresponding to minimum drag; in the
case of positive sweep minimum drag is obtained for a taper λ = 0.5, whereas for negative sweep a
higher value is necessary, λ = 0.78. The role of twist is essentially that of redistributing the span wise
loading so as to better approach the elliptic distribution; this explains the opposite sign of the twist
angle that is obtained when positive or negative swept wings are considered.
It can also be observed how higher values of twist are necessary in the latter case. Finally, the
behavior of the thickness at the root section appears less intuitive; only at the low-weight end of the
front a clear trend can be observed, with an almost linear increase of drag with the thickness. As
anticipated, after optimization of the planform a further improvement of the aerodynamic
characteristics has been obtained by modifying the shape of the wing section. One of the solutions
belonging to the Pareto front has been selected as starting point; attention has been focused on the
positive sweep angles, and the geometry chosen lies approximately at the center of the front, being
characterized by CD=C2L = 0.776 and W/Wo = 0.65. The wing section has been modified using the
same shape functions technique described in x3; 12 design variables have been used also in this case,
and for simplicity the wing profile has been maintained constant in the span wise direction. As
429 A. Vicini, D. Quagliarella, “Airfoil and Wing Design Through Hybrid Optimization Strategies”, AIAA-98-2729.
232
modifying the wing section may have a strong impact on the aerodynamic characteristics but not on
the structural weight, which is going to remain (almost) constant, the optimization problem in this
case has been formulated so as to reduce wave drag with control on pitching moment; the latter in
fact determines the level of trim drag.
Figure 7.28 Comparison Between the Pareto Fronts and the Results Obtained Through the Gradient
Based Method – Courtesy of [Vicini & Quagliarella]
The design objectives have been formulated as obj1 = CDw/C2L, obj2 = (CM−0.5)2; like in the previous
case, the lift coefficient has been fixed to CL = 0.5, and the maximum thickness has been maintained
at the value obtained by the previous run at each span wise station. The same GA parameters used
for the wing planform optimization have been adopted, except for the mutation rate which has been
reduced to Pm = 0.04, and for the population size which has been increased to 100. The Pareto front
obtained where some of the corresponding wing section shapes are also shown. Depending on the
actual design requirements, it is now possible to extract from this front the solution with the desired
characteristics. In particular, the drag rise curve (at CL = 0.5) of the wing before optimization of the
section is compared with those of three wings extracted from the front: the low-drag end of the front,
corresponding to an unconstrained optimization, and the two solutions characterized by the same
pitching moment and wave drag coefficients, respectively, of the initial wing. As can be seen the first
two of these solutions provide an overall improvement of the drag rise curve; at the design point M
= 0.85 the reduction of wave drag is 32 drag counts for the unconstrained solution, and 22 for the
fixed CM one. On the other hand, when the drag coefficient is kept constant so that a reduction of CM
ca be achieved, lower wave drag values are obtained for Mach numbers lower than the design one,
but a steeper increase at higher Mach numbers.
7.2.1.5 Conclusions
In most practical applications, design problems are governed by several criteria, most often deriving
from different disciplines; to approach such design tasks, robust and effective system-level
optimization tools are needed. Genetic algorithms are characterized by a number of favorable
features that make them attracting for this class of problems; besides, multi objective optimization,
which is a peculiar feature of GAs, appears particularly suited for multidisciplinary environments, as
it allows to determine sets of Pareto solutions in the design space where tradeoffs can be
conveniently examined a-posteriori. In fact the generated solutions (Pareto front) represent different
levels of compromise among the design goals or constraints. Therefore, the designer can make
233
his/her choice introducing an a-posteriori selection criteria. The flexibility of the design process can
thus be increased, as the need of interrelating criteria of different natures is avoided, and the effect
of changing constraints can be evaluated off-line. In this work an effective algorithm for multi
objective applications has been developed through a hybrid approach, by coupling a multi objective
genetic algorithm to a gradient based operator. Applications to multi objective airfoil and wing
design have been presented. The basic idea to use a gradient based routine in the fashion of a genetic
operator derives from the observation that mutation by itself is not very effective as a refinement
operator, leading to generally poor convergence speed; on the other hand, it has been demonstrated
how the beneficial effects of the HcO can be exploited even when its use is considerably limited, in
terms of number of elements processed and computations carried out for each element. For the class
of problems that has been investigated, significant improvements have been obtained both with
respect to simple genetic algorithms, in terms of computational efficiency, and with respect to
gradient based approaches in terms of robustness. In particular, it has been possible to use the HGA
with success even in the case of multi objective problems, when a weighting approach must be used
to compose a scalar objective function each time resort is made to the HcO. For full details, please
consult the [Vicini & Quagliarella]430.
➢ Design of Experiment (DoE) sample plan to generate initial sample points in the design space-
point selection;
➢ Numerical simulations are performed to compute the output/performance of each sample
point;
➢ Sample point data (input & output) are used by an approximation model to construct the
surrogate.
430A. Vicini, D. Quagliarella, “Airfoil and Wing Design Through Hybrid Optimization Strategies”, AIAA-98-2729.
431Zhihui Li, Xinqian Zheng, “Review of design optimization methods for turbomachinery aerodynamics”,
Progress in Aerospace Sciences, July 2017.
234
Replacing a particular problem analysis with a surrogate analysis does not affect the problem
formulation, but it will strongly influence the solutions identified. Therefore, once the surrogate
model is constructed it must be validated (sometimes considered a 4th step). This has the purpose of
establishing the predictive capabilities of the surrogate model in design regions away from known
sample data. There are both parametric and non-parametric alternatives in constructing a surrogate
model. Parametric approaches (such as kriging or polynomial regression) are model dependent
forming a functional relationship between the response variables and the design variable samples
that are known. Non-parametric approaches (such as Radial-Basis Functions(RBF) or Artificial
Neural Networks (ANN)) use local models in different regions of the sample data to build-up an over
fall frame work of the model. Furthermore, surrogates can also be classified into regression type
(polynomial regression, radial basis functions) which tend to be better suited to noisy functions, and
interpolation type, creating best response models.
Usage of any of these models is not straight forward as the quantity and quality of information the
user has to provide in the construction of the surrogate is not known a priori. Hence selection of the
most appropriate surrogate is considered problem dependent, as it will directly influence the
optimization algorithms decision making capability. There are no general rules leading to the choice
of type of surrogate, generation of sample data for training and validation, and indeed the
combination of surrogate model and optimization algorithm. Different surrogate models will be
better suited to different data sets and care must be taken to not over-generalize the problem or false
optimization my occur. As an example of Surrogate Modeling for Optimization, the work by [Du et
al.]432 where the results of aerodynamic design of a rectangular wing in subsonic inviscid flow using
surrogate-based local and global search algorithms, was investigated.
7.3.1 Case Study – Surrogate Based Aerodynamic Shape Optimization (SBO) of a Wing-Body of
Civilian Transport Aircraft Configuration
Aerodynamic shape optimization driven by high-fidelity (CFD) simulations is still challenging,
especially for complex aircraft configurations. The main difficulty is not only associated with the
extremely large computational cost, but also related to the complicated design space with many local
optima and a large number of design variables. Therefore, development of efficient global
optimization algorithms is still of great interest. This study by [Han et al.]433, focuses on
demonstrating surrogate-based optimization (SBO) for a wing-body configuration representative of
a modern civil transport aircraft parameterized with as many as 80 design variables, while most
previous SBO studies were limited to rather simple configurations with fewer parameters. The
freeform deformation (FFD) method is used to control the shape of the wing. A Reynolds-averaged
Navier-Stokes (RANS) flow solver is used to compute the aerodynamic coefficients at a set of initial
sample points. Kriging is used to build a surrogate model for the drag coefficient, which is to be
minimized, based on the initial samples. The surrogate model is iteratively refined based on different
sample infill strategies. For 80 design variables, the SBO type optimizer is shown to converge to an
optimal shape with lower drag based on about 300 samples. Several studies are conducted on the
influence of the resolution of the computational grid, the number and randomness of the initial
samples, and the number of design variables on the final result.
432 Xiaosong Du, Anand Amrit, Andrew Thelen, and Leifur Leifsson, Yu Zhang, and Zhong-Hua Han, and
Slawomir Koziel, “Aerodynamic Design of a Rectangular Wing in Subsonic Inviscid Flow by Surrogate-Based
Optimization” 35th AIAA Applied Aerodynamics Conference, 2017.
433 Zhong-Hua Han, Mohammad Abu-Zurayk, Stefan Görtz and Caslav Ilic, “Surrogate-Based Aerodynamic Shape
434 R.M. Hicks, E.M. Munnan, G.N. Vanderplaats, “An Assessment of Airfoil Design by Numerical Optimization”,
NASA TM X-3092 , NASA Ames Research Center, Moffett Field, 1974.
435 R.M. Hicks, P.A. Henne, “Wing design by numerical optimization”. J. Aircraft, (1978)
436 A. Jameson, “Aerodynamic design via control theory”. J. Sci. Computation, 1988.
437 A. Jameson, “”Optimum aerodynamic design using CFD and control theory,” AIAA Paper 95-1729 -1995.
438 D. Koo, D.W. Zingg, “Progress in aerodynamic shape optimization based on the Reynolds averaged navier-
stokes equations”, AIAA Paper 2016–1292, 54th AIAA Aerospace Science Meeting, 4–8, January 2016.
439 O. Chernukhin, D.W. Zingg, “Multimodality and global optimization in aerodynamic design”, AIAA J. 51(6),
1342–1354 (2013).
440 J.Wild, “On the Potential of Numerical Optimization of High-Lift Multi-Element Airfoils based on the Solution
of the Navier-Stokes-Equations”, Computational Fluid Dynamics (Springer, Berlin, 2002), pp. 267–272.
236
441 A.I.J. Forrester, A.J. Keane, “Recent advances in surrogate-based optimization”, 2009.
442 D.R. Jones, M. Schonlau, W.J. Welch, “Efficient global optimization of expensive black-box functions”, 1998.
443 D.D. Cox, S. John, SDO: A statistical method for global optimization, in Proceeding of the ICASE/NASA Langley
Workshop on Multidisciplinary Analysis and Optimization, vol. 2 (Reston, VA, 1998), pp. 738–748.
444 Y. Lian, M. Liou, “Multi objective optimization using coupled response surface model and evolutionary
using kriging model”, 46th AIAA Aerospace Sciences Meeting and Exhibit, 7–10 January 2008.
447 Y.A Tesfahunegn, S. Koziel, J.R. Gramanzini, S. Hosder, Z.-H Han, L. Leifsson, “Application of direct and
7.3.1.4 Initialization
Define the objective function and constraints (such as minimizing CD subject to constant CL as well as
wing thickness constraints); define the design variables and their range by parameterizing the
aerodynamic shape using freeform deformation (FFD), the class/shape transformation (CST) method
or any other methods.
450A.A. Giunta, S.F. Wojtkiewicz Jr., M.S. Eldred, “Overview of modern design of experiments methods for
computational simulations,” AIAA paper 2003-649 (2003).
238
candidate aerodynamic shape. Computational grids are generated (or a grid for the baseline shape is
deformed) for these candidate shapes and CFD computations are run to obtain the corresponding
aerodynamic data needed to evaluate the objective function and constraints. Thereafter, this data is
stored in a database. Note that, as an alternative to starting from scratch, it is possible to use and
append to a given database, which was generated, for example, during a previous optimization run.
This provides the flexibility to the user to terminate the optimization process any time and to restart
later or to change the optimization strategy.
451 N. Kroll, M. Abu-Zurayk, D. Dimitrov, T. Franz, T. Führer, T. Gerhold, S. Görtz, R. Heinrich, C. Ilic, J. Jepsen, J.
Jägersküpper, M. Kruse, A. Krumbein, S. Langer, D. Liu, R. Liepelt, L. Reimer, M. Ritter, A. Schwöppe, J. Scherer,
F. Spiering, R. Thormann, V. Togiti, D. Vollmer, J.-H.Wendisch, “DLR project digital-X: towards virtual aircraft
design and flight testing based on high-fidelity methods”. CEAS Aeronaut. J. 7(1), 3–27 (2016).
239
resulting in total number of 80 FFD nodes. The objective here is to minimize the drag, subject to lift
and sectional thickness constraint of wing. The optimization mathematical model is of the form:
Figure 7.30 Wing-Body Transport Aircraft Configuration and FFD box (8 control sections with 10 FFD
Nodes for each Section, Resulting in 40 nodes on upper and lower wing surfaces, respectively)
452 T. Gerhold, V. Hannemann, D. Schwamborn, “On the Validation of the DLR-TAU Code, in New Results in
Numerical and Experimental Fluid Mechanics”, vol. 72, Notes on Numerical Fluid Mechanics, ed. by W. Nitsche,
H.-J. Heinemann, R. Hilbig (Springer Vieweg, Berlin, 1999), pp. 426–433. ISBN 3-528-03122-0
453 N. Kroll, J.K. Fassbender (eds.), MEGAFLOW – “Numerical Flow Simulation for Aircraft Design, vol. 89, Notes
invited lecture, in Proceedings of the European Conference on Computational Fluid Dynamics” (ECCOMAS CFD
2006), eds by P.Wesseling, E. Oñate, J. Périaux, (TU Delft, The Netherlands, 2006)
240
problems. For example, in455 the first author of the current study conducted a thorough grid
convergence study for the baseline RAE 2822 airfoil and for the optimized airfoil to ensure that the
variation of the drag coefficient is less than 1 count. While a full-fledged grid convergence study is
beyond the scope of the present study, a set of grids of different size is generated to study the
characteristics of the baseline configuration. The grids are block-structured and are generated using
a DLR in-house tool, which automatically changes the CAD model of the aircraft and the associated
block topology, and applies Pointwise to create the surface and volume grid. Here we study the
aerodynamic performance of two representative grids: a grid with 0.78 M nodes and a finer grid with
3.8 M nodes. The comparison of pressure coefficient contour is shown in Figure 7.31. From the
pressure distribution, one can clearly see that there is a shock wave at around 65% of the chord-
wise location, and there is only very small difference between the pressure distributions on the fine
Figure 7.31 Comparison of Surface CP on Fine and Coarse Grids (M = 0.83, Re = 4.34E7, CL = 0.5)
455Y. Zhang, Z.-H. Han, L.-X. Shi, W.-P. Song, “Multi-round surrogate-based optimization for benchmark
aerodynamic design problems”, AIAA Paper 2016–1545, 54th AIAA Aerospace Science Meeting, 4–8.
241
and the coarse grids, where the differences can be seen near the shock wave for both grids.
Concerning the drag coefficient, there is only 1 count difference between the grids, which enables us
to use the coarse grid to run the intensive aerodynamic shape optimizations. Note that the reference
point of moment coefficient is at the nose of the fuselage.
➢ SBO can definitely be an alternative to gradient-based method in the case that fast adjoint
gradient is not available, such as the case of laminar wing design considering transition;
➢ Adjoint gradients can be incorporated in SBO to further improve the overall efficiency, which
lends it to the method of gradient-enhanced Kriging (GEK);
➢ It is straight forward to extend the SBO framework to robust design, i.e., to design under
uncertain flow conditions or random shape variations due to manufacturing tolerances.
Figure 7.32 CP of the Baseline and SBO Optimized Configurations (M = 0.83, Re = 4.34E7,CL = 0.5)
242
Ẑ(s0 ) = ∑ λi Z(si )
i=1
Eq. 7.19
where:
Z(si) = the measured value at the ith location
In IDW, the weight, λi, depends solely on the distance to the prediction location. However, with the
kriging method, the weights are based not only on the distance between the measured points and the
prediction location but also on the overall spatial arrangement of the measured points. To use the
spatial arrangement in the weights, the spatial autocorrelation must be quantified. Thus, in ordinary
kriging, the weight, λi,
depends on a fitted
model to the
measured points, the
distance to the
prediction location,
and the spatial
relationships among
the measured values
around the prediction
location. The
following sections
discuss how the
general kriging
formula is used to
create a map of the
prediction surface and
a map of the accuracy
of the predictions.
Regression models Figure 7.34 Example of 1D data Interpolation by Kriging, with Confidence
can be zeroth, first, or Intervals
second order
polynomials which globally approximates design space, while stochastic processes compose local
deviations. Kriging model provides a measure for the predictor error (Kriging variance) to establish
a confidence interval as shown in Figure 7.34 where the squares indicate the location of the data.
The kriging interpolation, shown in red, runs along the means of the normally distributed confidence
intervals shown in gray. The dashed curve shows a spline that is smooth, but departs significantly
from the expected intermediate values given by those means.
The response surface models (RSM) are generally second-order polynomial models that have limited
capability to accurately predict nonlinear results. Higher-order models are more accurate with
nonlinear problems, but may be unstable. Kriging models show greater promise for accurate global
approximations of the design space. They are flexible due to the wide range of correlation functions
which can be used to build the approximation framework. Thus, Kriging models can accurately
predict both linear and nonlinear functions. An advantage of the Kriging model is its ability to reduce
the number of needed parameters if the database is small. Thus, the model can be successful even if
there are less training members than design variables. Kriging models combine a global model with
localized departures as:
y(x) = f(x) + Z(x)
Eq. 7.20
245
where y(x) is the objective function, f(x) is an approximation function which is usually a polynomial
function and Z(x) is a stochastic function with zero variance and nonzero covariance. When f(x)
globally approximates the
design space, Z(x) creates
deviations where the Z(x) is
given by 457.
7.4.2 Gradient-Enhanced
Kriging
It is straightforward that the
computation cost can be
reduced with the gradient
information. There are several
ways of implementing
Gradient-Enhanced Kriging
(GEK). Error! Reference
source not found. shows the
prediction comparison of the
Rosenbrock function based on
Kriging model and GEK model
and it is apparent that GEK
outperforms Kriging for the
same number of samples. The
indirect GEK uses the gradient
information to provide
additional data to the Figure 7.35 Prediction Comparison of the Rosenbrock Function
observation results with Based on Kriging Model and GEK Model
defining a small but finite step
size. However, the direct GEK modifies the prior covariance or changing the prior covariance matrix
with appending the partial derivatives to observation results. The main advantages of direct GEK
over indirect GEK are:
The GEK has also been developed to solve the 2D airfoil drag minimization problems by [Yamazaki
et al.]458.
457 Zhihui Li, Xinqian Zheng, “Review of design optimization methods for turbomachinery aerodynamics”,
Progress in Aerospace Sciences, July 2017.
458 W. Yamazaki, M. Rumpfkeil, D. Mavriplis, “Design Optimization Utilizing Gradient/Hessian Enhanced
In the forward phase, the sums of the input contributions are connected to the nodes in the next layer
by a sigmoidal transfer function. (see Figure 8.3). The standard sigmoid transfer function, fs, is:
1
fS (x) =
1 + e−x
Eq. 8.1
248
Figure 8.2 Artificial Neural Networks (ANN) with 1 & 2 Hidden Layers
8.1.1 Learning
Learning is nothing else than determining network unknown parameters through some algorithm in
order to minimize network’s performance measure, typically a function of the difference between
predicted and target (desired) outputs. When ANN learning is iterative in nature, it consists of three
459 Zhihui Li, Xinqian Zheng, “Review of design optimization methods for turbomachinery aerodynamics”,
Progress in Aerospace Sciences · July 2017.
460 X.D. Wang, C. Hirsch, S. Kang, C. Lacor, “Multi-objective optimization of turbomachinery using improved NSGA-
II and approximation model”, Computer Methods Appl. Mech. Eng. 200 (2011) 883–895.
249
phases:
(i) training, (ii) validation, and (iii) testing. From previous knowledge, examples or data points
are selected to train the network, grouped in the so-called training dataset. During an iterative
learning, while the training dataset is used to tune network unknowns, a process of cross-validation
takes place by using a set of
data completely distinct from
the training counterpart (the
validation dataset), so that the
generalization performance of
the network can be attested.
Once ‘optimum’ network
parameters are determined,
typically associated to a
minimum of the validation
performance curve (called
early stop ; see Figure 8.4),
many authors still perform a
final assessment of model’s Figure 8.4 Assessing ANN’s Generalization Ability via Cross-
accuracy, by presenting to it a Validation
third fully distinct dataset
called ‘testing’. Heuristics suggests that early stopping avoids overfitting, i.e. the loss of ANN’s
generalization ability. [Quesada]461 provides an easy to understand algorithms to train a neural
network in machine learning blog, using optimizer procedure.
file:///C:/Users/Owner/Desktop/5%20algorithms%20to%20train%20a%20neural%20networ
461
k%20_%20Machine%20learning%20blog.html
462file:///C:/Users/Owner/Desktop/First%20neural%20network%20for%20beginners%20explained%20(
with%20code).html
250
The so-called activation function usually serves to turn the total value calculated before to a number
between 0 and 1 (done for example by a sigmoid function shown by Figure 8.3). Other function
exist and may change the limits of our function, but keeps the same aim of limiting the value. That’s
all a neuron does ! Take all values from connected neurons multiplied by their respective weight, add
them, and apply an activation function. Then, the neuron is ready to send its new value to other
neurons. After every neurons of a column did it, the neural network passes to the next column. In the
end, the last values obtained should be one usable to determine the desired output.
Figure 8.6 Network Diagram for a feed-forward NN with three inputs and one output
Figure 8.6 presents a sample ANN where a Network diagram for a feed-forward NN with three
inputs, two hidden layers, and one output. For this sample network, the values of the hidden nodes
z1,1 through z1,H1 would be constructed as
3 1
z1,i = a w i, jηi
1
Eq. 8.2
i =1
where a1 and w1i,j are the activation function and weights associated with the first hidden layer,
respectively. Similarly, the second layer of hidden nodes is constructed a
463 Anand Pratap Singh, Shivaji Medida, Karthik Duraisamy, “Machine Learning-augmented Predictive Modeling
H1 2
z 2, j = a w i, jz1,i
2
Eq. 8.3
i =1
Finally, the output is
H2 3
y f( ) = a w i, j z 2,i
3
Eq. 8.4
i=1
Given training data, error back-propagation algorithms464 are used to find wnij . Once the weights are
found, computing the output depends only on the number of hidden nodes, and not on the volume of
the training data. Hyper-parameters of the NN method include the number of hidden layers, the
number of nodes in each hidden layer, and the forms of the activation functions. Typically, 3 layers
and about 100 nodes were employed with a sigmoid activation function.
And the list goes on. These questions remain central to addressing the long-term viability of ANNs.
The good news is that such topics are currently being intensely investigated by academic researchers
and industry (Google, Facebook, etc.) alike. Undoubtedly, the next decade will witness significant
progress in addressing these issues. From a practical standpoint, the work of determine the number
of layers and nodes based upon prediction success, i.e. more layers and more nodes do not improve
performance. Additionally, cross-validation is imperative to suppress overfitting. As a general rule,
one should never trust results of a ANN unless rigorous cross-validation has been performed. Cross-
validation plays the same critical role as a convergence study of a numerical scheme. Given the
464 Zhang, Z. J. and Duraisamy, K., “Machine Learning Methods for Data-Driven Turbulence Modeling,” 22nd AIAA
Computational Fluid Dynamics Conference, AIAA Aviation, (AIAA 2015-2460), Dallas, TX, Jun 2015.
252
computational maturity of ANNs and how readily available they are (see Google’s open source
software called Tensor Flow), it is perhaps time for part of the turbulence modelling community to
adopt what has become an important and highly successful part of the machine learning culture:
challenge data sets.
Roxana M. Greenman, “Two-Dimensional High-Lift Aerodynamic Optimization Using Neural Networks”, Ames
465
with a carpet map. The neural network will be able to capture the design space with the small amount
of data that is generated. Next, the optimizer will be able to locate extreme as in the design space by
using the captured design space to calculate the path that must be followed to reach a maxima. The
agile artificial intelligence (AI) design space capture and surfing process is shown in Figure 8.7.
Figure 8.7 Agile AI-Enhanced Design Space Capture and Smart Surfing
Recently, neural networks have been applied to a wide range of problems in the aerospace industry.
For example, neural networks have been used in aerodynamic performance optimization of rotor
blade design466. The study demonstrated that for several rotor blade designs, neural networks were
advantageous in reducing the time required for the optimization. [Failer and Schreck]467 successfully
used neural networks to predict real-time three-dimensional unsteady separated flow fields and
aerodynamic coefficients of a pitching wing. It has also been demonstrated that neural networks are
capable of predicting measured data with sufficient accuracy to enable identification of
instrumentation system degradation. [Steck and Rokhsaz]468 demonstrated that neural networks can
be successfully trained to predict aerodynamic forces with sufficient accuracy or design and
modeling. [Rai and Madavan]469 demonstrated the feasibility of applying neural networks to
aerodynamic design of turbomachinery airfoils.
466 LaMarsh, W. J.; Walsh, J. L.; and Rogers, J. L.:, “Aerodynamic Performance Optimization of a Rotor Blade Using
a Neural Network as the Analysis”. AIAA Paper 92-4837, Sept. 1992.
467 Failer, W. E.; and Schreck, S. J., “Real-Time Prediction of Unsteady Aerodynamics: Application for Aircraft
Control and Maneuverability Enhancement”. IEEE Transactions on Neural Networks, vol. 6, no. 6, Nov. 1995.
468 Steck, J. E.; and Rokhsaz, K.: “Some Applications of Artificial Neural Networks in Modeling of Nonlinear
In this reading, an incompressible 2D Navier-Stokes solver is used to compute the flow field about
the three-element airfoil shown in Figure 8.9. The selected airfoil is a cross-section of the Flap-Edge
model that was tested in
the 7- by 10-Foot Wind
Tunnel No. 1 at the NASA
Ames Research Center.
Extensive wind-tunnel
investigations have been
carried out for the Flap-
Edge geometry shown in
Figure 8.9. The model is
a three-element un-swept
wing consisting of a 12%c
LB-546 slat, NACA 632-
215 Mod B main element
and a 30%c Fowler flap Figure 8.8 Illustration of AI-Enhanced Design Process
where c is chord and is
equal to c = 30.0 inches for the un-
deflected (clean, all high-lift
components stowed) airfoil section.
the CFD database for this flap
optimization problem, there are two
different slat deflection settings, six
and twenty-six degrees, and for each,
27 different flap riggings (refer to
Figure 8.9-b) are computed for ten
different angles of attack. [for details
see Greenman]470.
The neural networks are trained by
using the flap riggings and angles of
attack as the inputs and the
aerodynamic forces as the outputs.
The neural networks are defined to be
successfully trained to predict the
aerodynamic coefficients when given
a set of inputs that are not in the
training set, the outputs are predicted
within the experimental error.
Finally, the trained neural networks
are integrated with the optimizer to
allow the design space to be easily
searched for points of interest. It will Figure 8.9 Edge Geometry and definition of flap and slat
be shown that this agile, artificial high-lift rigging
Roxana M. Greenman, “Two-Dimensional High-Lift Aerodynamic Optimization Using Neural Networks”, Ames
470
intelligence enhanced design process minimizes the cost and time required to accurately optimize
the high lift flap rigging.
8.1.2.1.3 Summary
Multiple input, single output networks were trained using the NASA Ames variation of the Levenberg-
Marquardt algorithm. The neural networks were first trained with wind tunnel experimental data of
the three-element airfoil to test the validity of the neural networks. The networks did accurately
predict the lift coefficients of the individual main and flap elements. However, there was noticeable
error in predicting the slat lift coefficient. The prediction error is most likely caused by the sparse
training data since there were only five different used to train the neural networks. This results from
the fact that the errors are also summed and amplify in the prediction of the total lift coefficient.
Computational data is next used to train the neural networks to test if computational data can be
used to train the neural networks. The neural networks were used to create a computational data
base which may be used to impact design. Solutions were obtained by solving the two-dimensional
Reynolds-averaged incompressible Navier-Stokes equations. The flow field was assumed to be fully
turbulent and the Spalart-Allmaras turbulence model been used. The computational data set had to
be pre-processed to reduce the prediction error at or beyond maximum lift. In high-lift aerodynamics,
both experimentally and computationally, it is difficult to predict the maximum lift, and at which
angle of attack it occurs. In order to predict maximum lift and the angle of attack where it occurs, a
maximum lift criteria was needed. The pressure difference rule, which states that there exists a
certain pressure difference between the peak suction pressure and the pressure at the trailing edge
of the element at the maximum lift condition, was applied to all three elements. For this configuration,
it was found that only the pressure difference on the slat element was needed to predict maximum
lift. By applying the pressure difference rule, the prediction errors of the neural networks were
reduced.
The amount of data that is required to train the neural networks was reduced to allow computational
fluid dynamics to impact the design phase. Different subsets of the training methods were created by
removing entire configurations from the six-degree-deflected slat training set. The mean and
standard deviations of the root-mean-square prediction errors were calculated to compare the
different methods of training. Even though the entire computational data set was sparse, it was
reduced to only 70% of the entire data. It was found that the trained neural networks predicted the
aerodynamic coefficients within an acceptable accuracy defined to be the experimental error. The
aerodynamic data had to be represented in a nonlinear fashion so that the neural networks could
learn and predict accurately. By carefully choosing the training subset, the computational data set
was even further reduced to contain only 52% of the configurations.
These trained neural networks also predicted the aerodynamic coefficients within the acceptable
error. Thus, the computational data required to accurately represent the flow field of a multi-element
airfoil was reduced to allow computational fluid dynamics to be a usable tool for design. This same
procedure was followed in the twenty-six-degree-deflected slat computational data. This data set had
higher deflected flaps which were actually out of the normal flight envelope. The same trends were
found except that the prediction error was much higher in this training set than the previous one.
This was caused by the fact that the flow field was severely separated with the higher deflected flaps.
Thus, the training data representing the flow field was noisy which leads to prediction errors. The
computational design space needs to be easily searched for areas of interest such as maximums or
optimal points. An optimization study to search the design space was conducted by using neural
networks that were trained with computational data.
Artificial neural networks have been successfully integrated with a gradient based optimizer to
minimize the amount of data required to completely define the design space of a three-element
airfoil. The accuracy of the neural networks' prediction was tested for both the initial and modified
configurations by generating the grid and computing the INS2D-UP solution. The high-lift flap
256
aerodynamics were optimized for a three-element airfoil by maximizing the lift coefficient. The
design variables were flap deflection, gap, and overlap.
8.1.2.1.4 Conclusion
Overall, the neural networks were trained successfully to predict the high-lift aerodynamics of a
multi-element airfoil. The neural networks were also able to predict the aerodynamics successfully
when only 52-70% of the entire computational data set was used to train. The neural networks were
integrated with an optimizer thus allowing a quick way to search the design space for points of
interest. Optimization with neural networks reduced the turnaround time, CPU time, and cost of
multiple optimization runs. Therefore, neural networks are an excellent tool to allow computational
fluid dynamics to impact the design space. For a complete study of ANN in design and optimization
environments, reader encourage to consult the [Rai et al.]471 and [Timnak et al.]472 beside current
author.
8.1.2.2 Case Study 2 - Artificial Neural Networks (ANNs) Trained Through Deep Reinforcement
Learning Discover Control Strategies For Active Flow Control 473
We present the first application of an Artificial Neural Network (ANNs) trained through a Deep
Reinforcement Learning agent to perform active flow control. It is shown that, in a 2D simulation
of the Kármán vortex street at moderate Reynolds number (Re = 100), our Artificial Neural Network
is able to learn an active control strategy from experimenting with the mass flow rates of two jets on
the sides of a cylinder. By interacting with the unsteady wake, the Artificial Neural Network
successfully stabilizes the vortex alley and reduces drag by about 8%. This is performed while using
small mass flow rates for the actuation, on the order of 0.5% of the mass flow rate intersecting the
cylinder cross section once a new pseudo-periodic shedding regime is found. This opens the way to
a new class of methods for performing active flow control.
471 Man Mohan Rai, Nateri K. Madavan, and Frank W. Huber, “Improving the Unsteady Aerodynamic Performance
of Transonic Turbines Using Neural Networks”, NASA/TM-1999-208791.
472 N. Timnak, A. Jahangiriana and S.A. Seyyedsalehi, “An optimum neural network for evolutionary aerodynamic
Trained Through Deep Reinforcement Learning Discover Control Strategies For Active Flow Control”, Journal of
Fluid Mechanics, April 2019.
474 Brunton, Steven L & Noack, Bernd R 2015 Closed-loop turbulence control: progress and challenges. Applied
turbulent boundary layer), benefits in terms of reduced drag can also be achieved. Another way to
obtain drag reduction is by applying an active control to the flow. A number of techniques can be
used in active drag control and have been proven effective in several experiments, a typical example
being to use small jets.
Interestingly, it has been shown that effective separation control can be achieved with even quite
weak actuation, as long as it is used in an efficient way [Schoppa & Hussain]476. This underlines the
need to develop techniques that can effectively control a complex actuation input into a flow, in order
to reduce drag. Unfortunately, designing active flow control strategies is a complex endeavor [Duriez
et al]477. Given a set of point measurements of the flow pressure or velocity around an object, there
is no easy way to find a strategy to use this information in order to perform active control and reduce
drag. The high dimensionality and computational cost of the solution domain (set by the complexity
and non-linearity inherent to Fluid Mechanics) mean that analytical solutions, and real-time
predictive simulations (that would decide which control to use by simulating several control
scenarios in real time) seem out of reach. Despite the considerable efforts put into the theory of flow
control, and the use of a variety of analytical and semi-analytical techniques [Barbagallo et al.]478;
[Sipp & Schmid]479, bottom-up approaches based on an analysis of the flow equations face
considerable difficulties when attempting to design flow control techniques.
A consequence of these challenges is the simplicity of the control strategies used in most published
works about active flow control, which traditionally focus on either harmonic or constant control
input [Schoppa & Hussain]480. Therefore, there is a need to develop efficient control methods, that
perform complex active control and take full advantage of actuation possibilities. Indeed, it seems
that, as of today, the actuation possibilities are large, but only simplistic (and probably suboptimal)
control strategies are implemented. To the knowledge of the authors, only few published examples
of successful complex active control strategies are available with respect to the importance and
extent of the field [Pastoor et al.]481; [Gautier et al]482 ; [Li et al.]483 ; [Erdmann et al.]484 ; [Guéniat et
al.]485.
In the present work, we aim at introducing for the first time Deep Neural Networks and
Reinforcement Learning to the field of active flow control. Deep Neural Networks are
476 Schoppa, Wade & Hussain, Fazle 1998 A large-scale control strategy for drag reduction in turbulent boundary
layers. Physics of Fluids 10 (5), 1049–1051.
477 Duriez, Thomas, Brunton, Steven L & Noack, Bernd R 2016 Machine Learning Control-Taming Nonlinear
Closed-loop control of unsteadiness over a rounded backward-facing step. Journal of Fluid Mechanics 703.
479 Sipp, Denis & Schmid, Peter J 2016 Linear closed-loop control of fluid instabilities and noise-induced
perturbations: A review of approaches and tools. Applied Mechanics Reviews 68 (2), 020801.
480 Schoppa, Wade & Hussain, Fazle 1998 A large-scale control strategy for drag reduction in turbulent boundary
control for bluff body drag reduction. Journal of fluid mechanics 608, 161–196.
482 Gautier, Nicolas, Aider, J-L, Duriez, Thomas, Noack, Br, Segond, Marc & Abel, Markus,“ Closed-loop separation
car model by linear genetic programming control. Experiments in Fluids 58 (8), 103.
484 Erdmann, Ralf, Pätzold, Andreas, Engert, Marcus, Peltzer, Inken & Nitsche, Wolfgang 2011 On active control
loop control of fluid flows. Theoretical and Computational Fluid Dynamics 30 (6), 497–510.
258
revolutionizing large fields of research, such as image analysis [Krizhevsky et al.]486, speech
recognition, and optimal control. Those methods have surpassed previous algorithms in all these
examples, including methods such as genetic programming, in terms of complexity of the tasks
learned and learning speed. It has been speculated that Deep Neural Networks will bring advances
also to fluid mechanics [Kutz]487, but until this day those have been limited to a few applications, such
as the definition of reduced order models [Wang et al.]488, the effective control of swimmers, or
performing Particle Image Velocimetry (PIV). As Deep Neural Networks, together with the
Reinforcement Learning framework, have allowed recent breakthroughs in the optimal control of
complex dynamic systems [Lillicrap et al.]489; [Schulman et al.]490, it is natural to attempt to use them
for optimal flow control.
Artificial Neural Networks (ANNs) are the attempt to reproduce in machines some of the features
that are believed to be at the origin of the intelligent thinking of the brain (LeCun et al., 2015). The
key idea consists in performing computations using a network of simple processing units, called
neurons. The output value of each neuron is obtained by applying a transfer function on the weighted
sum of its inputs. When performing supervised learning an algorithm, such as stochastic gradient
descent, is then used for tuning the neurons weights so as to minimize a cost function on a training
set. Given the success of this training algorithm, ANNs can in theory solve any problem since they are
universal approximations: a large enough feed-forward neural network using a nonlinear activation
function can fit arbitrarily well any function, and the Recurrent Neural Network paradigm is even
turning complete. Therefore, virtually any problem or phenomenon that can be represented by a
function could be a field of experimentation with ANNs.
However, the problem of designing the ANNs, and designing the algorithms that train and use them,
is still the object of active research. While the case of supervised learning (i.e., when the solution is
known and the ANN should simply be trained at reproducing it, such as image labeling or PIV) is now
mostly solved owing to the advance of Deep Neural Networks and Deep Convolutional Networks (He
et al., 2016), the case of reinforcement learning (when an agent tries to lean through the feedback of
a reward function) is still the focus of much attention [Mnih et al.]491; [Gu et al.]492; [Schulman et al]493.
In the case of reinforcement learning, an agent (controlled by the ANN) interacts with an
environment through 3 channels of exchange of information in a closed-loop fashion. First, the agent
is given access at each time step to an observation ot of the state st of the environment. The
environment can be any stochastic process, and the observation is only a noisy, partial description of
the environment. Second, the agent performs an action, at, that influences the time evolution of the
environment.
Finally, the agent receives a reward rt depending on the state of the environment following the action.
The reinforcement learning framework consists in finding strategies to learn from experimenting
486 Krizhevsky, Alex, Sutskever, Ilya & Hinton, Geoffrey E 2012 Image net classification with deep convolutional
neural networks. In Advances in neural information processing systems, pp. 1097–1105.
487 Kutz, J. Nathan 2017 Deep learning in fluid dynamics. Journal of Fluid Mechanics 814, 1–4.
488 Wang, Z, Xiao, D, Fang, F, Govindan, R, Pain, Cc & Guo, Y 2018 Model identification of reduced order fluid
dynamics systems using deep learning. International Journal for Numerical Methods in Fluids 86 (4), 255–268.
489 Lillicrap, Timothy P, Hunt, Jonathan J, Pritzel, Alexander, Heess, Nicolas, Erez, Tom, Tassa, Yuval, Silver,
David & Wierstra, Daan 2015 Continuous control with deep reinforcement learning. preprint arXiv:1509.02971
490 Schulman, John, Levine, Sergey, Moritz, Philipp, Jordan, Michael I. & Abbeel, Pieter 2015 Trust region policy
Riedmiller, Martin 2013 Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602 .
492 Gu, Shixiang, Lillicrap, Timothy, Sutskever, Ilya & Levine, Sergey 2016 Continuous deep q-learningwith model-
with the environment, in order to discover control sequences a(t = 1, , , , , T) that maximize the reward.
The environment can be any system that provides the interface (ot, at, rt), either it is an Atari game
emulator, a robot acting in the physical world that should perform a specific task, or a fluid mechanics
system whose drag should be minimized in our case.
In the present work, we apply for the first time the Deep Reinforcement Learning (DRL, i.e.
Reinforcement Learning performed on a deep ANN) paradigm to an active flow control problem.
We use a Proximal Policy Optimization (PPO) method together with a Fully Connected Artificial
Neural Network (FCANN) to control two synthetic jets located on the sides of a cylinder immersed
in a constant flow in a 2D simulation. The geometry is chosen owing to its simplicity, the low
computational cost associated with resolving a 2D unsteady wake at moderate Re number, and the
simultaneous presence of the causes that make active flow control challenging (time dependence,
non-linearity, high dimensionality). The PPO agent manages to control the jets and to interact with
the unsteady wake to reduce the drag. We choose to release all our code as Open Source, to help
trigger interest in those methods and facilitate further developments. In the following, we first
present the simulation environment, before giving details about the network and reinforcement
framework, and finally we offer an overview of the results obtained.
8.1.2.2.2 Simulation Environment
The PPO agent performs active flow control in a 2D simulation environment. In the following, all
quantities are considered non-dimensionalized. The geometry of the simulation, adapted from the
2D test case of well-known benchmarks [Schäfer et al]494, consists of a cylinder of non-dimensional
diameter D = 1 immersed in a box of total non-dimensional length L = 22 (along the X-axis) and height
H = 4.1 (along Y-axis). The origin of the coordinate system is in the center of the cylinder. Similarly
to the benchmark of the cylinder is slightly off the centerline of the domain (a shift of 0.05 in the Y
direction is used), in order to help trigger the vortex shedding. The inflow profile (on the left wall of
the domain) is parabolic, following the formula (cf. 2D-2 test case in [Schäfer et al.]495):
H 2
6 [( ) − y 2 ]
2
U(y) =
H2
Eq. 8.5
Figure 8.10 Unsteady Non-Dimensional Pressure Wake Behind the Cylinder after Flow Initialization
without Active Control. The Location of the Velocity Probes is Indicated by the Black Dots While The
Location of the Control Jets is Indicated by the Red Dot
494 Schäfer, M., Turek, S., Durst, F., Krause, E. & Rannacher, R. 1996 Benchmark Computations of Laminar Flow
Around a Cylinder, pp. 547–566. Wiesbaden: Vieweg+Teubner Verlag.
495 See Previous.
260
where (U(y), V(y) = 0) is the non-dimensionalized velocity vector. Using this velocity profile, the
mean velocity magnitude is Ū = 2U(0)/3 = 1. A no-slip boundary condition is imposed on the top and
bottom walls and on the solid walls of the cylinder. An outflow boundary condition is imposed on the
right wall of the domain. The configuration of the simulation is visible in Figure 8.10.
The Reynolds number based on the mean velocity magnitude and cylinder diameter (Re = UD/ν, with
ν the kinematic viscosity) is set to Re = 100. Computations are performed on an unstructured mesh
generated with Gmsh. The mesh is refined around the cylinder and is composed of 9262 triangular
elements. A non-dimensional, constant numerical time step dt = 5.10-3 is used. The total
instantaneous drag on the cylinder C is computed following:
FD = ∫(σ. n) ex dS
C
Eq. 8.6
where _ is the Cauchy stress tensor, n is the unit vector normal to the outer cylinder surface, and ex =
(1. 0). In the following, the drag is normalized into the drag coefficient:
FD
CD =
(1/2)ρU̅ 2D
Eq. 8.7
where ρ = 1 is the non-dimensional volumetric mass density of the fluid. Similarly, the lift force FL
and
FL = ∫(σ. n) ey dS
C
Eq. 8.8
lift coefficient CL are defined as
FL
CL =
(1/2)ρU̅ 2D
Eq. 8.9
where ey = (0, 1). In the interest of short solution time, the governing Navier-Stokes equations are
solved in a segregated manner. More precisely, the Incremental Pressure Correction Scheme (IPCS
method with an explicit treatment of the nonlinear term is used. More details are available in
([Rabault]496, Appendix B). Spatial discretization then relies on the finite element method
implemented within the FEniCS framework. We remark that both the mesh density and the Reynolds
number could easily be increased in a later study, but are kept low here as it allows for fast training
on a laptop which is the primary aim of our proof-of-concept demonstration.
In addition, two jets (1 and 2) normal to the cylinder wall are implemented on the sides of the
cylinder, at angles θ1 = 90 and θ2 = 270 degrees relatively to the flow direction. The jets are controlled
through their non-dimensional mass flow rates, Qi, i = 1; 2, and are set through a parabolic-like
velocity profile going to zero at the edges of the jet, (see [Rabault]497, Appendix B) for the details. The
jet widths are set to 10 degree. Choosing jets normal to the cylinder wall, located at the top and
bottom extremities of the cylinder, means that all drag reduction observed will be the result of
496 Jean Rabault, Miroslav Kuchta, Atle Jensen, Ulysse Reglade and Nicolas Cerardi, “Artificial Neural Networks
Trained Through Deep Reinforcement Learning Discover Control Strategies For Active Flow Control”, Journal of
Fluid Mechanics, April 2019.
497 See Previous.
261
indirect flow control, rather than direct injection of momentum. In addition, the control is set up in
such a way that the total mass flow rate injected by the jets is zero, i.e. Q1 + Q2 = 0. This synthetic jets
condition is chosen as it is more realistic than a case when mass is added or subtracted from the flow,
and makes the numerical scheme more stable specially with respects to the boundary conditions of
the problem. In addition, it makes sure that the drag reduction observed is the result of actual flow
control, rather than some sort of propulsion phenomenon. In the following, the injected mass flow
rates are normalized following:
D2
Qi
Q∗i = ; Q ref = ∫ ρU(y)dy
Q ref
−D2
Eq. 8.10
where Qref is the reference mass flow rate intercepting the cylinder. During learning, we impose that
∣Q∣< 0.06. This helps in the learning process by preventing non-physically large actuation, and
prevents problems in the numeric of the simulation by enforcing the CFL condition close to the
actuation jets.
Finally, information is extracted from the simulation and provided to the PPO agent. A total of 151
velocity probes, which report the local value of the horizontal and vertical components of the velocity
field, are located in several locations in the neighborhood of the cylinder and in its wake (see Figure
8.10). This means that the network gets detailed information about the flow configuration, which is
our objective as this article focuses on finding the best possible control strategy of the vortex
shedding pattern. A different question would be to assess the ability of the network to perform
control with a partial observation of the system. To illustrate that this is possible with adequate
training, we provide some results with an input layer reduced to 11 and 5 probes in (see [Rabault]498,
Appendix E) but further parameter space study and sensitivity analysis is out of the scope of the
present paper and is let to future work.
An unsteady wake develops behind the cylinder, which is in good agreement with what is expected
at this Reynolds number. A simple benchmark of the simulation was performed by observing the
pressure fluctuations, drag coefficient and Strouhal number St = f D/Ū , where f is the vortex
shedding frequency. The mean value of CD in the case without actuation (around 3.205) is within 1%
of what is reported in the benchmark of [Schäfer et al.]499, which validates our simulations, and
similar agreement is found for St (typical value of around 0.30) . In addition, we also performed tests
on refined meshes, going up to around 30000 triangular elements, and found that the mean drag
varied by less than 1 % following mesh refinement. A pressure field snapshot of the fully developed
unsteady wake is presented in Figure 8.10.
498See Previous.
499Schäfer, M., Turek, S., Durst, F., Krause, E. & Rannacher, R. 1996 Benchmark Computations of Laminar Flow
Around a Cylinder, pp. 547–566. Wiesbaden: Vieweg+Teubner Verlag.
262
agent uses the Proximal Policy Optimization (PPO, Schulman et al.]500) method for performing
learning.
PPO is a reinforcement learning algorithm that belongs to the family of policy gradient methods. This
method was chosen for several reasons. In particular, it is less complex mathematically and faster
than concurring Trust Region Policy Optimization methods (TRPO), and requires little to no meta
parameter tuning. It is also better adapted to continuous control problems than Deep Q Learning
(DQN) and its variations. From the point of view of the fluid mechanist, the PPO agent acts as a black
box (though details about its internals are available in and the referred literature). A brief
introduction to the PPO method is provided in ([Rabault]501, Appendix C).
The PPO method is episode-based, which means that it learns from performing active control for a
limited amount of time before analyzing the results obtained and resuming learning with a new
episode. In our case, the simulation is first performed with no active control until a well-developed
unsteady wake is obtained, and the corresponding state is saved and used as a start for each
subsequent learning episode. The instantaneous reward function, rt, is computed following:
rt = −〈CD 〉T − 0.2|〈CL 〉T |
Eq. 8.11
where < . >T indicates the sliding average back in time over a duration corresponding to one vortex
shedding cycle. The ANN tries to maximize this function rt, i.e. to make it as little negative as possible
therefore minimizing drag and mean lift (to take into account long-term dynamics, an actualized
reward is actually used during gradient descent; (see the Appendix C in [Rabault]502, for more
details). This specific reward function has several advantages compared with using the plain
instantaneous drag coefficient. Firstly, using values averaged over one vortex shedding cycle leads to
less variability in the value of the reward function, which was found to improve learning speed and
stability. Secondly, the use of a penalization term based on the lift coefficient is necessary to prevent
the network from ’cheating’. Indeed, in the absence of this penalization, the ANN manages to find a
way to modify the configuration of the flow in such a way that a larger drag reduction is obtained (up
to around 18 % drag reduction, depending on the simulation configuration used), but at the cost of a
large induced lift which is damageable in most practical applications.
The ANN used is relatively simple, being composed of two dense layers of 512 fully connected
neurons, plus the layers required to acquire data from the probes, and generate data for the 2 jets.
This network configuration was found empirically through trial and error, as is usually done with
ANNs. Results obtained with smaller networks are less good, as their modeling ability is not sufficient
in regards to the complexity of the flow configuration obtained. Larger networks are also less
successful, as they are harder to train. In total, our network has slightly over 300000 weights. For
more details, readers are referred to the code implementation (see the [Rabault]503 Appendix A).
At first, no learning could be obtained from the PPO agent interacting with the simulation
environment. The reason for this was the difficulty for the PPO agent to learn the necessity to set
time-correlated, continuous control signals, as the PPO first tries purely random control and must
500 Schulman, John, Wolski, Filip, Dhariwal, Prafulla, Radford, Alec & Klimov, Oleg 2017 Proximal policy
optimization algorithms. arXiv preprint arXiv:1707.06347 .
501 Jean Rabault, Miroslav Kuchta, Atle Jensen, Ulysse Reglade and Nicolas Cerardi, “Artificial Neural Networks
Trained Through Deep Reinforcement Learning Discover Control Strategies For Active Flow Control”, Journal of
Fluid Mechanics, April 2019.
502 See Previous.
503 Jean Rabault, Miroslav Kuchta, Atle Jensen, Ulysse Reglade and Nicolas Cerardi, “Artificial Neural Networks
Trained Through Deep Reinforcement Learning Discover Control Strategies For Active Flow Control”, Journal of
Fluid Mechanics, April 2019.
263
observe some improvement on the reward function for performing learning. Therefore, we
implemented two tricks to help the PPO agent learn control strategies:
• The control value provided by the network is kept constant for a duration of 50 numerical
time steps, corresponding to around 7.5 % of the vortex shedding period. This means, in
practice, that the PPO agent is allowed to interact with the simulation and update its control
only each 50 time steps.
• The control is made continuous in time to avoid jumps in the pressure and velocity due to the
use of an incompressible solver. For this, the control at each time step in the simulation is
obtained for each jet as cs+1 = cs + α(a - cs), where cs is the control of the jet considered at the
previous numerical time step, cs+1 is the new control, a is the action set by the PPO agent for
the current 50 time steps, and α = 0.1 is a numerical parameter.
Using those technical tricks, and choosing an episode duration Tmax = 20.0 (which spans around 6.5
vortex shedding periods, and corresponds to 4000 numerical time steps, i.e. 80 actions by the
network), the PPO agent is able to learn a control strategy after typically about 200 epochs
corresponding to 1300 vortex
shedding periods or 16000
sampled actions, which
requires around 24 hours of
training on a modern desktop
using one single core. This
training time could be
reduced easily by at least a
factor of 10, using more cores
to parallelize the data
sampling from the epochs
which is a fully parallel
process. Fine tuning the policy
can take a bit longer time, and
up to around 350 epochs can
be necessary to obtain a fully
stabilized control strategy. A
training has also been
performed going up to over Figure 8.11 Illustration of the Robustness of the Learning Process
1000 episodes to confirm that
no more changes were
obtained if the network is let to train for a significantly longer time. Most of the computation time is
spent in the flow simulation. This setup with simple, quick simulations makes experimentation and
reproduction of our results easy, while being enough for a proof-of-concept in the context of a first
application of Reinforcement Learning to active flow control and providing an interesting control
strategy for further analysis.
The drag values reported are obtained at each training epoch (including exploration noise), for 10
different trainings using the same meta parameters, but different values of the random seed. Robust
learning takes place within 200 epochs, with fine converged strategy requiring a few more epochs to
stabilize. The drag reduction is slightly less than what is reported in the rest of the text, as these
results include the random exploration noise and are computed over the second half of the training
epochs, where some of the transient in the drag value is still present during training.
264
504 Illustration of the robustnessof the learning process. The drag values reported are obtained at each training
epoch (including exploration noise), for 10 different trainings using the same meta parameters, but different
values of the random seed. Robust learning takes place within 200 epochs, with fine converged strategy
requiring a few more epochs to stabilize. The drag reduction is slightly less than what is reported in the rest of
the text, as these results include the random exploration noise and are computed over the second half of the
training epochs, where some of the transient in the drag value is still present during training.
265
deterministic prediction, and therefore exploration noise is not present in the following figures and
results. The time series for the drag coefficient obtained using the active flow control strategy
discovered through training in the first run, compared with the baseline simulation (no active
control, i.e. Q1 = Q2 = 0), is presented in Figure 8.12 together with the corresponding control signal
(inset). Similar results and control Laws are obtained for all training runs, and the results presented
in Figure 8.12 are therefore representative of the learning obtained with all 10 realizations. In the
case without actuation (baseline), the drag coefficient CD varies periodically at twice the vortex
shedding frequency, as should be expected. The mean value for the drag coefficient is < CD >≈ 3.205,
and the amplitude of the fluctuations of the drag coefficient is around 0.034. By contrast, the mean
value for the drag coefficient in the case with active flow control is around < C’D > ≈2.95, which
represents a drag reduction of around 8%.
To put this drag reduction into perspective, we estimate the drag obtained in the hypothetical case
where no vortex shedding is present. For this, we perform a simulation with the upper half domain
and a symmetric boundary condition on the lower boundary (which cuts the cylinder through its
equator). More details about this simulation are presented in Appendix D of [Rabault]505. The steady-
state drag obtained on a full cylinder in the case without vortex shedding is then CDs = 2.93 (see
Appendix D) of [Rabault]506, which means that the active control is able to suppress around 93% of
the drag increase observed in the baseline without control compared with the hypothetical reference
case where the flow would be kept completely stable.
In addition to this reduction in drag, the fluctuations of the drag coefficient are reduced to around
0:0016 by the active control, i.e. a factor of around 20 compared with the baseline. Similarly,
fluctuations in lift are reduced, though by a more modest factor of around 5.7. Finally, a Fourier
analysis of the drag coefficients obtained shows that the actuation slightly modifies the characteristic
frequency of the system. The actively controlled system has a shedding frequency around 3.5% lower
than the baseline. Several interesting points are visible from the active control signal imposed by the
ANN presented in Figure 8.12. Firstly, the active flow control is composed of two phases. In the first
one, the ANN changes the configuration of the flow by performing a relatively large transient
actuation (non-dimensional time ranging from 0 to around 11). This changes the flow configuration,
and sets the system in a state in which less drag is generated. Following this transient actuation, a
second regime is reached in which a smaller actuation amplitude is used. The actuation in this new
regime is pseudo-periodic. Therefore, it appears that the ANN has found a way to both set the flow in
a modified configuration in which less drag is present, and keep it in this modified configuration at a
relatively small cost. In a separate simulation, the small actuation present in the pseudo-periodic
regime once the initial actuation has taken place was suppressed. This led to a rapid collapse of the
modified flow regime, and the original base flow configuration was recovered. As a consequence, it
appears that the modified flow configuration is unstable, though only small corrections are needed
to keep the system in its neighborhood.
Secondly, it is striking to observe that the ANN resorts to quite small actuations. The peak value for
the norm of the non-dimensional control mass flow rate Q*1 , which is reached during the transient
active control regime, is only around 0.02, i.e. a factor 3 smaller than the maximum value allowed
during training. Once the pseudo-periodic regime is established, the peak value of the actuation is
reduced to around 0.006. This is an illustration of the sensitivity of the Navier-Stokes
equations to small perturbations, and a proof that this property of the equations can be exploited to
actively control the flow configuration, if forcing is applied in an appropriate manner.
Understanding which strategy an ANN decides to use from the analysis of its weights is known to be
challenging, even on simple image analysis tasks. Indeed, the strategy of the network is encoded in
the complex combination of the weights of all its neurons. A number of properties of each individual
network, such as the variations in architecture, make systematic analysis challenging [Rauber et
al.]507. Through the combination of the neuron weights, the network builds its own internal
representation of how the flow in a given state will be affected by actuation, and how this will affect
the reward value. This is a sort of private, ’encrypted’ model obtained through experience and
interaction with the flow.
Therefore, it appears challenging to directly analyze the control strategy from the trained network,
which should be considered rather as a black box in this regard. Instead, we can look at macroscopic
flow features and how the active control modifies them. This pinpoints the effect of the actuation on
the flow and separation happening in the wake. Representative snapshots of the flow configuration
in the baseline case (no actuation), and in the controlled case when the pseudo-periodic regime is
reached (i.e., after the initial large transient actuation), are presented in Figure 8.13. As visible in
Figure 8.13, the active control leads to a modification of the 2D flow configuration. In particular, the
Kármán alley is altered in the case with active control and the velocity fluctuations induced by the
vortexes are globally less strong, and less active close to the upper and lower walls. More strikingly,
the extent of the recirculation area is dramatically increased. Defining the recirculation area as the
region in the downstream neighborhood of the cylinder where the horizontal component of the
velocity is negative, we observe a 130% increase in the recirculation area, averaged over the pseudo-
period. The recirculation area in the active control case represents 103% of what is obtained in the
hypothetical stable configuration of Appendix D of ([Rabault]508, (so, the recirculation area is slightly
larger in the controlled case than in the hypothetical stable case, though the difference is so small
that it may be due to a side effect such as slightly larger separation close to the jets, rather than a true
change in the extent of the developed wake), while the recirculation area in the baseline
Figure 8.13 Comparison of representative snapshots of the velocity magnitude in the case without
actuation (top), and with active flow control (bottom). The bottom figure corresponds to the established
pseudo-periodic modified regime, which is attained after the initial transient control.
507 Rauber, Paulo E, Fadel, Samuel G, Falcao, Alexandre X & Telea, Alexandru C 2017 Visualizing the hidden
activity of artificial neural networks. IEEE transactions on visualization and computer graphics 23 (1), 101–110.
508 Jean Rabault, Miroslav Kuchta, Atle Jensen, Ulysse Reglade and Nicolas Cerardi, “Artificial Neural Networks
Trained Through Deep Reinforcement Learning Discover Control Strategies For Active Flow Control”, Journal of
Fluid Mechanics, April 2019.
267
configuration with vortex shedding is only 44% of this same stable configuration value. This is,
similarly to what was observed for CD, an illustration of the efficiency of the control strategy at
reducing the effect of vortex shedding. To go into more details, we look at the mean and the Standard
Deviation (STD) of the flow velocity magnitude and pressure, averaged over a large number of vortex
shedding periods (in the case with active flow control, we consider the pseudo-periodic regime).
8.1.2.2.6 Conclusion
We show for the first time that the Deep Reinforcement Learning paradigm (DRL, and more
specifically the Proximal Policy Optimization algorithm, PPO) can discover an active flow control
strategy for synthetic jets on a cylinder, and control the configuration of the 2D Kármán vortex street.
From the point of view of the ANN and DRL, this is just yet-another-environment to interact with. The
discovery of the control strategy takes place through the optimization of a reward function, here
defined from the fluctuations of the drag and lift components experienced by the cylinder. A drag
reduction of up to around 8 % is observed. In order to reduce drag, the ANN decides to increase the
area of the separated region, which in turn induces a lower pressure drop behind the cylinder, and
therefore lower drag. This brings the flow into a configuration that presents some similarities with
what would be obtained from boat-tailing.
The value of the drag coefficient and extent of the recirculation bubble when control is turned on are
very close to what is obtained by simulating the flow around a half cylinder using a symmetric
boundary condition at the lower wall, which allows to estimate the drag expected around a cylinder
at comparable Reynolds number if no vortex shedding was present. This implies that the active
control is able to effectively cancel the detrimental effect of vortex shedding on drag. The learning
obtained is remarkable, as little meta parameter tuning was necessary, and training takes place in
about one day on a laptop. In addition, we have resorted to strong regularization of the output of the
DRL agent through under sampling of the simulation and imposing a continuous control for helping
the learning process. It could be expected that relaxing those constraints, i.e. giving more freedom to
the network, could lead to even more efficient strategies.
These results are potentially of considerable importance for Fluid Mechanics, as they provide a proof
that DRL can be used to solve the high dimensionality, analytically intractable problem of active flow
control. The ANN and DRL approach has a number of strengths which make it an appealing
methodology. In particular, ANNs allow for an efficient global approximation of strongly nonlinear
functions, and they can be trained through direct experimentation of the DRL agent with the flow
which makes it in theory easily applicable to both simulations and experiments without changes in
the DRL methodology. In addition, once trained, the ANN requires only few calculations to compute
the control at each time step. In the present case when two hidden layers of width 512 are used, most
of the computational cost comes from a matrix multiplication, where the size of the matrices to
multiply is [512, 512]. This is much less computationally expensive than the underlying problem.
Finally, we are able to show that learning takes place in a timely manner, requiring a reasonable
number of vortex shedding periods to obtain a converged strategy.
This work opens a number of research directions, including applying the DRL methodology to more
complex simulations, for example more realistic 3D LES/DNS on large computer clusters, or even
applying such an approach directly to a real world experiment. In addition, a number of interesting
questions arise from the use of ANNs and DRL. For example, can some form of transfer learning be
used between simulations and the real world if the simulations are realistic enough (i.e., can one train
an ANN in a simulation, and then use it in the real world)? The use of DRL for active flow control may
provide a technique to finally take advantage of advanced, complex flow actuation possibilities, such
as those allowed by complex jet actuator arrays.
268
269
Figure 9.1 Schematic of a Wind Turbine Generation System- (Courtesy of Gaurav Kapoor)
509 Gaurav Kapoor, “Exploration of a Computational Fluid Dynamics Integrated Design Methodology for Potential
Application to a Wind Turbine Blade”, Thesis Submitted to the Department of Aerospace Engineering, College
of Engineering, Embry-Riddle Aeronautical University, Daytona Beach, December 2014.
270
a design configuration that resulted in an increased blade length and surface area, thus leading to an
overall lift force augmentation producing a 25.26% increase in the power output. Both the optimized
candidates obtained were validated using the flow solver to verify the optimized design for
maximized power output. The coefficient of pressure plots at various span locations of the blade
bolster the claim that most of the mechanical power is produced in the outer 30-40% of the blade.
Figure 9.2 Schematic of Internal Components of a Modern HAWT- (Courtesy of Gaurav Kapoor)
the relative wind, the purpose of airfoils at the root of blade is mainly structural, having a minimal
contribution to the aerodynamic performance of the blade. Thus the root section of the wind turbine
blade is thicker and
stronger than its tip
section (Figure 9.3).
Wind turbine blades are
shaped to extract
maximum power from the
wind at the minimum cost
involved. Primarily the
blade design is driven by Figure 9.3 Profiles of Flat-back and Sharp Trailing Edge Airfoils
the aerodynamic and
performance requirements. But in true sense, the economics mean that the blade shape is a
compromise to keep the cost of construction, operation and maintenance to a minimum. The blade
design procedure starts with obtaining a solution set for both aerodynamic and structural efficiency.
The best blade design is a tradeoff between both aerodynamic performance and structural stiffness.
ωR
λ=
v0
Eq. 9.1
Where, 𝜔 is the angular velocity of the wind turbine rotor, 𝑅 is radius of the rotor and 𝑣0 is the free
wind speed. A higher tip speed ratio (TSR) induces the net aerodynamic force on the blade
(component of lift and
drag) to be
approximately parallel
to the rotor axis . The
lift to drag ratio can be
affected severely by
presence of dirt or
roughness on the blade
surfaces510. Low tip
speed ratio
unfortunately results
is lower aerodynamic
efficiency due to two
effects. Since the lift
force on the blade
generates torque,
according to the laws
of motion, it has an
equal but opposite
effect on the incident Figure 9.6 Swirling Flow in the Wind Turbine Wake
wind, tending to push
it around tangentially in the other direction. As a result, the air downwind of the turbine has a swirl,
i.e. it spins in the opposite direction to the blade rotation, as depicted in Figure 9.6. This swirl
represents lost power which reduces the available power that can be extracted from the incident
wind. Lower rotational speed requires higher torque to maintain the same power output, so lower
tip speed ratio results in
greater wake swirl
losses.
The other reason for the
reduction in
aerodynamic efficiency
at low tip speed ratio is
due to the tip losses,
where high-pressure
air from the upwind
side of the blade
escapes around the Figure 9.5 Typical Wind Turbine Blade Planform View
blade tip to the low-
pressure side, thereby wasting energy. Since power is a product of blade torque and rotational speed,
at slower rotational speed the blades need to generate more lift force to maintain the same power
output. In order to generate greater lift for a given length, the blade has to be wider, geometrically
510Nianxin Ren , Jinping Ou, “Dust Effect on the Performance of Wind Turbine Airfoils”, J.Electromagnetic
Analysis & Applications, 2009, 1: 102-107, doi:10.4236/jemaa.2009.12016, Published Online June 2009.
273
speaking, a greater proportion of the blade’s width is designed to be close to the tip (Figure 9.5).
The higher lift force on a wider blade translates to greater structural loads on the outer components
such as the hub and bearings. There are practical limits on the absolute tip speed ratio as well. At
these speeds, bird impacts and rain erosion starts to decrease the longevity of the blades and noise
increases dramatically with the tip speed.
• Cut-in speed - the minimum wind speed at which the turbine blades overcome frictional force
and begin to rotate.
• Operation mode - the range of wind speeds within which the wind turbine actively
generates power.
• Cut-out mode - the speed at which the turbine is brought to rest to avoid structural
damage due to high wind speeds.
For AOC 15/50 HAWT, if free stream wind speed is less than the cut in speed (4.9 m/s), the wind
turbine rotor will not rotate due to less available wind energy and insufficient torque produced to
overcome the inertia of the blade. The rotor begins to rotate at speed of 4.9 m/s and begins to
generate power. This region of the blade operation from the cut-in wind speed of 4.9 m/s to the cut-
out wind speed of 22.4 m/s is referred to as the operation mode or active mode of the wind turbine.
Ideal or rated wind speed is 12 m/s for the AOC 15/50 wind turbine. And if free stream wind speed
is above 22.4 m/s which is cut-out speed for the AOC 15/50 HAWT, rotor stops rotating to prevent
any damage or failure to wind turbine blade and other gear/bearing systems embedded in the
nacelle.
V0 − U
a=
V0
Eq. 9.2
CP = 4a (1 − a)2
dCP Figure 9.9 Actuator Disk Concept, Pressure and
= a(1 − a)(1 − 3a) Velocity Profiles
da
Eq. 9.5
The maximum value of 𝐶𝑃 = 16⁄27 =
~ 0.593 is obtained for 𝑎 = 1⁄3. This
theoretical maximum value is
known as the Betz Limit [18] and it
is not possible to design a wind
turbine that goes beyond this
theoretical limit. In other words,
according to the Betz's law, no
turbine can capture more than
16/27 (~ 59.3%) of the kinetic
energy in wind.
shape transition continues span-wise to a shape is based on a SERI 819 airfoil at 75% of the tip radius
and a shape that is based on a SERI 820 airfoil at 95% of the tip radius.
The blade was designed in the ANSYS® DesignModeler by using the curve generation function to
import the three different airfoil profiles (S819, S820 and S821) and then the 3D blade was modeled
by using the skin/loft feature. Since the hub does not hold any importance in this case study, it was
modeled to be a simple circular extrusion to which another two blades were duplicated at 120°
angular symmetry using the pattern feature. The blade root section was twisted towards the feather
at 1.54° and the blade tip was given a feather angle of -1.54°(away from the feather) to represent the
same blade geometric features as used in the Power Performance Test Report for AOC 15/50. Also
the blade was imparted a 6° of positive (downwind wind turbine) cone angle. Figure 9.11 shows
Various Radial Stations.
Figure 9.11 Turbine Blade Showing Various Radial Stations in ANSYS® DesignModeler
underlying response values, thereby introducing truncation errors in the predicted output function
value. A full quadratic polynomial (degree 2, order 3) approximation of F can be written as:
𝐲 = 𝚽𝛃
Eq. 9.9
Here y is the initial response matrix
[𝑦1, 𝑦2,…,]𝑇 and 𝚽 is the
[Vandermonde] matrix of size (N𝑠 ×
Nvar) given in [Kapoor]512. Figure
9.12 displays response surface
variation of chord and twist angle at
station 4 and 3 respectively, w.r.t
Torque.
9.1.13 CFD Domain Mesh and Numerical Model for Rotating Bodies
This subtopic gives an insight into the CFD numerical models for turbo-machinery applications. The
im of this paragraph is providing the numerical basis to perform CFD simulation of rotating bodies.
The main challenge in turbo-machinery applications is the introduction of a rotating body to apply
forces on the fluid (e.g. compression or expansion). From an analytical point of view the rotation
should be introduced into constitutive equations of motion, and there are mainly two approaches:
512 Gaurav Kapoor, “Exploration of a Computational Fluid Dynamics Integrated Design Methodology for Potential
Application to a Wind Turbine Blade”, Thesis Submitted to the Department of Aerospace Engineering, College
of Engineering, Embry-Riddle Aeronautical University, Daytona Beach, December 2014.
279
the Moving Reference Frame (MRF) and the Sliding Mesh (SLM). The first one consists of rewriting
N-S equations in a rotating frame, while the second one introduces rotation assigning a rotational
component of velocity to all nodes of the domain (physical grid rotation). It is immediately
understandable that SLM approach is more realistic that MRF, but also more CPU demanding as the
computational model needs re-meshing at every time advancement during the simulation procedure.
Since the rotation of grid intrinsically depends on time-evolution of simulation, this approach is not
recommended for steady state simulations as the solution obtained is not time-dependent. In other
words, a time steady calculation performed with MRF approach according to the evidence in most of
the turbo-machinery problems, does not compute a time-accurate solution.
vr = v − (ω × r)
Eq. 9.10
In the above equation, vr is the relative velocity (the velocity as viewed from the rotating frame) while
280
𝑣 is the absolute velocity (the velocity as viewed from the stationary frame). When the equations of
motion are solved in the rotating reference frame, the acceleration of the fluid is increased by the
additional terms that appear in the momentum equation. Moreover, the equations can be formulated
expressing absolute or relative velocity as dependent variable of momentum equation. Constitutive
N-S equations for which the solution is being calculated according to the relative velocity formulation
for continuity, momentum and energy respectively are as follows:
∂ρ
+ ∇. ρ𝐯r = 0
∂t
∂(ρ𝐯𝐫 )
+ ∇. (ρ𝐯𝐫 . 𝐯r ) + ρ(2𝛚
⏟ × 𝐯r + ⏟
𝛚 × 𝛚 × 𝐯r ) = −∇p + ∇. 𝛕 + 𝐅
∂t
Coriolis Centripetal
∂(ρEr )
+ ∇. (ρ𝐯𝐫 . Hr ) + ∇. (k∇T + 𝛕. 𝐯𝐫 ) + 𝐒H
∂t
Eq. 9.11
The momentum equation formulated above contains two additional acceleration terms, the Coriolis
component of acceleration (2 𝜔 x 𝑣𝑟 ) and the centripetal acceleration (𝜔 x 𝜔 x 𝑣𝑟 ). In addition, viscous
stress tensor does not change with respect to the MRF equation, except for the introduction of the
relative velocity. Energy equation is written in the form of internal energy Er, introducing the total
enthalpy Hr of the system in consideration. MRF model can be applied to different zones in the
domain (both rotating and nonrotating), solving RANS formulation of equations. Moreover
translational or rotational periodic boundaries can be applied wherever periodic surfaces are
present in the domain. For these reasons MRF model is widely used for industrial applications, being
one of the most versatile and low CPU-demanding approaches for turbo-machinery simulation.
➢ Velocity Inlet – The upstream surface of the domain is set to the velocity inlet condition as the
free stream velocity to be simulated in the computational domain is
known beforehand.
➢ Wall – The blade upper and lower surfaces are selected as wall with no-slip condition.
Periodic Boundary – The edges of the computational domain on either side of the wedge are
selected to be the periodic boundaries. The velocities going out from the left symmetry
boundary can enter the right boundary on the other side in an infinite loop.
➢ Pressure Outlet – The surface of the computational domain downstream from the blade is
set to a pressure outlet condition. This gives a better prediction of the exit pressure
distribution and thus results in better accuracy of the overall solution. The pressure at the
outlet was set to be atmospheric pressure.
➢ Symmetry – The curved surface of the computational domain is selected to be the symmetry
type. This just means that these boundaries do not affect the flow in any possible way.
the torque produced, a uniform inflow velocity profile was modelled for all CFD simulations for
parametric study and sensitivity analysis. All simulations were computed in steady state until
convergence or till the end of prescribed iterations to allow developed flows in the domain. Then in
order to maintain computational stability, the simulations were switched to transient solver.
Convergence was monitored looking at the thrust force time histories over different revolutions and
reached in a few cycles (about 2 to 3) for all wind conditions tested. Also the residual tolerance of
10-6 was reached for all velocity and energy terms to ascertain the robustness of the obtained flow
parameters. Furthermore, the difference in the mass flow at the inlet and the outlet of the
computational domain showed a negligible error (order 10-6). Additionally, a vertex point was
created on the symmetry axis at one blade length downstream of the blade to track the history of
average velocity at the vertex point over the course of simulations. The simulations were stopped
when the average velocity at this vertex was fairly constant and did not show any appreciable change.
All the above four conditions were satisfied as per the best practices to be followed in ANSYS® Fluent
for obtaining an accurate and converged solution. (See Table 9.2).
Power = 𝛕. 𝛚
Eq. 9.12
283
Coefficient of power (CP), a measure of how efficiently a wind turbine converts the energy available
in the wind to electricity.
9.1.18.1 Routine 1
The total length of the blade (7.5 m) and the maximum chord (0.749 m) occurring at Station 2 are
kept a constant (constrained) with an aim to optimize the existing blade within the length
requirements.
9.1.18.2 Routine 2
The starting point of this routine is taken as the best candidate point of Routine 1 to begin the search
on the response surface. The constraints applied are bounded by the design space spanning (+ -) 10%
from the base value of the design variables P3, P4 and P9. Further details are available in513.
9.1.19.2 Routine 2
There is a change in the values of
design variables P3, P4 and P9
after the optimization routine 2.
The optimized value of P3 turns
out to be 0.43578 m (+7.33%)
instead of the baseline value of
0.406 m. Also, the optimized
values of P4 and P9 are 5.214 m
(+10%) and 2.9549º (+10.87%)
respectively. This also indicates
that the total blade length has
been increased by 10%, which
results in the augmented power
of 1069.5 Nm (+25.26%). The Figure 9.15 Torque (P12) vs Twist_Station3 (P9) for
graph (obtained from what-if Optimization Routine 2
scenario study) in Figure 9.15
above is in agreement with the above optimization result. The graph clearly shows that if all the other
input parameters are held constant, the optimum design values for P3, P4 and P9 (Twist Station 3)
at a value of approximately 2.9 degrees gives the maximum blade torque output of about 1069 Nm.
The graph also verifies the optimization routine 2 carried out. (See Table 9.4).
9.1.20 Conclusion
In this research, the flow around the airfoils comprising the HAWT blade and the three dimensional
rotor blade is established using the commercial solver ANSYS® Fluent. A pressure based
compressible flow solver with k-ω SST turbulence model was used for all the flow simulations. To
study the dependence (sensitivity) of blade geometric/design parameters (what-if scenario) on the
power generated using ANSYS® Fluent, the Design of Experiments (DOE) approach of ANSYS
DesignXplorer was used. Parameter correlation study and sensitivity analysis conducted gave an
insight to how the changes in the blade geometry would affect the power output of the blade. The
blade aerodynamic optimization inclined toward the non-linear or quadratic relationship between
parameters, clearly indicated by the scatter plots and the quadratic determination matrix. This
parametric correlation study reveals that the blade design variables on the outer 40% of the blade
span have a predominant effect on the power output of the blade. Only the most sensitive design
variables are used for the blade optimization problem.
Using the results obtained from CFD simulations, a full quadratic polynomial response surface model
(RSM) is constructed, which is then optimized using the Nonlinear Programming by Quadratic
Lagrangian (NLPQL) technique to obtain the optimum values of the design variables. For
285
constructing the RSM, the Latin Hypercube Sampling (LHS) design is used to obtain the Design of
Experiments (DoE) plan. The main advantage of using this approach for shape optimization
problems is that values obtained from commercially available flow solvers can directly be used in the
optimization process, without making any changes to the solver’s code. Also the noise and non-
smoothness issues associated with CFD results are smoothened out by using the RSM which is
quadratic polynomial in terms of the design variables. Thus the optimization process can be
performed effectively and smoothly without any sudden divergence issues associated with the CFD
results. As evident from the CFD validations carried out on the optimum candidate point, the
optimization algorithm generated a design configuration that resulted in a localized optimum design
that had increased power output (+7.55%) at wind speed of 8.03 m/s only. The algorithm thus
resulted in a local optimum solution rather than a global optimum. Achieving a global optimum
solution to this problem would require several data points to be generated for obtaining a complete
and well established response surface spanning the entire operating wind spectrum of the turbine,
this is a costly affair in terms of the computational resources available. The Cp plots at various span
locations also bolster the claim that only the outer (from tip) 30-40% of the blade contributes most
towards the power output.
The results show that the power coefficient of the VAWT, equipped with the optimized blades,
improves at all TSRs from 0.4 to 1.5. Moreover, it was shown that and the maximum growth rate
occurs at TSR = 0.9, with a value of 26.82%.
514 Ning Ma, Hang Lei , Zhaolong Han, Dai Zhou , Yan Bao , Kai Zhang , Lei Zhou, Caiyong Chen, “Airfoil
optimization to improve power performance of a high-solidity vertical axis wind turbine at a moderate tip speed
ratio”, Energy 150 (2018) 236-252.
286
adaptability to turbulent or skewed flows and reduced level of noise515-516. For offshore applications,
due to the lower center of gravity, easier manufacturing, reduced installation & maintenance cost and
higher scalability to large sizes, VAWTs are regarded to be more promising than HAWTs.
515 Zamani M, Nazari S, Moshizi SA, Maghrebi MJ. Three dimensional simulation of J-shaped Darrieus vertical axis
wind turbine. Energy 2016.
516 Balduzzi F, Bianchini A, Carnevale EA, Ferrari L, Magnani S. Feasibility analysis of a Darrieus vertical-axis
vertical axis wind turbine using three dimensional CFD model. Energy 2017.
519 Bachant P, Wosnik M. Effects of Reynolds number on the energy conversion and near-wake dynamics of a high
around the blade521. Capture ratio means the ratio of the wind power obtained by the wind turbine
to the total amount of power carried by the air flowing through the turbine. Plenty of airfoil families
have been created and applied for aircrafts and HAWTs during the past decades. Although those
airfoil families are not customized for VAWTs, they can be used as original templates and then
optimized for VAWTs, which is much easier than designing brand new airfoil families for VAWTs. A
complete profile optimization process generally contains parametric modelling, calculation and
optimization parts. The widely used traditional methods for calculating the aerodynamics of VAWTs
include the Blade Element Momentum (BEM) theory, Vortex method and Cascade model. Although
assumptions and simplifications involved in these methods once brought much convenience, they
also result in intrinsic limitations which may cause large deviations between experimental outcomes
and the calculated ones. CFD method, supported by computer technologies and favored by
researchers nowadays, has the following advantages compared to traditional methods. On one hand,
it is a more accurate approach and the generated results are usually in good agreement with the
experimental data. On the other, it can reveal the phenomenon of dynamic stall better and even
visualize detailed information of the flow field including streamlines and vorticity evolution522. When
it comes to airfoil optimization, there has been some research using different algorithms. [Ribeiro et
al]523 utilized the Non-dominated Sorting GA-II (NSGA-II) in the airfoil optimization work for wind
turbines. CFD method combined with artificial neural networks could save about 50 % computational
time when compared with conventional CFD methods. To develop an automatic airfoil profile
optimization process for improving the power performance of a high solidity VAWT at a moderate
TSR. Specific objectives are as follows:
➢ Establishing an appropriate numerical model of a three bladed VAWT and proving its validity;
➢ Building an optimization system based on 3D CFD simulation to improve the power
coefficient of the VAWT.
To achieve these objectives, the present study involves the work of numerical modelling, verification
& validation of CFD model and the development of an optimization system. Findings are shown with
indicators like power coefficient, pressure coefficient and vorticity distribution.
521 Wang Q, Chen J, Pang X, Li S, Guo X. A new direct design method for the medium thickness wind turbine airfoil.
J Fluid Structure 2013.
522 Lei H, Zhou D, Lu J, Chen C, Han Z, Bao Y. The impact of pitch motion of a platform on the aerodynamic
that needs clarifying is that the blades are regarded as rigid in the simulation. But the absence of aero
elasticity has very small effect on the precision of generated power coefficients for the case used in
our paper, which is a 0.8m-diameter small-scale wind turbine operating under low wind speeds. The
effect of the aero elasticity is usually obvious for large, multi-megawatt and flexible wind turbines,
especially under high wind speeds. Based on the considerations above, we chose to perform the
simulation in the absence of aero elasticity. This could also effectively control the complexity of our
work.
Figure 9.16 Schematic diagram of an H-rotor type VAWT: (a) 3D view; (b) top view - [Courtesy of
Ma et al)
289
power loss of 5.5% was observed when shaft-to-turbine diameter ratio equals 16% compared to the
case without shaft. The similar model and corresponding data in the study of [Elkhoury et al.]525 will
be used for comparison and verification.
The parameterization of the airfoil profile is the first step of optimization as it can use variables
modified by the algorithm to generate new geometries. Five rules that an effective parameterization
method should satisfy have been mentioned in526. In this paper, Eq. 9.13 is used to represent the
upper or lower profile of the airfoil. Of the two terms, the former one pxa(c - b)b can be seen as the
mean camber line (MCL) while the latter one qxe(c –b)f can be seen as the thickness distribution.
Such an equation has sufficient flexibility to cover a large search space which allows the appearance
of non-routine profiles. Besides, this equation is highly adaptive and suitable for both symmetric and
cambered airfoil profiles.
Table 9.6 Ranges of variables and specific values of variables for NACA0018 airfoil - [Courtesy of
Ma et al.)
The value ranges of design variables during the optimization are also set according to the initial
values of variables for NACA0018. For example, the initial value of a1 for NACA0018 is 0.5112, then
the range of a1 is set to be 0.5112 ± s. The parameter s is related to the scope of the search space of
the optimization system. It would be well if 0.2 is assigned to s and the range of a1 would be
approximately 0.31 - 0.71 in this case. ‘s = 0.2’ is a moderate value which can not only ensure a
525 Elkhoury M, Kiwata T, Aoun E. Experimental and numerical investigation of a three-dimensional vertical-
axis wind turbine with variable-pitch. 2015..
526 Giannakoglou KC. Design of optimal aerodynamic shapes using stochastic optimization methods and
sufficient scope of
search space but also
save much
computational
resources. It also needs
to be noted that all the
variables are greater
than zero except q2.
9.2.4 Computational
Domain &
Meshing
The computational
domain can be divided
into an outer zone and a
rotating inner zone. All
Figure 9.17 Plan Views of the Computational Domain and Boundary
the sizes with respect
Conditions - [Courtesy of Ma et al.)
to rotor diameter D and
blade span h as well as
the boundary conditions are
presented in Figure 9.17. The
determination of these relative
sizes is based on the work of
previous researchers. Other studies
also provide useful guidelines on
settings of domain size despite that
their advice is mainly for 2D/2.5D
simulations. In the present study,
the center of the inner zone is
situated at 5 rotor diameters
downstream of the inlet boundary
and 15 rotor diameters upstream of
the outlet boundary, which can
allow for sufficient space for the
wake generation. The width of the
outer zone is 10 times the rotor
diameter while its height is 3 times
the blade span. For the inner zone,
its diameter is 2 times the rotor
diameter and its height is 1.5 times
the blade span.
The mesh setup is completed in
STAR-CCM+ which contains
components of meshing, calculating
and post-processing. The rotational
speed of the inner zone could be
adjusted in order to investigate the Figure 9.18 Mesh Setup (a) Topology; & BCs. (b) Grids at the
power performance at different Tip Mid-Height Plane - [Courtesy of Ma et al.)
Speed Ratios (TSRs). Figure 9.18
shows the grids of the
291
computational domain. Unstructured trimmed cells with varied sizes are adopted in different zones.
Prismatic boundary layer cells are used on the surface of the blades. The thickness of the boundary
layer is 0.01mwith a growth ratio of 1.25. The ‘all-y+ wall treatment’ is selected in STARCCM which
is a hybrid treatment and adaptable to both coarse and fine grids. The maximum value of y+ on the
blade surfaces is 0.7 which indicates acceptable resolution of boundary layer cells. There are roughly
50 cells along the chord length and 180 cells along the span wise direction.
A uniform velocity profile (V∞ = 8 m/s) is set as the inlet condition and the outlet boundary is
considered as pressure outlet with zero relative pressure [7]. The turbulence intensity is set as 1%
and the value of turbulent length scale is 0.014mat the inlet. The slip wall condition is assigned to the
remaining boundaries of the outer zone while the non-slip wall condition is assigned to blade
surfaces. Two sub-domains are linked with interface boundary condition at their contact surfaces.
Besides, the grid size of two subdomains near the interface should be consistent enough to ensure
the effective and smooth data exchange. The settings mentioned above are used in the auto-meshing
for each model generated during the optimization process. In this way, the number of grids and the
grid size around the blades are consistent, which ensures the reliability of the comparative analysis
later.
Qω
Cp = 3
0.5ρDhV∞
Eq. 9.14
where Q, u, r D, h and V∞ are torque of the VAWT, angular velocity of the rotor, air density, rotor
diameter, blade span and freestream velocity, respectively. The number of grids and predicted value
of CP for each case are shown in Table 9.7. The relative change of CP from the medium grid set to the
292
Figure 9.19 Time-varying power coefficients in a rotational period when adopting different time
steps - [Courtesy of Ma et al.)
293
527 Elkhoury M, Kiwata T, Aoun E. Experimental and numerical investigation of a three-dimensional vertical-axis
wind turbine with variable-pitch. J Wind Eng Ind Aerod 2015;139:111e23.
528 Lei H, Zhou D, Bao Y, Li Y, Han Z. Three-dimensional improved delayed detached Eddy simulation of a two-
532 Nemec M, Zingg DW, Pulliam TH. Multipoint and multi-objective aerodynamic shape optimization. AIAA 2004.
533 Rai MM, Madavan NK. Aerodynamic design using neural networks. AIAA J 2000.
534 Chehouri A, Younes R, Ilinca A, Perron J. Review of performance optimization techniques applied to wind
flows between different programs and to automate the operation of simulation process. As there are
no standard rules for integrating function components in ISIGHT, each optimization system is
designed according to special needs of the corresponding study. The complete scheme of the loop is
illustrated in Figure 9.21.
The modelling component firstly uses design variables (a1, a2, b1, b2, e1, e2, f1, f2, p1, p2, q1, q2)
and other constant parameters (c, D, h) to generate airfoil profiles with the help of NX (Siemens PLM
Software, Texas, United States). An initial population is needed to start the whole optimization
process. Apart from NACA0018 airfoil, other individuals of the initial population are randomly
generated within the scope which is defined by the ranges of variables shown in Table 9.6. Then
the geometry is imported into STARCCM where the meshing & calculation work is done. After that,
values of time varying dependent variables (e.g. the torque of the VAWT) are exported to MATLAB
for post processing in which the values of objectives and constraints are calculated for the
optimization. The running of the entire system is controlled by macros which can call applications
and issue the orders. After the input of the initial population, the process automatically operates until
the ultimate results are outputted.
535 Ribeiro AFP, Awruch AM, Gomes HM. An airfoil optimization technique for wind turbines. Applied Math, 2012.
536 Goçmen T, Ozerdem B. Airfoil optimization for noise emission problem and aerodynamic performance criterion
point where the power coefficient starts to fall off also changes. The reverse point of the initial case
is at TSR = 1.3 while that of the optimized case is at TSR = 1.2, which means the appropriate angular
speed to generate the maximum power output has dropped down when the freestream velocity
equals 8 m/s.
As the power coefficient CP is derived from the torque values, a comparison of the time-varying
torque values during one rotational period is conducted, which is illustrated in Figure 9.25 with the
solid line representing the optimized case and the dashed line representing the initial case. Here the
analysis is divided into
two parts. The first part
is about the variation
trend at a specific TSR
and TSR = 0.9 is selected
at which the largest
growth rate of the power
coefficient CP occurs.
The analysis at other two
TSRs is much alike and
will not be discussed in
detail. The second part
investigates the effects of
TSR on the torque values
and involves all the six
diagrams in Figure
9.25. At TSR = 0.9, the
peak torque value of one Figure 9.24 Comparison of average power coefficients and the growth
blade with the optimized rate of CP between the initial and optimized airfoils at different TSRs
profile is 3.87 Nm
appearing at ϴ = 82 degrees while that with the initial profile is 3.45 Nm occurring at ϴ = 76 degrees.
The maximum difference of the torque values takes place at ϴ = 90 degrees with a difference value
of 1.09 Nm. Besides, the torque value of one blade with either the optimized or the initial profile
keeps negative in most parts of the downstream area due to the increase of the drag force. In Figure
9.25 (b), the solid line is obviously above the dashed line which means the power performance has
been improved at every second during the operation of the VAWT. The peak torque values of the
VAWT equipped with three optimized blades during one rotational period appear at ϴ= 82, 202 and
322 degrees with an approximate value of 3.7 Nm while that of the VAWT equipped with initial blades
are approximately 3.2 Nm occurring at ϴ = 76, 196 and 316 degrees. The maximum differences of the
torque values of the VAWT appear at ϴ = 90, 210 and 330 degrees. Besides, the torque value is always
above zero for the VAWT with either optimized or initial blades. In fact, by comparing Figure 9.25
(b), it is obvious that the peak values of the VAWT appear at the same time when the peak value of
every single blade occurs.
When comparing the torque curves at different TSRs, the curves of the optimized cases are focused
on and there are two changing rules worth attention. On one hand, the position where the peak value
appears moves with the change of TSR. For the torque curve of one blade, the positions of peak value
are ϴ = 60 degrees at TSR = 0.6, ϴ = 82 degree at TSR = 0.9 and ϴ = 94 degree at TSR = 1.2. The peak
value occurs later during one rotational period when the TSR is larger. On the other, the interval in
which the relatively high torque values appear becomes larger with the increase of TSR. It can be
seen that the azimuthal interval is 44 degree at TSR = 0.6, or 52 deg at TSR = 0.9, or 60 degree at TSR
= 1.2 during 1/3 rotational period. This can also explain why the peak value at TSR = 0.9 is larger
than that at TSR = 1.2 while the average power coefficient at TSR = 0.9 is less than that at TSR = 1.2.
298
9.2.13 Discussion
This section discusses the findings from
figures mentioned above, which involves
discussions on the flow regime around
blades, airfoil profile change and (b) TSR = 0.9
pressure distribution on the blade. They
could be referred to when designing
airfoils of VAWTs in real applications.
Specific airfoil profile could reduce the
energy loss caused by vortices (related to
the flow regime), thus improving the
power performance. Besides,
understanding the change in pressure
distribution induced by the airfoil profile
evolution could help improve the
structural performance of blades.
A small vortex forms near the trailing
edge because of the circulation of flow
leaving blade surfaces at ϴ = 0. This (c) TSR=1.2
vortex moves along the inner side of
blade and induces a reversed pressure
wave transmitting towards the leading
edge. In fact, the adverse pressure
gradient could easily cause the formation
of bubbles on the inner side of blade and
drives them towards the leading edge.
This phenomenon is also reflected by the
pressure coefficient change on the inner
side from ϴ = 0 to 60. During such an
azimuthal range, the maximum negative
pressure on the inner side keeps Figure 9.25 Instantaneous torque values during one
increasing and its position continues to rotational period - [Courtesy of Ma et al.)
537 Ning Ma, Hang Lei , Zhaolong Han, Dai Zhou , Yan Bao , Kai Zhang , Lei Zhou, Caiyong Chen, “Airfoil
optimization to improve power performance of a high-solidity vertical axis wind turbine at a moderate tip speed
ratio”, Energy 150 (2018) 236-252.
299
move towards the leading edge. From ϴ = 90 to 120, a large vortex is detected as there is an obvious
trough of the negative pressure coefficient on the inner side, which is in accordance with the vorticity
distribution in Figure 9.25.
This vortex has started to shed from the inner side when ϴ = 90 degree and its moving speed is
smaller than that of the blade as the center of the trough of pressure coefficient changes from 0.22c
to 0.4c. Besides, the vortex center keeps moving away from the inner surface because the maximum
negative pressure coefficient decreases a lot from q = 90 to q= 120 degrees. When ϴ= 150 , this large
vortex has lost contact with the blade surface (Figure 9.25(b)) and consequently has little impact
on pressure distribution.
Figure 9.26 Instantaneous Contours of Vorticity Magnitude at TSR = 0.9 for the VAWT with
Optimized Airfoils - [Courtesy of Ma et al.)
300
It needs to be noted that the pressure distribution is obviously different between the initial and
optimized cases when ϴ = 90 degree. There isn't a trough of pressure coefficient on the inner side for
the optimized case, which is caused by the shape change of airfoil profile. As is stated before, the
radius of the leading edge increases and the slope of inner side profile also enlarges (Figure 9.22)
for the optimized case. The increased leading edge radius might delay the movement of leading edge
vortex towards the trailing edge at the beginning. But the larger slope of the inner side profile would
cause the vortex to move a bit faster towards the trailing edge later on. The key parameters like the
radius of the leading edge, angle of the trailing edge and slope of the profile should be determined
with care as they have impacts on the movement of vortices and pressure distribution.
Leading edge vortices could be easily found in Figure 9.26 (a)-(d), which also correspond to the
small troughs of pressure coefficient near the leading edge on the blade inner side. Such leading-
edge vortices are usually induced by the laminar-turbulent transition on the blade inner side in the
upstream area. In the downstream area, the flow reattaches to the inner side while small laminar
separation bubbles are induced on the outer side. The scale of such bubbles is relatively small
compared to those forming in the upstream area. They keep growing and moving towards the trailing
edge as is shown from ϴ =210 degree to ϴ = 330 degrees in Figure 9.26.
➢ The study focuses on the improvement of power performance which means the optimized
airfoil profile might not be the optimized choice for performance in other aspects like
structural performance.
➢ The research object is a high-solidity VAWT with the solidity of 0.75. The findings in this
paper may not be applicable to VAWTs with low or moderate solidity. But the optimization
system developed in this study could be used for VAWTs with various levels of solidity.
➢ The research is conducted at a moderate tip speed ratio (TSR =1.0). This means the
optimization results may change when adopting higher or lower TSRs. When the system is
used for actual design, the TSR could be adjusted to the real value and the system will also
work.
9.2.15 Conclusion
This work establishes a comprehensive optimization system to improve the power performance of a
three-bladed high-solidity VAWT at a moderate TSR. The system is highly adaptive featuring the
combination of 3D CFD simulation, Multi-Island GA and auto operation. Besides, the airfoil is studied
as part of a three-bladed VAWT rather than as a single isolated body. Following conclusions are
obtained:
➢ The maximum thickness of the optimized profile increases by 12.8% compared to NACA0018
profile and the position of the maximum thickness moves from 0.3c to 0.275c. The mean
camber line of the optimized airfoil is not aligned to the chord anymore and the offset of it
reaches the maximum downward at the position of 0.08c and upward at the position of 0.78c.
➢ The radius of the leading edge of the optimized profile enlarges while the angle of the trailing
edge shrinks compared to the initial profile. The increased leading edge radius might delay
the movement of leading edge vortex towards the trailing edge. But the larger slope of the
inner side profile would cause the vortex to move a bit faster towards the trailing edge later
on.
301
➢ The power performance of the VAWT equipped with the optimized blades improves sensibly.
In terms of the power coefficient, the maximum growth rate of it occurs at TSR = 0.9 with a
value of 26.82% and the growth rate keeps over 20% when the TSR is no larger than 1.0. The
power performance improvement mainly benefits from the increase of torque value in the
azimuthal angle range [60 degree, 120 deg.].
302
538 Saltelli, A., S. Tarantola, F. Campolongo and M. Ratto, “Sensitivity Analysis in Practice: A Guide to Assessing
Scientific Models”, John Wiley & Sons, Ltd, 2004.
539A. Saltelli, M. Ratto, T. Andres, F. Campolongo, J. Cariboni, D. Gatelli, M. Saisana and S. Tarantola “Global
Sensitivity Analysis: The Primer” , John Wiley & Sons, Ltd, 2008.
540 From Wikipedia, the free encyclopedia.
541 Same Source
303
A close third might be the extra effort involved in setting up gradient-based optimizations, due to the
need for smooth mesh deformation, preconditioning of design variables, and the lack of robustness
to the failure of any step of the process. This last is not true of e.g. genetic algorithms, for which a flow
code failure is simply a member of the population that is not evaluated. Issue (I) is an unavoidable
consequence of the locality of gradient-based methods: in a design space with many local optima they
will tend to find the nearest (with respect to the starting point), not the best, local optimum. Further
there is evidence to suggest that configurations such as multi-element high-lift devices have just such
highly oscillatory design spaces. This problem may be tackled with hybrid optimization algorithms
that combine non-deterministic techniques for broad searches, with gradient-based methods for
detailed optimization, a significant topic of current research542-543. However issue (II) is perhaps
more critical in general, and that with which this article is concerned. Inevitably the availability of
sensitivity analysis tools for codes lags behind the codes themselves, a situation exacerbated by the
considerable effort required to linearize complex solvers. A good example is the lack of reliable adjoin
solvers for Navier-Stokes problems until recently. The situation deteriorates when multi-disciplinary
problems are considered, and the sensitivities of coupled multi-code systems are required.
Nonetheless these problems are worth overcoming, as when they are available, gradient-based
algorithms combined with adjoin gradients are the only methods which can offer rapid optimizations
for extremely large numbers of design variables544, as needed for the aerodynamic shape design545.
Each technique has its own unique characteristics. For example, the Direct Differentiation, has the
advantage of being exact, due to direct differentiation of governing equations with respect to design
parameters, but limited in scope. There are two basic components in obtaining aerodynamic
sensitivity. They are:
• Obtaining the sensitivity of the governing equations with respect to the state variables,
• Obtaining the sensitivity of the grid with respect to the design parameters.
The sensitivity of the state variables with respect to the design parameters are described by a set of
linear-algebraic relation. These systems of equations can be solved directly by a LU decomposition
of the coefficient matrix. This direct inversion procedure becomes extremely expensive as the
problem dimension increases. A hybrid approach of an efficient banded matrix solver with influence
542 G. Lombardi, G. Mengali, F. Beux, A hybrid genetic based optimization procedure for aircraft conceptual
analysis, Optimization and Engineering, 2006.
543 V. Kelner, G. Grondin, O. Léonard, S. Moreau, Multi-objective optimization of a fan blade by coupling a genetic
of off-diagonal elements iterated can be implemented to overcome this difficulty546. Figure 10.1
displays different methods for obtaining the sensitivity derivatives.
Adjoint
Variable
(AV)
Finite Direct
(Analytical)
Differencing Differentiation
(FD) (DD)
Sensitivity
Derivatives
Automatic Symbolic
Differentiation Differentiation
(AD) (SD)
I Q
= R(Q) Q = ρ, ρu , ρv , ρw , E
T
J t
Eq. 10.1
Here, R is the residual and J is the Jacobean Transformation:
546 Sadrehaghighi, I., Smith, R.E., Tiwari, S., N., “Grid Sensitivity And Aerodynamic Optimization of Generic
Airfoils”, Journal of Aircraft, Volume 32, No. 6, Pages 1234-1239.
305
(ξ, , ζ)
J= Eq. 10.2
(x, y,z)
Where R is the residual vector, Q is a four-element vector of conserved flow variables. Navier-Stokes
equations are discretized in time using the Euler implicit method and linearized by employing the
flux Jacobean. This results in a large system of linear equations in delta form at each time step as
I R
n
+ ΔQ = −R n Eq. 10.3
Jt Q
𝐑(𝐐(𝐏), 𝐗(𝐏), 𝐏) = 0
Eq. 10.4
Where the explicit dependency of R on grid and vector of parameters P is evident. The parameters P
control the grid X as well as the solution Q. The fundamental sensitivity equation containing ∂Q/∂P
is obtained by direct differentiation of Eq. 10.4 as
R Q R X R
Q P + X P + P = 0
Eq. 10.5
It is important to notice that Eq. 10.5 is a set of linear algebraic equation and matrices ∂R/∂Q and
∂R/∂X is well understood. The flow sensitivity, ∂Q/∂P can now be directly obtained as
−1
Q R R X R
= + Eq. 10.6
P Q X P P
Further simplification could include the vector of grid sensitivity which is
X X XB
= Eq. 10.7
P XB P
547 Sadrehaghighi, I., Smith, R.E., Tiwari, S., N., “Grid Sensitivity and Aerodynamic Optimization Of Generic
Airfoils”, Journal of Aircraft, Volume 32, No. 6, Pages 1234-1239.
548 Régis Duvigneau, “A Sensitivity Equation Method for Unsteady Compressible Flows: Implementation and
to the adjoint equations, sensitivity equations are integrated forward in time and do not need a
complete storage of the solution. Nevertheless, their computational cost depends on the number of
parameters considered. Some works can be found in the literature on the sensitivity equation method
for incompressible flow problems, for different applications like optimum design, uncertainty
propagation or flow characterization. However, only a few works can be found related to
compressible flows. Here, the implementation of the method is described in the context of unsteady
compressible Navier-Stokes equations, for laminar and subsonic flow regimes.
∂𝐐
+ ∇. Ê = ∇. Ĝ where 𝐐 = {ρ, ρu, ρv, e}T
∂t
Eq. 10.8
and Ê = (Ex(Q), Ey(Q)) is the vector of the convective fluxes and Ĝ= (Gx(Q), Gy(Q)) the vector of the
diffusive fluxes. The pressure p is obtained from the perfect gas state equation, and the extension to
the 3D case is straightforward. The individual fluxes are given by;
ρu ρv
ρu + P 2 ρuv
Ê𝐱 (𝐐) = ρuv , Ê𝐱 (𝐐) = ρv 2 + P
p p
ρu (e + ) ρv (e + )
[ ρ ] [ ρ ]
Eq. 10.9
The viscous fluxes are written as:
0 0
τ𝑥𝑥 τ𝑥𝑦
Ĝ𝐱 (𝐐) = [ τ𝑦𝑥 ] , Ĝ𝐱 (𝐐) = [ τ𝑦𝑦 ]
uτ𝑥𝑥 + 𝑣𝜏𝑦𝑥 − 𝑞𝑥 uτ𝑥𝑦 + 𝑣𝜏𝑦𝑦 − 𝑞𝑦
Eq. 10.10
where the symmetric viscous stress tensor and heat flux is defined in [Duvigneau]549. We consider in
this section the differentiation of the previous flow equations with respect to a parameter, denoted
α, that could be either scalar or geometrical. α could be a control parameter for design optimization
purpose, or an uncertain parameter for uncertainty quantification. The objective of this section is to
derive a set of partial differential equations governing the sensitivity fields denoted:
∂𝐐 ∂𝐐́
𝐐́ = or + ∇. Ế = ∇. Ĝ́
∂α ∂t
Eq. 10.11
The sensitivity of the convective fluxes can be expressed as:
549Régis Duvigneau, “A Sensitivity Equation Method for Unsteady Compressible Flows: Implementation and
Verification”, Research Report No. 8739, June 2015.
307
́
(ρu) ́
(ρv)
́
(ρu)u + (ρu)ú + Ṕ ́
(ρv)u + (ρv)ú
Ế𝐱 (𝐐, 𝐐́) = ́ v + (ρu)v́
(ρu) , Ế𝐲 (𝐐, 𝐐́) = ́ v + (ρu)v́ + 𝑃́
(ρu)
p 𝑃́ p 𝑃́
́ (e + ) + (𝜌𝑢) [𝑒́ + ( )]
(ρu) ́ (e + ) + (𝜌𝑣) [𝑒́ + ( )]
(ρv)
[ ρ 𝜌 ] [ ρ 𝜌 ]
Eq. 10.12
The sensitivity of the diffusive fluxes reads:
0 0
τ ́ τ 𝑥𝑦́
Ĝ́𝐱 (𝐐, 𝐐́) = Ĝ́𝐱 (𝐐, 𝐐́) =
𝑥𝑥
τ𝑦𝑥 ́ , τ𝑦𝑦 ́
[ú τ𝑥𝑥 + 𝑢𝜏𝑥𝑥́ + 𝑣́ 𝜏𝑦𝑥 + 𝑣𝜏𝑦𝑥
́ − 𝑞𝑥́ ] [ú τ𝑥𝑦 + 𝑢𝜏𝑥𝑦
́ + 𝑣́ 𝜏𝑦𝑦 + 𝑣𝜏𝑦𝑦
́ − 𝑞𝑦́ ]
Eq. 10.13
where the sensitivity of the viscous stress tensor and heat flux can be found in [Duvigneau]550. The
boundary conditions for the sensitivity equations are obtained by differentiating the boundary
conditions for the flow. There are some important properties of sensitivity equation w.r.t connective
fluxes which can be summarized below:
10.3.2 Numerical Solutions using an Iterative procedure for Flow and Sensitivity Equation
10.3.2.1 Flow Equations
The flow Eq. 10.8 form a system of conservation laws, which is solved using a mixed finite
volume/finite-element formulation551. The temporal scheme with the implicit discretization of
convective and diffusive terms yield the following non-linear problem, to be solved at each step:
3 1
𝐃𝐢𝐢 = 𝛄𝐢 ( + )
2Δt Δτi (𝐐k )
Eq. 10.15
where γi is the volume of the cell i, ∆t the physical time-step and ∆τi a local pseudo time-step adjusted
according to stability criteria. This non-linear problem for the flow is solved using an implicit
approach based on an approximate Jacobian matrix552. More precisely, a conservative state
550 Régis Duvigneau, “A Sensitivity Equation Method for Unsteady Compressible Flows: Implementation and
Verification”, Research Report No. 8739, June 2015.
551 Alain Dervieux, Jean-Antoine Desideri. “Compressible flow solvers using unstructured grids”. [Research
553Régis Duvigneau, “A Sensitivity Equation Method for Unsteady Compressible Flows: Implementation and
Verification”, Research Report No. 8739, June 2015.
554 Alain Dervieux, Jean-Antoine Desideri. “Compressible flow solvers using unstructured grids”. [Research
555 Steffen K¨ammerer, J¨urgen F. Mayer, and Heinz Stetter, “Development of a Three-Dimensional Geometry
Optimization Method for Turbomachinery Applications”, International Journal of Rotating Machinery, 2004.
556 Alejandro Cesar Limache, “Aerodynamic Modeling Using Computational Fluid Dynamics and Sensitivity
Equations”, Dissertation submitted to the Faculty of the Virginia Polytechnic Institute, 2000.
310
determination of these aerodynamic forces (and aerodynamic moments) is not a trivial problem,
since these forces are not independent of the aircraft motion but they are coupled to it. Therefore, at
a given time the particular values of the aerodynamic forces will affect the future aircraft motion, and
conversely, the particular values of the aerodynamic forces at a given time will depend on the
particular maneuver that the aircraft is performing at that time.
10.3.4.1 Kinematics
The problem gets more complex because these forces depend not only on the instantaneous values
of the motion variables but also depend on
the past history of the motion. As a
consequence, even under the assumption
that the aerodynamic forces can be
evaluated (numerically or experimentally)
at any instant for a particular motion of the
aircraft, it is obviously materially
impossible to evaluate and record all the
aerodynamic force histories associated to
each one of the infinitely many possible
histories of motion. From these
observations, one can conclude that an
analysis of the aircraft behavior is not
viable unless a mathematical model for the
representation of the aerodynamic forces
is introduced. The formulation of an
aerodynamic mathematical model
probably started with the work of
[Bryan]557 at the beginning of the 20th
century. Bryan postulated that the
aerodynamic forces depend only on the
instantaneous values of the motion
Figure 10.3 Aircraft Flying in an Arbitrary Motion
variables and that this dependence was
linear. Considering a rigid aircraft of
arbitrary geometry moving in an arbitrary way in the atmospheric air as shown in Figure 10.3. The
air is assumed to be uniform and at rest with respect to an inertial reference frame denoted by S. The
unit vectors of the axes of the Cartesian coordinate system associated to S will be denoted by {exS,
eyS, ezS}.
Another reference frame is defined to be fixed to the aircraft and is denoted by R. A Cartesian
coordinate system associated to R is defined such that its origin is located at an arbitrary point c of
the aircraft and the unit vectors of its axes will be denoted by {exR, eyR, ezR}. The subscript S will be
attached to the vector components when the vector is resolved in the inertial coordinate system S.
Similarly, the subscripts R will be attached to the vector components when the vector is resolved in
the coordinate system R. For example, the vector position _xc of the aircraft with respect to an
observer in the reference frame S will be written as xc = x0S exS + y0S eyS + z0S ezS when resolved in
terms of the Cartesian components of reference frame S.
𝐔 = {V𝑐 , 𝛼, 𝛽, 0,0,0}
Eq. 10.21
Figure 10.6 Aircraft Flying in a Rectilinear Steady Motion with no-Angular Rates
As a consequence, by restricting the motions to the class of rectilinear, steady flights, the
mathematical representation of the aerodynamic forces has been simplified from a functional form
of six function arguments to a function F of three scalar arguments: {Vc, α, β}. Observe that in the
representation the Mach number M = Vc as could be used instead of the speed Vc, in such a case, Eq.
10.21 can be rewritten as:
𝐅 = F{M , α, β, 0, 0,0}
Eq. 10.22
Now consider a symmetric aircraft moving in rectilinear longitudinal steady motion (i.e. constant
speed and angle of attack, β = 0, no-angular rates) as the one shown in Figure 10.6. Eq. 10.22 for
the pitching moment coefficient can be written as:
𝐂𝐦 = Cm {Vc , α} β=0,p=0,q=o,r=0
Eq. 10.23
If a Taylor Expansion to first order in the motion variables Vc, α is performed in the above equation,
then it follows that, except for the rotary contribution due to non-zero q, and it is the linear
approximation to the more general. Note that in this case the stability derivatives: CmV and Cmα have
313
now a clear meaning in terms of the partial derivatives of Cm. This idea can be extended to the
sensitivity derivatives with respect to the sideslip angle β if one considers the general rectilinear
steady flight of Eq. 10.22. Be cautious that in the present work, the mathematical model is based
on the choice of aerodynamically steady motions and not CFD steady. The model allows one to
formally determine the dependence of the aerodynamic forces (and aerodynamic moments) with
respect to all the motion variables including the angular rates of the aircraft.
∂ρ
Continuity + ∇. (ρ𝐕R ) = 𝟎
∂t
∂(ρ𝐕𝐑 )
Momentum + ∇. (ρ𝐕R ⨂ 𝐕R − P𝐈 − 𝛕) = ρ[𝐟 + 𝛀 − 𝟐𝛚 × 𝐕𝐑 ]
∂t
∂(ρER )
Energy + ∇. [(ρER − P)𝐕𝐑 − 𝛕. 𝐕𝐑 − k∇T] = ρ[𝐟 + 𝛀]. 𝐕𝐑
∂t
Eq. 10.24
In Eq. 10.24 the variables are expressed in a Eulerian (local) way as seen from the non-Inertial
rotating frame R. In this sense, _ VR is the local velocity of the flow as seen from the rotating reference
frame; ρ, P , T and e are the density, the static pressure, the temperature and the internal energy of
the flow, respectively; ER is the total energy (per unit of mass) as seen from the rotating frame:
558 Owczarek, J.A. , “Fundamentals of Gas Dynamics”, International Textbook Company, Scranton, 1964.
314
1
ER = e + (𝐕𝐑 . 𝐕𝐑 )
2
Eq. 10.25
where f is net body force; Ω is the “pseudo force vector”:
dR 𝛚
𝛀 = −𝛚 × (𝛚 × 𝐱 𝐑 ) + × 𝐱 𝐑 − 𝐚 𝐑/𝐒
dR t
Eq. 10.26
and contains the effect of the non-inertial motion (except for the Coriolis term −2ω × VR). V ⊗ V is a
second order tensor defined in terms of its components in Cartesian coordinates, I is the identity
tensor and τ is the stress tensor; each of these is a second order tensor.
∂𝐐 ∂𝐅 ∂𝐆 ∂𝐇
+ + + =𝐖
∂t ∂x ∂y ∂z
ρ ρu ρv
ρu ρuu + P ρvu
where Q = ρv , 𝐅= ρuv , 𝐆 = ρvv + P
ρw ρuw ρvw
[ ρE ] [u(ρE + P)] [v(ρE + P)]
ρw 0
ρwu ρ(fx + Ωx ) − 2ρ(ωy w − ωz v)
𝐇= ρwv , 𝐰= ρ(fy + Ωy ) − 2ρ(ωz u − ωx w)
ρww + P ρ(fz + Ωz ) − 2ρ(ωx v − ωy u)
[w(ρE + P)]
[ρ(fx + Ωx )u + ρ(fy + Ωy )v + ρ(fz + Ωz )w]
Eq. 10.27
dS 𝐕c
𝐱 r/s = 𝐱 c , 𝐕R/S = 𝐕c , 𝐚R/S = 𝐚 c =
ds t
Eq. 10.28
315
Also, the angular velocity components ωx, ωy and ωz can be identified with the standard flight
mechanics angular velocity components p, q and r of the aircraft:
dS 𝐕c
𝐱 r/s = 𝐱 c , 𝐕R/S = 𝐕c , 𝐚R/S = 𝐚 c =
ds t
Eq. 10.29
Also, the angular velocity components ωx, ωy and ωz can be identified with the standard flight
mechanics angular velocity components p, q and r of the aircraft:
𝛚𝑥 = p , 𝛚y = q , 𝛚z = r
Eq. 10.30
Now, since the aircraft will be forced to move in an aerodynamically steady motion, the following
conditions apply:
d𝛚 ds Vc
=0 , = 𝛚 × 𝐕c
dt ds t
Eq. 10.31
Implementing, the above criteria into Euler equations (Eq. 10.27), everything remains the same
except that Ω and W are now given by:
𝛀 = − 𝛚 × (𝛚 × 𝐱) − 𝛚 × 𝐕c
0
ρ[𝐟x + 𝛀𝐱 ] − 2ρ(qw − rv)
𝐖=− ρ[𝐟y + 𝛀𝐲 ] − 2ρ(ru − pw)
ρ[𝐟z + 𝛀𝐳 ] − 2ρ(pv − qu)
[ρ[𝐟x + 𝛀𝐱 ]u + ρ[𝐟𝐲 + 𝛀y ]v + ρ[𝐟𝐳 + 𝛀z ]w]
Eq. 10.32
Further details can be obtained from [Limache]559.
Figure 10.8 Airfoil flying in rectilinear longitudinal steady motion (i.e. β = 0, no-angular rates)
559Alejandro Cesar Limache, “Aerodynamic Modeling Using Computational Fluid Dynamics and Sensitivity
Equations”, Dissertation submitted to the Faculty of the Virginia Polytechnic Institute, 2000.
316
Since the problem is two-dimensional the only aerodynamic forces and moments that have to be
considered are the lift L, the drag D and the pitching moment M . For the class of rectilinear motions
defined above these three aerodynamic forces and their corresponding non-dimensional coefficients
CL, CD and Cm can be represented in terms of the steady aerodynamic function F as F = F (Vc, α, 0, 0, 0,
0). Assuming the flow is inviscid, these aerodynamic forces can be determined by an appropriate
integration of the static pressure P along the airfoil’s surface σa:
Figure 10.9 Pressure Contours and Velocity Streamlines for the air passing around an Airfoil
in a Aerodynamically Steady Rectilinear Uniform Flight
The flow solutions were obtained using the Class Code provided by [Kyle Anderson]. The Class Code
solves the inviscid flow equations using a finite-volume formulation. The formulation uses an
implicit, time-marching, iterative, node-centered scheme. The fluxes can be evaluated at the faces
using Van Leer flux splitting or the Roe difference scheme. The method is second order, since it uses
a second-order approximation of the flow variables at the faces. Something that must be noted from
the obtained solutions is the behavior of the flow variables at the far field. For far-field it can be seen
that the pressure coefficient:
317
P − P∞ P − P∞
CP = =
1 2 1
ρ∞ V∞ γP M 2
2 2 ∞
Eq. 10.34
goes to zero at the far-field, i.e. the pressure P matches with the free-stream condition P∞. Also, as
expected, it can be seen that the streamlines tend to be straight lines since in the simulation the airfoil
is flying in a rectilinear motion. Finally, observe that the aerodynamic forces can be determined from
the obtained pressure distributions by using Eq. 10.33. From the symmetry, it follows that for the
cases shown in Figure 10.9, we have, CL = 0.0 and Cm = 0.0. By running the CFD code at different
Mach numbers M and different angles of attack α the three set of functions can be constructed:
CL = CL (M, α) , CD = CD (M, α) , CM = CM (M, α)
Eq. 10.35
Q2
∂ (
∂u Q1 ) Q 2 ∂Q1 1 ∂Q 2 Q2 1
= =− 2 + = − 2 Sη1 + S
∂η ∂η Q1 ∂η Q1 ∂η Q1 Q1 η2
Eq. 10.37
318
Taking the derivative d/dη of the flow equation with respect to the parameter η. Since the
coordinates x and the time t are independent of η, we can interchange partial derivatives (assuming
certain smoothness) in the above equation. Following the development in [Limache]560, we get
∂Sη 𝜕 𝜕 𝜕
= [∇𝑄 F. Sη ] + [∇𝑄 G. Sη ] + [∇𝑄 H. Sη ] = W𝐴 Sη + W𝐵
∂t 𝜕𝑥 𝜕𝑦 𝜕𝑧
Eq. 10.38
If the PDE defined by Eq. 10.38 is solved with the appropriate boundary conditions the flow
sensitivities can be determined. Observe that the Jacobians ∇QF, ∇QG, ∇QH are known functions of the
flow solution and they are independent of the sensitivities. Also since the pseudo force vector Ω and
the external forces components fx, fy and fz are independent of the state variables the matrices WA and
WB do not depend on Sη and their description in available in561. From these properties, it follows that
the flow sensitivity Eq. 10.38 is a linear equation. Furthermore, the same numerical techniques
applied to the flow solutions can be used to solve this linear equation.
560 Alejandro Cesar Limache, “Aerodynamic Modeling Using Computational Fluid Dynamics and Sensitivity
Equations”, Dissertation submitted to the Faculty of the Virginia Polytechnic Institute & State University, 2000.
561 See Previous.
319
(c) M = 0.8
Figure 10.10 Pitch-Rate Pressure Sensitivity and the Velocity Sensitivity Streamlines of a NACA 0012
Airfoil for Different Mach Number
adaptive grid technology one should be aware that an acceptable refinement for the flow problem
may not be acceptable for the sensitivity problem.
Figure 10.10 display some of the pitch-rate flow-sensitivities results for the case of a NACA 0012
airfoil at different Mach Numbers and at α = 0.0 and q = 0. The origin of coordinates of the body-fixed
reference frame was chosen to be at the leading edge. The fact that q = 0 implies that the flow
solutions required to calculate the source term WB, and the Jacobians ∇QF and ∇QG, can be simply
extracted from the Class Code. That is to say when q = 0 the flow solutions can be obtained from a
standard (inertial) CFD code for uniform flows. However, this approach is not valid in cases where q
320
is non-zero. In such cases, the flow solutions must be obtained using the A-NISFLOW Code. Figure
10.10-(a) the pitch-rate pressure sensitivity contours and the velocity sensitivity streamlines are
shown for a NACA 0012 airfoil at M = 0.2.
Note that the finite-volume sensitivity formulation allows one to treat the range of subsonic,
supersonic and transonic speeds. In Figure 10.10-(b-c) the pitch-rate pressure sensitivity and the
pitch-rate velocity sensitivity streamlines of the same airfoil are shown but at the higher Mach
numbers Mc = 0.5 and Mc = 0.8, respectively. The main features of the resulting pressure sensitivity
and velocity sensitivity streamlines for the case M = 0.5 are similar to the low subsonic case. The case
Mc = 0.8 corresponds to the transonic where at α = 0.0 a shock exists on both the upper and lower
surfaces of the airfoil and by symmetry the shocks are located at the same location along the chord.
From the corresponding pressure sensitivity shown in Figure 10.3-(c), it can be seen that on the
upper surface, between the leading edge and some distance before the shock location, the expected
change in pressure is a more or less uniform pressure increase. On the lower surface the expected
change is a more or less uniform pressure drop. On the other hand on the upper surface, near the
shock location the expected change is a large pressure increase. This pressure increase is an
indication that if the airfoil tends to pitch nose-down, the shock on the upper surface would tend to
move towards the leading edge. The opposite phenomenon occurs on the lower surface where the
pressure sensitivity is large but negative. This result indicates that if the airfoil tends to pitch nose-
down, the lower-surface shock would tend to move towards the trailing edge. These expected
motions of the shocks are in complete agreement with the shock motion phenomena observed in the
flow solutions. As mentioned earlier, S-NISFLOW can be used to calculate the sensitivity derivatives
with respect to the pitch rate q. As example, we consider again the NACA 0012 airfoil at M = 0.8 at α
= 0.0 and the results are shown in Figure 10.11.
Figure 10.11 Pitch-Rate Pressure Sensitivity for a NACA 0012 Airfoil at Mc = 0.8 and α = 0.0
10.3.4.10 Validating
To validate the results, Table 10.1 also shows the same aerodynamic derivatives calculated using
QUADPAN and VORLAX which are two panel methods developed at Lockheed. Both methods use
potential flow formulations to estimate the local velocity and they recover the pressure from
approximations to the isentropic, compressible Bernoulli equation. The results show good
agreement between QUADPAN and S-NISFLOW at Mc = 0.1 and M = 0.5. The difference in Cq is around
321
2%-3% at M = 0.1 and increases to 4%-5% at M = 0.5. For Cmq the difference is less than 2% at Mc =
0.1 and increases slightly to less than 4% at M = 0.5. This small difference may be due to inaccuracies
of the discretization and/or to the different approaches used in simulating the pitching motion effect.
The estimates at M = 0.8 are quite different. One explanation is that VORLAX and QUADPAN cannot
model embedded shocks, whereas it is clear from the flow solution that at M = 0.8 there is an
embedded shock. The comparisons with VORLAX, also shown in Error! Reference source not found., a
re somewhat worse. At the lower Mach numbers, the VORLAX values are around 6%/9% smaller in
magnitude than those obtained using QUADPAN/S-NISFLOW. Actually, we expect that the values
generated by VORLAX to be less accurate because VORLAX implements a lifting surface panel
method, i.e. it is based on the approximation that the airfoil has zero-thickness. The values produced
by VORLAX at low Mach numbers are close to the incompressible predictions: CLq = −3π/2 and Cmq =
−π/2 obtained from thin airfoil theory using the fictitious Etkin’s camber.
In addition, the S-NISFLOW results have been validated by comparisons to finite-difference estimates
based on nonlinear flow solutions obtained using A-NISFLOW. These finite-difference estimates
correspond to column FD A-NISFLOW of Table 10.1. In subsonic cases, these finite difference
estimates are within 0.3% of the S-NISFLOW values. For the case M = 0.8 the comparison degrades
to about 3% but some of this degrading may be explained by insufficient grid-refinement in solving
the linear sensitivity equation and/or the nonlinear flow equation. Observe that the negative values
of Cmq indicates that the moment produced is always in the opposite direction of the pitch rotation
(i.e. damping-in-pitch). It also follows from the results that both Cq and Cmq tend to increase in
magnitude when the Mach number increases.
10.3.4.11 Conclusions
A mathematical model for the determination of the aerodynamic forces and sensitivity derivatives
has been developed. The mathematical model is based on the idea aerodynamically steady motions
and as part of this work a complete mathematical characterization and generalization of all possible
aerodynamically steady motions was obtained. In particular, it was proved that in these generalized
steady motions the aircraft moves in a class of helical trajectories. For the case of longitudinal
322
motions the most general aerodynamically steady motions correspond to circular trajectories where
the aircraft is moving at a constant translational velocity, at a constant angle of attack and at a
constant pitch rate. One important use of these results is the determination of time invariant
aerodynamic forces and moments. While such forces could be determined experimentally, such
experiments can be quite complex.
An alternative approach is to determine these steady aerodynamic forces using computer simulation,
and this was the approach used in this work. The idea is to determine numerically the flow around
the aircraft and obtain the aerodynamic forces and moments by an appropriate integration of the
pressure and shear forces acting along the aircraft surface. The best reference frame to determine
these flow solutions is the body-fixed reference frame, since as seen from this reference frame the
flow is steady, i.e. time-invariant. However, for general aerodynamically steady motions the body-
fixed reference frame is non-inertial and standard (inertial) CFD formulations cannot be used. The
problem is solved by developing a CFD formulation for general non-inertial reference frames. This
CFD formulation was presented and it differs from the non-inertial CFD formulations used in
turbomachinery simulations by the fact the non-inertial frame is not only allowed to rotate but also
to translate with a non-zero acceleration.
The mathematical description of the aerodynamically steady motions was incorporated into the CFD
formulation. As a consequence, the generalized Navier-Stokes equations and the generalized Euler
equations were derived. These equations allow the determination of the flow around an aircraft
moving in any aerodynamically steady motion. The formulation is valid for all ranges of Mach
numbers including transonic flow. The method was implemented for the planar case using the
generalized Euler equations. The developed computer codes can be used to obtain numerical flow
solutions for the flow around any airfoil moving in general steady motions (i.e. circular trajectories).
To the best of our knowledge this type of numerical simulations have never been done. From these
numerical solutions it is possible to determine the variation of the lift, drag and pitching moment
with respect to the pitch rate q at different Mach numbers. In particular, it is shown that for the case
of a NACA 0012 airfoil at subsonic speeds there is a linear behavior of the lift and pitching moments
with respect to q. This linear behavior would explain why linear aerodynamic models work so well
in practice. One of the advantages of the mathematical model developed here, is that the sensitivity
(stability) derivatives with respect to the six motion variables can be obtained in a straightforward
manner. In particular, the model allows the determination of rotary stability derivatives in a
decoupled way. Also, it is important to note that the model makes explicit the notion of the
“sensitivity derivatives” as “partial derivatives” of the aerodynamic forces. The sensitivity derivatives
can be obtained either by finite differences or by using the sensitivity equation method.
The sensitivity equation method was applied to the mathematical formulation in order to handle the
computation of stability derivatives. The method was implemented numerically for the case of planar
motions. Pitch-rate derivatives Cq and Cmq were computed for a NACA 0012 airfoil. The results were
compared with two panel methods (VORLAX and QUADPAN) developed at Lockheed. Good
agreement was shown for the subsonic cases where these panel methods are valid. The method
presented here is able to obtain stability derivatives in transonic flows (where approximations based
on the linearized potential flow equations do not work well). The main virtue of the sensitivity
equation approach is that the stability derivatives can be obtained from a single solution of the
nonlinear fluid mechanics (plus relatively cheap solutions of a linear partial differential equation).
The natural continuation of this work, would be the numerical implementations of the 3D CFD
formulation and of the Sensitivity formulation presented here. Based on the successful results
obtained from the two-dimensional implementation, the 3D implementation should give results of
direct practical application for the evaluation of the aerodynamic forces and stability derivatives of
arbitrary aircraft. The implementation of the three-dimensional Euler Equations seems to be
straightforward. The implementation of the Navier-Stokes equations should also be direct as long as
the turbulence models do not introduce problems.
323
n N i,p (r) ωi
X (r) = R i,p (r) Di i = 0,........., n R i,p (r) = n Eq. 10.39
i =0
Ni,p (r) ωi
i =0
where X(r) is the vector valued surface coordinate in the r-direction, Di are the control points
(forming a control polygon), ωi are weights, Ni,p(r) are the p-th degree B-Spline basis function (see
Eq. 10.45), and Ri,p(r) are known as the Rational Basis Functions satisfying:
Tiller, W., “Rational B-Splines for Curve and Surface Representation, "Computer Graphics, 1983.
562
Gaetan K.W. Kenway, Joaquim R. R. A. Martins, and Graeme J. Kennedy, “Aero structural optimization of the
563
The procedure is easily applicable to 3D (see section 7.5.1), and as an example, the common wing &
fuselage as articulated in Figure 10.13 [Kenway et al.]564. The choice for number of control points
and their locations are best determined using an inverse B-Spline interpolation of the initial data. The
algorithm yields a system of linear equations with a positive and banded coefficient matrix.
Therefore, it can be solved safely using techniques such as Gaussian elimination without pivoting.
Figure 10.12 B-Spline Approximation of NACA0012 (left) and RAE2822 (right) Airfoils
Figure 10.13 Free Form Deformation (FFD) Volume with Associated Control Points - (Courtesy of
Kenway et al.)
The procedure can be easily extended to cross-sectional configurations, when critical cross-sections
are denoted by several circular conic sections, and the intermediate surfaces have been generated
using linear interpolation. Increasing the weights would deform the circular segments to other conic
segments (elliptic, parabolic, etc.) as desired for different flight regions. In this manner, the number
of design parameters can be kept to a minimum, which is an important factor in reducing costs. An
efficient gradient-based algorithm for aerodynamic shape optimization is presented by [Hicken and
Zingg]565 where to integrate geometry parameterization and mesh movement.
564Gaetan K.W. Kenway, Joaquim R. R. A. Martins, and Graeme J. Kennedy, “Aero structural optimization of the
Common Research Model configuration”, American Institute of Aeronautics and Astronautics.
565 Jason E. Hickenand David W. Zingg, “Aerodynamic Optimization Algorithm with Integrated Geometry
Parameterization and Mesh Movement”, AIAA Journal Vol. 48, No. 2, February 2010.
325
Where n is the iteration number, Sn the vector of search direction, and γ a scalar move parameter.
The first step is to determine a feasible search direction Sn, and then perform a one-dimensional
search in this direction to reduce the objective function as much as possible, subjected to the
constraints566.
566 Sadrehaghighi, I., Smith, R.E., Tiwari, S., N., “Grid Sensitivity and Aerodynamic Optimization Of Generic Airfoils”,
10.4.2 Case Study 3 - 2D Study of Airfoil Grid Sensitivity via Direct Differentiation (DD)
The structured grid sensitivity of a generic airfoil with respect to design parameters using the NURBS
parameterization is discussed in this section. The geometry, as shown in Figure 10.15 has six pre-
specified control points. The control points are numbered counter-clockwise, starting and ending
with control points (0 and 5), assigned to the tail of the airfoil. A total of 18 design parameters (i.e.,
three design parameters per control point) available for optimization purpose. Depending on desired
accuracy and degree of freedom for optimization, the number of design parameters could be reduced
for each particular problem. For the present case, such reduction is achieved by considering fixed
weights and chord-length. Out of the remaining four control points with two degrees of freedom for
each, control points 1 and 5 have been chosen as a case study. The number of design parameters is
now reduced to four with XD = {X1; Y1; X5; Y5}T, with initial values specified in. Non-zero contribution
to the surface grid sensitivity
coefficients of these control points
are the basis functions R1,3(r) and
R5,3(r). The sensitivity gradients
are restricted only to the region
influenced by the elected control
point. This locality feature of the
NURBS parameterization makes it
a desirable tool for complex design
and optimization when only a local
perturbation of the geometry is
warranted. Similar results can be
obtained for design control point 5
where the sensitivity gradients are
restricted to the lower portion of
domain. Figure 10.16 -
(Bottom) shows C-type dual
blocks structured grid its
sensitivity with respect to NURBS
input polygon Y1.
10.4.2.1 Airfoil Grid, Flow Figure 10.15 Seven Control Point Representation of a Generic
Sensitivity, & Airfoil
Optimization
The second phase of the problem is obtaining the flow sensitivity coefficients using the previously
obtained grid sensitivity coefficients. In order to achieve this, a converged flow field solution about a
fixed design point should be obtained. The 2D, compressible, thin-layer Navier-Stokes equations are
solved for a free stream Mach number of M = 0.7, Reynolds number Re = 106, and angle of attack = 0.
The solution is implicitly advanced in time using local time step-ping as a means of promoting
convergence toward the steady-state. The residual is reduced by ten orders of magnitude. All
computations are performed on NASA Langley's Cray-2YMP mainframe with a computation cost of
0.1209 x 10-3 CPU seconds/iteration/grid point. Due to surface curvature, the flow accelerates along
the upper surface to supersonic speeds, terminated by a weak shock wave behind which it becomes
subsonic. The average relative error has been reduced by three orders of magnitude. The sensitivities
of the aerodynamic forces, such as drag and lift coefficients with respect to design parameters, are
obtained and results are presented in Table 10.2. An inspection, indicates the substantial influence
of parameters Y1 and Y5 on the aerodynamic forces acting on the surface. The optimum design is
327
10.4.2.2 Discussions
Several observations should be made at
this point. First, although control points
1 and 5 demonstrated to have
substantial influence on the design of
the airfoil, they are not the only control
points affecting the design. In fact, Figure 10.16 Sample Grid and Grid Sensitivity
control points 2 and 4 near the nose
might have greater affect due to
sensitive nature of lift and drag forces on this region. The choice of control points 1 and 5 here was
largely based on their camber like behavior. A complete design and optimization should include all
the relevant control points
(e.g., control points 1, 2, 4, and
5). For geometries with large
number of control points, in
order to contain the
computational costs within a
reasonable range, a criteria
for selecting the most
influential control points for
optimization purposes should
be established. This decision
could be based on the already
known sensitivity
coefficients, where control
points having the largest
coefficients could be chosen
as design parameters.
Secondly, the optimum airfoil
of is only valid for this Figure 10.17 Optimization Cycle History
328
particular example and design range. As a direct consequence of the nonlinear nature of governing
equations and their sensitivity coefficients, the validity of this optimum design would be restricted
to a very small range of the original design parameters.
The best estimate for this range would be the finite-difference step size used to confirm the sensitivity
coefficients (i.e., 10-3 or less). All the airfoils with the original control points within this range should
conform to the optimum design of Table 10.3, while keeping the grid and flow conditions constant,
and indicates the percentage in design
improvement. Table 10.2 shows the
Aerodynamic sensitivity coefficients. It
is evident that grid sensitivity plays a
significant role in the aerodynamic
optimization process. The algebraic
grid generation scheme presented here
is intended to demonstrate the
elements involved in obtaining the grid Table 10.3 Design improvement for an Airfoil
sensitivity from an algebraic grid
generation system. Each grid generation formulation requires considerable analytical differentiation
with respect to parameters which control the boundaries as well as the interior grid. It is implied that
aerodynamic surfaces, such as the airfoil considered here, should be parameterized in terms of design
parameters. Due to the high cost of aerodynamic optimization process, it is imperative to keep the
number of design parameters as
low as possible. Analytical
parameterization, although
facilitates this notion, has the
disadvantage of being restricted to
simple geometries. A geometric
parameterization such as NURBS,
with local sensitivity, has been
advocated for more complex
geometries. Future investigations
should include the implementation Table 10.2 Aerodynamic Sensitivity Coefficient
of present approach using larger
grid dimensions, adequate to resolve full physics of viscous flow analysis. A grid optimization
mechanism based on grid sensitivity coefficients with respect to grid parameters should be included
in the overall optimization process.
An optimized grid applied to present geometry, should increase the quality and convergence rate of
flow analysis within optimization cycles. Other directions could be establishing a link between
geometric design parameters (e.g., control points and weights) and basic physical design parameters
(e.g., camber and thickness). This would provide a consistent model throughout the analysis which
could easily be modified for optimization. Also, the effects including all the relevant control points on
the design cycles should be investigated. Another contribution would be the extension of the current
algorithm to three-dimensional space for complex applications. For three-dimensional applications,
even a geometric parameterization of a complete aerodynamic surface can require a large number of
parameters for its designation. A hybrid approach can be selected when certain sections or skeleton
parts of a surface are specified with NURBS and interpolation formulas are used for intermediate
329
surfaces.
∑ni=0 ∑m
j=1 Ni.p (u)Nj,q (v)ωi,j 𝐏i,j
𝐬(u, v) = 0 ≤ u, v ≤ 1
∑ni=0 ∑m
j=1 Ni,p (u)Nj,q (v)ωi,j
Eq. 10.43
where Pi,j are control points, ωi,j are the corresponding weights. Ni,p(u) and Nj,q (v) are p-th and q-th
567 Xingchen Zhang, “CAD-based geometry parametrization for shape optimization using Non-uniform Rational
B-splines”, A thesis submitted for the degree of Doctor of Philosophy, September 2017.
568 L. Piegl and W. Tiller. The NURBS book. Springer Science & Business Media, 2nd edition, 1997.
330
{0,
⏟ , , ,0 up+1 , , , , , ui , , , , , , ur−p−1 , , , , , , 1,
⏟, , , ,1}
p+1 p+1
{0,
⏟ , , ,0 uq+1 , , , , , ui , , , , , , us−q−1 , , , , , , 1,
⏟, , , ,1}
q+1 q+1
Eq. 10.44
where r = n + p + 1 and s = m + q + 1. Ni,p(u) and Nj,q(v) are given by the following recursive expression:
n m
Ni,p (u)Nj,q (v)ωi,j
s(u, v) = ∑ ∑ R i,j (u, v) Pi,j , R i,j (u, v) =
∑nk=0 ∑m
l=1 Ni,p (u)Nj,q (v)ωk,l
i=0 j=1
Eq. 10.46
As can be seen from the definition of NURBS, three factors have impact on the shape of a NURBS
surface given the basis functions, namely position of the control point, weight of the control point
and the knot vector. Usually in practice, only the positions of control points are used. In this study,
the approach is extended to NURBS, i.e. weights can also be perturbed by using the homogeneous
form of NURBS.
Compared to B-splines, NURBS provide weight attached to each control point, thus having more
freedoms in controlling the shape. Figure 10.20 presents a NURBS curve with 7 control points.
Different weights are set to the control points P4, which clearly show how the weight affects the
331
curve. Put specifically, a larger weight pulls the curve closer to this control point, while a smaller
weight pushes the curve away from this control point. (see [Zhang]569.
569 Xingchen Zhang, “CAD-based geometry parametrization for shape optimization using Non-uniform Rational
B-splines”, A thesis submitted for the degree of Doctor of Philosophy, September 2017.
570 D. F. Rogers. An introduction to NURBS: with historical perspective. Elsevier, 2000.
571 See Previous.
332
a link between them is required such that the mesh can be deformed along with the perturbation of
geometry automatically. This is achieved by projecting the mesh nodes onto the NURBS patches and
assigning the corresponding parametric coordinates (u , v) to each surface mesh points. In this way,
during the optimization process, the surface mesh points can be easily mapped onto the deformed
NURBS surface to obtain new spatial coordinates with the parametric coordinates (u , v). This
guarantees that the surface grid points always remain on the NURBS surface. Good performance of
mesh projection is crucial for the shape optimization problem. The point inversion algorithm is
devoted to perform mesh projection, and to find the parametric coordinates (u , v) satisfying S(u , v)
= X, where X is a mesh node, S is a NURBS surface. Note that the parametric coordinates (u , v) are
maintained throughout the optimization once found.
Figure 10.23 3D Volume Grid and Grid Sensitivty w.r.t. Wing Root Chord
572 Bischof, C. H., Jones, W. T., Mauser, A., Samareh, J. A., “Experience with Application of ADIC Automatic
Differentiation tool to the CSCMDO 3-D Volume Grid generation Code”, AIAA 96-0716.
573 Charles A. Mader, Joaquim R. R. A. Martins, Juan J. Alonso, and Edwin van der Weide, “ADjoint: An Approach
for the Rapid Development of Discrete Adjoint Solvers”, AIAA Journal Vol. 46, No. 4, April 2008.
333
In another work published by the [Necci and Ceresola]574, the goal was to demonstrate the practical
applicability of AD to industrial problems dealing with complex configurations. It was shown that
automatic differentiation could be successfully used not only for research purposes but also to carry
out a design activity on complex industrial configurations. Progress has been made from that point
on and AD capabilities have been extended to compute stability and control dynamic derivatives for
civil and fighter aircrafts. A first conceptual comparison between AD and other methods will be
provided in order to explain AD peculiarities; then a deeper study of the AD numerical technology
will be carried out. Finally some results will be given.
I 0 0 u0 f0
[A B 0] [u1 ] = [f1 ]
C D E u2 f2
Eq. 10.47
where I is the identity operator, A, B, C, D and E are some operators arising from the discretization of
the time-dependent system, f0 is the initial condition for u0, and fn is the source term for the equation
for un. The temporal propagation of information forward in time is reflected in the lower-
triangular structure of the matrix. This reflects the fact that it is possible to time step the system,
and solve for parts of the solution uu at a time. If the discrete operator were not lower-triangular, all
time steps of the solution uu would be coupled, and would have to be solved for together. Notice again
that the value at the initial time u0 is prescribed, and the value at the final time u2 is diagnostic. Now
let us take the adjoint of this system. Since the operator has been assumed to be linear, the adjoint of
this system is given by the block-structured matrix:
I A∗ C ∗ λ0 ∂J/ ∂u0
[0 B ∗ ∗ ] [λ ] = [∂J/ ∂u ]
D 1 1
∗ λ ∂J/ ∂u2
0 0 E 2
Eq. 10.48
where λ is the adjoint variable corresponding to u. Observe that the adjoint system is now upper-
triangular: the adjoint propagates information from later times to earlier times, in the opposite
sense to the propagation of the forward system. To solve this system, one would first solve for λ2,
574Carlo Necci and Nicola Ceresola (2012). Unsteady Differentiation of Aerodynamic Coefficients: Methodology
and Application, Applied Computational Fluid Dynamics, Prof. Hyoung Woo Oh (Ed.), ISBN: 978-953-51-0271-
7, In Tech, Available from: http://www.intechopen.com/books/applied-computational-fluid-
dynamics/unsteadydifferentiation-of-aerodynamic-coefficients-methodology-and-application.
334
then compute λ1, and finally λ0. Notice once more that
the prescribed-diagnostic relationship applies. In the
forward model, the initial condition is prescribed, and
the solution at the final time is diagnostic. In the adjoint
model, the solution at the final time is prescribed (a so- Adjoint
called terminal condition, rather than an initial Forward
condition), and the solution at the beginning of time is
diagnostic. This is why when the adjoint of a
continuous system is derived, the formulation always
includes the specification of a terminal condition on the Figure 10.24 Propagation of Information
adjoint system. (See Figure 10.24).
10.5.3 Classical Formulation of the Adjoint Variable (AV) Approach to Optimal Design
Following the developments in [Jameson]576, a wing, for example, is a device to produce lift by
controlling the flow, and its design can be regarded as a problem in the optimal control of the flow
equations by variation of the shape of the boundary. If the boundary shape is regarded as arbitrary
within some requirements of smoothness, then the full generality of shapes cannot be defined with a
finite number of parameters, and one must use the concept of the Frechet derivative of the cost with
respect to a function. Clearly, such a derivative cannot be determined directly by finite differences of
the design parameters because there are now an infinite number of these. Using techniques of control
theory, however, the gradient can be determined indirectly by solving an adjoint equation which has
coefficients defined by the solution of the flow equations. The cost of solving the adjoint equation is
comparable to that of solving the flow equations. Thus the gradient can be determined with roughly
the computational costs of two flow solutions, independently of the number of design variables,
Dolfin-Adjoint file:///C:/Users/Owner/Desktop/Properties%20of%20the%20adjoint%20equations.html
575
Antony Jameson, “Optimization and Engineering, 6, 33–62, 2005”, Department of Mechanical and Aerospace
576
which may be infinite if the boundary is regarded as a free surface. For flow about an airfoil or wing,
the aerodynamic properties which define the cost function, I, are functions of the flow-field variables
(w) and the physical location of the boundary which may be represented by the function (x). Then
I T I T
I = I ( w , x) or δI = δw + δx
w x
Eq. 10.49
changes in the cost function. Using control theory, the governing equations of the flow field are
introduced as a constraint in such a way that the final expression for the gradient does not require
re-evaluation of the flow field. In order to achieve this, δw must be eliminated from Eq. 10.49.
Suppose that the governing equation R which expresses the dependence of w and x within the flow
field domain D can be written as
R R
R = R ( w , x) or δR = δw + δx Eq. 10.50
w x
∂I ∂I T ∂R ∂R
δI = [ ] δw + [ ] δx − ψT ([ ] δw + [ ] δx) =
∂w ∂x ∂w ∂x
∂I T T
∂R ∂IT ∂R
{ − ψ [ ]} δw + { − ψT [ ]} δx
⏟
∂w ∂w ⏟
∂x ∂x
≡0 G
Eq. 10.51
Choosing ψ to satisfy the adjoint equation:
R T I
ψ =
w w
Eq. 10.52
The first term is eliminated so we have
I T T R
δI = Gx where G= −ψ
x x
Eq. 10.53
This equation is independent of ∂w, with the result that the gradient of I with respect to an arbitrary
number of design variables can be determined without the need for additional flow-field evaluations.
The cost involved in calculating sensitivities using the adjoint method is therefore practically
independent of the number of design variables. The cost involved in calculating sensitivities using
the adjoint method is therefore practically independent of the number of design variables. After
In mathematical optimization, the method of Lagrange multipliers is a strategy for finding the local maxima
577
having solved the governing equations, the adjoint equations are solved only once for each I.
Moreover, the cost of solution of the adjoint equations is similar to that of the solution of the
governing equations since they are of similar complexity and the partial derivative terms are easily
computed. Therefore, if the number of design variables is greater than the number of functions for
which we seek sensitivity information, the adjoint method is computationally more efficient.
Otherwise, if the number of functions to be differentiated is greater than the number of design
variables, the direct method would be a better choice578.
In the case that Eq. 10.53 is a partial differential equation, the adjoint equation Eq. 10.52 is also a
partial differential equation and appropriate boundary conditions must be determined. After making
a step in the negative gradient direction, the gradient can be recalculated and the process repeated
to follow a path of steepest descent until a minimum is reached. In order to avoid violating
constraints, such as a minimum acceptable wing thickness, the gradient may be projected into the
allowable subspace within which the constraints are satisfied. In this way one can devise procedures
which must necessarily converge at least to a local minimum, and which can be accelerated by the
use of more sophisticated descent methods such as conjugate gradient or quasi-Newton
algorithms579. There is the possibility of more than one local minimum, but in any case the method
will lead to an improvement over the original design. Furthermore, unlike the traditional inverse
algorithms, any measure of performance can be used as the cost function. Similar information can be
found at [Oliviu Sugar-Gabor]580. All components just described are integrated into a computational
x w
dI/dxFx
Figure 10.25 Aerodynamic Shape Optimization Procedure using Adjoint Variable as Sensitivity
Analysis
578 Joaquim R.R.A. Martins, Joaquim R.R.A. Martins, And James J. Reuther, “A Coupled-Adjoint Sensitivity Analysis
Method for High-Fidelity Aero-Structural Design”, Optimization and Engineering, 6, 33–62, 2005.
579 A. Jameson, L. Martinelli and N.A. Pierce, “Optimum Aerodynamic Design Using the Navier–Stokes Equations”,
framework capable of optimizing the aerodynamic shape, as shown in Figure 10.25. Generally, the
adjoint problem is about as complex as a flow solution. If the number of design variables is large,
it becomes compelling to take advantage of the cost differential between one adjoint solution
and the large number of flow field evaluations required to determine the gradient by finite
differences. To treat the multipoint problem, a composite cost function (D) is developed as a
weighted sum of cost functions at each independent design point:
where λ1 and λ2 are the relative weights of the two cost functions at different design points. The
composite gradient is then obtained by taking the same weighted sum of the gradients developed by
the above procedure for each point. Readers are encourage to consult [Reuther, et al.]581 for further
details.
581 James Reuther, Antony Jameson, Juan Jose Alonso, Mark J. Rimlinger and David Saunders, “Constrained
Multipoint Aerodynamic Shape Optimization Using an Adjoint Formulation and Parallel Computers”, The
Research Institute of Advanced Computer Science is operated by Universities Space Research Association, The
American City Building, Suite 212, Columbia, MD 21044.
582 Michael B. Giles and Niles A. Pierce, “An Introduction to the Adjoint Approach to Design”, Flow, Turbulence
discontinuities, then again a stochastic search method may be more appropriate 583.
10.5.5 Case Study 3 – Adjoint Aero-Design Optimization for Multistage Turbomachinery Blades
The adjoint method for blade design optimization will be described. The main objective is to develop
the capability of carrying out aerodynamic blading shape design optimization in a multistage
turbomachinery environment. To this end, an adjoint mixing-plane treatment has been proposed. In
the first part, the numerical elements pertinent to the present approach will be described. Attention
is paid to the exactly opposite propagation of the adjoint characteristics against the physical flow
characteristics, providing a simple and consistent guidance in the adjoint method development and
applications. The adjoint mixing-plane treatment is formulated to have the two fundamental features
of its counterpart in the physical flow domain: conservation and non-reflectiveness across the
interface. The adjoint solver is verified by comparing gradient results with a direct finite difference
method and through a 2D inverse design. The adjoint mixing-plane treatment is verified by
comparing gradient results against those by the finite difference method for a 2D compressor stage.
The redesign of the 2D compressor stage further demonstrates the validity of the adjoint mixing-
plane treatment and the benefit of using it in a multi-blade row environment. Results for pressure
contours are presented in Figure 10.26. In summary, method ingredients includes584:
Original Optimized
➢ A continuous Adjoint method has been developed based on a 3D time-marching finite volume
RANS solver. This enables the performance gradient sensitivities to be calculated very
See Previous.
583
584 Osney Thermo-Fluids Laboratory, “Adjoint Aerodynamic Design Optimization for Multi-stage
Turbomachinery Blades”, University of Oxford.
339
efficiently, particularly for situations with a large number of design variables 585.
➢ A novel adjoint ‘mixing-plane’ model has been developed. This model makes it possible to
carry out the adjoint design optimization in a multi-stage environment.
➢ The optimization approach with the above capabilities has been implemented in a parallel
mode, with which a simultaneous multi-point optimization can be conducted586.
585 D. X. Wang and L. He, “Adjoint Aerodynamic Design Optimization for Blades in Multi-Stage Turbomachines:
Part I – Methodology and Verification”, ASME Journal of Turbomachinery, Vol.132, No.2, 021011, 2010.
586 D. X. Wang, L. He, Y.S Li and R.G. Wells, “Adjoint Aerodynamic Design Optimization for Blades in Multi-Stage
Turbomachines: Part II – Verification and Application”, ASME Journal of Turbomachinery, Vol.132, 2010.
587 Mohamad Hamed Hekmat and Masoud Mirzaei, “Development of Discrete Adjoint Approach Based on the
Lattice Boltzmann Method”, Hindawi Publishing Corporation, Advances in Mechanical Engineering, Volume
2014, Article ID 230854, 16 pages-http://dx.doi.org/10.1155/2014/23085.
588 M. H. Hekmat, Aerodynamic optimization of an airfoil using adjoint equations approach [Dissertation], K. N.
optimization,” Proceedings of the 38th Aerospace Sciences Meeting and Exhibit, American Institute of
Aeronautics and Astronautics, Reno, Nev, USA, 2000, AIAA paper 2000-0667.
590 M. B. Giles and N. A. Pierce, “An introduction to the adjoint approach to design,” Flow, Turbulence and
on the adjoint solution can be applied to reduce the numerical error at a fixed computational cost.
There are two approaches to develop the adjoint equation: continuous and discrete.592 Continuous
adjoint approach utilizes the differential forms of the governing equations of the flow field and cost
function. But, in the discrete adjoint approach, the adjoint equation is directly extracted from a set of
discrete flow field equations and the cost function is gained from a numerical approximation of the
equations.
The importance of the discrete version is that the exact gradient of the discrete cost function is
obtained593. This ensures that the optimization process can converge fully. Moreover, creation of the
adjoint program is conceptually straightforward. This benefit includes the iterative solution process
since the transposed matrix has the same eigenvalues as the original linear matrix, and so the same
iterative solution method is guaranteed to converge. But, a major disadvantage of this version is the
complexity of the adjoint equation derived from the discrete flow field (Navier-Stokes) equations, so
that the complete extraction of all the discrete terms in the adjoint equation and the gradient vector
requires a lot of algebraic manipulations. In addition, the viscous flux in the NS equations further
increases the complexity of deriving the terms in viscous flows. The discrete adjoint equation
becomes very complicated when the flow field (NS) equations are discretized with higher order
schemes and using flux limiters594. Therefore the implementation cost of the discrete adjoint
equation based on the NS equations is usually higher in comparison with the LB-based equation. In
comparison with the conventional flow field (NS) equations, the LB method uses the Boltzmann
transport equation to solve the flow field. Due to the lower complexity of the LB equation, the
implementation of this method would be considerably easier. The importance of this benefit will
become more clear when a gradient-based technique is used in optimization problems involving flow
field equations (irrespective of the flow field solution) in optimization process.
The simplicity of the Boltzmann equation, as an alternative to the conventional flow field equations,
can be very helpful in facilitating the process of extracting the continuous adjoint equation (and
particularly the discrete adjoint equation). Moreover, the LB method uses an inherently immersed
boundary grid system. This unique feature of the LB method enables the modeling of complicated
geometries and the analysis of flows around them in a straightforward manner without the need for
conventional body-fitted grid generation.
So, by exploiting the LB method, the optimization of special geometries is done with lower cost595.
Furthermore, it can remove the drawbacks of non-gradient techniques, which necessitate a lot of flow
solutions (due to the selective random design variables) and consequently the application of grid
generation. Additionally, in gradient methods, there is no need to consider some grid modification
techniques. Besides, the inherent parallel processing nature of the LB method, owing to data transfer
from the nearest neighbor in terms of streaming and fully localized calculations of collision phase,
distinguishes it from the conventional methods of computational fluid dynamics596.
This capability of the LB method seems to be very effective in the analysis of complex unsteady flows
and optimization problems with a large number of design variables or constraints that demand long
computational times. As a result, it is possible to employ some optimization techniques which are not
592 S. Nadarajah, The discrete adjoint approach to aerodynamic shape optimization [Dissertation], Stanford
University, 2003.
593 S. Nadarajah, The discrete adjoint approach to aerodynamic shape optimization [Dissertation], Stanford
University, 2003.
594 S. Nadarajah, “A comparison of the discrete and continuous adjoint approach to automatic aerodynamic
optimization,” Proceedings of the 38th Aerospace Sciences Meeting and Exhibit, American Institute of
Aeronautics and Astronautics, Reno, Nev, USA, 2000, AIAA paper 2000-0667.
595 G. Pingen, Optimal design for fluidic systems: topology and shape optimization with the Lattice Boltzmann
method,” Structural and Multidisciplinary Optimization, vol. 34, no. 6, pp. 507–524, 2007.
341
efficient for traditional computational fluid dynamics due to the high computational cost (e.g., adjoint
method in problems with high number of constraints or non-gradient methods in problems with high
number of design variables).
Taking into account all the aforementioned advantages, the significance of flow optimization utilizing
the LB method becomes obvious compared to the other conventional techniques. These unique
advantages of the LB method form the main idea of its application to optimization problems.
Execution of the adjoint optimization approach by means of conventional NS-based computational
methods has been extensively studied in recent decades, for instance, through the works of
[Jameson]597-598-599, [Elliot and Peraire]600, Nadarajah]601, [Hekmat et al.]602-603, [Tonomura et al.]604,
[ Jacques and Dwight]605, [Hicken and Zingg]606. However, there are few researches in combination
with the adjoint method and the LB method. [Tekitek et al.]607, for the first time, performed the
discrete adjoint approach via the LB method. They extracted the discrete adjoint equation based on
the LB equation to identify the unknown parameters of the LB method without representing a general
framework and evaluating the details of the adjoint variables and gradients. In addition, [Pingen et
al.]608 explored the ability of the incompressible LB method in topology optimization problems. Their
major work was focused on the extraction of steady discrete adjoint equation and gradient of cost
function using the steady LB equation and its implicit solution. Moreover, they examined the
performance of mathematical algorithms (e.g., iterative and direct algorithms) to solve these
equations implicitly. Their research was restricted to optimization in steady flows and did not offer
an overall framework for implementation and computation.
In this research, for the first time, both macroscopic and microscopic discrete adjoint equations are
extracted based on the LB method using a general procedure with low implementation cost. The
proposed approach can be performed similarly for any optimization problem with the corresponding
cost function and design variables vector. Moreover, this approach is not limited to flow fields and
can be employed for steady as well as unsteady flows. The main focus is to derive the macroscopic
and microscopic discrete adjoint equations using the LB method. The validity of this development is
examined for an unsteady laminar poiseuille flow in an infinite periodic channel with uniform and
597 A. Jameson, “Aerodynamic design via control theory,” Journal of Scientific Computing, vol. 3, no. 3, 1988.
598 A. Jameson, “Computational aerodynamics for aircraft design,” Science, vol. 245, no. 4916, 1989.
599 A. Jameson, “Automatic design of transonic airfoils to reduce the shock induced pressure drag,” Proceedings
of the 31st Annual Conference on Aviation and Aeronautics, Tel Aviv, Israel, 1990.
600 J. Elliot and J. Peraire, “Aerodynamic design using unstructured meshes,” Proceedings of the 27th Fluid
Dynamics Conference, New Orleans, La, USA, 1996, AIAA paper 96-1941.
601 S. Nadarajah, The discrete adjoint approach to aerodynamic shape optimization [Dissertation], Stanford
University, 2003.
602 M. H. Hekmat, M. Mirzaei, and E. Izadpanah, “Constrained and non-constrained aerodynamic optimization
using the adjoint equations approach,” Journal of Mechanical Science and Technology, vol. 23, no. 9, pp. 2479–
2491, 2009.
603 M. H. Hekmat, M. Mirzaei, and E. Izadpanah, “Numerical investigation of adjoint method in aerodynamic
optimization,” Mechanical & Aerospace Engineering, vol. 5, no. 1, pp. 75–86, 2009.
604 O. Tonomura, M. Kano, and S. Hasebe, “Shape optimization of micro-channels using CFD and adjoint method,”
Proceedings of the 20th European Symposium on Computer Aided Process Engineering, 2010.
605 J. E.V. Peter and R. P. Dwight, “Numerical sensitivity analysis for aerodynamic optimization: a survey of
approaches,” Computers & Fluids, vol. 39, no. 3, pp. 373–391, 2010.
606 J. E. Hicken and D. W. Zingg, “Aerodynamic optimization algorithm with integrated geometry
parameterization and mesh movement,” AIAA Journal, vol. 48, no. 2, pp. 400–413, 2010.
607 M. M. Tekitek, M. Bouzidi, F. Dubois, and P. Lallemand, “Adjoint lattice Boltzmann equation for parameter
identification,” Computers & Fluids, vol. 35, no. 8-9, pp. 805–813, 2006.
608 G. Pingen, A. Evgrafov, and K. Maute, “Adjoint parameter sensitivity analysis for the hydrodynamic lattice
Boltzmann method with applications to design optimization,” Computers & Fluids, vol. 38, no. 4, 2009.
342
constant body forces, using an inverse optimization problem. Moreover, the accuracy of the cost
function gradients is evaluated by comparison to the results of the finite difference approach.
∂𝐟 1
+ 𝐜. ∇𝐟 = (𝐟 eq − 𝐟)
∂t λ
Eq. 10.55
where 𝑓 = 𝑓(𝑥, 𝑐, 𝑡) is the density distribution function, 𝑐 is the particle velocity vector, 𝑡 is time, and
the dot (⋅) denotes the scalar product. Here, the local equilibrium distribution function is denoted by
𝑓eq and 𝜆 is called the relaxation time. In the LB method, Eq. 10.55 is discretized in velocity or phase
space and assumed to be valid along specific directions. Hence, the discrete Boltzmann equation can
be written along 𝑖 direction as
∂𝐟i 1 eq
+ 𝐜i . ∇𝐟𝐢 = (𝐟i − 𝐟i )
∂t λ
Eq. 10.56
where 𝑓i is the distribution function along 𝑖 direction and 𝑓eq 𝑖 is the corresponding local equilibrium
distribution function. The left-hand side of the above equation represents the streaming or
propagation process, where the distribution function streams along the lattice link 𝑖 with velocity 𝑐𝑖
= Δ𝑥/Δ𝑡 (Δ𝑥 and Δ𝑡 are the displacement and time steps, resp.) and the right-hand side represents
the rate of change of the distribution function in the collision process. Eq. 10.56 is now discretized
with respect to space and time, using a first order upwind finite difference approximation in time and
space, resulting in the discretized LB equation:
𝐞𝐪
𝐟𝐢 (𝐜 + 𝐜𝐢 ∇t , t + ∇t) = (1 − β)𝐟𝐢 (𝐱 , t) + β𝐟𝐢 (𝐱 , t)
Eq. 10.57
Where 𝛽 = Δ𝑡/𝜆 is the inverse of the dimensionless relaxation time. To facilitate the implementation
of the LB method, it is common to choose time step Δ𝑡 and lattice spacing Δ𝑥 equal to 1 (𝑐 = 1). In an
incompressible flow, ideally, so that the divergence of the velocity is zero, the density is assumed to
be equal to the initial density; that is, 𝜌 = 𝜌0. Therefore, assuming an order 𝑂(M2) density fluctuation,
the equilibrium distribution function is obtained using Taylor series expansion of the Maxwell-
Boltzmann equilibrium distribution and ignoring 𝑂(M3) terms or higher609-610:
609 W. Gladrow and A. Dieter, Lattice-Gas Cellular Automata and Lattice Boltzmann Models: an Introduction,
Springer, Heidelberg, Germany, 2nd edition, 2000.
610 L. Sh. Luo, “The lattice-gas and lattice Boltzmann methods: past, present, and future,” Proceedings of the
(0,0) i=0
(i − 1)π (i − 1)π
(cos [ ] , sin [ ]) c i = 1,2,3,4
ci = 2 2
(i − 5)π π (i − 5)π π
√2 (cos [ + ] , sin [ + ]) c i = 5.6.7.8
{ 2 4 2 4
Eq. 10.59
For the D2Q9 model, the lattice weight parameters
are as 611
4
𝑖=0
9
1
ωi = 𝑖 = 1,2,3,4
9
1
{36 𝑖 = 5,6,7,8
Eq. 10.60
For low Mach number (M ≪ 1) or low density
fluctuation (Δ𝜌/𝜌 ≪ 1) flows, it is possible to recover
the macroscopic incompressible NS equations from
the LB equation using the Chapman-Enskog
expansion and choosing the kinematic viscosity as: Figure 10.27 Lattice arrangement of the
D2Q9 model
1 ∆𝑥 2
ν= (2 − β)
6𝛽 ∆𝑡
Eq. 10.61
Also, the macroscopic properties such as density 𝜌 and velocity vector 𝑢 = [𝑢𝑥, 𝑢𝑦]𝑇 in lattice units can
be evaluated by taking statistical moments of the distribution function, leading to the following
equations:
8 8
1
𝛒 = ∑ fi (x, t) , 𝐮 = ∑ ci fi (x, t)
ρ0
i=0 i=0
Eq. 10.62
Furthermore, the speed of sound (𝑐𝑠) is 𝑐/√3 for the D2Q9 model 612.
611 W. Gladrow and A. Dieter, Lattice-Gas Cellular Automata and Lattice Boltzmann Models: an Introduction,
Springer, Heidelberg, Germany, 2nd edition, 2000.
612 L. Sh. Luo, “The lattice-gas and lattice Boltzmann methods: past, present, and future,” Proceedings of the
Since the state variables depend on the design variables, 𝜃 = 𝜃(𝜅), a variation in 𝜅 changes 𝜃 and
finally these variations result in a change:
∂𝐈 T ∂𝐈 T
δ𝐈 = ( ) δ𝛉 + ( ) δ𝛋
∂𝛉 I ∂𝛋 II
Eq. 10.64
Here, the subscripts 𝐼 and 𝐼𝐼 are used to distinguish the contributions due to 𝛿𝜃 from 𝛿𝜅 in the cost
function variation 𝛿𝐼. In fact, changes in the design variables either directly or indirectly can lead to
changes in the cost function. The first term is the indirect effect of the design variables on the cost
function that appears as the change in the state variables, and the second term is the direct effect of
the design variables on the cost function. Suppose that the governing equation 𝑅 which expresses the
dependence of 𝜃 and 𝜅 within the flow field domain can be written as
𝐑(𝛉, 𝛋) = 0
Eq. 10.65
correlation between the state and the design variables (the state and the design variables are only
correlated by the governing equation of the flow field). Therefore, 𝛿𝜃 can be determined as
∂𝐑 ∂𝐑
δ𝐑 = ( ) δ𝛉 + ( ) δ𝛋 = 0
∂𝛉 I ∂𝛋 II
Eq. 10.66
Since variation 𝛿𝑅 is zero, it can be multiplied by a Lagrange multiplier (adjoint variable) 𝜓 and
subtracted from the variation 𝛿𝑅 with no change in the result. Thus, Eq. 10.64 can be replaced by
∂I T ∂I T ∂R ∂R
δ𝐈 = ( ) δθ + ( ) δκ − ψT [( ) δθ + ( ) δκ ]
∂θ ∂κ ∂θ ∂κ
T T
∂I ∂R ∂I ∂R
δ𝐈 = {( ) − ψT ( )} δθ + {( ) δκ − ψT ( )} δκ
∂θ ∂θ I ∂κ ∂κ II
Eq. 10.67
Now, the adjoint variable 𝜓 is determined such that the dependency of the cost function variation on
the state variables variation is eliminated. Therefore, choosing 𝜓 to satisfy the adjoint equation
∂I T ∂R ∂I T ∂R
δI = {( ) − ψ ( )} δθ + {( ) − ψT ( )} δκ
T
⏟ ∂θ ∂θ I ∂κ ∂κ II
0
Eq. 10.68
the first term is eliminated, and we find that
∂I ∂R 𝑇
( ) = ψ( )
∂θ ∂θ
Eq. 10.69
And the results in
345
∂I T ∂R
δI = ( ) − ψT ( ) δκ
∂κ ∂κ
Eq. 10.70
According to Eq. 10.70, 𝛿𝐼 is independent of 𝛿𝜃 and, as a result, for a large number of design
variables, we can compute the cost function gradient only with one flow solution in addition to one
adjoint solution in each optimization cycle. In fact, the computational cost of the adjoint equation
solution is independent of the number of the design variables.
tf tf
1 2 σ
Itx = |𝛚t − 𝛚dt | + |𝛋 |2
2 2
Eq. 10.74
where 𝑤𝑡 = [𝜌𝑡, 𝑢𝑥𝑡, 𝑢𝑦𝑡]𝑇 ∈ 𝑅3 is the flow field variables (macroscopic properties) vector at time 𝑡 and
𝑤𝑑𝑡 is the desired flow field variables vector at time 𝑡. The first term in the cost function measures the
distance of 𝑤𝑡 and 𝑤𝑑𝑡 and the second term is a regularization term with 𝜎 ≥ 0 (typically, 𝜎 ∈ [10-5 ,
10− 3], which leads to improved smoothness properties of the optimization problem for 𝜎 > 0.
346
V𝑖 = 𝑐𝑠2 (c𝑖 . 𝛋)
Eq. 10.75
is added to the right-hand side of Eq. 10.56. In the above equation, 𝜅 = [𝜅𝑥 , 𝜅𝑦]𝑇 ∈ 𝑅2 is the 2D body
forces (design variables) vector. Therefore, considering the force term Eq. 10.75, the LB equation
Eq. 10.57 can be rewritten in the following vector form:
eq
𝐅t+1 = (1 − β)𝐅t + β𝐅t + 𝐕
Eq. 10.76
where 𝐹𝑡 ≡ 𝐹(𝑥,𝑡) = [𝑓0,𝑡, . . . , 𝑓8,𝑡]𝑇 ∈ 𝑅9 is the distribution functions (microscopic properties) vector
at time 𝑡 and 𝑉 = [V0, , , ,V8]𝑇. It can be shown that the existence of the force term Eq. 10.75 in the LB
equation results in the recovery of the NS equations in the presence of the body force 𝜅 613.
𝛚t = 𝚽t (𝛚t−1 , 𝛋) for t = 1, , , , , , t f
Eq. 10.77
According to Eq. 10.76, it can be deduced that the distribution functions 𝑓𝑖 at time 𝑡 are functions of
the flow field variables in previous time step, 𝑤𝑡−1, and the design variables 𝜅. Finally, the constraint
𝑅𝑡 in the problem Eq. 10.71 based on the LB method can be expressed as
613 Z. Guo, C. Zheng, and B. Shi, “Discrete lattice effects on the forcing term in the lattice Boltzmann method,”
Physical Review E, vol. 65, no. 4, Article ID046308, pp. 1–6, 2002.
614 Mohamad Hamed Hekmat and Masoud Mirzaei, “Development of Discrete Adjoint Approach Based on the
Lattice Boltzmann Method”, Hindawi Publishing Corporation, Advances in Mechanical Engineering, Volume
2014, Article ID 230854, 16 pages-http://dx.doi.org/10.1155/2014/23085.
347
of the lattice space, and any direction of the lattice lines. Since the flow field variables vector depends
on the design variables vector implicitly, 𝑤𝑡 = 𝑤𝑡(𝜅), a variation 𝛿𝜅 in the design variables vector
causes a variation 𝛿𝑤𝑡 in the flow field variables vector, and consequently, these variations lead to a
variation 𝛿𝐼 in the cost function. Again, following closely the development in [Hekmat & Mirzaei]615,
we
T
∂𝐈tx ∂𝚽t
𝐆t = ∇𝛋 𝐈t = ∑ [ +𝛙Tt ] t = 1, , , , , , t f
∂𝛋 ∂𝛋
x
Eq. 10.79
where 𝐺𝑡 is the cost function gradient vector with respect to the design variables vector at time
Comparing the discrete LB equation with the MADA equation, it is observed that, as was expected,
the MADA equation is quite similar to the discrete LB equation and this equation has the inherent
aspects of the original LB equation (e.g., linearity, simplicity, etc.). Therefore, it is expected that the
computational cost for solving this equation is nearly equal to that of the LB equation. The main
difference between these equations is in their solving procedure. In the LB method, the flow field
variables vector is evaluated by solving the discrete LB equation at time interval [1, 𝑡𝑓] as forward in
time, while, in the adjoint method, the macroscopic adjoint variables vector is evaluated by solving
the MADA equation at the same interval as backward in time. Evaluation of the Gradient Vector. After
evaluating the present derivatives in the MADA equation, the macroscopic adjoint variables vector is
eventually achieved. Now, to evaluate the gradient vector using Eq. 10.79, first, the derivatives
𝜕𝐼𝑡𝑥/𝜕𝜅 and 𝜕Φ𝑡/𝜕𝜅 should be evaluated. Considering the cost function Eq. 10.74, we have
𝐅t = 𝚽t (𝐅t−1 , 𝛋) t = 1, , , , , , t f where
615Mohamad Hamed Hekmat and Masoud Mirzaei, “Development of Discrete Adjoint Approach Based on the
Lattice Boltzmann Method”, Hindawi Publishing Corporation, Advances in Mechanical Engineering, Volume
2014, Article ID 230854, 16 pages-http://dx.doi.org/10.1155/2014/23085.
348
𝑒𝑞
𝚽t = [Φ𝑡0 (F𝑡−1 , κ), . . . . . ., Φ𝑡8 (F𝑡−1 , κ)]𝑇 = (1 − β)𝐅 𝑡−1 + βF𝑡 + 𝐕
Eq. 10.82
And constraint 𝑅𝑡 in problem Eq. 10.71 based on the LB method can be expressed as follows:
10.6.7 Optimization Process Using the Adjoint Approach Based on the LB Method
Consider the following:
616 J. Nocedal and S. J. Wright, Numerical Optimization, Springer, New York, NY, USA, 2nd edition, 2006.
349
In this study, we have used a convenient and simplified form of the forcing term as the design variable
in the optimization problem. Different representations of the forcing term have also been proposed.
Thus, more investigations may be needed to examine the effect of such forcing term on accuracy and
stability of the LB method with regard to satisfying the incompressible NS equation. Also, for the
optimization problem, the study was limited to a design variable that is constant throughout the
space and time. Other cost functions may also be considered. Moreover, we have to note that changing
the cost function will change the adjoint equation. In addition, the continuous adjoint equations can
be derived using the continuous Boltzmann equation in space and time and the results can be
compared to our findings. Besides, due to the local nature of the LB method, it is well suited for
parallelization and GPU computing. This capability of the LB method seems to be effective in
efficiency of the adjoint methods and can be examined for (unsteady) optimization problems with a
lot of constraints, [Hekmat and Mirzaei]617. .
617Mohamad Hamed Hekmat and Masoud Mirzaei, “Development of Discrete Adjoint Approach Based on the
Lattice Boltzmann Method”, Hindawi Publishing Corporation, Advances in Mechanical Engineering, Volume
2014, Article ID 230854, 16 pages-http://dx.doi.org/10.1155/2014/23085.
351
Phase 1
field. This should give the same
solution as the original analysis at the
design point. However, the results of
performance testing and, in particular,
measured velocity and pressure
profiles in the vicinity of the blade row
may be used to test the theoretical
analysis and design. Where
differences are found, corrections to
the mathematical models or to the
hardware may be required.
In summary, the design and analysis
of a turbomachine requires an
Phase 2
accurate description of the internal
flow field. A numerical simulation of
the turbomachine, including the
various components of the
equations of motion in either exact
or approximate (empirical) form,
must be constructed. This
simulation consists of two basic
parts, a meridional plane or
through-flow solution and a blade
design method that includes an
analytical blade-to-blade flow
Phase 3
solution. Each of these solutions
affects the other and must be
developed iteratively to produce a
consistent model of the flow. Once a
model is generated, a check against
the desi, “an requirements must be
made to assure that no part of it fails
the design requirements. Once Figure 11.2 Turbomachine Design Process
complete, the model must be
compatible with mechanical limitations and manufacturing capabilities. An additional iteration
with the design constraints may be necessary to finally arrive at an acceptable engineering
solution to the design of the turbomachine. Figure 11.2 illustrated these steps require in
turbomachinery design and analysis618.
618M. V. Casey, “Computational Methods for Preliminary Design and Geometry Definition in Turbomachinery”,
Fluid Dynamics Laboratory, Sulzer Innotec AG, CH-8401, Winterthur, Switzerland.
353
Traditional
Gradient Method
Adjoint Variable
(AV)
Gradient Based
Algorithms
Conjugate Gradient
Method
Quasi-Newton
Techniques
Direct Methods
Genetic Algorithms
Turbomachinery
Evolutionary
Optimization Algorithms
Process Stochastic
Algorithms
Simulated
Annealing (SA)
Inverse Methods Set of Equations
Particle Swarm
Optimization (PSO)
619Zhihui Li, Xinqian Zheng, “Review of design optimization methods for turbomachinery aerodynamics”,
Progress in Aerospace Sciences, July 2017.
354
or adjoint formulations. These methods are efficient and can find a true optimum as long as the
objective function is differentiable and convex. However, the optimization process can sometimes
lead to a local, not necessarily a global, optimum close to the starting point. Furthermore, such
computations can easily get bogged down when many constraints are considered.
Genetic Algorithms and Evolutionary algorithms are typical stochastic optimization algorithms.
These methods are robust optimization algorithms that can cope with noisy, multimodal functions,
but are also computationally expensive in terms of the necessary number of flow analyses required
for convergence. They start with multiple points sprinkled over the entire design space and search
for true optimums based on the objective function instead of the local gradient information by using
selection, recombination, and mutation operations. Figure 11.3 shows a typical optimization
process for turbomachinery application.
620 Wu, C. H., "A General Theory of Three Dimensional Flow in Subsonic and Supersonic Turbomachines of Axial,
Radial, and Mixed Flow Types," National Advisory Committee on Aeronautics, NACA TN 2604, 1952.
621 Wislicenus, C. F., Fluid Mechanics of Turbomachinery, New York, N. Y., Dover Publications Inc., 1965.
622 Smith, L. H., Jr., "The Radial Equilibrium Equation of Turbomachinery," Trans. ASME, J. Eng. Power, 88A, 1966.
623 Novak, R. A., "Streamline Curvature Computing Procedures for Fluid Flow Problems," Trans. ASME, J. Eng.
an accurate analysis technique. [Kansan’s]625 was one first to successfully compute the velocity and
pressure distribution on the blade-to-blade plane. None of the above methods incorporates sufficient
modeling of the turbulent boundary layer flow associated with turbomachine blade rows. [Raj and
Lakshminarayana]626 conducted experiments which gave insight into the nature of the blade
boundary layers and the structure of the wake shed from the trailing edge of a blade. These data will
help in the formulation of more accurate models of this flow phenomenon. The availability of the
various through-flow and blade-to-blade solutions leads to the possibility of synthesizing a three-
dimensional model of the turbomachine flow field. The interaction of the flow on the meridional
plane and on the blade-to-blade planes becomes important in this case. The result of blade-to-blade
analysis is that forces due to the geometry of the blading may be determined. [Novak and Hearsey]627
utilized the Streamline Curvature Method in a similar manner to generate a quasi-three-dimensional
analysis. It should be stressed that the above techniques are for analysis of already designed blade
rows and do not apply to the actual determination of a blade shape, i.e., the design problem. The aim
here to address the design problem by using the Streamline Curvature Method to construct an
averaged through flow picture that satisfies the general design requirements. Then two methods, the
Mean Streamline Method and the Streamline Curvature Method, will be used to actually define
blading that generates the flow field prescribed by the through-flow analysis. Combining the
through-flow and the blade-to-blade analysis, a quasi-three-dimensional analytical
representation of the flow field is generated.
625 Katsanis, T., "Use of Arbitrary Quasi-Orthogonal for Calculating Flow Distribution on a Blade-to-Blade Surface
in a Turbomachine." NASA TN D-2809, May 1965.
626 Raj, R., and B. Lakshminarayana, "Characteristics of the Wake Behind a Cascade of Airfoils," J. Fluid Mechanics,
Turbomachinery," Part I and TI, Trans. ASIT4, J. Fluids Eng., March 1977.
628 Benini, E., “Advances in Aerodynamic Design of Gas Turbines Compressors “, University of Padova Italy.
356
➢ Objectives: Maximize Adiabatic Efficiency (η), Maximize stall margin (SM), both at nominal
condition.
➢ Boundary conditions: inlet conditions (pressure P, temperature T), flight Mach number, M,
(in the case of an aero-engine).
➢ Decision (Design) variables: number of stages, compressor and stage geometry parameters.
➢ Functional constraints: mass flow rate, m, (based on engine Power or Thrust requirements),
Pressure Ratio, PR, (from Cycle analysis), correct compressor component matching (i.e.
intake-compressor, compressor-combustor, and above all compressor-turbine) as
determined by a Matching Index (MI).
➢ Side Constraints: each decision variables must be chosen within feasible lower and upper
bounds (sides).
➢ Multi-disciplinary constraints: structural and vibrational, weight, costs, manufacturability,
accessibility, reliability.
Where n is the number of decision variable. Table 11.1 summarizes these decision making criteria.
It is worth underlying that this might not be the general formulation, as some constraints could be
turned into objectives, mainly depending on compressor’s final destination and/or manufacturer’s
strategies.
intuitively conflicting each other, while cost function is inevitably different to the one assigned to the
civil application.
Finally, a constraint based on the static thrust to be delivered by the overall engine at sea level is set
which inevitably influences compressor design. From the problem formulations given above,
remarkable importance is attributed to maximize or minimize some compressor performance
indexes or figures of merit. Therefore, before examining how to deal with such problems, it is worth
analyzing how performance can be significantly affected by the choice in the design variables. For
instance, maximizing adiabatic efficiency requires a deep understanding of the physics governing
stage losses, which have to be minimized either in design and off-design conditions. This, in turn, will
have an important impact on the choice of stage geometrical and functional variables. On the other
hand, maximizing stall margin involves acquiring a proper insight of stall physics and minimizing
stall losses. Again, such problem can be tackled if proper stage geometry is foreseen. Lastly,
minimizing compressor weight (at least from the aerodynamic point of view) implicates reducing the
number of compressor stages and increasing individual stage loading, a fact which ultimately affects
the choice of the blade shape, particularly cascade parameters. Based on the arguments above, in the
following a brief summary of basic and advanced compressor aerodynamics is given629.
Figure 11.4 Sketch of a Compressor Stage (left) and Cascade of Geometries at Mid- Span (right)
629 Benini, E., “Advances in Aerodynamic Design of Gas Turbines Compressors “, University of Padova Italy.
630 Same source – see above.
358
restrictions on weight and cost, play an important role that must be properly accounted for. In this
framework, some early choices could be revisited and subject to aerodynamic criteria checking, so
that an iterative process occurs until a satisfactory preliminary design is obtained. A second
preliminary step, distinct from the 1D procedure, is the two-dimensional design (2D), which include
both cascade and through flow models, from which a characterization of both design and off-design
multi-stage compressor performance can be carried out after some iterations, if necessary. In this
case, both direct and inverse design methodologies have been successfully applied.
Numerical optimization strategies may be of great help in this case as the models involved are
relatively simple to run on a computer. Often an optimization involves coupling a prediction tool, e.g.
a blade to blade solver and/or a through flow code, and an optimization algorithm which assists the
designer to explore the search space with the aim of obtaining the desired objectives. Finally, a fully
three-dimensional (3D) design is carried out including all the details necessary to build the
aerodynamic parts of the compressor. In this phase, some design intervention is needed to account
for the real three-dimensional, viscous flows in the stages, especially tip clearance, secondary flows
and casing treatment for stall delay. This is usually carried out using CFD models, where the running
blade is modeled in its actual deformed shape, analyzed and, if necessary optimized. While
traditional 3D analyses are aimed at evaluating and improving compressor performance of a single
stage, the recent availability of powerful computers makes the analyses of multistage compressors
an affordable task for most industries. Most advanced CFD computations include evaluation of
complex unsteady effects due to successive full-span rotor-stator interaction631.
631 Benini, E., “Advances in Aerodynamic Design of Gas Turbines Compressors “, University of Padova Italy.
359
compressor stages can be estimated (see Figure 11.6). With this respect, the designer can use
statistical indications based on typical values of admissible peripheral speeds and stage loading. This
is very useful for estimating the preliminary stage pressure ratio once the range for the other
functional parameters has been settled. Result of stage stacking consists in the flow path definition,
from which the distribution of stage
parameters along the mean radii can
be obtained. Because the stacking
procedure is intrinsically iterative, a
loop is required to satisfy all the
design objectives and constraints. As
a first check, the axial Mach
distribution along the stages must be
calculated and a value not exceeding
0.5 is tolerated for both subsonic and
transonic stages. By imposing such a
constraint, the values of stage area
passage can be derived from the
continuity equation632. Next, the
values of the hub-to-tip ratios must be
defined. To this end, it is worth
recalling that such value comes from
a trade-off between aerodynamic,
technological and economic Figure 11.6 Preliminary Estimation of Number of Stages in
Compressor
constraints. For inlet stages, values
between 0.45 and 0.66 can be
assigned, while outlet stages often are given a higher value, say from 0.8 to 0.92, in order not to
increase the exit Mach number (a condition that is detrimental for pneumatic combustor losses).
Despite its relative simplicity, mean line 1D methods based on stage-stacking techniques still play an
important role in the design of compressor stages. Recent works includes a numerical methodology
used for optimizing a stator stagger setting in a multistage axial-flow compressor environment
(seven-stage aircraft compressor), based on a stage-by-stage model to 'stack' the stages together with
a dynamic surge prediction model. The absolute inlet and exit angles of the rotor are taken as design
variables. Analytical relations between the isentropic efficiency and the flow coefficient, the work
coefficient, the flow angles and the degree of reaction of the compressor stage were obtained.
Numerical examples were provided to illustrate the effects of various parameters on the optimal
performance of the compressor stage633.
The SCM has the advantage of simulating individual streamlines, making it easier to be implemented
because properties are conserved along each streamline but is typically lower compared to the other
methods. On the other hand, MTFM uses a fixed geometrical grid, so that streamline conservation
properties cannot be applied. However, despite stream function values must be interpolated
throughout the grid, the MTFM is numerically more stable than SCM. Finally STFM is a hybrid
approach which combines advantages of accuracy of SCM with stability of MTFM. These methods
have recently been made more realistic by taking account of end-wall effects and span wise mixing
by four aerodynamic mechanisms: turbulent diffusion, turbulent convection by secondary flows,
span wise migration of airfoil boundary layer fluid and span wise convection of fluid in blade wakes.
As a result of the application of through flow codes, the compressor map in both design and off design
operation can be obtained exhibiting high accuracy635. Remarkable developments in the design
techniques have been obtained using such codes. Among others, [Massardo] described a technique
for the design optimization of an axial-flow compressor stage. The procedure allowed for
optimization of the complete radial distribution of the geometry, being the objective function
obtained using a through flow calculation (see Figure 11.7). Some examples were given of the
possibility to use the procedure both for redesign and the complete design of axial-flow compressor
stages.
direct methods are used636. These are usually very expensive procedures in terms of computational
cost such that they can be profitably used in the final stages of the design, when a good starting
solution, obtained using a combination of 1D and/or 2D methods, is already available. Moreover,
large computational resources are necessary to obtain results within reasonable industrial times.
Examples of 3D designs of both subsonic and transonic compressor blading’s are today numerous in
the open literature. For numerical optimization, searching direction was found by the steepest decent
and conjugate direction methods, and it was used to determine optimum moving distance along the
searching direction. The object of present optimization was to maximize efficiency. An optimum
stacking line was also found to design a custom-tailored 3D blade for maximum efficiency with the
other parameters fixed. The method combined a parametric geometry definition method, a powerful
blade-to-blade flow solver and an optimization technique (breeder genetic algorithm) with an
appropriate fitness function. Particular effort has been devoted to the design of the fitness function
for this application which includes non-dimensional terms related to the required performance at
design and off-design operating points. It has been found that essential aspects of the design (such as
the required flow turning, or mechanical constraints) should not be part of the fitness function, but
need to be treated as so-called "killer" criteria in the genetic algorithm. Finally, it has been found
worthwhile to examine the effect of the weighting factors of the fitness function to identify how these
affect the performance of the sections637.
A multi-objective design optimization method for 3D compressor rotor blades was developed by
[Benini, 2004], where the optimization problem was to maximize the isentropic efficiency of the rotor
and to maximize its pressure ratio at the design point, using a constraint on the mass flow rate. Direct
objective function calculation was performed iteratively using the 3D Navier-Stokes equations and a
multi-objective evolutionary algorithm featuring a special genetic diversity preserving method was
used for handling the optimization problem. In this work, blade geometry was parameterized using
three profiles along the span (hub, mid span and tip profiles), each of which was described by camber
and thickness distributions, both defined using Bezier polynomials. The blade surface was then
obtained by interpolating profile coordinates in the span direction using spline curves. By specifying
a proper value of the tangential coordinate of the first mid span and the tip profiles control point with
respect to the hub profile, the effect of blade lean was achieved. Results of tip profiles control point
with respect to the hub profile and the effect of blade lean was achieved. Performance enhancement
severe shock losses (Figure 11.7). Computational time was enormous, involving about 2000 CPU
hours on a 4-processor machine.
Two generic transonic designs have been presented, one of which referred to compressor rotor,
where loss reductions in the region of 20 per cent have been achieved by imposing a proper target
surface Mach number which resulted in a modified blade shape. Figure 11.8 comparison of blade
loading distributions of an original supersonic blade, a new design (prescribed by inverse mode), and
a reference blade (R2-56 blade) for a given pressure ratio (left); comparison of passage Mach number
distributions at 95% span. Results
showed that an optimum combination
of pressure-loading tailoring with
surface objective can lead to a
minimization of the amount of sucked
flow required for a net performance
improvement at design and off-design
operations. By prescribing a desired
loading distribution over the blade the
placement of the passage shock in the
new design was about the same as the
original blade. However, the passage
shock was weakened in the tip region
where the relative Mach number is
high.
638 Benini, E.,” Three-Dimensional Multi-Objective Design Optimization of a Transonic Compressor Rotor”. Journal
of Propulsion and Power, Vol. 20, No. 3 (May/June), pp.559-565, ISSN: 0748-4658-2004.
363
11.4 Case Study 2 – Turbine Airfoil Optimization using Quasi 3D Analysis Codes
Turbine airfoil design has long been a domain of expert designers who use their knowledge and
experience along with analysis codes to make design decisions. The turbine aerodynamic design is a
three-step process that is pitch line analysis, through-flow analysis, and blade-to-blade analysis, as
depicted in Figure 11.9. In the pitch line analysis, flow equations are solved at the blade pitch, and
a free vortex assumption is used to get flow parameters at the hub and the tip. Using this analysis the
flow path of the turbine is optimized, and number of stages, work distribution across stages, stage
reaction, and number of airfoils in each blade row are determined. In the through-flow analysis, the
calculation is carried out on a series of meridional planes where the flow is assumed to be
axisymmetric and the boundary conditions of each stage are determined. The axisymmetric through-
flow method allows for variation in flow parameters in the radial direction without using the free
vortex assumption and accounts for interactions between multiple stages. In the blade-to-blade
analysis, airfoil profiles are designed on quasi-3D surfaces using a computational fluid dynamics
code.
The design of airfoil profiles involves slicing the blade on quasi-3D surfaces, designing each section
separately, and stacking the sections together to obtain a sooth radial geometry. The objective of
airfoil design is to define the airfoil shape so as to ensure structural integrity and minimize losses.
The primary sources of losses in an airfoil are profile loss, shock loss, secondary flow loss, tip
clearance loss, and end-wall loss. Profile loss is associated with boundary layer growth over the blade
profile causing viscous and turbulent dissipation. This also includes loss due to boundary layer
separation because of conditions such as extreme angles of incidence and high inlet Mach number.
Shock losses arise due to viscous dissipation within the shock wave which results in increase in static
pressure and subsequent thickening of the boundary layer, which may lead to flow separation
downstream of the shock. End-wall loss is associated with boundary layer growth on the inner and
outer walls on the annulus. Secondary flow losses arise from flows, which are present when a wall
boundary layer is turned through an angle by an adjacent curved surface. Tip clearance loss is caused
by leakage flows in the tip clearance region of the rotor blade, where the leaked flow fails to
contribute to the work output and also interacts with the end-wall boundary layer. The objective of
364
the design is to create the most efficient airfoil by minimizing these losses. This often requires
trading-off one loss versus another such that the overall loss is minimized.
To compute all these losses a 3D viscous analysis is required; however, due to the computational load
of such a code, a quasi-3D analysis code is often used in the design process. Thus the impact of the
blade geometry on 3D losses cannot be determined and only 2D losses can be minimized, that is,
profile and shock losses. A viscous quasi-3D analysis though less computationally intense is still too
expensive for use in design optimization, and an inviscid quasi-3D code is used instead. Consequently,
viscous losses are not computed from the analysis code and airfoil performance is gauged by the
characteristics of the Mach number distribution on the blade surface. The most practical formulation
for low-speed turbine airfoil designs still remains the direct optimization formulation based on 2D
inviscid blade-to-blade solvers. This work automates the direct design process as described in the
next section639.
on the mean line of the airfoil as shown in Figure 11.10 (B). In this representation, the mean line
and the thickness distribution can be varied independently, and good control of the thickness
distribution is obtained. Here we discusses optimization of low-pressure turbine blades with the
following parameters in Table 11.2.
• The curvature change on the unguided portion of the airfoil (unguided turning)
• The difference between the blades mean line angle and the flow angle at the trailing edge
(over turning), and
366
• The difference between the inlet angle and the metal angle at the inlet (Δ1).
640 Goel, Sanjay,” Turbine Airfoil Optimization Using Quasi-3D Analysis Codes”, University at Albany, USA.
367
and heuristic search. In the current investigation, the BFGS variable metric method implemented in
an optimization code ADS was used. A one-dimensional search technique was used in which the
search was bounded followed by use of polynomial interpolation.
641 Goel, Sanjay,” Turbine Airfoil Optimization Using Quasi-3D Analysis Codes”, University at Albany, USA.
368
respectively. The vertical lines represent the edges of the blade rows and the location of the frame.
The turbine has six stages, each stage composed of two blade rows. The first blade row consists of
nozzles and the second blade row consists of buckets. The stages are numbered from 1 to 6 in the
Figure 11.12.
In the current investigation, stage 5
nozzle was designed using sections
from five streamlines equally spaced
along the blade span (hub to tip).
Figure 11.13 Shows the approximate
locations of the streamlines for an
airfoil in which the first and the last
streamlines are shown at the hub and
tip. In reality however streamlines at
5% and 95% span were used instead of
streamlines directly on the hub and tip
because Mach number distributions
very close to the end walls are distorted
by the end wall effects and not Figure 11.13 Schematics of an airfoil showing
representative of the flow away from stream lines along the radial direction
the walls. The starting solution for the
test case was obtained by estimating the airfoil shape based on
shapes of similar airfoils designed in the past. All the Mach number
and airfoil geometry plots use the same reference radial and axial
locations as shown in Figure 11.14. To ensure slope and
curvature smoothness of the geometry, second- order polynomials
were used to represent the radial distribution of geometry
parameters.
Thus there are three design variables for each geometry
parameter, that is, C0, C1, and C2. These are the coefficients of the
2nd polynomial representing the geometry parameter. The
efficient of the fit match well with the starting design since the
design is based on a previously designed airfoil. Subsequently the
smoothness is maintained since the parameters are not changed Figure 11.14 3D model of
directly but rather the coefficients of the polynomials are varied. an airfoil showing the
The geometry parameters which describe the low-pressure passage between adjacent
turbine airfoil geometry are Stagger, Tmaxx, C1, C2, C3, Ratu, Ratl, airfoils
Pcttle, ti, and E. These geometry parameters are varied within limits typically prescribed in design
practice and on the basis of prior experience and manufacturing limitations. The limits for these
parameters are described along with the results for each specific test case.
• A numerical metric for emulating designer judgment in evaluation of airfoil Mach number
Distribution.
• An optimization formulation for design of airfoil sections.
369
Designer heuristics are computed using curve fits and error norms. A set of penalty functions has
been defined which allows for flexible constraint boundaries and influence constrained variables
even within constraint limits. In the new approach multiple two-dimensional sections of the airfoil
are designed with constraints on radial smoothness using polynomial fits on the parametric
geometry variables in the radial direction (Figure 11.14)642. Airfoil design is a labor intensive,
repetitive, and cumbersome task for the designers and is a bottleneck for both the design cycle and
rapid generation of inputs for complex multistage analyses. Automating the design process
significantly cuts down the design cycle time and facilitates the task of running multistage analysis
by rapidly generating airfoil geometries. While designing an airfoil, it is hard to establish the
existence of a unique optimum. Multiple evaluation criteria which are weighted together to define
the objective function and the relative importance of these are determined based on designer
experience. Furthermore, the analysis codes are not exact, and even with precisely defined quality
metrics, a significant margin of error remains. In manual design the evaluation criteria are implicitly
considered by the designer, with weighting factors based on past experience and individual biases.
Subjectivity is introduced into the design process since the evaluation criteria for the design are
partially based on heuristics abstracted from designer experiences. Thus in order to completely
understand the results of airfoil optimization, an evaluation of the qualitative changes to the design
is essential after the optimization is completed. Over time as the metrics to evaluate airfoil design
become more acceptable, a standard metric will emerge, till such time designers will need to tinker
with the weights to suit their own preferences 643
642 Goel, Sanjay,” Turbine Airfoil Optimization Using Quasi-3D Analysis Codes”, University at Albany, USA.
643 Goel, Sanjay,” Turbine Airfoil Optimization Using Quasi-3D Analysis Codes”, University at Albany, USA.
644 A. Rubinoa, M. Pini, P.Colonna, T. Albring, S. Nimmagadda, T. Economon, J. Alonso, “ Adjoint-based fluid
dynamic design optimization in quasi-periodic unsteady flow problems using a harmonic balance method”,
https://doi.org/10.1016/j.jcp.2018.06.023.
645 P. Stadtmüller, L. Fottner, “A test case for the numerical investigation of wake passing effects on a highly loaded
LP turbine cascade blade”, ASME Turbo Expo 2001, Power for Land, Sea, and Air, ASME, 2001.
646 F. Menter, M. Kuntz, R. Langtry, “Ten years of industrial experience with the SST turbulence model, in:
The terminology OptC1 and OptC2 is used to refer to the first and the second shape optimization
problem, respectively. Here, we investigate the Non-Uniform case (optC1). Readers are encourage to
consult the [Rubino et al.]648 for details on the second case.
647A. Prasad, “Calculation of the mixed-out state in turbomachine flows”, J. Turbomachinery. 127(3) (2004).
648A. Rubinoa, M. Pini, P.Colonna, T. Albring, S. Nimmagadda, T. Economon, J. Alonso, “ Adjoint-based fluid
dynamic design optimization in quasi-periodic unsteady flow problems using a harmonic balance method”,
https://doi.org/10.1016/j.jcp.2018.06.023.
371
= [0, ± ω0], ωN5 = [0, ± ω0, ± 2ω0] and ωN7 = [0, ± ω0 , ± 2ω0, ± 3ω0]. The resolved frequencies are,
therefore, multiples of the fundamental blade passing frequency only. The total pressure loss
coefficient, defined in Eq. 11.2, as function of time, and is obtained by spectral interpolation of the
harmonic balance result. The RMSE of the total pressure loss coefficient for the solution obtained
with 5 time instances is equal to 0.010. The harmonic balance solution obtained with 5 time instances
is about 9x faster than the time-accurate solution calculated over a total simulation time of five
periods, which includes the initial transient before reaching convergence to a periodic flow field
solution. 5 time instances are used for shape optimization, as a trade-off between accuracy and
computational cost. In Figure 11.18-a , b, and c, the Mach number contours from the HB simulation
are reported for 3 different time instances with the simulation period given by T = 1/f1. The results
show the bar wakes entering the cascade and a separation area occurring at about x/l = 0.7.
11.5.2 Optimization Problem and Results
The shape optimization problem of the cascade configuration is considered. It can be expressed as
⏟
Minimize ζP (𝐔𝐧 , 𝐗 𝐧 , 𝛂)
α
Subject to α out < α out,0 + 4 and δt = δt0
𝐔n < Gn n = 1 , 2, , , , , , N and 𝐗 𝑛 = M𝑛
Eq. 11.4
where the time-averaged total pressure loss coefficient ζP, obtained from Eq. 11.2, is selected as
objective function. Inequality constraints on the absolute exit flow angle (α out) and trailing edge
thickness (δt) are imposed. The optimization is performed using an ensemble of 16 geometrical
design parameters αbased on a free-form deformation (FFD) approach649. The gradients of the
Figure 11.16 Shape Optimization History of the Total Pressure Loss Coefficient and Comparison
Between Baseline and Optimized Blade Profile (OptC1)
649J. Samareh, Aerodynamic shape optimization based on free-form deformation, in: 10th AIAA/ISSMO
Multidisciplinary Analysis and Optimization Conference, 2004.
372
Figure 11.18 Mach Number Contours calculated at Three Different Time Instances with the HB
Method, Based on the OptC1 test case, for both the Baseline (a), (b), (c) and the Optimized (d), (e), (f)
Blade Profile
650A. Rubinoa, M. Pini, P.Colonna, T. Albring, S. Nimmagadda, T. Economon, J. Alonso, “ Adjoint-based fluid
dynamic design optimization in quasi-periodic unsteady flow problems using a harmonic balance method”,
https://doi.org/10.1016/j.jcp.2018.06.023.
373
evaluations, although satisfying the constraint requires more evaluations. Figure 11.17-b highlights
that the performance of the optimized blade is significantly improved, as the total pressure loss
coefficient is approximately 38%lower, while the constraint on the absolute outlet flow angle is
satisfied. The separation area, as seen in Figure 11.18, is considerably smaller with the optimized
blade shape. The unsteady optimization leads to a decrease in the peak of the total pressure loss
coefficient of 44% and a reduction of 54% of the signal amplitude in Figure 11.17. Furthermore, the
objective function spectrum obtained from a URANS simulation of the optimized blade (Figure
11.17) does not contain additional frequencies when compared with the baseline configuration.
11.6 Case Study 4 - Using Shock Control Bumps To Improve Transonic Compressor
Blade Performance651
Shock control bumps can help to delay and weaken shocks, reducing loss generation and shock-
induced separation and delaying stall inception for transonic turbomachinery components, as
described by [John et al.]652. The use of shock control bumps on turbomachinery blades is
investigated here for the first time using 3D analysis. The aerodynamic optimization of a modern
research fan blade and a highly loaded compressor blade are carried out using shock control bumps
to improve their performance. Both the efficiency and stall margin of transonic fan and compressor
blades may be increased through the addition of shock control bumps to the geometry. It is shown
how shock induced separation can be delayed and reduced for both cases. A significant efficiency
improvement is shown for the compressor blade across its characteristic, and the stall margin of the
Figure 11.19 Tip Speed) With Pronounced Negative Camber. From Prince – Courtesy of [Prince]
651 Alistair John, Ning Qin, and Shahrokh Shahpar, “Using Shock Control Bumps To Improve Transonic
Fan/Compressor Blade Performance”, GT2018-77065.
652 See Previous.
374
fan blade is increased by designing bumps that reduce shock-induced separation near to stall. Adjoint
surface sensitivities are used to highlight the critical regions of the blade geometries, and it is shown
how adding bumps in these regions improves blade performance. Finally, the performance of the
optimized geometries at conditions away from where they are designed is analyzed in detail.
653 Ginder, R., and Calvert, W., 1987. “The design of an advanced civil fan rotor”. Journal of turbomachinery,
109(3), pp. 340–345.
654 Cumpsty, N. A., 1989. Compressor aerodynamics. Longman Scientific & Technical.
655 Prince, D. C., 1980. “Three-dimensional shock structures for transonic/supersonic compressor rotors”. Journal
Figure 11.20 Schematic Of Shock Structures (A) Datum, (B) S-Shaped Design. Courtesy of [John et
al.]
658 Ashill, P., and Fulker, J., 1992. “92-01-022 a novel technique for controlling shock strength of laminar-flow
aerofoil sections”. DGLR BERICHT, pp. 175–175.
659 Drela, M., and Giles, M. B., 1987. “Viscous-inviscid analysis of transonic and low Reynolds number airfoils”.
reduction”. Proceedings of the Institution of Mechanical Engineers, Part G: Journal of Aerospace Engineering,
222(5), pp. 619–629.
376
complexity. They showed that 3D bump configurations were more robust than 2D bump designs
(where a 2D bump is extended continuously along the span).
The only use of a shock control bump on turbomachinery blades found in the literature is by
[Mazaheri and Khatibirad]663, who tested a 2D shock control bump on a (mid-span) section of the
NASA rotor 67 geometry. They added a bump modelled using the Hicks-Henne function. It was shown
how the interaction of the bump with the original wave structure resulted in a more desirable
pressure gradient, with a weaker compression wave fan and a more isentropic compression field.
The bump design was optimized and was shown to reduce the separation area at an off-design
condition. They describe how this may have the potential to improve the stall properties of the blade
section. Two optimizations were carried out, one at the design condition and another at 4% higher
rotational speed. Optimal bumps were produced for each condition, with an increase in efficiency of
0.67% for the on-design case and 2.9% in the off design case reported. The optimized geometry for
the design condition is shown in Figure 11.21.
The work by [Mazaheri & Khatibirad]664 demonstrated the benefit that bumps may provide at both
on and off-design conditions, and their potential to improve stall margin. The simplified 2D analysis
lacks accuracy
however as the
complex
behavior of
radial and
separated flow
cannot be
predicted. For a
thorough
understanding
of the potential
for the use of
shock control
bumps, 3D
analysis and the
design of 3D
Figure 11.21 Datum Geometry and Optimized Shock Control Bumps on The Mid-
bumps is
Section of Nasa Rotor 67- From Mazaheri et al..
needed to truly
assess their effect.
663 Mazaheri, K., and Khatibirad, S., 2017. “Using a shock control bump to improve the performance of an axial
compressor blade section”. Shock Waves, 27(2), pp. 299–312.
664 See Previous.
377
on the original experimental values) is specified. The inlet turbulence intensity is 1%. At the outlet, a
value for circumferentially mixed-out and radially mean-mass capacity (non-dimensional mass flow)
is used. Periodic boundaries are used to represent full annulus flow. Stationary walls are treated as
adiabatic viscous walls and the rotational speed of the non-stationary portions of the domain is
1800.01rads1, as specified in the experiment. Rolls-Royce CFD solver Hydra is used for all of the
simulations presented here, using the Spalart-Allmaras turbulence model (fully turbulent). The 4.27
M cell mesh is generated by PADRAM, has y+ of the order of one on all surfaces with 30 cells in the
tip gap.
Figure 11.22 The R37 CFD Domain Used – Courtesy of [John et al.]
11.6.5 Validation
As previously alluded to, many studies have struggled when matching simulations of Rotor 37 to the
experiment. A wide range of work has been undertaken to investigate the discrepancy found between
simulation and experiment, with the primary work being the 1994 ASME/IGTI blind test case study
in which a range of codes were used to simulate the rotor, with no knowledge of the experimental
values. A large variation was seen between the different predictions, prompting analysis by
[Denton]665. Recent work has also been carried out by [Chima]666 and [Hah]667. The differences are
usually attributed to uncertainty in the experimental measurements, the lack of real geometry in the
simulations (e.g. the upstream hub cavity is usually missing) and also the difficulty in fully resolving
665 Denton, J., 1997. “Lessons from rotor 37”. Journal of Thermal Science, 6(1), pp. 1–13.
666 Chima, R., 2009. “Swift code assessment for two similar transonic compressors”. In 47th AIAA Aerospace
Sciences Meeting including The New Horizons Forum and Aerospace Exposition, p. 1058.
667 Hah, C., 2009. “Large eddy simulation of transonic flow field in NASA rotor 37”. 47th AIAA Aerospace
Sciences Meeting including The New Horizons Forum and Aerospace Exposition, p. 1061.
378
the complex flows. The pressure ratio agreement is reasonable across the characteristic, but the
efficiency prediction is about 2% below the experimental value at the design point (98% choke). This
matches the trend of previous results, where the better the PR prediction, the worse the efficiency
match. This ’trade-off’ has been seen in a range of previous simulations. Figure 11.23 gives the
radial profiles of total PR and efficiency at 98% of simulated choke compared to the experimental
values at 98% experimental choke. The radial trends have been captured fairly well, although there
is an offset from the experiment for both. The choke mass flow found in the simulations was
20.91kg/s, matching quite closely the experimental of 20.93kg/s.
11.6.7 Validation
Due to experimental data for this geometry not being available, simulation validation was carried out
using a similar fan blade geometry that has experimental data available. The related blade has very
similar performance parameters, and the simulation set up is identical. The results are given here. A
379
comparison of the simulations of this related blade against experimental data can be seen in 668. Both
the pressure ratio and efficiency curves match the experimental data well, though there is a slight
offset to the overall values and stall margin. The radial curves show good comparison to experimental
data, although the radial variation in efficiency is under predicted compared to the experiment.
Overall, the simulation compares well, lying within 1% across the range of flow rates.
Figure 11.24 (a) 3d Separation (Orange) On The R37 Geometry (Flow Right To Left), (b) Rel. Mach
No. Contour At 60% Span – Courtesy of [John et al.]
668 Alistair John, Ning Qin, and Shahrokh Shahpar, “Using Shock Control Bumps To Improve Transonic
Fan/Compressor Blade Performance”, GT2018-77065.
669 Alistair John, Ning Qin, and Shahrokh Shahpar, “Using Shock Control Bumps To Improve Transonic
Figure 11.25 Shock Region Flow Features For RR-Fan At Points A) A, B) B, C) C, D) D, E) E, F) F. Flow
Direction – Courtesy of [John et al.]
RIGHT TO LEFT.
670 Duta, M. C., Shahpar, S., and Giles, M. B., 2007. “Turbomachinery design optimization using automatic
differentiated adjoint code”. ASME Turbo Expo 2007: Power for Land, Sea, and Air, American Society of
Mechanical Engineers, pp. 1435–1444.
381
efficiency, and therefore if shock control bumps are applied here some benefit should be found. For
complete details, please consult the [John et al.]671
671Alistair John, Ning Qin, and Shahrokh Shahpar, “Using Shock Control Bumps To Improve Transonic
Fan/Compressor Blade Performance”, GT2018-77065.
382
During this work, a study was carried out (not detailed here for brevity) to compare the benefit of
using individual bumps (where a series of discrete bumps is added to the datum geometry in the
radial direction) with a continuous bump (note continuous bumps are still ’3D’ and their shape,
position and amplitude can vary in the radial direction). It was concluded that, for these cases, the
individual bumps needed to have greater amplitude than the continuous bumps to offer the same
benefit, leading to increased separation downstream of the bump position. The continuous bumps
tested offered greater benefit, and therefore only results using the ’continuous’ bump geometry
approach are presented here.
width localized between 40 and 60% span. This makes sense as the strongest shock location, largest
separation and maximum adjoint sensitivity occur around mid-span for Rotor 37, and therefore
greater shock control is needed in this region. The resulting variation from hub to tip of the geometry
demonstrates the benefit provided by optimizing the geometry. Without optimization it would be
difficult to manually specify the bump position, width, amplitude and asymmetry, which would result
in reduced benefit.
Figure 11.30 Datum (Left) And Optimized (Right) Rotor 37 Static Pressure Contours. Flow Direction
Right To Left – Courtesy of [John et al.]
Figure 11.32 Datum (Left) And Optimized (Right) Rotor 37 Separated Flow Contours (Orange). Flow
Direction Right To Left.
11.6.15 Conclusion
This work has demonstrated how
shock control bumps can be used Figure 11.33 Lift Plots For The Datum and Optimized
to improve the performance of Geometries at 60% Span – Courtesy of [John et al.]
transonic fan/compressor blades.
Blade geometries that
incorporate shock control
bumps have the ability to
reduce shock loss and
reduce/eliminate shock-
induced separation and
increase both efficiency
and stall margin. Shock
control bumps have the
benefit that only small
modifications to the blade
geometry are required to
achieve these
improvements, compared
to the large changes
required by blade designs
that make use of negative
camber or similar shock
control approaches. It has
been demonstrated that
both the efficiency and Figure 11.34 R37 Optimized Characteristic Vs Datum – Courtesy of
pressure ratio of a highly [John et al.]
loaded compressor blade
can be increased across a
range of flow rates by delaying the shock and significantly reducing the separation and wake. For a
modern fan blade the optimized bump design eliminated the majority of separation, reduced the
thickness of the wake and extended the stall margin. For further and complete info, please consult
386
And many others. An extensive evaluation of select frameworks has been performed at NASA Langley
Research Center. The optimization problem is often divided or decomposed into separate sub-
optimizations managed by an overall optimizer that strives to minimize the global objectives.
Examples of these techniques are Concurrent Optimization675, Collaborative Optimization676, and
Bi-Level System Synthesis 677. Simpler optimization techniques, such as All-In-One optimization (in
which all design variables are varied simultaneously) and sequential disciplinary optimization (in
which each discipline is optimized sequentially) can lead to sub-optimal design and lack of
robustness.
Systems", Proceedings, 2nd NASA/USAF Symposium on Recent Advances in Multidisciplinary Analysis and
Optimization, Hampton, Virginia, 1988.
676 Braun, R.D., "Collaborative Optimization: An Architecture for Large-Scale Distributed Design", Ph.D. thesis,
increase at a super linear rate, the computational cost of MDO is usually much higher than the sum of
the costs of the single-discipline optimizations for the disciplines represented in the MDO.
Additionally, even if each discipline employs linear analysis methods, the combined system may
require costly nonlinear analysis. For example, linear aerodynamics may be used to predict pressure
distribution a wing, and linear structure analysis may be then used to predict is placement so waver,
but the dependent pressure displacements may not be linear. Finally, for each disciplinary
optimization we may be able to use as single-objective function, but for the MDO problem we may
need to have multiple objective with an attendant increasing cost of optimizations678.
Therefore, the choice of approach should be made only after careful consideration of all the factors
pertaining to the problem at handClarification of Some Terminology
Although the terms Multi-Physics and Multi-Disciplinary are used interchangeably, but there are
distinctive different. While Multi-Physics refers to the cases when one solver is used in different
physics, multi-disciplinary is referred to the cases when two or more solver is used in different
physics, and the information in shared coupling data from separate analysis packages. In essence,
difference is the way data is obtained for optimization process.
Symposium on Multidisciplinary Analysis and Optimization. Long Beach, California, USA, 2000.
389
Figure 12.1 a) Single-level optimization method with integrated analyses (first generation MDO
methods). b) Single-level optimization method with distributed analyses (second generation MDO
methods). c) Multi-level optimization method (third generation MDO methods)
1. The first category includes problems with two or three interacting disciplines where a single
analyst can acquire all the required expertise (multi-physics). This may lead to MDO where
design variables in several disciplines have to be obtained simultaneously to ensure efficient
design. Most of the papers in this category represent a single group of researchers or
practitioners working with a single computer program, so that organizational challenges
were minimized. Because of this, it is easier for researchers working on problems in this
category to deal with some of the issues of complexity of MDO problems, such as the need for
multi-objective optimization.
2. The second category includes works where the MDO of an entire system is carried out at the
conceptual level by employing simple analysis tools. For aircraft design, the ACSYNT and
FLOPS programs represent this level of MDO application. Because of the simplicity of the
analysis tools, it is possible to integrate the various disciplinary analyses in a single, usually
modular, computer program and avoid large computational burdens. As the design process
moves on, the level of analysis complexity employed at the conceptual design level increases
uniformly throughout or selectively. Therefore, some of these codes are beginning to face
some of the organizational challenges encountered when MDO is practiced at a more
advanced stage of design process.
3. The third category of MDO research includes works that focus on the organizational and
computational challenges and develop techniques that help address these challenges. These
include decomposition methods and global sensitivity techniques that permit overall system
optimization to proceed with minimum changes to disciplinary codes. These also include the
development of tools that facilitate efficient organization of modules or that help with
organization of data transfer. Finally, approximation techniques are extensively used to
address the computational burden challenge, but they often also help with the organizational
challenge. This, accordingly includes sections on Mathematical Modeling, Design-oriented
Analysis, Approximation Concepts, Optimization Procedures, System Sensitivity,
Decomposition, and Human Interface.
Joaquim R. R. A. Martins] 682 of UM, and of course the one envisioned by Wikipedia. They are listed
and characterized in following section.
682 Joaquim R. R. A. Martins and Andrew B. Lambe, ”Multidisciplinary Design Optimization: Survey of
Architectures”, AIAA Journal, September 2013.
683 Jaroslaw Sobieszczanski-Sobieski, “Multidisciplinary an Emerging New Discipline Design Optimization
example, in many aircraft companies, the structural loads are calculated by a simpler aerodynamic
model than the one used for calculating aerodynamic drag. Finally, models of various levels of
complexity may be used for the same response calculation in an approximation procedure or fast re-
analysis described in the next two sections.
Recent Aerospace industry emphasis on economics will, undoubtedly, spawn generation of a new
category of mathematical models to simulate man-made phenomena of manufacturing and aerospace
vehicle operation with requisite support and maintenance. These models will share at least some of
their input variables with those used in the product design to account for the vehicle physics. This
will enable one to build a system mathematical model encompassing all the principal phases of the
product life cycle: desired formulation of product design, manufacturing, and operation. Based on
such an extended model of a system, it will be possible to optimize the entire life cycle for a variety
of economic objectives, e.g., minimum cost or a maximum return on investment. There are numerous
references that bring the life cycle issues into the MDO domain and discussion on the role of MDO in
the Integrated Product and Process Development (IPPD), also known as Concurrent Engineering (CE).
Mathematical modeling of an aerospace vehicle critically depends on an efficient and flexible
description of geometry.
objective function and constraints. Second, often the different disciplinary analyses are executed on
different machines, possibly at different sites, and communication with a central DSS program may
become unwieldy. Third, some disciplines may produce noisy or jagged response as a function of the
design variables685.
If we do not use a smooth approximation to the response in this discipline we will have to degrade
the DSS to less efficient non-gradient methods. For all of the above reasons, most optimizations of
complex engineering systems couple a DSS to easy-to-calculate approximations of the objective
function and/or constraints. The optimum of the approximate problem is found and then the
approximation is updated by the full analysis executed at that optimum and the process repeated.
This process of sequential approximate optimization is popular also in single-discipline
optimization, but its use is more critical in MDO as the principal cost control measure. Most often the
approximations used in engineering system optimization are local approximations based on the
derivatives. Linear and quadratic approximations are frequently used, and occasionally intermediate
variables or intermediate response quantities686 are used to improve the accuracy of the
approximation. A procedure for updating the sensitivity derivatives in a sequence of approximations
using the past data was formulated for a general case in Scotti687. Global approximations have also
been extensively used in MDO. Simpler analysis procedures can be viewed as global approximations
when they are used temporarily during the optimization process, with more accurate procedures
employed periodically during the process. For example, Unger et al. 688 developed a procedure where
both the simpler and more sophisticated models are used simultaneously during the optimization
procedure. The sophisticated model provides a scale factor for correcting the simpler model where
the scale factor is updated periodically during the design process. Another global approximation
approach that is particularly suitable for MDO is the response-surface technique. This technique
replaces the objective and/or constraints functions with simple functions, often polynomials, which
are fitted to data at a set of carefully selected design points. Neural networks are sometimes used to
function in the same role. The values of the objective function and constraints at the selected set of
points are used to “train” the network. Like the polynomial fit, the neural network provides an
estimate of objective function and constraints for the optimizer that is very inexpensive after the
initial investment in the net training has been made.
689 Sobieszczanski-Sobieski, J., “Sensitivity of Complex, Internally Coupled Systems”, AIAA Journal, Vol. 28, No. 1,
1990, pp. 153–160.
690 Sobieszczanski-Sobieski, J.; Barthelemy, J.-F.M.; and Riley, K. M., “Sensitivity of Optimum Solutions to Problems
Parameters”, AIAA Journal, Vol. 20, No. 9, September 1982, pp. 1291–1299.
691 See previous.
692 Ide, H.; Abdi, F. F.; and Shankar, V. J., “CFD Sensitivity Study for Aerodynamic/Control Optimization Problems”,
Parameters”, AIAA Journal, Vol. 20, No. 9, September 1982, pp. 1291–1299.
694 Barthelemy, J.-F; and Sobieszczanski - Sobieski, J., “Optimum Sensitivity Derivatives of Objective Functions in
Nonlinear Programming”, AIAA Journal, Vol. 21, No. 6, June 1983, pp. 913–915.
695 Braun, R. D., Kroo, I. M., and Gage, P. Y.,”Post-Optimality Analysis in Aircraft Design”, Proceedings of the AIAA
Aircraft, Design, Systems, and Operations Meeting, Monterey, California”, AIAA Paper No. 93-3932, 1993.
696 Sobieszczanski-Sobieski, J.,”Multidisciplinary Design Optimization: An Emerging, New Engineering Discipline”.
In Advances in Structural Optimization, Herskovits, J., (ed.), pp. 483–496, Kluwer Academic, 1995.
394
Systems”, Hampton, VA, September 28–30, 1988. NASA TM-101494. NASA CP-3031, Part 1, 1989.
395
matched. Each subspace optimization operates on its own design variables, some of which
correspond to the targets treated as the subspace optimization parameters, and uses a specialized
analysis to satisfy its own constraints. The objective function to be minimized is a cumulative
measure of the discrepancies between the design variables and their targets. The ensuing system-
level optimization satisfies all the constraints and adjusts the targets so as to minimize the system
objective and to enforce the matching. This optimization is guided by the above optimum sensitivity
derivatives.
Each of the above procedures applies also to hierarchic systems. A hierarchic system is defined as
one in which a subsystem exchanges data directly with the system only but not with any other
subsystem. Such data exchange occurs in analysis of structures by sub structuring. One iteration of
the procedure comprises the system analysis from the assembled system level down to the individual
system components level and optimization that proceeds in the opposite direction. The analysis data
passed from above become constant parameters in the lower level optimization. The optimization
results that are being passed from the bottom up include sensitivity of the optimum to these
parameters. The coordination problem solution depends on these sensitivity data. The current
practice relies on the engineer's insight to recognize whether the system is hierarchic, non-hierarchic,
or hybrid and to choose an appropriate decomposition scheme698.
Survey of Recent Developments”, AIAA 96-0711, 34th Aerospace Sciences Meeting and Exhibit, Reno, NV, 1995.
396
12.5.2.2 Constraints
A constraint is a condition that must be satisfied in order for the design to be feasible. An example of
a constraint in aircraft design is that the lift generated by a wing must be equal to the weight of the
aircraft. In addition to physical laws, constraints can reflect resource limitations, user requirements,
or bounds on the validity of the analysis models. Constraints can be used explicitly by the solution
algorithm or can be incorporated into the objective using Lagrange multipliers.
12.5.2.3 Objective
An objective is a numerical value that is to be maximized or minimized. For example, a designer may
wish to maximize profit or minimize weight. Many solution methods work only with single objectives.
When using these methods, the designer normally weights the various objectives and sums them to
form a single objective. Other methods allow multi-objective optimization, such as the calculation of
a Pareto front.
12.5.2.4 Models
The designer must also choose models to relate the constraints and the objectives to the design
variables. These models are dependent on the discipline involved. They may be empirical models,
such as a regression analysis of aircraft prices, theoretical models, such as from computational fluid
dynamics, or reduced-order models of either of these. In choosing the models the designer must trade
off fidelity with analysis time. The multidisciplinary nature of most design problems complicates
model choice and implementation. Often several iterations are necessary between the disciplines in
order to find the values of the objectives and constraints. As an example, the aerodynamic loads on a
wing affect the structural deformation of the wing. The structural deformation in turn changes the
shape of the wing and the aerodynamic loads. Therefore, in analyzing a wing, the aerodynamic and
structural analyses must be run a number of times in turn until the loads and deformation converge.
Minimizing F(𝐱) w. r. t. 𝐱
S. T. 𝐠(𝐱) ≤ 0 , 𝐡(𝐱) = 0 and 𝐱 𝑳𝑩 ≤ 𝐱 ≤ 𝐱 𝑼𝑩
Eq. 12.1
Where F is an objective, x is a vector of design variables, g is a vector of inequality constraints, h is a
vector of equality constraints, and xLB and xUB are vectors of lower and upper bounds on the design
variables. Maximization problems can be converted to minimization problems by multiplying the
397
objective by -1. Constraints can be reversed in a similar manner. Equality constraints can be replaced
by two inequality constraints.
Figure 12.2 Parallel Jacobi that exchanges sub problem solutions at the end of an iteration (left),
Sequential Gauss-Seidel that exchanges solutions as soon as they become available (center), and Hybrid
(right) – (Courtesy of Ryberg et al.)
Choosing the most appropriate architecture for the problem can significantly reduce the solution
time. These time savings come from the selected methods for solving each discipline, the coupling
scheme used in the architecture, and the degree to which operations are carried out in parallel. The
latter consideration becomes especially important as the design becomes more detailed and the
701 Joaquim R. R. A. Martins and Andrew B. Lambe, “Multidisciplinary Design Optimization: Survey of
Architectures”, AIAA Journal, 2013.
702 S. Tosserams, L. F. P. Etman, J. E. Rooda, “A classification of methods for distributed system optimization based
number of variables and/or constraints increases. The purpose here is to survey the available MDO
architectures and present them in a unified notation to facilitate understanding and comparison.
Furthermore, we propose the use of a new standard diagram to visualize the algorithm of a given
MDO architecture, how its components are organized, and its data flow. We pay particular attention
to the newer MDO architectures that have yet to gain widespread use. For each architecture, we
discuss its features and expected performance. We also present a new classification of MDO
architectures and show how they relate mathematically. This classification is especially novel as it is
able to draw similarities between architectures that were developed independently.
Symbol Definition
x Vector of design variables
yt Vector of coupling variable targets (inputs to a discipline analysis)
y Vector of coupling variable responses (outputs from a discipline analysis)
ŷ Vector of state variables (variables used inside only one discipline analysis)
f Objective function
c Vector of design constraints
cc Vector of consistency constraints
R Governing equations of a discipline analysis in residual form
N Number of disciplines
n() Length of given variable vector
m() Length of given constraint vector
( )0 Functions or variables that are shared by more than one discipline
( )i Functions or variables that apply only to discipline i
( )* Functions or variables at their optimal value
( )∼ Approximation of a given function or vector of functions
Table 12.1 Mathematical Notation for MDO Problem Formulations (Courtesy of Martins & Lambe)
This is not a comprehensive list; additional notation specific to particular architectures is introduced
when the respective architectures are described. We also take this opportunity to clarify many of the
terms we use that are specific to the field of MDO. A design variable is a variable in the MDO problem
that is always under the explicit control of an optimizer. In traditional engineering design, values of
these variables are selected explicitly by the designer or design team. Design variables may pertain
only to a single discipline, i.e., local, or may be shared by multiple disciplines. We denote the vector
of design variables local to discipline i by xi and shared variables by x0. The full vector of design
variables is given by
𝐱 = {x0 x1 … … xN }
Eq. 12.2
The subscripts for local and shared data are also used in describing objectives and constraints. A
discipline analysis is a simulation that models the behavior of one aspect of a multidisciplinary
system. Running a discipline analysis consists in solving a system of equations such as the Navier–
Stokes equations in fluid mechanics, or the static equilibrium equations in structural mechanics
which compute a set of discipline responses, known as state variables. State variables may or may
399
not be controlled by the optimization, depending on the formulation employed. We denote the vector
of state variables computed within discipline i by ŷi. We denote the associated set of disciplinary
equations in residual form by Ri, so that the expression Ri = 0 represents the solution of these
equations with respect to ŷi. In a multidisciplinary system, most disciplines are required to exchange
coupling variables to model the interactions of the whole system. Often, the number of variables
exchanged is much smaller than the total number of state variables computed in a particular
discipline. For example, in aircraft design, state information about the entire flow field resulting from
the aerodynamics analysis is not required by the structural analyses. Instead, only the aerodynamic
loads on the aircraft surface are passed. The coupling variables supplied by a given discipline i are
denoted by yi. Another common term for yi is response variables, since they describe the response
of the analysis to a design decision. In general, a transformation is required to compute yi from yi for
each discipline. Similarly, a transformation may be needed to convert input coupling variables into
a usable format within each discipline703. In this work, the mappings between yi and ŷi are lumped
into the analysis equations Ri. This simplifies our notation with no loss of generality. In many
formulations, copies of the coupling variables must be made to allow discipline analyses or
optimizations to run independently and in parallel. These copies are known as target variables, which
we denote by a superscript t. For example, the copy of the response variables produced by discipline
i is denoted yti . These variables are used as the input to disciplines that are coupled to discipline i
through yi. In order to preserve consistency between the coupling variable inputs and outputs at the
optimal solution, we define a set of consistency constraints, cci = yt i- yi which we add to the problem
formulation.
703 Cramer, E. J., Dennis Jr, J. E., Frank, P. D., Lewis, R. M., and Shubin, G. R., “Problem Formulation for
Multidisciplinary Optimization,” SIAM Journal on Optimization,” Vol. 4, No. 4, 1994, pp. 754–776.
704 Lambe, A. B. and Martins, J. R. R. A., “Extensions to the Design Structure Matrix for the Description of
Multidisciplinary Design, Analysis, and Optimization Processes,” Structural and Multidisciplinary Optimization.
705 Steward, D. V., “The Design Structure Matrix: A Method for Managing the Design of Complex Systems,” IEEE
Review and New Directions,” IEEE Transactions on Engineering Management, Vol. 48, No. 3, 2001, pp. 292–306.
707 Lambe, A. B. and Martins, J. R. R. A., “Extensions to the Design Structure Matrix for the Description of
Multidisciplinary Design, Analysis, and Optimization Processes,” Structural and Multidisciplinary Optimization.
400
simple examples. Figure 12.3, the first example, shows a Gauss–Seidel multidisciplinary analysis
(MDA) procedure for three disciplines, which is described in Table 12.2.
Figure 12.3 A block Gauss–Seidel Multidisciplinary Analysis (MDA) Process to Solve a Three-Discipline
Coupled System – (Courtesy of Martins & Lambe )
The components consist of the discipline analyses themselves and a special component, known as a
driver, which controls the iteration. The interfaces between these components consist of the data
that is exchanged. Following the DSM rules, the components are placed on the main diagonal of a
matrix while the interfaces are placed in the off-diagonal locations such that inputs to a component
are placed in the same column and outputs are placed in the same row. External inputs and outputs
may also be defined and are placed on the outer edges of the diagram. In the case of Figure 12.3,
external input consists of the design variables and an initial guess of the system coupling variables.
Each discipline analysis computes its own set of coupling variables which is passed to other discipline
analyses or back to the driver. At the end of the MDA process, each discipline returns the final set of
coupling variables computed. Thick gray lines are used to show the data flow between components.
A numbering system is used to show the order in which the components are executed. The algorithm
starts at component zero and proceeds in numerical order. Loops are denoted using the notation j ≠
k for k < j so that the algorithm must return to step k until a looping condition is satisfied before
proceeding. The data nodes are also labeled with numbers to denote the time at which the input data
is retrieved. As an added visualization aid, consecutive components in the algorithm are connected
by a thin black line. Thus, following the procedure in Figure 12.3 yields Table 12.2. The second
example, illustrated in Figure 12.4 is the solution process for an optimization problem using
gradient based optimization. The problem has a single objective and a vector of constraints. Figure
12.4 shows separate components to compute the objective, constraints, and their gradients and a
driver to control the iteration. Notice that in this example, multiple components are evaluated at step
one of the algorithm. This numbering denotes parallel execution. In some cases, it may be advisable
to lump components together to reflect underlying problem structures, such as lumping together the
objective and constraint components. In the following sections, we have done just that in the
401
architecture diagrams to simplify presentation. We will also make note of other simplifications as we
proceed.
T T
x = [x0T , x1T , , , , , xN ]
Eq. 12.4
to concatenate the disciplinary variable groups. In future problem statements, we omit the local
objective functions fi except when necessary to highlight certain architectural features. Problem (1)
is known as the “all-at-once” (AAO) problem. Figure 12.5 shows the XDSM for solving this
problem.
Figure 12.5 XDSM for Solving the AAO Problem (Courtesy of Martins & Lambe)
To keep the diagrams compact, we adopt the convention that any block referring to discipline i
represents a repeated pattern for every discipline. Thus, in Figure 12.5 a residual block exists for
every discipline in the problem and each block can be executed in parallel. As an added visual cue in
the XDSM, the “Residual i” component is displayed as a stack of similar components. There is a conflict
with the established literature when it comes to the labeling of Problem (1). What most authors refer
to as AAO, following the lead of [Cramer et al.], others label as the simultaneous analysis and design
(SAND) problem. Our AAO problem is most like what [Cramer et al.] refer to as simply “the most
403
general formulation”708. Here, we classify the formulation (1) as the AAO problem because it includes
all design, state, and input and output coupling variables in the problem, so the optimizer is
responsible for all variables at once. The SAND architecture uses a different problem statement and
is presented next.
The AAO problem is never solved in practice because the consistency constraints, which are linear in
this formulation, can be eliminated quite easily. Eliminating these constraints reduces the problem
size without compromising the performance of the optimization algorithm. As we will see,
eliminating the consistency constraints from Problem (1) results in the problem solved by the SAND
architecture. However, we have presented the AAO problem first because it functions as a common
starting point for deriving both the SAND problem and the Individual Discipline Feasible (IDF)
problem and, subsequently, all other correctly-formulated MDO problems. Depending on which
equality constraint groups are eliminated from Problem (1), we can derive the other three monolithic
architectures: Multidisciplinary Feasible (MDF), Individual Discipline Feasible (IDF), and
Simultaneous Analysis and Design (SAND). In the next three subsections, we describe how each
architecture is derived and the relative advantages and disadvantages of each. We emphasize that in
all cases, in spite of the elements added or removed by each architecture, we are always solving the
same MDO problem.
Minimize f0 = (x, y) w. r. t x, y, ŷ
S. T. c0 (x, y) ≥ 0 , ci (x0 , xi , yi ) ≥ 0 for i = 1, ,, , N
R i (x0 , xi , y, ŷ0 ) = 0 , for i = 1, ,, , N
Eq 12.5
The XDSM for SAND is shown in Figure 12.6. [Cramer et al.] refer to this architecture as “All-at-
Once”. However, we use the name SAND to reflect the consistent set of analysis and design variables
chosen by the optimizer. The optimizer, therefore, can simultaneously analyze and design the system.
Several features of the SAND architecture are noteworthy. Because we do not need to solve any
discipline analysis explicitly or exactly, the optimization problem can potentially be solved very
quickly by letting the optimizer explore regions that are infeasible with respect to the analysis
constraints. The SAND methodology is not restricted to multidisciplinary systems and can be used in
single discipline optimization as well. In that case, we only need to define a single group of design
constraints. If the disciplinary residual equations are simply discretized partial differential
equations, the SAND problem is just a PDE-constrained optimization problem like many others in the
literature.
Two major disadvantages are still present in the SAND architecture. First, the problem formulation
still requires all state variables and discipline analysis equations, meaning that large problem size
and potential premature termination of the optimizer at an infeasible design can be issues in practice.
Second, and more importantly, the fact that the discipline analyses equations are treated as explicit
constraints means that the residual values and possibly their derivatives need to be available to the
optimizer. In engineering design, many discipline analysis codes operate in a “black-box” fashion,
708 Cramer, E. J., Dennis Jr, J. E., Frank, P. D., Lewis, R. M., and Shubin, G. R., “Problem Formulation for
Multidisciplinary Optimization,” SIAM Journal on Optimization, Vol. 4, No. 4, 1994, pp. 754–776.
709 Haftka, R. T., “Simultaneous Analysis and Design,” AIAA Journal, Vol. 23, No. 7, July 1985, pp. 1099–1103.
404
directly computing the coupling variables while hiding the discipline analyses residuals and state
variables in the process. Even if the source code for the discipline analysis can be modified to return
residuals, the cost and effort required often eliminates this option from consideration. Therefore,
most practical MDO problems require an architecture that can take advantage of existing discipline
analysis codes. The following two monolithic architectures address this concern.
Figure 12.6 Diagram for the SAND Architecture (Courtesy of Martins & Lambe)
710Cramer, E. J., Dennis Jr, J. E., Frank, P. D., Lewis, R. M., and Shubin, G. R., “Problem Formulation for
Multidisciplinary Optimization,” SIAM Journal on Optimization, Vol. 4, No. 4, 1994, pp. 754–776.
405
Within the optimization iteration, specialized software for solving the discipline analyses can now be
used to return coupling variable values to the objective and constraint function calculations. The net
effect is that the IDF problem is both substantially smaller than the SAND problem and requires
minimal modification to existing discipline analyses. In the field of PDE-constrained optimization, the
IDF architecture is exactly analogous to a reduced-space method. In spite of the reduced problem
size when compared to SAND, the size of the IDF problem can still be an issue. If the number of
coupling variables is large, the size of the resulting optimization problem might still be too large to
solve efficiently. The large problem size can be mitigated to some extent by careful selection of the
partitioning strategy or aggregation of the coupling variables to reduce information transfer between
disciplines.
Figure 12.7 Diagram of the IDF Architecture (Courtesy of Martins & Lambe)
A bigger issue in IDF concerns gradient computation. If gradient-based optimization software is used
which is likely because it is generally more efficient for large problems evaluating the objective and
constraint function gradients becomes a costly part of the optimization procedure. This is because
the gradients themselves must be discipline-feasible, i.e., the changes in design variables cannot
cause the output coupling variables to violate the discipline analysis equations to first order. The
errors caused by inaccurate gradient values can severely impact the performance of gradient-based
optimization.
In practice, gradients are often calculated using some type of finite-differencing procedure, where
the discipline analysis is evaluated for each design variable. While this approach preserves
disciplinary feasibility, it is costly and unreliable. If the discipline analysis code allows for the use of
complex numbers, the complex-step method is an alternative approach which gives machine-
precision derivative estimates. If the analysis codes require a particularly long time to evaluate, the
use of automatic differentiation or analytic derivative calculations (i.e. direct or adjoint methods) can
be used to avoid multiple discipline analysis evaluations. While the development time for these
methods can be long, the reward is accurate derivative estimates and massive reductions in
computational cost, especially for design optimization based on high-fidelity models.
406
Figure 12.8 Diagram for the MDF Architecture with a Gauss–Seidel Multidisciplinary Analysis
(Courtesy of Martins & Lambe)
711 Cramer, E. J., Dennis Jr, J. E., Frank, P. D., Lewis, R. M., and Shubin, G. R., “Problem Formulation for
Multidisciplinary Optimization,” SIAM Journal on Optimization, Vol. 4, No. 4, 1994, pp. 754–776.
712 Alexandrov, N. M. and Lewis, R. M., “Analytical and Computational Aspects of Collaborative Optimization for
Multidisciplinary Design,” AIAA Journal, Vol. 40, No. 2, 2002, pp. 301–309.
713 Balling, R. J. and Sobieszczanski-Sobieski, J., “Optimization of Coupled Systems: A Critical Overview of
methods715 . Note that due to the sequential nature of the Gauss–Seidel iteration, we cannot evaluate
the disciplines in parallel and cannot apply our convention for compacting the XDSM. Using a
different MDA method results in a different XDS An obvious advantage of MDF over the other
monolithic architectures is that the optimization problem is as small as it can be for a monolithic
architecture, since only the design variables and design constraints are under the direct control of
the optimizer. Another benefit is that MDF always returns a fully consistent system design, even if
the optimization process is terminated early. This is advantageous in an engineering design context
if time is limited and we are not as concerned with finding a mathematically optimal design as with
finding an improved design. Note, however, that design constraint satisfaction is not guaranteed if
the optimization is terminated early; that depends on whether the optimization algorithm maintains
a feasible design point or not. In particular, methods of feasible directions require and maintain a
feasible design point while many robust sequential quadratic programming and interior point
methods do not.
The main disadvantage of MDF is that a consistent set of coupling variables must be computed and
returned to the optimizer every time the objective and constraint functions are re-evaluated. In other
words, the architecture requires a full MDA to be performed for every optimizer iteration. Instead of
simply running each individual discipline analysis once per optimizer iteration, as we do in IDF, we
need to run every discipline analysis multiple times until a consistent set of coupling variables is
found. This task requires its own specialized iterative procedure outside of the optimization.
Developing an MDA procedure can be time consuming if one is not already in place. Gradient
calculations are also much more difficult for MDF than for IDF. Just as the gradient information in IDF
must be discipline-feasible, the gradient information under MDF must be feasible with respect to all
disciplines. Fortunately, research in the sensitivity of coupled systems is fairly mature, and semi-
715Kennedy, G. J. and Martins, J. R. R. A., “Parallel Solution Methods for Aero structural Analysis and Design
Optimization,” Proceedings of the 13th AIAA/ISSMO Multidisciplinary Analysis Optimization Conference”, Fort
Worth, TX, September 2010, AIAA 2010-9308.
408
analytic methods are available to drastically reduce the cost of this step by eliminating finite
differencing over the full MDA. There is also some preliminary work towards automating the
implementation of these coupled sensitivity methods. The required partial derivatives can be
obtained using any of the methods described in Section C for the individual disciplines in IDF. As an
example of multidisciplinary feasible (MDF) method, consider the fluid/solid coupling of an aero-
structural problem expatriate in Figure 12.9, [Martins]716. It was argued that this would be superior
to Sequential Optimization with final result always being an elliptic lift distribution.
Minimize ∑ fi (xi ) w. r. t x1 , , , x𝑁
i=1
S. T. c0 (x1 , , , , x𝑁 ) ≤ 0 x1 , , , x𝑁
c1 (x1 ) ≤ 0 , , , , , , c𝑁 (x𝑁 ) ≤ 0
Eq. 12.8
In this problem, there are no shared design variables, x0, and the objective function is separable, i.e.
it can be expressed as a sum of functions, each of which depend only on the corresponding local
design variables, xi. On the other hand, the constraints include a set of constraints, c0, that depends
on more than one set of design variables. This problem is referred to as a complicating constraints
problem; if c0 did not exist, we could simply decompose this optimization problem into N
independent problems. Another possibility is that a problem includes shared design variables and a
separable objective function, with no complicating constraints, i.e.,
Minimize ∑ fi (𝑥0 , xi ) w. r. t x0 , x1 , , , x𝑁
i=1
S. T. c1 (x0 , x1 ) ≤ 0 , , , , , c(x0 , xN ) ≤ 0
Eq. 12.9
This is referred to as a problem with complicating variables. In this case, the decomposition would
be straightforward if there were no shared design variables, x0, and we could solve N optimization
problems independently and in parallel. Specialized decomposition methods were developed to
reintroduce the complicating variables or constraints into these problems with only small increases
in time and cost relative to the N independent problems. Examples of these methods include
[Dantzig–Wolfe] decomposition and Benders decomposition for Problems (5) and (6), respectively.
However, these decomposition methods were designed to work with the simplex algorithm on linear
716Joaquim R. R. A. Martins, “Multidisciplinary Design Optimization”, 7th International Fab Lab Forum and
Symposium on Digital Fabrication Lima, Peru, 2011, (Remote presentation).
409
programming problems. In the simplex algorithm, the active set changes only by one constraint at a
time so decomposition is the only way to exploit the special problem structure. However, algorithms
for nonlinear optimization that are based on Newton’s method, such as sequential quadratic
programming and interior point methods, may also use specialized matrix factorization techniques
to exploit sparsity structures in the problem.
While nonlinear decomposition algorithms were later developed, to the best of our knowledge, no
performance comparisons have been made between these decomposition algorithms and Newton-
like algorithms employing sparse matrix factorization. Intuition suggests that the latter should be
faster due to the ability of Newton methods to exploit second-order problem information. Thus, while
decomposition methods do exist for nonlinear problems, problem structure is not the primary
motivation for their development. The primary motivation for decomposing the MDO problem comes
from the structure of the engineering design environment. Typical industrial practice involves
breaking up the design of a large system and distributing aspects of that design to specific
engineering groups. These groups may be geographically distributed and may only communicate
infrequently. More importantly, however, these groups typically like to retain control of their own
design procedures and make use of in-house expertise, rather than simply passing on discipline
analysis results to a central design authority . Decomposition through distributed architectures allow
individual design groups to work in isolation, controlling their own sets of design variables, while
periodically updating information from other groups to improve their aspect of the overall design.
This approach to solving the problem conforms more closely with current industrial design practice
than the approach of the monolithic architectures. The structure of disciplinary design groups
working in isolation has a profound effect on the timing of each discipline analysis evaluation. In a
monolithic architecture, all discipline analysis programs are run exactly the same number of times,
based on requests from the optimizer or MDA program. In the context of parallel computing, this
approach can be thought of as a synchronous algorithm. In instances where some analyses or
optimizations are much more expensive than others, such as the case of multi fidelity optimization,
the performance suffers because the processors performing the inexpensive analyses and
optimizations experience long periods of inactivity while waiting to update their available
information. In the language of parallel computing, the computation is said to exhibit poor load
balancing. Another example of this case is aero structural optimization, in which a nonlinear
aerodynamics solver may require an order of magnitude more time to run than a linear structural
solver. By decomposing the optimization problem, the processor workloads may be balanced by
allowing disciplinary analyses with lower computational cost to perform more optimization on their
own. Those disciplines with less demanding optimizations may also be allowed to make more
progress before updating nonlocal information. In other words, the whole design process occurs not
only in parallel but also asynchronously. While the asynchronous design process may result in more
total computational effort, the intrinsically parallel nature of architecture allows much of the work
to proceed concurrently, reducing the wall clock time of the optimization.
717 Balling, R. J. and Sobieszczanski-Sobieski, J., “Optimization of Coupled Systems: A Critical Overview of
Approaches,” AIAA Journal, Vol. 34, No. 1, 1996, pp. 6–17.
718 Alexandrov, N. M. and Lewis, R. M., “Comparative Properties of Collaborative Optimization and Other
Approaches to MDO,” 1st ASMO UK/ISSMO Conference on Engineering Design Optimization, 1999.
719 See Previous.
410
example, the MDF architecture is closed with respect to both analysis and consistency constraints,
because their satisfaction is determined through the process of converging the MDA. Similarly, IDF is
closed analysis but open consistency since the consistency constraints can be satisfied by the
optimizer adjusting the coupling targets and design variables. [Tosserams et al. ]720 expanded on this
classification scheme by discussing whether or not distributed architectures used open or closed
local design constraints in the system sub problem. Closure of the constraints is an important
consideration when selecting an architecture because most robust optimization software will permit
the exploration of infeasible regions of the design space. Such exploration can result in faster
solutions via fewer optimizer iterations but this must be weighed against the increased optimization
problem size and the risk of terminating the optimization at an infeasible point.
The central idea in our classification is that distributed MDO architectures can be classified based on
their monolithic analogues: either MDF, IDF, or SAND. This stems from the different approaches to
handling the state and coupling variables in the monolithic architectures. It is similar to the previous
classifications in that an equality constraint must be removed from the optimization problem i.e.,
closed for every variable removed from the problem statement. However, using a classification based
on the monolithic architectures makes it much easier to see the connections between distributed
architectures, even when these architectures are developed in isolation from each other. In many
cases, the problem formulations in the distributed architecture can be derived directly from that of
the monolithic architecture by adding certain elements to the problem, by making certain
assumptions, and by applying a specific decomposition scheme. This classification can also be viewed
as a framework in which we can develop new distributed architectures, since the starting point for a
distributed architecture is always a monolithic architecture. This classification of architectures is
represented in Figure 12.10.
Known relationships between the architectures are shown by arrows. Due to the large number of
adaptations created for some distributed architectures, such as the introduction of surrogate models
and variations to solve multi objective problems, we have only included the “core” architectures in
our diagram. Details on the available variations for each distributed architecture are presented in the
720 Tosserams, S., Etman, L. F. P., and Rooda, J. E., “A Classification of Methods for Distributed System Optimization
based on Formulation Structure,” Structural and Multidisciplinary Optimization, Vol. 39, No. 5, 2009.
411
relevant sections. Note that none of the distributed architectures developed to date have been
considered analogues of SAND. As discussed in Section III, the desire to use independent “black-box”
computer codes for the disciplinary analyses necessarily excluded consideration of the SAND
problem formulation as a starting point. Nevertheless, the techniques used to derive distributed
architectures from IDF and MDF may also be useful when using SAND as a foundation.
Our classification scheme does not distinguish between the different solution techniques for the
distributed optimization problems. For example, we have not focused on the order in which the
distributed problems are solved. Coordination schemes are partially addressed in the Distributed IDF
group, where we have classified the architectures as either “penalty” or “multilevel”, based on
whether penalty functions or a problem hierarchy is used in the coordination. This grouping follows
from the work of [de Wit and van Keulen]721. One area that is not well explored in MDO is the use of
hybrid architectures. By hybrid, we mean an architecture that incorporates elements of two or more
other architectures in such a way that different disciplinary analyses or optimizations are treated
differently. For example, a hybrid monolithic architecture could be created from MDF and IDF by
resolving the coupling of some disciplines within an MDA, while the remaining coupling variables are
resolved through constraints. Some ideas for hybrid architectures have been proposed by [Marriage
and Martins]722 and [Geethaikrishnan et al.]723. Such architectures could be especially useful in
specific applications where the coupling characteristics vary widely among the disciplines. However,
general rules need to be developed to say under what conditions the use of certain architectures is
advantageous. As we note in Section V, much work remains in this area.
In the following sections, we introduce the distributed architectures for MDO. We prefer to use the
term “distributed” as opposed to “hierarchical” or “multilevel” because these architectures do not
necessarily create a hierarchy of problems to solve. In some cases, it is better to think of all
optimization problems as being on the same level. Furthermore, neither the systems being designed
nor the design team organization need to be hierarchical in nature for these architectures to be
applicable. Our focus here is to provide a unified description of these architectures and explain some
advantages and disadvantages of each. Along the way, we will point out variations and applications
of each architecture that can be found in the literature. We also aim to review the state-of-the-art in
architectures, since the most recent detailed architecture survey in the literature dates back from
more than a decade ago. More recent surveys, such as that of [Agte et al.]724, discuss MDO more
generally without detailing the architectures themselves.
721 de Wit, A. J. and van Keulen, F., “Overview of Methods for Multi-Level and/or Multi-Disciplinary Optimization,”
51st AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference, No. April,
Orlando, FL, April 2010.
722 Marriage, C. J. and Martins, J. R. R. A., “Reconfigurable Semi-Analytic Sensitivity Methods and MDO
Architectures within the _MDO Framework,” Proceedings of the 12th AIAA/ISSMO Multidisciplinary Analysis and
Optimization Conference, Victoria, BC, Canada, Sept. 2008.
723 Geethaikrishnan, C., Mujumdar, P. M., Sudhakar, K., and Adimurthy, V., “A Hybrid MDO Architecture for Launch
Vehicle Conceptual Design,” 51st AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials
Conference, No. April, Orlando, FL, April 2010.
724 Agte, J., de Weck, O., Sobieszczanski-Sobieski, J., Arendsen, P., Morris, A., and Spieck, M., “MDO: Assessment
and Direction for Advancement - an Opinion of One International Group,” Structural and Multidisciplinary
Optimization, Vol. 40, 2010, pp. 17–33.
412
formulation725-726 decomposes the system problem into independent sub problems with disjoint sets
of variables. Global sensitivity information is calculated at each iteration to give each sub problem a
linear approximation to a multidisciplinary analysis, improving the convergence behavior. At the
system level, a coordination problem is solved to recomputed the “responsibility”, “tradeoff”, and
“switch” coefficients assigned to each discipline to provide information on design variable
preferences for nonlocal constraint satisfaction. Using these coefficients gives each discipline a
certain degree of autonomy within the system as a whole.
Several variations of this architecture have been developed to incorporate meta models and higher
order information sharing among the disciplines. More recently, the architecture has also been
adapted to solve multi objective problems. [Parashar and Bloebaum]727 extended a multi objective
CSSO formulation to handle robust design optimization problems. An application of the architecture
to the design of high-temperature aircraft engine components is presented by [Tappeta et al.]728. The
version we consider here, due to [Sellar et al.]729, uses met model representations of each disciplinary
analysis to efficiently model multidisciplinary interactions. Using our unified notation, the CSSO
system sub problem is given by
Structural Optimization,” Engineering Optimization, Vol. 19, No. 3, 1992, pp. 171–186.
727 Parashar, S. and Bloebaum, C. L., “Robust Multi-Objective Genetic Algorithm Concurrent Subspace Optimization
(RMOGACSSO) for Multidisciplinary Design,” 11th AIAA/ISSMO Multidisciplinary Analysis and Optimization
Conference, Sept. 2006.
728 Tappeta, R. V., Nagendra, S., and Renaud, J. E., “A Multidisciplinary Design Optimization Approach for High
Temperature Aircraft Engine Components,” Structural Optimization, Vol. 18, No. 2-3, Oct. 1999, pp. 134–145.
729 Sellar, R. S., Batill, S. M., and Renaud, J. E., “Response Surface Based, Concurrent Subspace Optimization for
Multidisciplinary System Design,” 34th AIAA Aerospace Sciences and Meeting Exhibit, 1996.
413
et al.]730, and [Tedford and Martins]731 all show CSSO requiring many more analysis calls than other
Algorithm 2 CSSO
Input: Initial design variables x
Output: Optimal variables x*, objective function f *, and constraint values c*
0: Initiate main CSSO iteration
repeat
1: Initiate a design of experiments (DOE) to generate design points
for Each DOE point do
2: Initiate an MDA that uses exact disciplinary information
repeat
3: Evaluate discipline analyses
4: Update coupling variables y
until 4 ⇒ 3: MDA has converged
5: Update the disciplinary meta models with the latest design
end for 6 ⇒ 2
7: Initiate independent disciplinary optimizations (in parallel)
for Each discipline i do
repeat
8: Initiate an MDA with exact coupling variables for discipline i and approximate coupling
variables for the other disciplines
repeat
9: Evaluate discipline i outputs yi, and meta models for the other disciplines, y∼ j≠I until 10 ⇒ 9:
MDA has converged
11: Compute objective f and constraint functions c using current data until 12 ⇒ 8: Disciplinary
optimization i has converged
end for
13: Initiate a DOE that uses the sub problem solutions as sample points for Each sub problem
solution i do
14: Initiate an MDA that uses exact disciplinary information
repeat
15: Evaluate discipline analyses.
until 16 ! 15 MDA has converged
17: Update the disciplinary meta models with the newest design
end for 18 ⇒ 14
19: Initiate system-level optimization
repeat
20: Initiate an MDA that uses only meta model information
repeat
21: Evaluate disciplinary meta models
until 22 ⇒21: MDA has converged
23: Compute objective f, and constraint function values c
until 24 ⇒ 20: System level problem has converged
until 25 ⇒ 1: CSSO has converged
730 Perez, R. E., Liu, H. H. T., and Behdinan, K., “Evaluation of Multidisciplinary Optimization Approaches for
Aircraft Conceptual Design,” 10th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, No.
September, Albany, NY, Aug. 2004.
731 Tedford, N. P. and Martins, J. R. R. A., “Benchmarking Multidisciplinary Design Optimization Algorithms,”
architectures to converge to an optimal design. The results of [de Wit and van Keulen] showed that
CSSO was unable to reach the optimal solution of even a simple minimum-weight two-bar truss
problem. Thus, CSSO seems to be largely ineffective when compared with newer MDO architectures.
Figure 12.11 Diagram for the CSSO Architecture (Courtesy of Martins & Lambe)
732 Braun, R. D., “Collaborative Optimization: An Architecture for Large-Scale Distributed Design”, Ph.D. thesis,
Stanford University, Stanford, CA 94305, 1996.
733 Braun, R. D., Gage, P., Kroo, I. M., and Sobieski, I. P., “Implementation and Performance Issues in Collaborative
Optimization,” 6th AIAA, NASA, and ISSMO Symposium on Multidisciplinary Analysis and Optimization, 1996.
734 Braun, R. D., Collaborative Optimization: An Architecture for Large-Scale Distributed Design, Ph.D. thesis,
are only made if those variables directly influence the objective. Mathematically speaking, that means
∂f0/∂xi ≠ 0. In CO1, the quadratic equality constraints are replaced with linear equality constraints
for each target-response pair. In either case, post-optimality sensitivity analysis, i.e. computing
derivatives with respect to an optimized function, is required to evaluate the derivatives of the
consistency constraints J*i . The discipline i sub-problem in both CO1 and CO2 is minimize
𝑡
Minimize Ji [x̂0i , xi , y𝑖 (x̂0𝑖 , x𝑖 , 𝑦𝑗≠𝑖 )] w. r. t x̂0i , x𝑖
𝑡
S, T. [x̂0i , xi , y𝑖 (x̂0𝑖 , x𝑖 , 𝑦𝑗≠𝑖 )] ≥ 0
Eq. 12.13
Thus the system-level problem is responsible for minimizing the design objective, while the
discipline level problems minimize system inconsistency. [Braun] showed that the CO problem
statement is mathematically equivalent to the IDF problem statement (3) and, therefore, equivalent
to the original MDO problem (1) as well. CO is depicted in Figure 12.12 and the corresponding
procedure is detailed in Algorithm 3.
Figure 12.12 Diagram for the CO Architecture (Courtesy of Martins & Lambe)
In spite of the organizational advantage of having fully separate disciplinary sub problems, CO has
major weaknesses in the mathematical formulation that lead to poor performance in practice. In
particular, the system problem in CO1 has more equality constraints than variables, so if the system
cannot be made fully consistent, the system sub problem is infeasible. This can also happen in CO2,
but it is not the most problematic issue. The most significant difficulty with CO2 is that the constraint
gradients of the system problem at an optimal solution are all zero vectors. This represents a
breakdown in the constraint qualification of the Karush–Kuhn–Tucker optimality conditions, which
slows down convergence for most gradient-based optimization software. In the worst case, the CO2
formulation may not converge at all. These difficulties with the original formulations of CO have
inspired several researchers to develop modifications to improve the behavior of the architecture. In
416
a few cases, problems have been solved with CO and a gradient-free optimizer, such as a genetic
algorithm, or a gradient-based optimizer that does not use the Lagrange multipliers in the
termination condition to handle the troublesome constraints. While such approaches do avoid the
obvious problems with CO, they bring other issues. Gradient-free optimizers are computationally
expensive and can become the bottleneck within the CO architecture.
Gradient-based optimizers that do not terminate based on Lagrange multiplier values, such as
feasible direction methods, often fail in nonconvex feasible regions. As pointed out by [DeMiguel]735,
the CO system sub problem is set-constrained, i.e., nonconvex, because of the need to satisfy
optimality in the disciplinary sub problems. The approach taken by [DeMiguel and Murray] to fix the
problems with CO is to relax the troublesome constraints using an L1 exact penalty function with a
fixed penalty parameter value and add elastic variables to preserve the smoothness of the problem.
This revised approach is called Modified Collaborative Optimization (MCO). This approach satisfies
the requirement of mathematical rigor, as algorithms using the penalty function formulation are
known to converge to an optimal solution under mild assumptions. However, the test results of
[Brown and Olds] show strange behavior in a practical design problem. In particular, they observed
that for values of the penalty parameter above a threshold value, the problem could not improve the
initial design point. Below a lower threshold value, the architecture showed very poor convergence.
Finally, a penalty parameter could not be found which produced a final design close to those
computed by other architectures. In light of these findings, the authors rejected MCO from further
testing.
Another idea, proposed by [Sobieski and Kroo], uses surrogate models, also known as meta models,
to approximate the post-optimality behavior of the disciplinary sub-problems in the system sub-
problem. This both eliminates the post-optimality sensitivity calculation and improves the treatment
of the consistency constraints. While the approach does seem to be effective for the problems they
solve, to our knowledge, it has not been adopted by any other researchers to date. The simplest and
most effective known fix for the difficulties of CO involves relaxing the system sub problem equality
constraints to inequalities with a relaxation tolerance, which was originally proposed by [Braun et
al.]. This approach was also successful in other test problems, where the choice of tolerance is a small
fixed number, usually 106. The effectiveness of this approach stems from the fact that a positive
inconsistency value causes the gradient of the constraint to be nonzero if the constraint is active,
eliminating the constraint qualification issue. Nonzero inconsistency is not an issue in a practical
design setting provided the inconsistency is small enough such that other errors in the computational
model dominate at the final solution. [Li et al.] build on this approach by adaptively choosing the
tolerance during the solution procedure so that the system-level problem remains feasible at each
iteration. This approach appears to work when applied to the test problems but has yet to be verified
on larger test problems.
Despite the numerical issues, CO has been widely implemented on a number of MDO problems. Most
of applications are in the design of aerospace systems. Examples include the design of launch vehicles
, rocket engines, satellite constellations, flight trajectories , flight control systems, preliminary design
of complete aircraft, and aircraft family design. Outside aerospace engineering, CO has been applied
to problems involving automobile engines, bridge design, railway cars, and even the design of a
scanning optical microscope.
735 DeMiguel, A.-V. and Murray, W., “An Analysis of Collaborative Optimization Methods,” 8th
AIAA/USAF/NASA/ISSMO Symposium on Multidisciplinary Analysis & Optimization, Long Beach, CA, 2000.
417
2 N
2
Minimize J0 = ∑‖x̂0i − x0 ‖22 + ‖yit − y(x̂0i , xi , yj≠i
t
‖ w. r. t x0 , y t
2
2 i=1
Eq. 12.14
Note that this sub-problem is unconstrained. Also, unlike CO, post-optimality sensitivities are not
required by the system sub-problem because the disciplinary responses are treated as parameters.
The system sub-problem chooses the shared design variables by averaging all disciplinary
preferences. The ith disciplinary sub problem is where wCi and wFi are penalty weights for the
consistency and nonlocal design constraints, and s is a local set of elastic variables for the constraint
models. The wFi penalty weights are chosen to be larger than the largest Lagrange multiplier, while
the wCi weights are chosen to guide the optimization toward a consistent solution. Theoretically,
each wCi must be driven to infinity to enforce consistency exactly. However, smaller finite values are
used in practice to both provide an acceptable level of consistency and explore infeasible regions of
Figure 12.13 XDSM for the ECO Architecture (Courtesy of Martins & Lambe)
736Roth, B. D., Aircraft Family Design using Enhanced Collaborative Optimization, PHD thesis, Stanford
University, 2008.
418
the design space. The main new idea introduced in ECO is to include linear models of nonlocal
constraints, represented by ~ci≠i, and a quadratic model of the system objective function in each
disciplinary sub problem, represented by ~ f0. This is meant to increase each discipline’s “awareness”
of their influence on other disciplines and the global objective as a whole. The construction of the
constraint models deserves special attention because it strongly affects the structure of Figure
12.13. The constraint models for each discipline are constructed by first solving the optimization
problem that minimizes the constraint violation with respect to local elastic and design variables.
(Note that shared design variables and coupling targets are treated as parameters). A post-optimality
sensitivity analysis is then completed to determine the change in the optimized local design variables
with respect to the change in shared design variables. Combining these post-optimality derivatives
with the appropriate partial derivatives yields the linear constraint models. The optimized local
Eq. 12.15
419
design variables and elastic variables from Problem (13) are then used as part of the initial data for
Problem (12). The full algorithm for ECO is listed in Algorithm 4.
Based on the Roth’s results , ECO is effective in reducing the number of discipline analyses compared
to CO. The trade-off is in the additional time required to build and update the models for each
discipline, weighed against the simplified solution to the decomposed optimization problems. The
results also show that ECO compares favorably with the Analytical Target Cascading architecture
(wh. While ECO seems to be effective, CO tends to be an inefficient architecture for solving MDO
problems. Without any of the fixes discussed in this section, the architecture always requires a
disproportionately large number of function and discipline evaluations, assuming it converges at all.
When the system-level equality constraints are relaxed, the results from CO are more competitive
with other distributed architectures but still compare poorly with the results of monolithic
architectures.
Eq. 12.16
737 Sobieszczanski-Sobieski, J., Agte, J. S., and Sandusky Jr, R. R., “Bilevel Integrated System Synthesis,” AIAA
Journal, Vol. 38, No. 1, Jan. 2000, pp. 164–172.
738 Joaquim R. R. A. Martins and Andrew B. Lambe, “Multidisciplinary Design Optimization: Survey of
that we present here is due to [Tosserams et al.]739. This formulation conforms to our definition of
an MDO problem by explicitly including system wide objective and constraint functions. Like all other
ATC problem formulations, it is mathematically equivalent to Problem (1).
739 Tosserams, S., Etman, L. F. P., and Rooda, J. E., “Augmented fLagrangiang Coordination for Distributed Optimal
Design in fMDOg,” International Journal for Numerical Methods in Engineering, Vol. 73, 2008, pp. 1885–1910.
740 Haftka, R. T. and Watson, L. T., “Multidisciplinary Design Optimization with Quasi-separable Subsystems,”
successfully applied QSD with surrogate models to a structural optimization problem. However, they
made no comparison of the performance to other architectures, not even QSD without the surrogates.
A version of QSD without surrogate models was benchmarked by [de Wit and van Keulen].
Unfortunately, this architecture was the worst of all the architectures tested in terms of disciplinary
evaluations. A version of QSD using surrogate models should yield improved performance, due to the
smoothness introduced by the model, but this version has not been benchmarked to our knowledge.
741 Chittick, I. R. and Martins, J. R. R. A., “An Asymmetric Sub-optimization Approach to Aero-structural
Optimization,” Optimization and Engineering, Vol. 10, No. 1, March 2009, pp. 133–152.
742 Y. Panchenko, H. Moustapha, S. Mah, K. Patel, M.J. Dowhan, D. Hall, “Preliminary Multi-Disciplinary
European Congress on Computational Methods in Applied Sciences and Engineering, September 11-14, 2000.
744 Lytle, J.K., "The Numerical Propulsion System Simulation: A Multidisciplinary Design System for Aerospace
Preliminary Multi-Disciplinary Design Optimization (PMDO). The aim is to explore the conceptual
phase of the design process which involves the exploration of different concepts that satisfy engine
design specifications and requirements. The interaction that takes place among the disciplines is a
series of feedback loops and trades between conflicting requirements imposed on the system.
The optimizer used is iSIGHT©, developed by Engineous Software746. The iSIGHT software is a generic
shell environment that supports multidisciplinary optimization. The shell represents and manages
multiple elements of a particular design problem in conjunction with the integration of one or more
simulation programs. In essence, iSIGHT automates the execution of the different codes (in-house or
commercial), data exchange and iterative adjustment of the design parameters based on the problem
formulation and a specified optimization plan.
747 Raw, J. A. and Weir, G. C., "The Prediction of Off - Design Characteristics of Axial and Axial / centrifugal
Compressors ", SAE Technical Paper Series 800628, April, 1980.
748 Kacker, S.C. and Okapuu, U., “A Mean Line Prediction Method for Axial Flow Turbine Efficiency”, Journal of
• Diffusion factor
• Swirl angle at stator trailing edges
• Exit Mach number
• Ratio of hub to tip radius
• Blade angles
• Pressure ratio
• Choked flow
The objective of the optimization was to maximize efficiency. The optimization was run for
approximately 1000 iterations which took about 1 hour on an HPC-class workstation using a Genetic
Algorithm followed by a Direct Heuristic Search. The number of iterations required to achieve an
optimum seems excessive and several opportunities are being explored to reduce the iteration count:
The iSIGHT optimizer has a suite of explorative and gradient-based optimization methods that can be
applied in any sequence. Different combinations of optimization methods will be investigated in an
attempt to improve the efficiency of the optimization process. The design variables used by the
optimizer are expected to have a significant influence on the robustness and speed of optimization.
In the current axial compressor mean line application, the optimizer alters the gas path shape by
varying the coefficients of splines representing the hub and shroud curves. The dependence of the
compressor pressure ratio and efficiency on the spline coefficients is not direct. An improved set of
"physical" optimization variables has been suggested in which the optimizer varies axial
distributions of mean radius and area. The advantage of this formulation is that area and radius are
"physical" variables that have a direct link to the pressure ratio and efficiency predicted by the mean
line program. This direct link should result in a "cleaner" design space, a reduced number of
iterations to converge to an optimal solution, and improved robustness of the optimization
procedure.
750 Y. Panchenko, H. Moustapha, S. Mah, K. Patel, M.J. Dowhan, D. Hall, “Preliminary Multi-Disciplinary
Optimization in Turbomachinery Design”, ADA415759.
751 Sobieszczanski-Sobieski, J. and Haftka, R. T., "Multidisciplinary Aerospace Design Optimization: Survey of
Systems", Proceedings, 2nd NASA/USAF Symposium on Recent Advances in Multidisciplinary Analysis and
Optimization, Hampton, Virginia, 1988.
753 Braun, R.D., "Collaborative Optimization: An Architecture for Large-Scale Distributed Design", Ph.D. thesis,
755 Ann-Britt Ryberg, Rebecka Domeij Bäckryd, Larsgunnar Nilsson, “Meta-model-Based Multidisciplinary
Design Optimization for Automotive Applications”, Technical Report LIU-IEI-R-12/003, 2012.
756 Simpson, T., Peplinski, J., Koch, P. N., and Allen, J. , “Meta-models for computer-based engineering design:
numerical noise and hence make it easier to find the global optimum. Meta-models could also make
it possible to use advanced optimization algorithms which are better suited for finding global optima
but require many evaluations. In addition, meta-models render a view of the entire design space and
might also make it easier to detect errors in the simulation model since the entire design region is
analyzed. When the meta-model are built, it is also inexpensive to rerun optimizations, e.g. with
changed constraint limits. This makes it possible to investigate multiple scenarios almost without
any additional cost. One further benefit with meta-models, when used in multidisciplinary design
optimization, is the possibility for disciplinary autonomy. Simply put, a meta-model is a
mathematical approximation of a detailed and usually computationally costly simulation
model, i.e. a model of a model. The meta-models can be used as surrogates for the detailed model
when a large number of evaluations are needed, as in optimization, and when it is too time-
consuming to run the detailed model for each evaluation.
Figure 12.17 The concept of meta-modelling for a response depending on two design variables –
(Courtesy of Ryberg et al.)
761 Ann-Britt Ryberg, Rebecka Domeij Bäckryd, Larsgunnar Nilsson, “Meta-model-Based Multidisciplinary
Design Optimization for Automotive Applications”, Technical Report LIU-IEI-R-12/003, 2012.
762 See Previous.
428
of all parts of the automotive structure. As reflected by the former Saab Automobile organization,
simulations can roughly be divided into two different categories. The first one supports certain
design areas, e.g. body, chassis, or interior design. The other one evaluates disciplinary performance,
such as safety or aerodynamics, which depends on more than one design area, see Figure 12.18.
The former is consequently evaluating many different aspects, e.g. stiffness, strength, and durability,
for a certain area of the vehicle, while the latter focuses on one performance area which often
depends on the complete vehicle. In conjunction with the different simulation areas there is, in most
cases, also a corresponding test organization performing the hardware validation at the end of the
project. The division of the simulation work into design area simulations (division by object) and
performance area simulations (division by aspect) reflects the different types of decompositions
proposed for MDO problems.
Many parts and subsystems in a car are developed by suppliers and consequently simulations on
these parts and subsystems are first done by the suppliers. The integration of these systems into the
vehicle is then checked by the car manufacturer. One large such system, where extensive detailed
simulations normally are done separately, is the powertrain system, i.e. engine and gearbox. Many
different load cases are evaluated within each simulation area. Regarding safety, e.g. front, side, and
rear end crashes are studied and for each of these crash directions various impact speeds, barriers,
and occupants are considered. Optimization can be used to guide the design within one discipline,
maybe find the balance between contradicting load cases, or be multidisciplinary and consider load
cases from more than one discipline or simulation area.
Figure 12.18 Schematic illustration of Simulation Areas within the Automotive Industry, example from
Saab Automobile (Courtesy of Ryberg et al.)
automotive industry. However, there are some differences between the aerospace and automotive
industries that might influence which methods that are suitable and to what extent they might be
used. The aerospace industry has long product and design cycles and produces few but very
expensive products compared to the automotive industry. In addition, the aerospace industry usually
has a military branch, which mainly is state funded, and where there may be more time and resources
available to develop new processes and methods. The development in the aerospace industry is
rigorously ruled by standards and regulations while passenger cars are designed to fulfil a number
of market requirements and expectations in addition to the legislative requirements. The number of
large automotive manufacturers is also greater than the number of large aerospace manufacturers.
This might lead to stronger competition in the automotive industry, and the strive for better products
as well as shorter and less expensive product development may therefore be more pronounced. Thus,
it is logical that some methods and processes are developed within the aerospace industry, which
might have the time and resources available, and that these methods subsequently are adopted, and
maybe become even more used within the automotive industry, which constantly is seeking
improvements due to the fierce competition.
The aerospace industry has to follow methods and processes during the product development that
are approved by governmental safety agencies such as FAA (Federal Aviation Administration) in the
USA and EASA (European Aviation Safety Agency) in Europe. The development process has therefore
become rather conservative. The product development within the automotive industry, on the other
hand, is not as rigorously controlled. The requirements are more related to the performance of the
vehicle, like safety or CO2-emission, and specific methods are not prescribed for the product
development. These facts might be additional reasons for the faster introduction of new processes
and methods within the automotive industry compared to the aerospace industry. The question is
then whether the MDO methods developed specifically for the aerospace industry, are suitable for
automotive applications as is, or if some characteristics of the automotive applications requires the
methods to be adjusted. Another possibility could be that the methods found insufficient for
aerospace applications are better suited for automotive applications.
One of the differences between the development processes in the aerospace and the automotive
industries is the development of the structural parts, i.e. the wings and fuselage of the aero plane and
the body of the car. The wings are for example typically dimensioned with respect to fatigue, and
although there is considerable movement of the wings during flight, the stresses are kept within the
elastic region. The car body is, to a large extent, dimensioned by crashworthiness requirements, and
the problem then becomes highly non-linear with large plastic deformations. However, during
normal operation, the car body has small deformations compared to the aero plane structure. Thus,
when studying the aerodynamics of an aero plane, it is essential to take the deformations induced by
the aerodynamic forces into account, while this is not as important when studying the aerodynamics
of a passenger car. The forces induced by the deformations are one of the major loads that the wing
structure should carry, while the corresponding forces on a car are negligible compared to the forces
applied during a crash event. The coupling between disciplines, e.g. aerodynamics and structural
performance, is in this example thus much stronger in the aero plane case compared to the passenger
car case.
As a consequence of the coupling between disciplines, an iterative approach is needed to find a
consistent solution, i.e. a solution in balance. This might be done by first estimating the aerodynamic
loads for the structural simulation. The deflections obtained are then applied in the aerodynamic
simulation to find the aerodynamic forces. The iteration is continued until the forces and deflections
match each other. There are examples of coupled disciplines in the automotive industry as well, e.g.
vehicle dynamics and chassis structural performance, which both depend on the chassis stiffness, but
they are not dominating the product development. Incorporating MDO into the automotive design
process is therefore presumably simpler than in the aerospace industry since the disciplines are
430
more loosely coupled, as stated by [Agte et al.]763. It could be said that automotive designs are created
in a multi-attribute environment rather than in a truly multi-disciplinary environment, and aspects,
such as NVH and crashworthiness, are only coupled by shared system level variables. The absence of
strong coupling between disciplines makes it easier to incorporate meta-models in the optimization
process and consequently also possible to include very computationally expensive simulations more
conveniently.
It is often possible to use direct optimization methods for linear simulations, since the computational
cost for every simulation is low and the studied responses do not include many local minima and
maxima. Non-linear simulations are often computationally costly and the responses complex, and
consequently more advanced optimization methods are required. These methods, however, demand
more evaluations to find the optimum, and therefore the use of meta-models becomes interesting.
Another difference between the aerospace and automotive industries is how the development is
done. In the aerospace industry, different parts of the aero plane are developed by different
companies in a joint project, and the different companies have fixed input with which they should
fulfil certain requirements. Although there are system suppliers in the automotive industry with
responsibility for the performance of their parts, the responsibility of the complete vehicle is still left
to the vehicle manufacturer. Thus, there might be longer communication paths in aerospace
development projects, which result in a stronger need for the different parties to work independently
also when doing full scale MDO. Hence, the need for autonomy, as offered by multilevel optimization
methods, is even more obvious. Some of the results of the differences mentioned, e.g. coupling of
variables, can be seen later in this chapter when the experiences from the automotive industry
regarding MDO are presented and compared with the experiences from the aerospace industry.
763 Agte, J., de Weck, O., Sobieszczanski-Sobieski, J., Arendsen, P., Morris, A., and Spieck, M. (2010). MDO:
assessment and direction for advancement - an opinion of one international group. Structural and
Multidisciplinary Optimization, 40(1-6), 17-33.
764 Schematic description of simultaneous optimization of aerodynamic and structural performance of
aerospace structures. a) Coupling between disciplines in the original problem. b) Two-level optimization
431
structural analysis usually is much cheaper than the aerodynamic analysis. This sequential technique
works when structural deformations are approximately constant throughout the main part of the
flight time, as is the case for most conventional transport aircrafts. However, it does seldom lead to
the optimal design of the global system, as noted by [Kroo]765 (1997). In addition, when aerodynamic
performance is important for multiple design conditions with different structural deformations, a
completely integrated structural and aerodynamic optimization may be necessary to obtain high-
performance designs. The wish for disciplinary autonomy and parallelization of work have resulted
in the development of different multi-level optimization methods that have been tested on academic
examples. Although showing promising results, the regular use of these methods within the industry
has not yet been realized and incorporation of MDO methodology in efficient multi-level strategies,
rather than only single level approaches, is still missing according to [Agte et al.].
strategy for the case where the structure is designed to deform to a predefined shape. The aircraft weight can
be constrained during the upper level aerodynamic
optimization to be lesser than or equal to its optimum value from the structural optimization.
765 Kroo, I. (1997). MDO for large-scale design. In N. Alexandrov, and M. Hussaini (Ed.), Multidisciplinary design
optimization: state of the art, Proceedings of the ICASE/NASA Langley Workshop 1995, (pp. 22-44), USA.
432
industry but have not yet been fully implemented within the product development process.
According to [Duddeck], the difficulties so far have mainly been related to the overall computational
time and the accuracy of the meta-models.
766Song, S.-I., and Park, G.-J. “Multidisciplinary optimization of an automotive door with a tailored blank.”
Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering, 2006.
433
automotive applications. Further on, an MDO method must be chosen. It must be determined
whether a single- or multi-level method should be used. Using a multi-level method increases the
complexity of the optimization process considerably compared to using a single-level method.
Therefore, in order to motivate the use of a multi-level method, the benefits must be greater than
cost. [Ryberg et al.]767.