Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Numerical Solution of Cauchy Problems for Elliptic Equations in Rectangle-like Geometries

Fredrik Berntsson and Lars Eldn e Department of Mathematics, Linkping University o Oktober 2005

Femlab Conference 2005

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Overview
Introduction, Motivating Example Illposedness, Stabilization Numerical Test problem in FEMLAB Transformation to rectangular geometry. Mapping of Normal derivatives. Numerical solution of Test problem Conclusions, Future Work

Femlab Conference 2005

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Motivating example: Ilmenite iron melting furnace

Thermocouple

Electrode

Level K to level D thermocouples with level K highest

Center

Under electrodes

Between electrodes

The furnace material properties are temperature dependent. Problem: nd the inner shape of the furnace. Nonlinear, and (rather) complex geometry
PhD thesis: I M Skaar, Monitoring the Lining of a Melting Furnace, NTNU, Trondheim, 2001

Femlab Conference 2005

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Irregular geometry - non-constant coecient

Map the region to a rectangle:

liquid steel

liquid steel

Cauchy data

Find lining: the interface between iron and furnace

Femlab Conference 2005

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

The Cauchy Problem for Laplaces equation

Classical ill-posed problem (Hadamard, 1900-1930) uxx + uyy = 0, u(x, 0) = 0, x , u 1 (x, 0) = sin(nx), x . x n 1 u(x, y) = 2 sin(nx) sinh(ny). n Large n: arbitrarily small data, arbitrarily large solution. The solution does not depend continuously on the data!
Femlab Conference 2005 4

Solution:

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Illposed Cauchy Problem on Unit Square

y 1

ux = 0

(a(u)ux )x +(a(u)uy )y

ux = 0

u = g,

uy = h

Femlab Conference 2005

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Fourier analysis
By separation of variables:

T (x, y) = A0 y + B0 +
k=1

(Ak eky + Bk eky ) cos(kx),

Fourier coecients {Ak } and {Bk } satisfy (for k > 0): 1 k 1 k Ak Bk = gk hk ,

where {gk } and {hk } are the Fouriercosine coecients of g(x) and h(x) Since eky as k the problem is severely illposed!
Femlab Conference 2005 6

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Rewrite as an Initial value problem: u auy =

y

0 x a x

a1 0

u auy

0 y 1,

u(x, 0) uy (x, 0)

g h

The unbounded x-derivative makes the problem ill-posed! Approximation of derivative: v a(u) x x A diagonal matrix, depends on u D bounded dierentiation matrix, spectral or otherwise Resulting stable problem is solved using standard code (ode45) in Matlab.
Femlab Conference 2005 7

D(ADV )

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Approximate u(x) by a least squares cubic spline s(x). Then set u(x) s(x).

3 The basis functions Bj (x), for j = 1, . . . , 5.

Choose a coarse grid

bounded dierentiation operator.

Advantage: exible at boundaries.

Femlab Conference 2005 8

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

A Numerical Model Problem

The Heat Equation (a(u)ux)x + (a(u)uy )y = 0, in

Boundary Conditions u = g, u n = 0,
u n

= h, on L1, on L2 and L3,

Femlab Conference 2005

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Test problem generated in FEMLAB

Sides: insulated Upper boundary: 1450o Lower boundary: 800o Temperature-dependent conductivity

Femlab Conference 2005

10

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Thermal conductivity
18 16 Thermal conductivity [ W/moC ] 14 12 10 8 6 4 2 0

200

400

600 800 1000 Temperature [ oC ]

1200

1400

1600

The thermal conductivity for magnesia brick, as a function of temperature.

Femlab Conference 2005

11

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Constructed solution (direct problem solved using FEMLAB):

Temperature and Heat-Flux data along the lower boundary is used to identify the upper boundary!
Femlab Conference 2005 12

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Transformation of the Problem

The Conformal Mapping (x, y) = (, ) = [x(, ), y(, )] The problem is transformed into (a(v)v ) + (a(v)v ) = 0, v(, 0) = g((, 0)), v (, 0) = ||h((, 0)), v (0, ) = v (1, ) = 0, 0 < < 1, 0 < < 1 0 < < 1, 0 < < 1, 0 < < L ,

Schwarz-Christoel Matlab toolbox: polygonal regions

Femlab Conference 2005 13

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Transformation of the area

3.5 3 2.5 2 1.5 1 0.5 0 0.5 1 1.5 2

Conformal mapping of the auxilliary domain onto a rectangle. Actual grid is ner!

Femlab Conference 2005

14

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Computed ux data on the square

4600 4400

4200

4000

3800

3600

3400

3200 2

1.5

0.5

0.5

1.5

Transformed using Conformal mapping (blue curve). No noise added!

Femlab Conference 2005

15

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Computed heat-ux on original Domain

3500 3000

2500

2000

1500

1000

500

u The Heat-Flux on L1 computed by FEMLAB n

Femlab Conference 2005 16

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Transformation of normal derivatives

By the Conformal mapping u 1 v |L1 = |( ) | (, 0) n The SchwarzChristoel mapping function has a singularity at every corner of the polygonal domain. The Singularities in |(1)| almost cancel out the spikes in FEMLABs normal derivative.

Femlab Conference 2005

17

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Numerical Solution of the Test problem

liquid steel

liquid steel

Cauchy data

Solve Cauchy problem. Find the isotherm v = 1450 oC. Conformal mapping of the identied curve.

Femlab Conference 2005

18

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Noisy ux data
4500

4000

3500

3000 2

1.5

0.5

0.5

1.5

Normally distributed noise added. Realistic noise level!

Femlab Conference 2005

19

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Computed temperatures at 0.6 and 0.8

Numerical solution at y=0.60 Numerical solution at y=0.80

1340 1320 1300 1280

Temperature [ oC ]

1460 1440 1420 1400

Temperature [ oC ]

1260 1240 1220 1200 1180 1160 1140 0

1380 1360 1340 1320 1300 1280 1260 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Derivatives computed using 11 B-splines. Dierential equation only valid when temperature is below 1450 oC.
Femlab Conference 2005 20

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Noisy data, identied boundary

4 3

Derivatives computed using 11 B-splines.

Femlab Conference 2005

21

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Less regularization, temperature at 0.62 and 0.81

Numerical solution at y=0.62
1400 1460 1440 1350 1420 1400
Temperature [ C ]

Numerical solution at y=0.81

1300

Temperature [ oC ]

1380 1360 1340 1320

1250

1200

1300 1280

1150 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1260 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Derivatives computed using 19 B-splines! Impossible to identify boundary when the temperature is not smooth

Femlab Conference 2005

22

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Conclusions
Ecient solution: replace a derivative by bounded approximation, solve Cauchy problem stably using a method of lines General geometries: transform rectangle-like region to rectangle (equivalent to creating a curvilinear mesh). Solve the problem on the rectangle Transformation cheap to compute for rather general geometries (approximated by polygon) Nonlinear problems can be solved rather easily Stability theory: A stability estimate for a Cauchy problem for an elliptic partial dierential equation, Inverse Problems, vol. 21, no 5, pp. 16431653, October 2005.
Femlab Conference 2005 23

Presented at the COMSOL Multiphysics User's Conference 2005 Boston

Future Work

More detailed FEMLAB model of the Furnace. More realistic simulation of measurements. Identify boundaries given other conditions (e.g. insulated boundary). Applications: The Melting Furnace Use FEMLAB to creat Test problem. Finding good test problems is very important!

Femlab Conference 2005

24