Documente Academic
Documente Profesional
Documente Cultură
AS COMPUTATION
FINITE ELEMENT ANALYSIS AS COMPUTATION
Gangan Prathap
Preface
These course notes are specially prepared for instruction into the principles
underlying good finite element analysis practice. You might ask: Why another
book on FEA (Finite Element Analysis) and the FEM (Finite Element Method) when
there are hundreds of good textbooks in the market and perhaps thousands of
manuals, handbooks, course notes, documentation manuals and primers available on
the subject.
Some time earlier, I had written a very advanced monograph on the analytical and
philosophical bases for the design and development of the finite elements which
go into structural analysis software packages. Perhaps a hundred persons work on
element development in the world and out of these, maybe a dozen would probe
deep enough to want to explore the foundational bases for the method. This was
therefore addressed at a very small audience. In the world at large are several
hundred thousand engineers, technicians, teachers and students who routinely use
these packages in design, analysis, teaching or study environments. Their needs
are well served by the basic textbooks and the documentation manuals that
accompany installations of FEM software.
However, my experience with students and teachers and technicians and engineers
over two decades of interaction is that many if not most are oblivious to the
basic principles that drive the method. All of them can understand the
superficial basis of the FEM - what goes into the packages; what comes out; how
to interpret results and so on. But few could put a finger on why the method
does what it does; this deeper examination of the basis of the method is
available to a very few.
The FEM is now formally over thirty-five years old (the terminology "finite
element" having been coined in 1960). It is an approximate method of solving
problems that arise in science and engineering. It in fact originated and grew
as such, by intuition and inspired guess, by hard work and trial and error. Its
origins can be traced to aeronautical and civil engineering practice, mainly
from the point of view of structural engineering. Today, it can be used, with
clever variations, to solve a wide variety of problems in science and
engineering. There are billions of dollars worth of installed software and
hardware dedicated to finite element analysis all over the world and perhaps
billions of dollars are spent on analysis costs using this software every year.
It is therefore a remarkable commercial success of human imagination, skill and
craft.
We shall close these notes with a telescopic look into the present scenario and
the shape of things to come in the future.
iii
Acknowledgements: I am extremely grateful to Dr T S Prahlad, Director, N.A.L.
for his encouragement and support. Many colleagues were associated at various
stages of writing this book. They include Dr B P Naganarayana, Dr S Rajendran
and Mr Vinayak Ugru, Dr Sudhakar Marur and Dr Somenath Mukherjee. I owe a great
deal to them. Mr Arjun provided secretarial help and also prepared the line
drawings and I am thankful to him. I am also thankful to Latha and Rahul for
their patience and understanding during the preparation of this manuscript.
Gangan Prathap
Bangalore
2001
iv
Contents
Preface iii
6. Shear locking 57
8. Stress consistency 97
9. Conclusion 111
v
Chapter 1
1
especially the triumph of the infinitesimal calculus of continuum behavior in
physics. The development and application of complex mathematical tools led to
the growth of the branch of mathematical physics. This therefore encouraged the
study of the properties of elastic materials and of elasticity - the description
of deformation and internal forces in an elastic body under the action of
external forces using the same mathematical equipment that was been used in
other classical branches of physics. Thus, from great mathematicians such as
Cauchy, Navier, Poisson, Lagrange, Euler, Sophie Germain, came the formulation
of basic governing differential equations which originated from the application
of the infinitesimal calculus to the behavior of structural bodies. The study is
thus complete only if the solutions to such equations could be found. For nearly
a century, these solutions were made by analytical techniques however, these
were possible only for a very few situations where by clever conspiracy, the
loads and geometries were so simplified that the problem became tractable.
However, by ingenuity, the engineer could use this limited library of simple
solutions to construct meaningful pictures of more complicated situations.
1.3.1 Introduction
We have seen earlier in this chapter how a body of knowledge that governs the
behavior of materials, solids and structures came to exist. Gifted
mathematicians and engineers were able to formulate precise laws that governed
the behavior of such systems and could apply these laws through mathematical
models that described the behavior accurately. As these mathematical models had
to take into account the continuum nature of the structural bodies, the
description often resulted in what are called differential equations. Depending
on the sophistication of the description, these differential equations could be
very complex. In a few cases, one could simplify the behavior to gross
relationships - for a bar, or a spring, it was possible to replace the
differential equation description with a simple relation between forces and
displacements; we shall in fact see that this is a very simple and direct act
of the discretization process that is the very essence of the finite element
process. For a more general continuum structure, such discretizations
or simplifications were not obvious at first. Therefore, there was no recourse
2
Fig. 1.1 A simple bar problem under axial load
It was recognized quite early that where analytical techniques fail or were
difficult to implement, approximate techniques could be devised to compute the
answers. In fact, much before algebra and analysis arrived, in contexts more
general than solid or structural mechanics, simple approximation and
computational schemes were used to solve engineering and scientific problems. In
this chapter, we shall review what we mean by an analytical solution for a very
simple problem and then proceed to examine some computational solutions. This
will reveal to us how the discretization process is logically set up for such a
problem.
The problem is categorized as a boundary value problem. We presume here that for
this problem, the reader is able to understand that the governing differential
equation describing the situation is
EA u,xx = 0 (1.1)
3
the bar. In general, governing differential equations belong to a category
called partial differential equations but here we have a simpler form known as
an ordinary differential equation. Equation (1.1) is further classified as a
field equation or condition as one must find a solution to the variable u which
must satisfy the equation over the entire field or domain, in this case, for x
ranging from 0 to L. A solution is obviously
u = ax + b (1.2)
where a and b are as yet undetermined constants. To complete the solution, i.e.
to determine a and b in a unique sense for the boundary (support) and loading
conditions shown in Fig. 1.1, we must now introduce what are called the boundary
conditions. Two boundary conditions are needed here and we state them as
u = 0 at x = 0 (1.3a)
EA u,x = P at x = L (1.3b)
The first relates to the fact that the bar is fixed at the left end and the
second denotes that a force P is applied at the free end. Taking these two
conditions into account, we can show that the following description completely
takes stock of the situation:
where ε(x), σ(x) and P(x) are the strain, stress and axial force (stress
resultants) along the length of the bar. These are the principal quantities that
an engineer is interested in while performing stress analysis to check the
integrity of any proposed structural design.
4
Fig. 1.2 Grid for finite difference scheme
For a problem where the differential equations are already known, the technique
known as the finite difference method can be used to obtain an approximate or
numerical solution. In fact, before the finite element method was established,
it was this method which was used to obtain solutions for very complicated
problems in structural mechanics, fluid dynamics and heat transfer.
Figure 1.2 shows a very simple uniformly spaced mesh or grid of nodal
points that can be used to discretize the problem described by Equations (1.1)
and (1.3). The larger the number of nodal points used (i.e. the smaller the grid
spacing h), the greater will be the accuracy involved. The field variable is now
taken to be known (or very strictly, unknown, at this stage of the computation)
at these nodal points. Thus, u1,..., ui,..., un, are the unknown variables or
degrees of freedom. The next step is to rewrite the governing differential
equations and boundary conditions in finite difference form. We see that the
finite difference forms for u,x and u,xx are required. It is easy to show
(again, details are omitted, as these can be found in most books on numerical
analysis or fem) that at a grid point i.
5
We now attempt a solution in which four points are used in the finite difference
grid. The governing differential equations and boundary conditions are replaced
by the following discrete set of linear, algebraic equations:
u1 = 0 (1.6a)
u3 – 2u2 + u1 = 0 (1.6b)
u4 - 2u3 + u2 = 0 (1.6c)
u4 - u3 = Ph/EA (1.6d)
where h=L/3 and Equations (1.6a) and (1.6d) represent the boundary conditions at
x = 0 and L respectively. The reader can easily work out that the solution
obtained from this set of simultaneous algebraic equations is, u2 = PL/3EA,
u3 = 2PL/3EA and u4 = PL/EA. Comparison with Equation (1.4a) will confirm that
we have obtained the exact answers at the grid points. Other quantities of
interest like strain, stress, etc. can be computed from the grid point values
using the finite difference forms of these quantities.
The second method is one that originates from a more fundamental statement
of the principles involved. It recognizes that one of the most basic and elegant
principles known to man is the law of least action, or minimum potential energy,
or virtual work. It is a statement that Nature always chooses the path of
minimum economy. Thus, the law of least action or minimum total potential is as
axiomatic as the basic laws of equilibrium are. One can start with one axiom and
derive the other or vice-versa for all of mechanics or most of classical
physics. In dynamics, the equivalent principle is known as Hamilton's principle.
6
general statement known as the principle of virtual work would apply. From these
brief general philosophical reflections we shall now proceed to the special case
at hand to see how the energy principle applies.
L
U = ò 1/ 2 EA u 2 ,x dx
0 (1.7a)
V = P ux=L (1.7b)
Then,
Π ( υ(x )) =
L
ò0 1/ 2 EA u 2 ,x dx - P ux = L
(1.8)
δΠ = 0 (1.9)
There are several ways in which an approximation can be applied at the level of
energy, virtual work or weighted residual statements, e.g. the Rayleigh-Ritz
(R-R) procedure, the Galerkin procedure, the least-squares procedure, etc. We
first choose a set of independent functions which satisfy the geometric or
essential (also called kinematic) boundary conditions of the problem. These
functions are called admissible functions. In this case, the condition specified
in Equation (1.3a) is the geometric boundary condition. The other condition,
Equation (1.3b) is called a nonessential or natural or force boundary condition.
The admissible functions need not satisfy the natural boundary conditions (in
fact, it can be shown that if they are made to do so, there can even be a
deterioration in performance). The approximate admissible configuration for the
field variable, say u (x ) , is obtained by a linear combination of these functions
using unknown coefficients or parameters, also known as degrees of freedom or
generalized coordinates. These unknown constants are then determined to satisfy
the extremum or stationary statement. Methods like the Rayleigh-Ritz procedure
7
(the one going to be used now), the Galerkin method, collocation methods, least
square methods, etc. all proceed with approximating fields chosen in this way.
For our problem, let us choose only one simple linear function, using a
polynomial form for the purpose. The advantages of using polynomial functions
will become clear later in the book. Thus, only one constant is required, i.e.
u (x ) = a1x
(1.10)
P (x ) = Aσ(x) = P (1.11d)
Thus, for this problem, the approximate solution coincides with the exact
solution. Obviously, we cannot draw any conclusions here as to the errors
involved in approximation by the R-R procedure. In more complicated problems,
approximate solutions will be built up from several constants and admissible
functions and convergence to the exact solution will depend on the number of
terms used and the type of functions used.
The R-R procedure, when applied in a piecewise manner, over the element
domains that constitute the entire structure, becomes the finite element method.
Thus, an understanding of how the R-R method works will be crucial to our
understanding of the finite element method.
8
Fig. 1.3 A two-node bar element
u (x ) = u1N 1 + u2 N 2 (1.12)
where N1 =(1-ξ)/2 and N2 =(1+ξ)/2,where ξ=x/l. N1 and N2 are called the shape
functions. Note that this form is exactly the same as the linear function used
in our R-R approximation earlier, Equation (1.10), with generalized coordinates
ai being now replaced by nodal quantities ui. It is possible therefore to
compute the energy stored in a beam element as
é 1 - 1 ù ìïu1 üï
{u1 u2 } EA
ê ú í ý (1.13)
2l êë- 1 1úû ïîu2 ïþ
{u1 u2 } ìíPP1 üý
î 2þ (1.14)
Thus, if a variation is taken over u1 and u2, we can write the equation of
equilibrium as
EA é 1 - 1ù ìu1 ü ìP1 ü
ê ú í ý = í ý
2l ë- 1 1û îu2 þ îP2 þ (1.15)
This matrix representation is typical of the way finite element equations are
assembled and manipulated automatically on the digital computer. We shall now
attempt to solve our bar problem using a finite element model comprising just
one element. Node 1 is therefore placed at the fixed end and node 2 coincides
with the free end where the load P2 = P is applied. The fixity condition at node
1 is simply u1 = 0 and P1 indicates the reaction at this end. It is very simple
to show from these that u2 = PL/EA. We can see that the exact solution has been
obtained.
9
and three dimensional problems permit complex plate, shell and three dimensional
elasticity problems to be routinely handled by ready made software packages.
10
Chapter 2
We also saw earlier that the Rayleigh-Ritz (RR) and finite element (fem)
approaches offer ways in which approximate solutions can be achieved without the
need to solve the differential equations or boundary conditions. This is managed
by performing the discretization operation directly on the functional. Thus, a
real problem with an infinitely large number of degrees of freedom is replaced
with a computational model having a finite number of degrees of freedom. In the
RR procedure, the solution is approximated by using a finite number of
admissible functions ƒi and a finite number of unknown constants ai so that the
approximate displacement field is represented by a linear combination of these
functions using the unknown constants. In the fem process, this is done in a
piecewise manner - over each sub-region (element) of the structure, the
displacement field is approximated by using shape functions Ni within each sub-
region and nodal degrees of freedom ui at nodes strategically located so that
they connect the elements together without generating gaps or overlaps. The
functional now becomes a function of the degrees of freedom (ai or ui as the
case may be). The equilibrium configuration is obtained by applying the
criterion that Π must be stationary with respect to the degrees of freedom.
11
scenario, the displacement fields are computed from these "best-fit" stresses as
a consequence.
12
2.3 Bar under uniformly distributed axial load
(
u (x ) = (q 0 EA ) Lx − x 2 2 ) (2.1a)
σ (x ) = (q 0 A ) (L − x ) (2.1b)
ur (x ) = (q 0 EA ) (Lx 2 ) (2.2a)
(
σ r (x ) = q 0 A ) (L 2 ) (2.2b)
We now compare this with an fem solution based on a two-noded linear element.
The solution obtained will be
uf (x ) = (q 0 EA ) (Lx 2 ) (2.3a)
(
σ f (x ) = q 0 A ) (L 2 ) (2.3b)
We see that the RR and fem solutions are identical. This is to be expected
because the fem solution is effectively an RR solution. We can note also the
curious coincidence where all three solutions predict the same displacement at
the tip. However, from Fig. 2.1 we can see that the ur and uf are approximate
solutions to u. It is also clear from Fig. 2.1 that σ r = σ f = σ at the mid-point
of the beam. It is possible to speculate that σ r and σ f bear some unique
relationship to the true variation σ.
Consider next what will happen if two equal length linear (i.e., two-
noded) bar elements are used to model the bar. It is left to the reader to work
out that the solution described in Fig. 2.2 will be obtained. First, we must
note that the distributed axial load is consistently lumped at the nodes. Thus
the physical load system that the fem equations are solving is not that
described in Fig. 2.1 or Fig. 2.2 as σ. Instead, we must think of a substitute
stairstep distribution σf, produced by the consistent load lumping process (see
Fig. 2.2) which is sensed by the fem stiffness matrix. Now, a solution of the
set of algebraic equations will result in σ f and uf as the fem solution.
We see once again that the nodal predictions are exact. This is only a
coincidence for this particular type of problem and nothing more can be read
into this fact. More striking is the observation that the stresses computed by
13
the finite element system now approximates the original true stress in a
stairstep fashion.
It also seems reasonable to conclude that within each element, the true
state of stress is captured by the finite element stress in a "best-fit" sense.
In other words, we can generalize from Figs. 2.1 and 2.2, that the finite
element stress magnitudes are being computed according to some precise rule.
Also, there is the promise that by carefully understanding what this rule is, it
will be possible to derive some unequivocal guidelines as to where the stresses
are accurate. In this example, where an element capable of yielding constant
stresses is used to model a problem where the true stresses vary linearly, the
centroid of the element yields exact predictions. As we take up further examples
later, this will become more firmly established.
A cursory comparison of Figs. 2.1 and 2.2 also indicates that in a general
sense, the approximate displacements are more accurate than the approximate
stresses. It seems compelling now to argue that this is so because the
approximate displacements emerge as "discretized" integrals of the stresses or
strains and for that reason, are more accurate than the stresses.
We now take up a slightly more difficult problem. The cantilever bar has a load
distributed in a linearly varying fashion (Fig. 2.3). It is easy to establish
that the exact stress distribution in this case will be quadratic in nature. We
can write it as
σ (x ) = æç q 0 L2 8A ö÷ æç 4 3 − 2ξ - æç 1 - 3ξ 2 ö÷ 3 ö÷ (2.4)
è ø è è ø ø
Some interesting features about this equation can be noted. A dimensionless
coordinate, ξ = 2x/L-1 has been chosen so that it will also serve as a natural
coordinate system taking on values -1 and 1 at the ends of the single three-node
bar element shown as modeling the entire bar in Fig.2.3. We have also very
curiously expanded the quadratic variation using the terms 1, ξ, (1-3ξ ). These
2
can be identified with the Legendre polynomials and its relevance to the
treatment here will become more obvious as we proceed further.
Fig. 2.2 Two element model of bar under uniformly distributed axial load.
14
Fig. 2.3 Bar under linearly varying axial load
This is plotted on Fig. 2.3 as the dashed line. A comparison of Equations (2.4)
and (2.5) reveals an interesting fact - only the first two Legendre polynomial
terms are retained. Taking into account the fact that the Legendre polynomials
are orthogonal, what this means is that in this problem, we have obtained σ r in
a manner that seems to satisfy the following integral condition:
ò-1 δσ r (σ r − σ ) dξ = 0
1
(2.6)
That is, the RR procedure has determined a σ r that is a "best-fit" of the true
state of stress σ in the sense described by the orthogonality condition in
Equation (2.6). This is a result anticipated from our emerging results to the
various exercises we have conducted so far. We have not been able to derive it
from any general principle that this must be so for stronger reasons than shown
here up till now.
15
( )
Let us now proceed to an fem solution. It is logical to start here with
the three-node element that uses the quadratic shape functions, N = ξ ξ − 1 2 ,
1
(
N2 = 1 − ξ
2
) and ( )
N 3 = ξ ξ + 1 2. We first compute the consistent loads to be placed
at the nodes due to the distributed load using Pi = ò N iq dx . This results in the
following scheme of loads at the three nodes identified in Fig. 2.3: P1 = q 0 L 6 ,
2
( )
(4 3 − 2ξ )
σ f x = æç q 0 L 8A ö÷
2 (2.7)
è ø
This turns out to be exactly the same as the σ r computed by the RR
process in Equation (2.5). At first sight, this does not seem to be entirely
unexpected. Both the RR process and the fem process here have started out with
quadratic admissible functions for the displacement fields. This implies that
both have the capability to represent linear stress fields exactly or more
complicated stress fields by an approximate linear field that is in some sense a
best approximation. On second thought however, there is some more subtlety to be
taken care of. In the RR process, the computed σ r was responding to a
quadratically varying system σ (see Fig. 2.3). We could easily establish
through Equation (2.6) that σ r responded to σ in a "best-fit" manner. However,
in the fem process, the load system that is being used is the σf system, which
varies in the stairstep fashion. The question confronting us now is, in what
manner did σ f respond to σf - is it also consistent with the "best-fit"
paradigm?
δ σ æç σ − 2q 0 L2 3 ö÷ dξ + æ ö
0 1
ò−1 è ø ò0 δ σ çè σ − 0 ÷ø dξ = 0 (2.8)
(
σ (x ) = q 0 L2 8A ) (4 3 − 2ξ ) , (2.9)
16
which is identical to the result obtained in Equation (2.7) by carrying out the
finite element process. In other words, the fem process follows exactly the
"best-fit" description of computing stress fields. Another important lesson to
be learnt from this exercise is that the consistent lumping process preserves
the "best-fit" nature of the stress representation and subsequent prediction.
Thus, σ f is a "best-fit" of both σ and σf! Later, we shall see how the best-fit
nature gets disturbed if the loads are not consistently derived.
Before we pass on, it is useful for future reference to note that the
approximate solutions σ r or σ f intersect the exact solution σ at two points. A
comparison of Equation (2.4) with Equations (2.5) and (2.7) indicates that these
are the points where the quadratic Legendre polynomial, (1 − 3ξ )
2
vanishes, i.e.,
at ξ = ± 1 3 . Such points are well known in the literature of the finite element
method as optimal stress points or Barlow points, named after the person who
first detected such points. Our presentation shows clearly why such points
exist, and why in this problem, where a quadratic stress state is sought to be
approximated by a linear stress state, these points are at ξ = ± 1 3 . Curiously,
these are also the points used in what is called the two-point rule for Gauss-
Legendre numerical integration. Very often, the two issues are confused. Our
interpretation so far shows that the fact that such points are sampling points
of the Gauss-Legendre rule has nothing to do with the fact that similar points
arise as points of optimal stress recovery. The link between them is that in
both, the Legendre polynomials play a decisive role - the zeros of the Legendre
polynomials define the optimal points for numerical integration and these points
also help determine where the "best-fit" approximation coincides with a
polynomial field which is exactly one order higher.
Again, the algebra is very simple and is omitted here. One can show that
the one-term approximate solution would lead to the following computed stress:
(
σ r (x ) = q 0 L2 8A ) (4 3 ) (2.10)
17
What becomes obvious by comparing this with the true stress σ(x) in Equation
(2.4) and the computed stress from the two-term solution, σ r (x ) in Equation
(2.5) is that the one-term solution corresponds to only the constant part of the
Legendre polynomial expansion! Thus, given the orthogonal nature of the Legendre
polynomials, we can conclude that we have obtained the "best-fit" state of
stress even here. Also, it is clear that the optimal stress point is not easy to
identify to coincide with any of the points corresponding to the various Gauss
integration rules. The optimal point here is given by ξ = 1 − 4 3 .
We have made out a very strong case for the best-fit paradigm. Let us now
examine the merits, if any, of the competing paradigms. The argument that fem
procedures look to satisfy the differential equations and boundary conditions
does not seem compelling enough to warrant further discussion. However, the
belief that finite elements seek to determine nodal displacements accurately was
the basis for Barlow's original derivation of optimal points [2.3] and is also
the basis for what is called the "aliasing" paradigm [2.4].
The term aliasing is borrowed from sample data theory where it is used to
describe the misinterpretation of a time signal by a sampling device. An
original sine wave is represented in the output of a sampling device by an
altered sine wave of lower frequency - this is called the alias of the true
signal. This concept can be extended to finite element discretization - the
sample data points are now the values of the displacements at the nodes and the
alias is the function, which interpolates the displacements within the element
from the nodal displacements. We can recognize at once that Barlow [2.3]
developed his theory of optimal points using an identical idea - the term
substitute function is used instead of alias.
Let us now use the aliasing concept to derive the location of the optimal
points, as Barlow did earlier [2.3], or as MacNeal did more recently [2.4]. We
assume here that the finite element method seeks discretized displacement
fields, which are substitutes or aliases of the true displacement fields by
sensing the nodal displacements directly. We can compare this with the
"best-fit" interpretation where the fem is seen to seek discretized
strain/stress fields, which are the substitutes/aliases of the true
strain/stress fields in a "best fit" or "best approximation" sense. It is
instructive now to see how the alternative paradigm, the "displacement aliasing"
approach leads to subtle differences in interpreting the relationship between
the Barlow points and the Gauss points.
3 ±1/3, ± 1 ξ4 ξ
3
ξ3 ξ2 0, ± 3 5 0, ± 3 5 0, ± √5/3
1, ξ..,ξ
4
indicate polynomial orders
18
2.5.1 A one dimensional problem
We again take up the simplest problem, a bar under axial loading. We shall
assume that the bar is replaced by a single element of varying polynomial order
for its basis function (i.e. having varying no. of equally spaced nodes). Thus,
from Table 2.1, we see that p=1,2,3 correspond to basis functions of linear,
quadratic and cubic order, implying that the corresponding elements have 2,3,4
nodes respectively. These elements are therefore capable of representing a
constant, linear and quadratic state of strain/stress, where strain is taken to
be the first derivative of the displacement field. We shall adopt the following
notation: The true displacement, strain and stress field will be designated by
u, ε and σ . The discretized displacement, strain and stress field will be
designated by u , ε and σ . The aliased displacement, strain and stress field
will be designated by u , εa and σa. Nodal displacements will be represented by
a
ui.
We shall now take for granted that the best-fit rule operates according to
the orthogonality condition expressed in Equation (2.6) and that it can be used
interchangeably for stresses and strains. We shall designate the optimal points
determined by the aliasing algorithm as ξa, the Barlow points (aliasing), and
the points determined by the best-fit algorithm as ξb, the Barlow points (best-
fit). Note that ξa are the points established by Barlow [2.3] and MacNeal [2.4],
while ξb will correspond to the points given in References 2.6 and 2.7. Note
that the natural coordinate system ξ is being used here for convenience.
ò δε (ε − ε ) dV = 0
T
(2.11)
ò δε
T
D (ξ ) (ε − ε ) dV = 0 (2.12)
Note that we can consider Equation (2.11) as a special case of the orthogonality
condition in Equation (2.12) with the weight function D(ξ)=1. This case
corresponds to one in which a straightforward application of Legendre
polynomials can be made. This point was observed very early by Moan [2.5]. In
this case, one can determine the points where ε = ε as those corresponding to
points, which are the zeros of the Legendre polynomials. See Table 2.2 for a
list of unnormalised Legendre polynomials. We shall show below that in Equation
19
(2.11), the points of minimum error are the sampling points of the Gauss
Legendre integration rule only if ε is exactly one polynomial order lower than
ε. And in Equation (2.12), the optimal points no longer depend on the nature of
the Legendre polynomials, making it difficult to identify the optimal points.
Scenario a
Linear element (p = 1)
u = quadratic = bo + b1 ξ + b2 ξ
2
(2.13)
p =1
ε = linear = u,ξ = b1 + 2b2 ξ = å ε s Ps (ξ ) (2.14)
s=0
Note that we have written ε in terms of the Legendre polynomials for future
convenience. Note also that we have simplified the algebra by assuming that
strains can be written as derivatives in the natural co-ordinate system. It is
now necessary to work out how the algebra differs for the aliasing and best-fit
approaches.
1 ξ
(1-3ξ )
2 2
3
(3ξ-5ξ )
3
4
(3-30ξ +35ξ
2 4
)
20
Quadratic element (p = 2)
u = cubic = b0 + b1 ξ + b2 ξ + b3 ξ
2 3
Note that in these two examples, i.e. for the linear and quadratic elements, the
Barlow points from both schemes coincide with the Gauss points (the points where
the corresponding Legendre polynomials vanish). In our next example we will find
that this will not be so.
Cubic element (p = 3)
ε = cubic = u,ξ
( )
= (b1 + b3 ) + (2b2 + 12b4 5 )ξ − b3 1 − 3ξ 2 − 4b4 5 3ξ − 5ξ 3 ( )
p =3
= å ε sPs (ξ ) (2.18)
s=0
(
Best-fit: u = cubic. Let ε = c0 + c1ξ + c2 1 − 3ξ 2 , as this element is capable of )
representing a quadratic strain. Equation (2.11) will now give
( )
ε = (b1 + b3 ) + (2b2 + 12b4 5 ) ξ − b3 1 − 3ξ 2 . Thus, the Barlow points (best-fit) for
21
this example are ξ b = 0, 3 5 , the points where the Legendre polynomial
(
P3 (ξ ) = 3ξ − 5ξ 3
) vanishes.
Therefore, we have an example where the aliasing paradigm does not give
the correct picture about the way the finite element process computes strains.
However, the best-fit paradigm shows that as long as the discretized strain is
one order lower than the true strain, the corresponding Gauss points are the
optimal points. Table 2.1 summarizes the results obtained so far.
The experience of this writer and many of his colleagues is that the best
-fit model, is the one that corresponds to reality. If one were to actually
solve a problem where the true strain varies cubically using a 4-noded element,
which offers a discretized strain which is of quadratic order, the points of
optimal strain actually coincide with the Gauss points.
Scenario b:
So far, we have examined simple scenarios where the true strain is exactly one
polynomial order higher than the discretized strain with which it is replaced.
If P p (ξ ) , denoting the Legendre polynomial of order p, describes the order by
which the true strain exceeds the discretized strain, the simple rule is that he
optimal points are obtained as P p (ξ b ) = 0 . These are therefore the set of p Gauss
points at which the Legendre polynomial of order p vanishes. Consider now a case
where the true strain is two orders higher than the discretized strain - e.g. a
quadratic element (p = 2) modeling a region where the strain and stress field
vary cubically. Thus, we have,
( ) (
ε = (b1 + b3 ) + (2b2 + 12b4 5 )ξ − b3 1 − 3ξ 2 − 4b4 5 3ξ − 5ξ 3 )
ε = c0 + c1ξ (2.19)
a representation made very easy by the fact that the Legendre polynomials area
orthogonal and that therefore ε can be obtained from ε by simple inspection.
It is not however a simple matter to determine whether the optimal points
coincide with other well-known points like the Gauss points. In this example, we
have to seek the zeros of
(
b3 1 − 3 ξ 2 ) + 4b 5 (3ξ − 5ξ )
4
3
(2.21)
22
element procedure yields strains ε , which are the most reasonable one can
obtain in the circumstances.
Scenario c
Our investigation here will be complete in all respects if the basis for the
best-fit paradigm can be derived logically from a basic principle. In fact, some
recent work [2.6,2.7] shows how by taking an enlarged view of the variational
basis for the displacement type fem approach we will be actually led to the
conclusion that strains or stresses are always sought in the best-fit manner.
Let the continuum linear elastic problem have an exact solution described
by the displacement field u, strain field ε and stress field σ. (We project that
the strain field ε is derived from the displacement field through the strain-
displacement gradient operators of the theory of elasticity and that the stress
field is derived from the strain field through the constitutive laws.) Let us
now replace the continuum domain by a discretized domain and describe the
computed state to be defined by the quantities u , ε and σ , where again, we
take that the strain fields and stress fields are computed from the strain-
displacement and constitutive relationships. It is clear that ε is an
approximation of the true strain field ε. What the Hu-Washizu theorem does,
following the interpretation given by Fraejis de Veubeke [2.9], is to introduce
a dislocation potential to augment the usual total potential. This dislocation
potential is based on a third independent stress field σ which can be
considered to be the Lagrange multiplier removing the lack of compatibility
appearing between the true strain field ε and the discretized strain field ε .
23
Note that, σ is now an approximation of σ . The three-field Hu-Washizu theorem
can be stated as
δπ = 0 (2.22)
where
ì T T ü
π = ò íσ ε 2 + σ (ε − ε ) + P ý dv (2.23)
î þ
In the simpler minimum total potential principle, which is the basis for the
derivation of the displacement type finite element formulation in most
textbooks, only one field (i.e., the displacement field u) is subject to
variation. However, in this more general three field approach, all the three
fields are subject to variation and leads to three sets of equations which can
be grouped and classified as follows:
T
σ Orthogonality ò δσ (ε − ε ) dv = 0 (2.24b)
(Compatibility )
æ σ − σ ö dv = 0
ε ò δε
T
Orthogonality ç ÷ (2.24c)
(Equilibrium ) è ø
Equation (2.24a) shows that the variation on the displacement field u requires
that the independent stress field σ must satisfy the equilibrium equations (∇
signifies the operators that describe the equilibrium condition. Equation
(2.24c) is a variational condition to restore the equilibrium imbalance between
σ and σ . In the displacement type formulation, we choose σ = σ . This satisfies
the orthogonality condition seen in Equation (2.24c) identically. This leaves us
with the orthogonality condition in Equation (2.24b). We can now argue that this
tries to restore the compatibility imbalance between the exact strain field ε
and the discretized strain field ε . In the displacement type formulation this
can be stated as,
ò δσ
T
(ε − ε ) dV =0 (2.25)
Thus we see very clearly, that the strains computed by the finite element
procedure are a variationally correct (in a sense, a least squares correct)
`best approximation' of the true state of strain.
24
2.7 What does the finite element method do?
2.8 Conclusions
An important part of this exercise depended on our careful choice and use
of various stress terms. Thus terms like σ and σf was actual physical states
that were sought to be modeled. The stress terms σ r and σ f were the quantities
that emerged in what we can call the "first-order tradition" analysis in the
language of Sir Karl Popper [2.10] - where the RR or fem operations are
mechanically carried out using functional approximation and finite element
stiffness equations respectively. We noticed certain features that seemed to
relate these computed stresses to the true system they were modeling in a
predictable or repeatable manner. We then proposed a mechanism to explain how
25
this could have taken place. Our bold conjecture, after examining these
numerical experiments, was to propose that it is effectively seeking a best-fit
state.
Another important step we took was to prove that the "best-fit" paradigm
was neither gratuitous nor fortuitous. In fact, we could also establish that
this could be derived from more basic principles - in this regard, the
generalized theorem of Hu was found valuable to determine that the best-fit
paradigm had a rational basis. In subsequent chapters, we shall use this
foundation to examine other features of the finite element method.
One important conclusion we can derive from the best-fit paradigm is that,
an interpolation field for the stresses σ (or stress resultants as the case may
be) which is of higher order than the strain fields ε. On which it must `do
work' in the energy or virtual work principle is actually self-defeating because
the higher order terms cannot be `sensed'. This is precisely the basis for de
Veubeke's famous limitation principle [2.9], that ‘it is useless to look for a
better solution by injecting additional degrees of freedom in the stresses.’ We
can see that one cannot get stresses, which are of higher order than are
reflected in the strain expressions.
2.9 References
2.3. J. Barlow, Optimal stress locations in finite element models, Int. J. Num.
Meth. Engng. 10, 243-251 (1976).
26
2.7. G. Prathap, A variational basis for the Barlow points, Comput. Struct. 49,
381-383, 1993.
27
Chapter 3
We can see from our study so far that the quality of approximation we can
achieve by the RR or fem approach depends on the admissible assumed trial, field
or shape functions that we use. These functions can be chosen in many ways. An
obvious choice, and the one most universally preferred is the use of simple
polynomials. It is possible to use other functions like trigonometric functions
but we can see that the least squares or best-fit nature of stress prediction
that the finite element process seeks in an instinctive way motivates us to
prefer the use of polynomials. When this is done, interpretations using Legendre
polynomials become very convenient. We shall see this happy state of affairs
appearing frequently in our study here.
An important question that will immediately come to mind is - How does the
accuracy or efficiency of the approximate solution depends on our choice of
shape functions. Accuracy is represented by a quality called convergence. By
convergence, we mean that as we add more terms to the RR series, or as we add
more nodes and elements into the mesh that replaces the original structure, the
sequence of trial solutions must approach the exact solution. We want quantities
such as displacements, stresses and strain energies to be exactly recovered,
surely, and if possible, quickly. This leads to the questions: What kind of
assumed pattern is best for our trial functions? What are the minimum qualities,
or essences or requirements that the finite element shape functions must show or
meet so that convergence is assured. Two have been accepted for a long time:
continuity (or compatibility) and completeness.
3.2 Continuity
28
this requires that the trial function φ and its derivatives through order m-1
must be continuous across element edges. In most natural formulations in solid
mechanics, strains are defined by first derivatives of the displacement fields.
In such cases, a simple continuity of the displacement fields across element
0
edges suffices - this is called C continuity. Compatibility between adjacent
0
elements for problems which require only C -continuity can be easily assured if
the displacements along the side of an element depend only on the displacements
specified at all that nodes placed on that edge.
3.3 Completeness
We have understood so far that in the finite element method, or for that matter,
in any approximate method, we are trying to replace an unknown function φ(x),
which is the exact solution to a boundary value problem over a domain enclosed
by a boundary by an approximate function φ (x ) which is constituted from a set of
trial, shape or basis functions. We desire a trial function set that will ensure
that the approximation approaches the exact solution as the number of trial
functions is increased. It can be argued that the convergence of the trial
function set to the exact solution will take place if φ (x ) will be sufficient to
represent any well behaved function such as φ(x) as closely as possible as the
number of functions used becomes indefinitely large. This is called the
completeness requirement.
In the finite element context, where the total domain is sub-divided into
smaller sub-regions, completeness must be assured for the shape functions used
within each domain. The continuity requirements then provide the compatibility
of the functions across element edges.
29
what we desire the strains or stresses to be like. Of paramount importance now
is the idea that the approximate strains derived from the set of trial functions
chosen must be capable in the limit of approximation (i.e. as the number of
terms becomes very large) to describe the true strain or stress fields exactly.
This becomes very clear from the orthogonality condition derived in Equation
2.25.
30
Consider now, the case of the element being subjected to a rigid body motion so
that at each node a displacement ui=1 is applied. It is obvious from physical
considerations that every point in the element must have displacements u=1. Thus
from Eq 3.1 we have
ii) Σ Ni = 1.
This offers a very simple check for shape functions. Another useful check comes
again from considerations of rigid body motion - that a condition of zero strain
must be produced when rigid body displacements are applied to the nodes. The
reader is encouraged to show that this entails conditions such as
Note that for C elements, where derivatives of φ are also used as nodal degrees
1
of freedom, these rules apply only to those Ni which are associated with the
translational degrees of freedom.
Let us now return to the uniform bar carrying an axial load P as shown in
Fig. 2.1. The exact solution for this problem is given by a state of constant
strain. Thus, the polynomial series u (x ) = a0 + a1x (see Section 2.3) will contain
enough terms to ensure completeness. We could at the outset arrange for a0 to be
zero to satisfy the essential boundary condition. With u1 (x ) = a1x we were able
to obtain the exact solution to the problem. We can now investigate what will
happen if we omit the a1x term and start out with the trial function set
u2 (x ) = a2 x 2 . It is left to the reader to show that this results in a computed
stress variation σ 2 = 3Px 2EAL ; i.e. a linear variation is seen instead of the
actual constant state of stress. The reader is now encouraged to experiment with
RR solutions based on other incomplete sets like u3 (x ) = a3 x 3 and
u2 +3 (x ) = a2 x 2
+ a3 x . Figure 3.1 shows how erroneous the computed stresses σ2, σ3
3
The reason for this is now apparent. The term that we have omitted is the
one needed to cover the state of constant strain. In this case, this very term
also provides the exact answer. The omission of this term has led to the loss of
completeness. In fact, the reader can verify that as long as this term is left
out of any trial set containing polynomials of higher order, there will not be
any sort of clear convergence to the true state of strain. In the present
example, because of the absence of the linear term in u (the constant strain
term therefore vanishes), the computed strain vanishes at the origin!
31
Fig. 3.1 Stresses from RR solutions based on incomplete polynomials
We have now laid down some specific qualities (continuity and completeness) that
trial functions must satisfy to ensure that approximate solutions converge to
correct results. It is also hoped that if these requirements are met, solutions
from coarse meshes will be reasonably accurate and that convergence will be
acceptably rapid with mesh refinement. This was the received wisdom or reigning
paradigm around 1977. In the next chapter, we shall look at the twin issues of
errors and convergence from this point of view.
32
Chapter 4
Error analysis is that aspect of fem knowledge, which can be said to belong to
the "second-order tradition" of knowledge. To understand its proper place, let
us briefly review what has gone before in this text. We have seen structural
engineering appear first as art and technology, starting as the imitation of
structural form by human fabrication. The science of solid and structural
mechanics codified this knowledge in more precise terms. Mathematical models
were then created from this to describe the structural behavior of simple and
complex systems analytically and quantitatively. With the emergence of cheap
computational power, techniques for computational simulation emerged. The
finite element method is one such device. It too grew first as art, by
systematic practice and refinement. In epistemological terms, we can see this
as a first-order tradition or level or inner loop of fem art and practice.
Here, we know what to do and how to do it. However, the need for an outer loop
or second-order tradition of enquiry becomes obvious. How do we know why we
should do it this way? If the first stage is the stage of action, this second
stage is now of reflection. Error analysis therefore belongs to this tradition
of epistemological enquiry.
Let us now translate what we mean by error analysis into simpler terms in
the present context. We must first understand that the fem starts with the
basic premises of its paradigms, its definitions and its operational procedures
to provide numerical results to physical problems which are already described
by mathematical models that make analytical quantification possible. The answer
may be wrong because the mathematical model wrongly described the physical
model. We must take care to understand that this is not the issue here. If the
mathematical model does not explain the actual physical system as observed
through carefully designed empirical investigations, then the models must be
refined or revised so that closer agreement with the experiment is obtained.
This is one stage of the learning process and it is now assumed that over the
last few centuries we have gone through this stage successfully enough to
accept our mathematical models without further doubt or uncertainty.
Since our computational models are now created and manipulated using
digital computers, there are errors which occur due to the fact that
information in the form of numbers can be stored only to a finite precision
(word length as it is called) at every stage of the computation. These are
called round-off errors. We shall assume here that in most problems we deal
with, word length is sufficient so that round-off error is not a major
headache.
The real issue that is left for us to grapple with is that the
computational model prepared to simulate the mathematical model may be faulty
and can lead to errors. In the process of replacing the continuum region by
finite elements, errors originate in many ways. From physical intuition, we can
argue that this will depend on the type and shape of elements we use, the
number of elements used and the grading or density of the mesh used, the way
distributed loads are assigned to nodes, the manner in which boundary
33
conditions are modeled by specification of nodal degrees of freedom, etc. These
are the discretization errors that can occur.
34
with ten elements in a one-dimensional structure, errors must not be more than
a few percent. This is the usual range of problems where the continuity and
completeness paradigms explain completely the performance of finite elements.
In subsequent chapters, we shall discover however that a class of problems
exist where errors are much larger - the discretization fail in a dramatic
fashion. Convergence and error analysis must now be founded on a more complex
conceptual framework - new paradigms need to be introduced and falsified. This
will be postponed to subsequent chapters.
The order of error analysis proceeds from the understanding that the
finite element method seeks approximations for the displacement fields. The
errors are therefore interpreted using Taylor Series expansions for the true
displacement fields and truncations of these to represent the discretized
fields. The simple example below will highlight the essential features of this
approach.
Let us now go back to the case of the cantilever bar subjected to a uniformly
distributed axial load of intensity q0 (Section 2.3). The equilibrium equation
for this problem is
u,xx + q 0 AE = 0 (4.1)
We shall use the two-node linear bar elements to model this problem. We have
seen that this leads to a solution, which gives exact displacements at the
nodes. It was also clear to us that this did not mean that an exact solution
had been obtained; in fact while the true solution required a quadratic
variation of the displacement field u(x), the finite element solution u (x) was
piecewise linear. Thus, within each element, there is some error at locations
between the nodes.
Fig. 4.1 shows the displacement and strain error in the region of an
element of length 2l placed with its centroid at xi in a cantilever bar of
length L. If for convenience we choose q 0 AE = 1 , then we can show that
u(x) = Lx − x 2 2 (4.2a)
u (x) = Lx −(l 2
+ 2xxi − xi2 ) 2 (4.2b)
35
u
∈ (x )
∈ (x )
∈
Fig. 4.1 Displacement and strain error in a uniform bar under distributed axial
load q 0
ε(x)=L–x (4.2c)
ε(x) = L − xi (4.2d)
If we denote the errors in the displacement field and strain field by e(x) and
e’(x) respectively, we can show that
e (x ) = (x − xi + l ) (x − xi − l ) 2 (4.3a)
e ′(x ) = (x − xi ) (4.3b)
From this, we can argue that in this case, the strain error vanishes at the
element centroid, x=xi, and that it is a maximum at the element nodes. This is
clear from Fig. 4.1. It is also clear that for this problem, the displacement
error is a maximum at x=xi. What is more important to us is to derive measures
for these errors in terms of the element size l, or more usefully, in terms of
the parameter h = l L , which is the dimensionless quantity that indicates the
mesh division relative to the size of the structure. It will also be useful to
have these errors normalized with respect to typical displacements and strains
occurring in the problem. From Equations (4.2) and (4.3), we can estimate the
maximum normalized errors to be of the following orders of magnitude, where “O”
stands for "order",
e u max = O h 2( ) (4.4a)
36
e′ ε max
= O (h ) (4.4b)
Note that the order of error we have estimated for the displacement field is
that for a location inside the element as in this problem, the nodal
deflections are exact. The strain error is O(h) while the displacement error is
2
O(h ). At element centroids, the error in strain vanishes. It is not clear to us
in this analysis that the discretized strain ε (x ) is a "best-fit" of the true
strain ε (x ) . If this is accepted, then it is possible to show that the error in
2
the strain energy stored in such a situation is O(h ) as well.
We have argued earlier that the finite element method actually proceeds to
compute strains/stresses in a `best-fit' manner within each element. We shall
now use this argument to show how convergence and errors in a typical finite
element model of a simple problem can be interpreted. We shall see that a much
better insight into error and convergence analysis emerges if we base our
treatment on the `best-fit' paradigm.
The example that is ideally suited for this purpose is a uniform beam
with a tip load as shown in Fig. 4.2. We shall model it with linear (two-node)
Timoshenko beam elements which represent the bending moment within each element
by a constant. Since the bending moment varies linearly over the beam for this
problem, the finite element will replace this with a stairstep approximation.
Thus, with increase in number of elements, the stress pattern will approach the
true solution more closely and therefore the computed strain energy due to
bending will also converge. Since the applied load is at the tip, it is very
easy to associate this load with the deflection under the load using
Castigliano's theorem. It will then be possible to discern the convergence for
the tip deflection. Our challenge is therefore to see if the best-fit paradigm
can be used to predict the convergence rate for this example from first
principles.
37
Fig. 4.2 Cantilever beam under tip load
The dimensions of a cantilever under tip load (Fig. 4.2) are chosen such that
the tip deflection under the load will be w=4.0. The example chosen represents
a thin beam so that the influence of shear deformation and shear strain energy
is negligible.
We shall now discretize the beam using equal length linear elements based
on Timoshenko theory. We use this element instead of the classical beam element
for several reasons. This element serves to demonstrate the features of shear
locking which arise from an inconsistent definition of the shear strains (It is
therefore useful to take up this element for study later in Chapter 6). After
correcting for the inconsistent shear strain, this element permits constant
bending and shears strain accuracy within each element - the simplest
representation possible under the circumstances and therefore an advantage in
seeing how it works in this problem.
Figure 4.3 shows the bending moment diagrams for a 1, 2 and 4 element
idealizations of the present problem using the linear element. The true bending
moment (shown by the solid line) varies linearly. The computed (i.e.
discretized) bending moments are distributed in a piecewise constant manner as
shown by the broken lines. In each case, the elements pick up the bending
moment at the centroid correctly - i.e. it is doing so in a `best-fit' manner.
What we shall now attempt to do for this problem is to relate this to the
accuracy of results. We shall now interpret accuracy in the conventional
sense, as that of the deflection at the tip under the load. Table 4.1 shows the
Table 4.1 Normalized tip deflections of a thin cantilever beam, L/t = 100
No. of elements 1 2 3
Predicted rate .750 .938 .984
Element without RBF .750 .938 .984
Element with RBF 1.000 1.000 1.000
38
normalized tip deflection with increasing idealizations (neglecting a very
small amount due to shear deformation). An interesting pattern emerges. If
error is measured by the norm {w − w (fem ) } w , it turns out that this is given
exactly by the formula 1 4N 2 where N is the number of elements. It can now be
seen that this relationship can be established by arguing that this feature
emerges from the fact that strains are sought in the `best-fit' manner shown in
Fig. 4.3.
Consider a beam element of length 2l (Fig. 4.4). Let the moment and shear force
at the centroid be M and V. Thus the true bending moment over the element
region for our problem can be taken to vary as M+Vx. The discretized bending
moment sensed by our linear element would therefore be M. We shall now compute
the actual bending energy in the element region (i.e. from a continuum
analysis) and that given by the finite element (discretized) model. We can show
that
Fig. 4.3 Bending moment diagrams for a one-, two- and four-element dealizations
of a cantilever beam under tip load.
39
Energy in continuum model = (l EI) (M 2 + V 2l 2 3) (4.5)
Energy in discretized model = (l EI) (M 2 ) (4.6)
40
We also notice that convergence in this instance is from below. This can
be deduced from the fact that the discretized potential energy U is less than
the actual potential energy U for this problem. It is frequently believed that
the finite element displacement approach always underestimates the potential
energy and a displacement solution is consequently described as a lower bound
solution. However, this is not a universally valid generalization. We can see
briefly below (the reader is in fact encouraged to work out the case in detail)
where the cantilever beam has a uniformly distributed load acting on it using
the same linear Timoshenko beam element for discretization that this is not the
case. It turns out that tip rotations converge from above (in a 1 2N 2 rate)
while the tip deflections are fortuitously exact. The lower bound solution
nature has been disturbed because of the necessity of altering the load system
at the nodes of the finite element mesh under the `consistent load' lumping
procedure.
We see from the foregoing analysis that using linear Timoshenko beam
elements for the tip loaded cantilever, the energy for this problem converges
2
as O(h ) where h=2l=L/N is the element size. We also see that this order of
convergence carries over to the estimate of the tip deflections for this
problem. Many text-books are confused over such relationships, especially those
that proceed on the order of error analysis. These approaches arrive at
conclusions such as strain error is proportional to element size, i.e. O(h) and
2
displacement error proportional to the square of the element size, i.e. O(h )
for this problem. We can see that for this problem (see Fig. 4.3) this estimate
is meaningful if we consider the maximum error in strain to occur at element
nodes (at centroids the errors are zero as these are optimal strain points). We
also see that with element discretization, these errors in strain vanish as
41
O(h). We can also see that the strain energies are now converging at the rate
2
of O(h ) and this emerges directly from the consideration that the discretized
strains are `best-fits' of the actual strain. This conclusion is not so readily
arrived at in the order of error analysis methods, which often argue that the
2
strains are accurate to O(h), then strain energies are accurate to O(h ) because
strain energy expressions contain squares of the strain. This conclusion is
valid only for cases where the discretized strains are `best-fit'
approximations of the actual strains, as observed in the present example. If
the `best-fit' paradigm did not apply, the only valid conclusion that could be
drawn is that the strains that have O(h) error will produce errors in strain
energy that are O(2h).
MacNeal [4.2] describes the mathematics of the RBF correction using an original
cubic lateral displacement field (corresponding to a linearly varying bending
moment field) and what is called a aliased linearly interpolated field (giving
a constant bending moment variation for the discretized field). We might recall
that we examined the aliasing paradigm in Section 2.5 earlier. We find that a
quicker and more physically insightful picture can be obtained by using
Equations (4.5) and (4.6) above. Consider a case where only one element of
42
length 2l is used to model a cantilever beam of length L with a tip load P.
Applying Castigliano's theorem, we can show very easily that the continuum and
discretized solutions will differ by,
(
w L = P L2 3EI + 1 kAG ) (4.7)
w L = P (L2
4EI + 1 kAG ) (4.8)
Note the difference in bending flexibilities. This describes the inherent
approximation involved in the discretization process if all variational norms
are strictly followed. The RBF correction proposes to manipulate the shear
flexibility in the discretized case so that it compensates for the deficiency
*
in the bending flexibility for this problem. Thus if k is the compensated shear
correction factor, from Eq (4.7) and (4.8), we have
It is not accurate to say here that the correction term is derived from the
bending flexibility. The bending flexibility of an element that can represent
only a constant bending moment (e.g. the two-noded beam element used here) is
L2 4EI . The missing L2 12EI (or l 2 3EI ) is now brought in as a compensation
*
through the shear flexibility, i.e. k is changed to k . This "fudge factor"
therefore enhances the performance of the element by giving it a rate of
convergence that is not variationally permissible. Two wrongs make a right
here; or do they? Table 4.1 shows how the convergence in the problem shown in
Fig. 4.2 is changed by this procedure.
We have shown that the best-fit paradigm is a useful starting point for
deriving estimates about errors and convergence. Using a simple example and
this simple concept, we could establish that there is a variationally correct
rate of convergence for each element. This can be improved only by taking
liberties with the formulation, i.e. by introducing extra-variational steps.
The RBF is one such extra-variational device.
43
such errors has been one of the most challenging problems that the finite
element procedure has faced in its history. Most of the early techniques to
overcome these difficulties were ad hoc, more in the form of `art' or `black
magic'. In the subsequent chapters, our task will be to identify the principles
that establish the methodology underlying the finite element procedure using
critical, rational scientific criteria.
4.7 References
44
Chapter 5
By and large, the elements work without difficulty. However, there were
spectacular failures as well. These are what are now called the ‘locking’
0
problems in C finite elements - as the element libraries of GPPs stabilized
these elements came to be favored for reasons we shall discuss shortly. By
locking, we mean that finite element solutions vanish quickly to zero (errors
saturating quickly to nearly 100%!) as certain parameters (the penalty
multipliers) become very large. It was not clear why the displacement type
method, as it was understood around 1977, should produce for such problems,
answers that were only a fraction of a per cent of the correct answer with a
practical level of discretization. Studies in recent years have established that
an aspect known as consistency must be taken into account.
45
5.2 From C1 to C0 elements
As the second generation of GPPs started evolving around the late 70s and early
1
80s, their element libraries replaced what were called the C elements with what
0
were known as the C elements. The former were based on well known classical
theories of beams, plates and shells, reflecting the confidence structural
analysts had in such theories for over two centuries - namely the Euler-
Bernoulli beam theory, the Kirchhoff-Love plate theory and the equivalent shell
theories. These theories did not allow for transverse shear strain and permitted
the modeling of such structures by defining deformation in terms of a single
field, w, the transverse deflection of a point on what is called the neutral
axis (in a beam) and neutral surface of a plate or shell. The strains could then
be computed quite simply from the assumption that normal to the neutral surface
remained normal after deformation. One single governing differential equation
resulted, although of a higher order (in comparison to other theories we shall
discuss shortly), and this was considered to be an advantage.
There were some consequences arising from such an assumption both for the
mathematical modeling aspect as well as for the finite element (discretization)
aspect. In the former, it turned out that to capture the physics of deformation
of thick or moderately thick structures, or the behavior of plates and shells
made of newly emerging materials such as high performance laminated composites,
it was necessary to turn to more general theories accounting or transverse shear
deformation as well - these required the definition of rotations of normals
which were different from the slopes of the neutral surface. Some of the
1
contradictions that arose as a result of the old C theories - e.g. the use of
the fiction of the Kirchhoff effective shear reactions, could now be removed,
restoring the more physically meaningful set of three boundary conditions on the
edge of a plate or shell (the Poisson boundary conditions as they are called) to
be used. The orders of the governing equations were correspondingly reduced. A
salutary effect that carried over to finite element modeling was that the
elements could be designed to have nodal degrees of freedom which were the six
basic engineering degrees of freedom - the three translations and the three
rotations at a point. This was ideal from the point of view of the organization
of a general purpose package. Also, elements needed only simple basis functions
requiring only the continuity of the fields across element boundaries - these
0 1
are called the C requirements. In the older C formulations, continuity of
slope was also required and to achieve this in arbitrarily oriented edges, as
would be found in triangular or quadrilateral planforms of a plate bending
element, it was necessary to retain curvature degrees of freedom (w,xx, w,xy,
w,yy) at the nodes and rely on quintic polynomials for the element shape or
basis functions. So, as general purpose packages ideal for production run
0
analyses and design increasingly found favour in industry, the C beam, plate
1
and shell elements slowly began to replace the older C equivalents. It may be
instructive to note that the general two-dimensional (i.e. plane stress, plane
strain and axisymmetric) elements and three-dimensional (solid or brick as they
0
are called) elements were in any case based on C shape functions - thus this
0
development was welcome in that universally valid C shape functions and their
derivatives could be used for a very wide range of structural applications.
46
However, life was not very simple - surprisingly dramatic failures came to
be noticed and the greater part of academic activity in the late seventies, most
of the eighties and even in the nineties was spent in understanding and
eliminating what were called the locking problems.
When locking was first encountered, efforts were made to associate it with the
rank or non-singularity of the stiffness matrix linked to the penalty term (e.g.
the shear stiffness matrix in a Timoshenko beam element which becomes very large
as the beam becomes very thin, see the discussion below). However, on
reflection, it is obvious that these are symptoms of the problem and not the
cause. The high rank and non-singularity is the outcome of certain assumptions
made (or not made, i.e. leaving certain unanticipated requirements unsatisfied)
during the discretization process. It is therefore necessary to trace this to
the origin. The consistency approach argues that it is necessary in such
problems to discretize the penalty-linked strain fields in a consistent way so
that only physically meaningful constraints appear.
Π =
L
ò (1 2 EI w,2xx − qw ) dx (5.1)
0
47
in terms of a single dependent variable, the transverse deflection w of points
on the neutral axis of the beam. An application of the principle of minimum
total potential allows the governing differential equations and boundary
conditions to be recovered, but this will not be entered into here. A simple
exercise will establish that the exact solution satisfying the governing
differential equations and boundary conditions is,
( )
w (x ) = qL2 4EI x 2 − (qL 6EI ) x 3 − (q 24EI ) x 4 (5.2a)
EI w,xx = (q 2 ) (L − x )2 (5.2b)
EI w,xxx = − q (L − x ) (5.2c)
anticipating the boundary conditions at the fixed end. One can easily work out
that the approximate solution will emerge as,
( )
w (x ) = 5qL2 24EI x 2 − (qL 12EI ) x 3 (5.3a)
so that the approximate bending moment and shear force determined in this Ritz
process are,
(
EI w ,xx = 5qL2 12 − (qL 2 ) x ) (5.3b)
EI w ,xxx = − qL 2 (5.3c)
If the expressions in Equations (5.2) and (5.3) are written in terms of the
natural co-ordinate system ξ, where x = (1 + ξ ) L 2 so that the ends of the beam
are represented by ξ = -1 and +1, the exact and approximate solutions can be
expanded as,
( )(
EI w,xx = qL2 8 4 3 − 2ξ − 1 3 (1 − 3ξ 2 ) ) (5.4)
48
5.3.2 The Timoshenko beam theory and its penalty function form
The Timoshenko beam theory [5.1] offers a physically more general formulation of
beam flexure by taking into account the transverse shear deformation. The
description of beam behavior is improved by introducing two quantities at each
point on the neutral axis, the transverse deflection w and the face rotation θ
so that the shear strain at each point is given by γ = θ − w,x , the difference
between the face rotation and the slope of the neutral axis.
The total strain energy functional is now constructed from the two
independent functions for w(x) and θ(x), and it will now account for the bending
(flexural) energy and an energy of shear deformation.
Π =
0
ò
L
(
1 2 EI θ ,2x + 1 2 α (θ − w,x )2 − q w ) dx (5.6)
where the curvature κ=θ,x, the shear strain γ=θ-w,x and α=kGA is the shear
rigidity. The factor k accounts for the averaged correction made for the shear
strain distribution through the thickness. For example, for a beam of
rectangular cross-section, this is usually taken as 5/6.
The Timoshenko beam theory will asymptotically recover the elementary beam
theory as the beam becomes very thin, or as the shear rigidity becomes very
large, i.e. α→∞. This requires that the Kirchhoff constraint θ-w,x→0 must
emerge in the limit. For a very large α, these equations lead directly to the
simple fourth order differential equation for w of elementary beam theory. Thus,
this is secured very easily in the infinitesimal theory but it is this very same
point that poses difficulties when a simple Ritz type approximation is made.
Consider now a two term Ritz approximation based on θ = a1x and w = b1x . This
permits a constant bending strain (moment) approximation to be made. The shear
strain is now given by
γ = a1x − b1 (5.7)
and it would seem that this can represent a linearly varying shear force. The
Ritz variational procedure now leads to the following set of equations,
ìï éL 0 ù é L3 3 − L2 2 ù üï ìa1 ü ìï 0 üï
íEI ê ú + α ê ú ý í ý=í 2 ý (5.8)
ïî ë0 0 û êë − L2 2 L úû þï î b1 þ ïîqL 2 ïþ
Solving, we get
− 3qL2 qL 1.5qL3
a1 = ; b1 = − − (5.9)
12EI + αL2 2α 12EI + αL2
49
As α→∞, both a1 and b1→0. This is a clear case of a solution `locking'. This
could have been anticipated from a careful examination of Equations (5.6) and
(5.7). The penalty limit α→∞ in Equation (5.6) introduces a penalty condition
on the shear strain and this requires that the shear strain must vanish in the
Ritz approximation process - from Equation (5.7), the conditions emerging are
a1→0 and b1→0. Clearly, a1→0 imposes a zero bending strain θ,x→0 as well -
this spurious constraint produces the locking action on the solution. Thus, a
meaningful approximation can be made for a penalty function formulation only if
the penalty linked approximation field is consistently modeled. We shall see how
this is done next.
Consider now a three term approximation chosen to satisfy what we shall call the
consistency condition. Choose θ = a1x and w = b1x + b2 x 2 . Again, only a constant
bending strain is permitted. But we now have shear strain of linear order as,
Note that as α→∞, the constraint α1-2b2→0 is consistently balanced and will
not result in a spurious constraint on the bending field. The Ritz variational
procedure leads to the following equation structure:
ì é L3 3 − L2 2 − 2L3 3 ù ü ì 0 ü
ï éL 0 0 ù ê ú ï ìa1 ü
ï ê ú ï ï ï ïï 2 ïï
í EI ê0 0 0 ú + α ê − L 2 L2 ú í b1 ý = íqL 2 ý
2
L ý (5.11)
ï ê ú ï ïb ï ï 3 ï
êë0 0 0 úû ê − 2L 3
3
L2 4L3 3 ú î 2þ ïîqL 3 ïþ
ïî ë û ïþ
It can be easily worked out from this that the approximate solution is given by,
qL2
θ =− x (5.12a)
6EI
qL q 2 qL2
w = x− x + x2 (5.12b)
α 2α 12EI
− EI θ ,x = qL2 6 (5.12c)
( )
α θ − w ,x = q (L − x ) (5.12d)
There is no locking seen at all - the bending moment is now a least squares
correct constant approximation of the exact quadratic variation (this can be
seen by comparing Equation (5.12c) with Equations (5.4) and (5.5) earlier). The
shear force is now correctly captured as a linear variation - the consistently
represented field in Equation (5.12) being able to recover this even as a α→∞!
50
A comparison of the penalty linked matrices in Equations (5.8) and (5.11)
shows that while in the former, we have a non-singular matrix of rank 2, in the
latter we have a singular matrix of rank 2 for a matrix of order 3. It is clear
also that the singularity (or reduced rank) is a result of the consistent
condition represented by (α1-2b2) in the linear part of the shear strain
definition in Equation (5.10) - as a direct consequence, the third row of the
penalty linked matrix in Equation (5.11) is exactly twice the first row. It is
this aspect that led to a lot of speculation on the role the singularity or rank
of the matrix plays in such problems. We can further show that non-singularity
of penalty linked matrix arises in an inconsistent formulation only when the
order of approximation is low, as seen for the two-term Ritz approximation. We
can go for a quadratic inconsistent approximation (with linear bending strain
variation) in the next section to show that there is no locking' of the solution
and that the penalty linked matrix is not non-singular - however the effect of
inconsistency is to reduce the performance of the approximation to a sub-optimal
level.
Note now that the condition α→∞ forces a2→0 this becomes a spurious constraint
on the bending strain field. We shall now see what effect this spurious
constraint has on the approximation process. The Ritz variational procedure
leads to the following set of equations:
ì éL L2 0 0ù
ï ê ú
ï ê 2 ú
ï êL 4L3 3 0 0ú
í EI +
ê ú
ï ê0 0 0 0ú
ï ê ú
ï
î ëê0 0 0 0 ûú
(5.14)
é L3 3 L4 4 − L2 2 − 2L3 3 ù ü ì 0 ü
ê ú ï ìa1 ü
ï ï ï ï
ê L4 4 L5 5 − L3 3 − L4 2 ú ïï ïa2 ï ï 0 ï
α ê ú ý í ý = í 2 ý
ê − L2 2 − L3 3 L L2 ú ï ï b1 ï ïqL 2 ï
ê ú ï ï b2 ï ï 3 ï
î þ
ëê − 2 L 3 − L4 2 îqL 3 þ
3
L2 4L3 3 ûú ïþ
It can be seen that the penalty linked 4×4 matrix is singular as the fourth row
is exactly twice the second row - this arises from the consistent representation
(a1 − 2b2 ) of the linear part of shear strain in Equation (5.13). The rank of the
51
matrix is therefore 3 and the solution should be free of "locking" - however the
inconsistent constraint forces a2 → 0 and this means that the computed bending
strain κ → a1 ; i.e. it will have only a constant bending moment prediction
capability. What it means is that this four term inconsistent approach will
produce answers only as efficiently as the three term consistent Ritz
formulation. Indeed, one can work out from Equation (5.14) that,
qL2 15qL2
a1 = − − (5.15a)
6EI 60EI + αL2
15qL2
a2 = (5.15b)
60EI + αL2
qL 5qL3
b1 = −
α
−
(
2 60EI + αL2 ) (5.15c)
qL2 q
b2 = − + (5.15d)
6EI 2α
As α→∞, the bending moment and shear force variation are given by,
− EI θ ,x = qL2 6 (5.16a)
( )
α θ − w ,x = q (L − x ) + 2.5qL (6 (x L )2 − 6 (x L ) + 1 ) (5.16b)
The solution can sense only a constant bending moment in the thin beam limit.
There are now violent quadratic oscillations in the shear force and these
oscillations can be shown to vanish at the points (
x L = 1 2 1+1 3 ) and
( )
1 2 1 − 1 3 , or ξ = ± 1 3 , which are the points corresponding to the 2 point
Gauss-Legendre integration rule. The effect of the inconsistent representation
has been to reduce the effectiveness of the approximation. We shall next see how
the effectiveness can be improved by making the approximation consistent before
the variational process is carried out.
52
γ = − b1 + (a1 − 2b2 )x + a2 Lx − L2 6( )
(
= − b1 − a2 L2 6 )+ (a1 + a2 L - 2b2 )x (5.17)
In fact, it can be proved using the generalized (mixed) variational theorem such
as the Hellinger-Reissner and the Hu-Washizu principles [5.2] that the
variationally correct way to determine γ from the inconsistent γ is,
T
ò δ γ æç γ − γ ö÷ dx = 0
è ø
and this will yield precisely the form represented in Equation (5.17). The Ritz
variational procedure now leads to,
ì éL L2 0 0ù
ï ê ú
ï ê 2 ú
ï êL 4L3 3 0 0ú
í EI ê +
ú
ï ê0 0 0 0ú
ï ê ú
ï êë0 0 0 0 úû
î
(5.18)
é L3 3 L4 4 − L2 2 − 2L3 3 ù ü ì 0 ü
ê ú ï ìa1 ü
ï ï ï ï
ê L4 4 7L5 36 − L3 3 − L4 2 ú ïï ïa2 ï ï 0 ï
α ê ú ý í ý = í 2 ý
ê − L2 2 − L3 3 L L2 ú ï ï b1 ï ïqL 2 ï
ê ú ï ï b2 ï ï 3 ï
î þ
ëê − 2 L 3 − L4 2 îqL 3 þ
3
L2 4L3 3 ûú ïþ
It is important to recognize now that since the penalty linked matrix emerges
( )
from the terms - b1 + a2 L2 6 and (a1 + a2 L - 2b2 ) there would only be two linearly
independent rows and therefore the rank of the matrix is now 2. The approximate
solution is then given by,
5 qL
− EI κ = qL2 − x (5.19a)
12 2
α γ = q (L - x ) (5.19b)
Comparing Equation (5.19a) with Equation (5.3b) we see that the approximate
solution to the Timoshenko equation for this problem with a consistent shear
strain assumption gives exactly the same bending moment as the Ritz solution to
the classical beam equation could provide. We also see that Equation (5.19b) is
identical to Equation (5.12b) so that the shear force is now being exactly
computed. In other words, as α→∞, the terms a1, a2, b1 and b1 all yield
physically meaningful conditions representing the state of equilibrium
correctly.
53
5.4 Consistency and C0 displacement type finite elements
We shall see in the chapters to follow that locking, poor convergence and
0
violent stress oscillations seen in C displacement type finite element
formulations are due to a lack of consistent definition of the critical strain
fields when the discretization is made - i.e. of the strain-fields that are
constrained in the penalty regime.
The foregoing analysis showed how the lack of consistency translates into
a non-singular matrix of full rank that causes locking in low-order Ritz
approximations of such problems. It is also seen that in higher order
approximations, the situation is not as dramatic as to be described as locking,
but is damaging as to produce poor convergence and stress oscillations. It is
easy to predict all this by examining the constrained strain-field terms from
the consistency point of view rather than performing a post-mortem examination
of the penalty-linked stiffness matrix from rank or singularity considerations
as is mostly advocated in the literature.
54
strain-field, the constraint that appears in the limit can be correctly
enforced. We shall call such a representation field-consistent. The constraints
thus enforced are true constraints. Where the strain-field has coefficients in
which the contributions from some of the field variables are absent, the
constraints may incorrectly constrain some of these terms. This field-
inconsistent formulation is said to enforce additional spurious constraints. For
simple low order elements, these constraints are severe enough to produce
solutions that rapidly vanish - causing what is often described as locking.
55
strains appear. These examples show the universal need of the consistency aspect
and also the power of the general Hu-Washizu theorem to establish the complete,
correct and consistent variational basis for the displacement type finite
element method.
5.6 References
56
Chapter 6
Shear locking
In the preceding chapter, we saw a Ritz approximation of the Timoshenko beam
problem and noted that it was necessary to ensure a certain consistent
relationship between the trial functions to obtain accurate results. We shall
now take up the finite element representation of this problem, which is
essentially a piecewise Ritz approximation. Our conclusions from the preceding
chapter would therefore apply to this as well.
An element based on elementary theory needs two nodes with 2 degrees of freedom
at each node, the transverse deflection w and slope dw/dx and uses cubic
1
interpolation functions to meet the C continuity requirements of this theory
(Fig. 6.1). A similar two-noded beam element based on the shear flexible
0
Timoshenko beam theory will need only C continuity and can be based on simple
linear interpolations. It was therefore very attractive for general purpose
applications. However, the element was beset with problems, as we shall
presently see.
(
ò 1 2 EI χ χ + 1 2 kGA γ γ dx
T T
) (6.1)
where
χ = θ ,x (6.2a)
γ = θ − w,x (6.2b)
(a) 1 2
w w
w,x w,x
(b) 1 2
w w
θ θ
Fig. 6.1 (a) Classical thin beam and (b) Timoshenko beam elements.
57
In the conventional procedure, linear interpolations are chosen for the
displacement field variables as,
N 1 = (1 − ξ ) 2 (6.3a)
N 2 = (1 + ξ ) 2 (6.3b)
For the beam element shown in Fig. 6.1, for a length h the stiffness
matrix can be split into two parts, a bending related part and a shear related
part, as,
é0 0 0 0ù é 1 h2 −1 h 2ù
ê ú ê ú
ê0 1 0 − 1ú êh 2 h 2 3 - h 2 h2 6 ú
kb = EI h ê ú ks = kGt h ê ú
ê0 0 0 0ú ê -1 - h 2 1 - h 2 úú
ê ú ê
ê0 ê ú
ë − 1 0 1 úû 2
êëh 2 h 6 - h 2 h 2 3 úû
We shall now model a cantilever beam under a tip load using this element,
considering the case of a "thin" beam with E=1000, G=37500000, t=1, L=4, using a
fictitiously large value of G to simulate the "thin" beam condition. Table 6.1
shows that the normalized tip displacements are dramatically in error. In fact
with a classical beam element model, exact answers would have been obtained with
one element for this case. We can carefully examine Table 6.1 to see the trend
as the number of elements is increased. The tip deflections obtained, which are
several orders of magnitude lower than the correct answer, are directly related
to the square of the number of elements used for the idealization. In other
words, the discretization process has introduced an error so large that the
2
resulting answer has a stiffness related to the inverse of N . This is clearly
unrelated to the physics of the Timoshenko beam and also not the usual sort of
discretization errors encountered in the finite element method. It is this very
phenomenon that is known as shear locking.
0.800 × 10
-5
2
0.320 × 10
-4
4
0.128 × 10
-3
8
0.512 × 10
-3
16
58
The error in each element must be related to the element length, and
therefore when a beam of overall length L is divided into N elements of equal
length h, the additional stiffening introduced in each element due to shear
2
locking is seen to be proportional to h . In fact, numerical experiments showed
that the locking stiffness progresses without limit as the element depth t
decreases. Thus, we now have to look for a mechanism that can explain how this
2
spurious stiffness of (h/t) can be accounted for by considering the mathematics
of the discretization process.
é0 0 0 0 ù é1 h 2 −1 h 2ù
ê ú ê ú
ê0 1 0 − 1ú êh 2 h 2 4 − h 2 h 2 4 ú
kb = EI h ê ú ks = kGt h ê ú
ê0 0 0 0 ú ê− 1 − h 2 1 − h 2 úú
ê ú ê
ê0 ê ú
− 1 0 1úû
ë ëêh 2 h 4 − h 2 h 4 úû
2 2
It is clear from the formulation of the linear Timoshenko beam element using
exact integration (we shall call it the field-inconsistent element) that
ensuring the completeness and continuity conditions are not enough in some
problems. We shall propose a requirement for a consistent interpolation of the
constrained strain fields as the necessary paradigm to make our understanding of
the phenomena complete.
w = a0 + a1 (x l ) (6.4a)
θ = b0 + b1 (x l ) (6.4b)
59
If the strain-fields are now derived from the displacement fields given in
Equation (6.4), we get
χ = (b1 l ) (6.5a)
γ = (b0 − a1 l ) + b1 (x l ) (6.5b)
An exact evaluation of the strain energies for an element of length h=2l will
now yield the bending and shear strain energy as
b0 − a1 l → 0 (6.7a)
b1 → 0 (6.7b)
60
6.1.4 Functional re-constitution
We have postulated that the error of shear locking originates from the spurious
shear constraint in Equation (6.7b). We must now devise an error model for the
case where the inconsistent element is used to model a beam of length L and
depth t. The strain energy for such a beam can be set up as,
Π=
0
ò
L
{1 2 EI θ ,2x + 1 2 kGA (θ − w,x )2 dx} (6.8)
Thus the constants in the discretized strain energy functional have been re-
constituted into an equivalent continuum or infinitesimal form. From this re-
constituted functional, we can argue that an idealization of a beam region of
length 2l into a linear displacement type finite element would produce a
modified strain energy density within that region of,
(
π e = 1 2 EI + kGAl2 3 ) (θ,x )2 + 1 2 (kGA ) (θ − w,x )2 (6.10)
This strain energy density indicates that the original physical system has been
altered due to the presence of the inconsistent term in the shear strain field.
Thus, we can postulate that a beam of length L modeled by equal elements of
length 2l will have a re-constituted functional
Π =
0
ò {
1 2 (EI + kGAl2 3 ) (θ ,x )2 + 1 2 (kGA ) (θ − w,x )2 }dx
L
(6.11)
We now understand that the discretized beam is stiffer in bending (i.e. its
flexural rigidity) by the factor kGAl2 3EI . For a thin beam, this can be very
large, and produces the additional stiffening effect described as shear locking.
61
We must now show through a numerical experiment that this estimate for the
error, which has been established entirely a priori, starting from the
consistency paradigm and introducing the functional re-constitution technique,
anticipates very accurately, the behavior of a field-inconsistent linearly
interpolated shear flexible element in an actual digital computation. Exact
solutions are available for the static deflection W of a Timoshenko cantilever
beam of length L and depth t under a vertical tip load. If W fem is the result
from a numerical experiment involving a finite element digital computation using
elements of length 2l, the additional stiffening can be described by a parameter
as,
From Equation (6.12), we already have an a priori prediction for this factor as,
We can now re-interpret the results shown in Table 6.1 for the thin beam
case. Using Equations (6.13) and (6.14), we can argue a priori that the
inconsistent element will produce normalized tip deflections (W fem W ) = 1 (1 + e ) .
Since e>>1, we have
( )
W fem W = N 2 5 × 10 −5 (6.15)
for the thin beam. Table 6.2 shows how the predictions made thus compare with
the results obtained from an actual finite element computation using the field-
inconsistent element.
This has shown us that the consistency paradigm can be scientifically verified.
Traditional procedures such as counting constraint indices, or computing the
rank or condition number of the stiffness matrices could offer only a heuristic
picture of how and why locking sets in.
Table 6.2 - Normalized tip deflections for the thin beam (Case 2) computed from
fem model and predicted from error model (Equation (6.15)).
62
large. The percentage error norms therefore saturate quickly to a value
approaching 100% and do not sensibly reflect the relationship between error and
the structural parameter even on a logarithmic plot. A new error norm called the
additional stiffening parameter, e can be introduced to recognize the manner in
which the errors of locking kind can be blown out of proportion by a large
variation in the structural parameter. Essentially, this takes into account, the
fact that the spurious constraints give rise to a spurious energy term and
consequently alters the rigidity of the system being modeled. In many other
examples (e.g. Mindlin plates, curved beams etc.) it was seen that the rigidity,
I, of the field consistent system and the rigidity, I’, of the inconsistent
system, were related to the structural parameters in the form, I’/I = α(l/t)
2
2 2
Fig. 6.2 Variation of error norms e, E with structural parameter kGL /Et for a
cantilever beam under tip shear force.
63
by this definition, we shall anticipate again, the results above. If we
represent the conventional error norm in the form E = (W − W fem ) W , and plot both
E and the new error norm e from the results for the same problem using 4 FI
2
elements against the penalty multiplier (l/t) on a logarithmic scale, the
dependence is as shown in Fig. 6.2. It can be seen that E saturates quickly to a
value approaching 100% and cannot show meaningfully how the error propagates as
the penalty multiplier increases indefinitely. On the other hand, e captures
this relationship, very accurately.
V = V + kGA b1 (x l ) (6.16b)
We see that V has a linear term that relates directly to the constant that
appeared in the spurious constraint, Equation (6.7b). We shall see below from
Equation (6.17) that b1 will not be zero, in fact it is a measure of the bending
moment at the centroid of the element. Thus, in a field-inconsistent
formulation, this constant will activate a violent linear shear force variation
when the shear forces are evaluated directly from the shear strain field given
in Equation (6.5b). The oscillation is self-equilibrating and does not
contribute to the force equilibrium over the element. However, it contributes a
finite energy in Equation (6.9) and in the modeling of very slender beams, this
spurious energy is so large as to completely dominate the behavior of the beam
and cause a locking effect.
64
Fig. 6.3 Shear force oscillations in element nearest the root, for N element
models of a cantilever of length L = 60.
Consider a straight cantilever beam with a tip shear force Q at the free
end. This should produce a linearly varying bending moment M and a constant
shear force Q in the beam. An element of length 2l at any station on the beam
will now respond in the following manner. Since, a linear element is used, only
the average of the linearly varying bending moment is expected in each finite
element. If the element is field-consistent, the constant b1 can be associated
after accounting for discretization, to relate to the constant bending moment M0
at the element centroid as,
M 0 = EI b1 l or
b1 = M 0l EI (6.17)
b1′ = M 0l EI ′
= {M 0l EI (1 + e )}
= b1 (1 + e ) (6.18)
65
same factor. Also, for a very thin beam where e>>1, the centroidal moment M0
2
predicted by a field-consistent element diminishes in a t rate for a beam of
rectangular cross-section. These observations have been confirmed through
digital computation.
V = Q + (3M 0 l ) (x l ) (6.19)
These are the violent shear force linear oscillations within each element, which
originate directly from the field-inconsistency in the shear strain definition.
We can see that reduced integration ensures that the inconsistent constraint
does not appear and so is effective in producing a consistent element, at least
in this instance. We must now satisfy ourselves that such a modification did not
violate any variational theorem.
We explain now why the use of a trick like the reduced integration
technique, or the use of assumed strain methods allows the locking problem to be
overcome. It is obvious that it is not possible to reconcile this within the
ambit of the minimum total potential principle only, which had been the starting
point of the conventional formulation.
66
To eliminate problems such as locking, we look for a consistent
constrained strain field to replace the inconsistent kinematically derived
strain field in the minimum total potential principle. By closely examining the
strain gradient operators, it is possible to identify the order up to which the
consistent strain field must be interpolated. In this case, for the linear
displacement interpolations, Equations (6.5b), (6.7a) and (6.7b) tell us that
the consistent interpolation should be a constant. At this point we shall still
not presume what this constant should be, although past experience suggests it
is the same constant term seen in Equation (6.7a). Instead, we bring in the
Hellinger-Reissner theorem in the following form to see the identity of the
consistent strain field clearly. For now, it is sufficient to note that the
Hellinger-Reissner theorem is a restricted case of the Hu-Washizu theorem. In
this theorem, the functional is stated in the following form,
(
ò − 1 2 EI χ χ + EI χ χ − 1 2 kGA γ
T T T
γ + kGA γ T
)
γ dx (6.20)
where χ and γ are the new strain variables introduced into this multi-field
principle. Since we have difficulty only with the kinematically derived γ we can
have χ = χ and recommend the use of a γ which is of consistent order to replace
γ. A variation of the functional in Equation (6.20) with respect to the as yet
undetermined coefficients in the interpolation for γ yields
ò δγ
T
(γ − γ ) dx = 0 (6.21)
So far we have seen the linear beam element as an example to demonstrate the
principles involved in the finite element modeling of a constrained media
problem. We have been able to demonstrate that a conceptual framework that
includes a condition that specifies that the strain fields which are to be
constrained must satisfy a consistency criterion is able to provide a complete
scientific basis for the locking problems encountered in conventional
displacement type modeling. We have also shown that a correctness criterion
(which links the assumed strain variation of the displacement type formulation
67
to the mixed variational theorems) allows us to determine the consistent strain
field interpolation in a unique and mathematically satisfying manner.
It will be useful now to see how these concepts work if a quadratic beam
element is to be designed. This is a valuable exercise as later, the quadratic
beam element shall be used to examine problems such as encountered in curved
beam and shell elements and in quadrilateral plate elements due to non-uniform
mapping.
We shall now very quickly see how the field-consistency rules explain the
behavior of a higher order element. We saw in Chapter 5 that the conventional
formulation with lowest order interpolation functions led to spurious
constraints and a non-singular assembled stiffness matrix, which result in
locking. In a higher order formulation, the matrix was singular but the spurious
constraints resulted in a system that had a higher rank than was felt to be
desirable. This resulted in sub-optimal performance of the approximation. We can
now use the quadratic beam element to demonstrate that this is true in finite
element approximations as well.
68
Fig. 6.4 A uniform cantilever beam with tip shear force -
convergence trends of linear and quadratic elements.
w = a0 + a1 (x l ) + a2 (x l )2
θ = b0 + b1 (x l ) + b2 (x l )2
(
γ = (b0 + b2 3 − a1 l ) + (b1 − 2a2 l )ξ − b2 3 1 − 3ξ 2 ) (6.22)
Again, we emphasize the usefulness of expanding the strain field in terms of the
Legendre polynomials. When the strain energies are integrated, because of the
orthogonal nature of the Legendre polynomials the discretized energy expression
becomes the sum of the squares of the coefficients multiplying the Legendre
polynomials. Indeed, the strain energy due to transverse shear strain is,
Therefore, when we introduce the penalty limit condition that for a thin
beam the shear strain energies must vanish, we can argue that the coefficients
of the strain field expanded in terms of the Legendre polynomials must vanish
separately. In this case, three constraints emerge:
(b0 + b2 3 − a1 l ) → 0 (6.24a)
69
(b1 − 2a2 l ) → 0 (6.24b)
b2 → 0 (6.24c)
The b2 term that appears in the bending energy also makes an appearance in
the shear strain energy, reflecting its origin through the spurious constraint.
We can argue that this accounts for the poor behavior of the field-inconsistent
quadratic beam element (the 3I of Fig. 6.4). Ref. 6.1 derives the effect more
precisely, demonstrating that the following features can be fully accounted for:
70
6.2.3 The consistent formulation of the quadratic element
γ = c0 + c1ξ (6.25)
A very large part of structural analysis deals with the estimation of stresses
and displacements in thin flexible structures under lateral loads using what is
called plate theory. Thus, plate elements are the most commonly used elements in
general purpose structural analysis. At first, most General Purpose Packages
(GPPs) for structural analysis used plate elements based on what are called the
1
C theories. Such theories had difficulties and limitations and a1so attention
0
turned to what are called the C theories.
The Mindlin plate theory [6.2] is now the most commonly used basis for the
development of plate elements, especially as they can cover applications to
moderately thick and laminated plate and shell constructions. It has been
estimated that in large scale production runs using finite element packages, the
simple four-node quadrilateral plate element (the QUAD4 element) may account for
as much as 80% of all usage. It is therefore important to understand that the
evolution of the current generation of QUAD4 elements from those of yester-year,
over a span of nearly three decades was made difficult by the presence of shear
locking. We shall now see how this takes place.
71
The history behind the discovery of shear locking in plate elements is
quite interesting. It was first recognized when an attempt was made to represent
the behavior of shells using what is called the degenerate shell approach [6.3].
In this the shell behavior is modeled directly after a slight modification of
the 3D equations and shell geometry and domain are represented by a 3D brick
element but its degrees of freedom are condensed to three displacements and two
section rotations at each node. Unlike classical plate or shell theory, the
transverse shear strain and its energy is therefore accounted for in this
formulation. Such an approach was therefore equivalent to a Mindlin theory
formulation. These elements behaved very poorly in representing even the trivial
example of a plate in bending and the errors progressed without limit, as the
plates became thinner. The difficulty was attributed to shear locking. This is
in fact the two-dimensional manifestation of the same problem that we
encountered for the Timoshenko beam element; ironically it was noticed first in
the degenerate shell element and was only later related to the problems in
designing Timoshenko beam and Mindlin plate elements [6.4]. The remedy proposed
at once was the reduced integration of the shear strain energy [6.5,6.6]. This
was only partially successful and many issues remained unresolved. Some of these
were,
i) the 2×2 rule failed to remove shear locking in the 8-node serendipity plate
element,
ii) the 2×2 rule in the 9-node Lagrangian element removed locking but introduced
zero energy modes,
iii) the selective 2×3 and 3×2 rule for the transverse shear strain energies
from γxz and γyz recommended for a 8-node element also failed to remove shear
locking,
iv) the same selective 2×3 and 3×2 rule when applied to a 9-noded element is
optimal for a rectangular form of the element but not when the element was
distorted into a general quadrilateral form,
v) even after reduced integration of the shear energy terms, the degenerate
shell elements performed poorly when trying to represent the bending of curved
shells, due to an additional factor, identified as membrane locking [6.7],
originating now from the need for consistency of the membrane strain
interpolations. We shall consider the membrane-locking phenomenon in another
section.
We shall confine our study now to plate elements without going into the
complexities of the curved shell elements.
72
difficulty when a finite element model is made. The shear strains are now
computed as the difference between the section rotations and the slopes of the
neutral surfaces, thus,
γ xz = θ x − w,x
γ yz = θ y − w,y (6.26)
The stiffness matrix of a Mindlin plate element will now have terms from the
bending strain energy and the shear strain energy. It is the inconsistent
representation of the latter that causes shear locking.
The 4-node bi-linear element is the simplest element based on Mindlin theory
that could be devised. We shall first investigate the rectangular form of the
element [6.4] as it is in this configuration that the consistency requirements
can be easily understood and enforced. In fact, an optimum integration rule can
be found which ensures consistency if the element is rectangular. It was
established in Ref. 6.4 that an exactly integrated Mindlin plate element would
lock even in its rectangular form. Locking was seen to vanish for the
rectangular element if the bending energy was computed with a 2×2 Gaussian
integration rule while a reduced 1-point rule was used for the shear strain
energy. This rectangular element behaved very well if the plate was thin but the
results deteriorated as the plate became thicker. Also, after distortion to a
quadrilateral form, locking re-appeared. A spectral analysis of the element
stiffness matrix revealed a rank deficiency - there were two zero energy
mechanisms in addition to the usual three rigid body modes required for such an
element. It was the formation of these mechanisms that led to the deterioration
of element performance if the plate was too thick or if it was very loosely
constrained. It was not clear why the quadrilateral form locked even after
reduced integration. We can now demonstrate from our consistency view-point why
the 1-point integration of the shear strain energy is inadequate to retain all
the true Kirchhoff constraints in a rectangular thin plate element. However, we
shall postpone the discussion on why such a strategy cannot preserve consistency
if the element was distorted to a later section.
Following Ref. [6.4], the strain energy for an isotropic, linear elastic
plate element according to Mindlin theory can be constituted from its bending
and shear energies as,
U = UB + US
) {ò ò [θ
Et2
+θ y2 ,y +2ν θ x ,x θ y , y
=
(
2
x ,x
24 1 − ν 2
(
+ (1 − ν ) 2 θ x,y + θ y,x )2 ] dx dy (6.27)
+
6k (1 − ν )
t 2
[
ò ò (θ x − w,x ) + (θ y − w,y ) dx dy
2 2
] }
73
Fig. 6.5 Cartesian and natural coordinate system for a four-node rectangular
plate element.
Let us now examine the field-consistency requirements for one of the shear
strains, γxz, in the Cartesian system. The admissible displacement field
interpolations required for a 4-node element can be written in terms of the
Cartesian co-ordinates itself as,
w = a0 + a1x + a2 y + a3 xy (6.28a)
θ = b0 + b1x + b2 y + b3 xy (6.28b)
The shear strain field derived from these kinematically admissible shape
functions is,
As the plate thickness is reduced to zero, the shear strains must vanish. The
discretized constraints that are seen, to be enforced as γ xz → 0 in Equation
(6.29) are,
b0 − a1 → 0 (6.30a)
b2 − a3 → 0 (6.30b)
b1 → 0 (6.30c)
b3 → 0 (6.30d)
74
The constraints shown in Equations (6.30a) and (6.30b) are physically meaningful
and represent the Kirchhoff condition in a discretized form. Constraints (6.30c)
and (6.30d) are the cause for concern here - these are the spurious or
`inconsistent' constraints which lead to shear locking. Thus, in a rectangular
element, the requirement for consistency of the interpolations for the shear
strains in the Cartesian co-ordinate system is easily recognized as the
polynomials use only Cartesian co-ordinates. Let us now try to derive the
optimal element and also understand why the simple 1-point strategy of Ref. 6.4
led to zero energy mechanisms.
It is clear from Equations (6.29) and (6.30) that the terms b1x and b3xy
are the inconsistent terms which will contribute to locking in the form of
spurious constraints. Let us now look for optimal integration strategies for
removing shear locking without introducing any zero energy mechanisms. We shall
consider first, the part of the shear strain energy contributed by γxz. We must
integrate exactly, terms such as (b0-a1), (b2-a3)y, b1x, and b3xy. We now
identify terms such as (b0-a1), (b2-a3), b1, and b3 as being equivalent to the
quantities (θx-w,x)0, (θx-w,x),y0, (θx,x)0, and (θx,xy)0 where the subscript ‘0’
denotes the values at the centroid of the element (for simplicity, we let the
centroid of the element lie at the origin of the Cartesian co-ordinate system).
In the penalty limit of a thin plate, these four quantities act as constraints.
The first two reproduce the true Kirchhoff constraints and the remaining two act
as spurious constraints that cause shear locking by enforcing θx,x→0 and θx,xy→0
in the element.
Thus, only the true constraints are retained and all spurious constraints are
removed. This strategy can also be seen to be variationally correct in this
case; we shall see later that in a quadrilateral case, it is not possible to
ensure variational correctness exactly. By a very similar argument, we can show
that the part of the shear strain energy from γyz will require a 2×1 Gaussian
integration rule. This element would be the optimal rectangular bi-linear
Mindlin plate element.
Let us now look at the 1-point integration strategy used in Ref. 6.4. This
will give shear energy terms such as,
75
We have now only one true constraint each for the shear energy from γxz
and γyz respectively while the other Kirchhoff constraints (θ x − w,x ), y0 → 0 and
(θ y − w,y ),x0 → 0
are lost. This introduces two zero energy modes and accounts for
the consequent deterioration in performance of the element when the plates are
thick or are very loosely constrained, as shown in Ref. 6.4.
76
recovery in the Mindlin plate elements. For a field-consistent rectangular 4-
node element, the points are very easy to determine [6.8] (note that in a field-
inconsistent 4-node element, there will be violent linear oscillations in the
shear stress resultants corresponding to the inconsistent terms). Thus, Ref. 6.8
shows that bending moments and shear stress resultants Qxz and Qyz are accurate
at the centroid and at the 1×2 and 2×1 Gauss points in a rectangular element for
isotropic or orthotropic material. It is coincidental, and therefore fortuitous,
that the shear stress resultants are most accurate at the same points at which
they must be sampled in a selective integration strategy to remove the field-
inconsistencies! For anisotropic cases, it is safest to sample all stress
resultants (bending and shear) at the centroid.
Such rules can be extended directly to the 9-node rectangular element. The
bending moments are now accurate at the 2×2 Gauss points and the shear stress
resultants in an isotropic or orthotropic problem are optimal at the same 2×3
and 3×2 Gauss points which were used to remove the inconsistencies from the
strain definitions. However, accurate recovery of stresses from the 8-node
element is still a very challenging task because of the difficulty in
formulating a robust element. The most efficient elements known today are
variationally incorrect even after being made field-consistent and need special
filtering techniques before the shear stress resultants can be reliably sampled.
We can conclude this section on shear locking by noting that the available
understanding was unable to resolve the phenomena convincingly. The proposed
improvement, which was the consistency paradigm, together with the functional
re-constitution procedure, allowed us to derive an error estimate for a case
under locking and we could show through numerical (digital) experiments that
these estimates were accurate. In this way we are convinced that a theory with
the consistency paradigm is more successful from the falsifiability point of
view than one without.
6.5 References
77
6.6 O. C. Zienkiewicz, R. L. Taylor and J. M. Too, Reduced integration technique
in general analysis of plates and shells, Int. J. Num. Meth. Engng. 43, 275-
290, 1971.
6.7 H. Stolarski and T. Belytschko, Membrane locking and reduced ingression for
curved elements, J. Appl. Mech. 49, 172-178, 1982.
6.8 G. Prathap and C. R. Babu, Accurate force evaluation with a simple bi-linear
plate bending element, Comp. Struct. 25, 259-270, 1987.
78
Chapter 7
The shear locking phenomenon was the first of the over-stiffening behavior that
was classified as a locking problem. Other such pathologies were noticed in the
behavior of curved beams and shells, in plane stress modeling and in modeling of
three dimensional elastic behavior at the incompressible limit (ν→0.5). The
field-consistency paradigm now allows all these phenomena to be traced to
inconsistent representations of the constrained strain fields. A unifying
pattern is therefore introduced to the understanding of the locking problems in
constrained media elasticity - whether it is shear locking in a straight beam or
flat plate element, membrane locking in a curved beam or shell element,
parasitic shear in 2D plane stress and 3D elasticity or incompressible locking
in 2D plane strain and 3D elasticity as the Poisson's ratio ν→0.5.
This chapter will now briefly summarize this interpretation. Unlike the
chapter on shear locking, a detailed analysis is omitted as it would be beyond
the scope of the present book and readers should refer to the primary published
literature.
Earlier, we argued that the most part of structural analysis deals with the
behavior of thin flexible structures. One popular and efficient form of
construction of such thin flexible structures is the shell - enclosing space
using one or more curved surfaces. A shell is therefore the curved form of a
plate and its structural action is a combination of stretching and bending. It
is possible to perform a finite element analysis of a shell by using what is
called a facet representation - i.e. the shell surface is replaced with flat
triangular and/or quadrilateral plate elements in which a membrane stiffness
(membrane element) is superposed on a bending stiffness (plate bending element).
Such a model is understandably inaccurate in that with very coarse meshes, they
do not capture the bending-stretching coupling of thin shell behavior. Hence,
the motivation for designing elements with mid-surface curvature taken into
account and which can capture the membrane-bending coupling correctly. There are
two ways in which this could be done. One is to use elements based on specific
shell theories (e.g. the Donnell, Flugge, Sanders, Vlasov theories, etc.). There
are considerable controversies regarding the relative merits of these theories
as each has been obtained by carrying out approximations to different degrees
when the 3-dimensional field equations are reduced to the particular class of
shell equations. The second approach is called the degenerate shell approach -
three dimensional solid elements can be reduced (degenerated) into shell
elements having only mid-surface nodal variables - these are no longer dependent
on the various forms of shell theories proposed and should be simple to use.
They are in fact equivalent to a Mindlin type curved shell element.
However, accurate and robust curved shell elements have been extremely
difficult to design - a problem challenging researchers for nearly three
decades. One key issue behind this difficulty is the phenomenon of membrane
locking - the very poor behavior of curved elements, as they become thin. This
79
Fig. 7.1 The classical thin curved beam element.
was neither recognized nor understood for a very long time. It was believed at
first that this was because curved elements could not satisfy the strain-free
rigid body motion condition because of their inherent curved geometry. In this
section, we shall apply the field-consistency paradigm to the problem to
recognize that a consistent representation of membrane strains is desired to
avoid the phenomenon called membrane locking and that it is the need for
consistency rather than the requirement for strain-free rigid body motion which
is the key factor. This is most easily identified by working with the one-
dimensional curved beam (or arch) elements.
We shall now see that a similar problem appears in curved finite elements
in which both flexural and membrane deformation take place. Such elements
perform very poorly in cases where inextensional bending of the curved structure
was predominant, i.e. the membrane strains become vanishingly small when
compared to the bending strains. This phenomenon has now come to be called
membrane locking.
In this section, we shall examine in turn, the classical thin curved beam
element and the linear and quadratic shear flexible curved beam elements in a
curvilinear co-ordinate system.
Figure 7.1 describes the simplest curved beam element of length 2l and radius of
curvature R based on classical thin beam theory. The displacement degrees of
freedom required are the circumferential displacement u and the radial
displacement w. The co-ordinate s follows the middle line of the curved beam.
The membrane strain ε and the bending strain χ are described by the strain-
displacement relations
ε = u,s + w R (7.1a)
χ = u,s R − w,ss (7.1b)
0
A C description for u and a C1 description w is required. Kinematically
admissible displacement interpolations for u and w are
80
Fig. 7.2 Geometry of a circular arch
u = a0 + a1ξ (7.2a)
w = b0 + b1ξ + b2ξ 2
+ b3 ξ 3
(7.2b)
( ) (
χ = a1 Rl − 2b2 l 2 − 6b3 l 2 ξ ) (7.3a)
If a thin and deep arch (i.e. having a large span to rise ratio so that L/t>>1
and R/H is small, see Fig. 7.2) is modeled by the curved beam element, the
physical response is one known as inextensional bending such that the membrane
strain tends to vanish. From Equation (7.3b) we see that the inextensibility
condition leads to the following constraints:
a1 l + b0 R + b2 3R → 0 (7.4a)
b1 + 3b3 5 → 0 (7.4b)
b2 → 0 (7.4c)
b3 → 0 (7.4d)
81
i.e. it is a true constraint. However, the three remaining constraint (7.4b) to
(7.4d) have no participation from the u field. Let us now examine in turn what
these three constraints imply for the physical problem. From the three
constraints we have the conditions b1 → 0, b2 → 0 and b3 → 0. Each in turn
implies the conditions w,s → 0, w,ss → 0 and w,sss → 0 . These are the spurious
constraints. This element has been studied extensively and the numerical
evidence given there confirms that membrane locking is mainly determined by the
first of these three constraints. Next, we shall briefly describe the functional
re-constitution analysis that will yield semi-quantitative error estimates for
membrane locking in this element.
The analysis is a little more difficult than for the shear locking case
and would not be described in full here. For a curved beam of length 2l, moment
of inertia I and area of cross-section A, the strain energy can be written as,
(
π e = ò 1 2 EI χ T χ + 1 2 EA ε T ε ds ) (7.5)
where ε and χ are as defined in Equation (7.1). For simplicity assume that the
cross-section of the beam is rectangular so that if the depth of the beam is t,
then I = At2 12 . After discretisation, and reconstitution of the functional, we
can show that the inconsistent membrane energy terms disturb the bending energy
terms - the principal stiffening term is physically equivalent to a spurious in-
(
plane or membrane stiffening action due to an axial load N = EAl2 3R 2 . )
This action is quite complex and it is not as simple to derive useful
universally valid error estimates as was possible with the linear Timoshenko
beam element. We therefore choose a problem of simple configuration which can
reveal the manner in which the membrane locking action takes place. We consider
a shallow circular arch of length L and radius of curvature R, simply supported
at the ends. The loading system acting on it is assumed to induce nearly
inextensional bending. We also assume that the transverse deflection w can be
approximated in the form w=csin(πs/L). If the entire arch is discretized by
curved beam elements of length 2l each, the strain energy of the discretized
arch (again assuming that χ=-w,ss and neglecting the consistently modeled part
of the membrane energy) one can show that due to membrane locking, the bending
rigidity of the arch has now been altered in the following form:
(EI )′ (EI ) = {1 + (4 )
π 2 (Ll Rt )2 } (7.6)
82
Fig. 7.3 The Mindlin curved beam element.
The exercise above showed that the locking behavior of curved beams could be
traced to the inconsistent definition of the membrane strain field. The curved
beam element used to demonstrate this was based on classical thin beam theory
and used a curvilinear co-ordinate description. Of interest to us now is to
understand how the membrane locking phenomenon can take place in a general
curved shell element. Most general shell elements which have now been accepted
into the libraries of general purpose finite element packages are based on what
is called the degenerate shell theory which is equivalent to a shear deformable
theory and use quadratic interpolations (i.e. three nodes on each edge to
capture the curved geometry in a Cartesian system).
83
strain χ and the transverse shear strain γ. For an element of length 2l and
radius R,
U M = (1 2 ) (EA ) ò ε 2 ds where ε = u,s + w R (7.7a)
u = a0 + a1ξ (7.8a)
w = b0 + b1ξ (7.8b)
θ = c0 + c1ξ (7.8c)
(1 2 ) (EAl2 ) òw , 2
3R 2 s ds (7.10)
84
The membrane locking action is therefore identical to that predicted for the
classical thin curved beam. A variationally correct field-consistent element can
easily be derived using the Hu-Washizu theorem. This is a very accurate element.
The quadratic Mindlin curved beam element can also be interpreted in the
light of our studies of the classical curved beam earlier. Spurious constraints
at both the linear (i.e. w,s→0) and quadratic (i.e w,ss→0) levels lead to
errors that propagate at a (Ll Rt )2 rate and l 2 Rt (
rate respectively. We )
2
expect now that in a quadratic Mindlin curved beam element, the latter
constraint is the one that is present and that these errors are large enough to
make the membrane locking effect noticeable. This also explains why the
degenerate shell elements, which are based on quadratic interpolations, suffer
from membrane locking.
(
ε = (a1 l + b0 R + b2 3R ) + (2a2 l + b1 R ) ξ − b2 3R 1 − 3ξ 2 ) (7.12)
ò
é
ë
( )
2ù
U B = (1 2 ) (EI ) ê1 + (4 15 ) l 2 Rt ú w,ss
û
2
ds (7.13)
It is interesting to compare the error norm for locking for this element,
i.e. (l 2
Rt2 ) with the corresponding error norm for locking in the inconsistent
85
( )
linear Mindlin curved beam element, which is L l Rt 2 . It is clear that since
L>>l when an arch is discretized with several elements, the locking in the
quadratic element is much less than in the linear element. One way to look at
this is to examine the rate at which locking is relieved as more elements are
added to the mesh, i.e. as l is reduced. In the case of the inconsistent linear
2
element, locking is relieved only at an l rate whereas in the case of the
4
quadratic element, locking is relieved at the much faster l rate. However,
locking in the quadratic element is significant enough to degrade the
performance of the quadratic curved beam and shell elements to a level where it
is impractical to use in its inconsistent form. This is clearly seen from the
fact that to reduce membrane-locking effects to acceptable limits, one must use
element lengths l << Rt . Thus for an arch with R=100 and t=1, one must use
about a 100 elements to achieve an accuracy that can be obtained with two to
four field-consistent elements.
Our experience with the linear and quadratic straight beam elements showed that
inconsistency in the constrained strain-field is accompanied by spurious stress
oscillations corresponding to the inconsistent constraints. We can now
understand that the inconsistent constraint b1→0 will induce linear axial force
oscillations in the linear curved beam element. Similarly, the inconsistent
quadratic constraint will trigger off quadratic oscillations in the quadratic
element.
So far, we have examined the phenomenon of membrane locking in what were called
the simple curved beam elements. These were based on curvilinear geometry and
the membrane locking effect was easy to identify and quantify. However, curved
beam and shell elements which are routinely used in general purpose finite
element packages are based on quadratic formulations in a Cartesian system. We
shall call these the general curved beam and shell elements. Membrane locking is
not so easy to anticipate in such a formulation although the predictions made in
this section have been found to be valid for such elements.
Shear locking and membrane locking are phenomena seen in finite element modeling
using structural elements based on specific theories, e.g. beam, plate and shell
theories. These are theories obtained by simplifying a more general continuum
description, e.g. the three dimensional theory of elasticity or the two
dimensional theories of plane stress, plane strain or axisymmetric elasticity.
The simplifications are made by introducing restrictive assumptions on strain
and/or stress conditions like vanishing of shear strains or vanishing of
transverse normal strains, etc. In the beam, plate and shell theories, this
allows structural behavior to be described by mid-surface quantities. However,
these restrictive assumptions impose constraints which make finite element
modeling quite difficult unless the consistency aspects are taken care of. We
shall now examine the behavior of continuum elements designed for modeling
continuum elasticity directly, i.e. 2D elements for plane strain or stress or
axisymmetric elasticity and solid elements for 3D elasticity. It appears that
86
Fig. 7.4 (a) Constant strain triangle, (b) Rectangular bilinear element,
(c) Eight-node rectangle and (d) Nine-node rectangle.
However, in some applications, e.g. the plane stress modeling of beam flexure or
in plane strain and axisymmetric applications where the Poisson's ratio
approaches 0.5 (for nearly incompressible materials or for materials undergoing
plastic deformation), considerable difficulties are seen. Parasitic shear occurs
in the plane stress and 3D finite element modeling of beam flexure. Under pure
bending loads, for which the shear stress should be zero, the elements yield
large values of shear stress except at certain points (e.g. centroids in linear
elements). In the linear elements, the errors progress in the same fashion as
for shear locking in the linear Timoshenko beam element, i.e. the errors
progress indefinitely as the element aspect ratio increases. Reduced integration
of the shear strain energy or addition of bubble modes for the 4-noded
87
Fig. 7.5 Plane stress model of beam flexure.
U = U E + UG
)( )
L T é ù
Eb
( )
= ò0 ò0 ê
(
êë 2 1 - ν 2
u 2 ,x + v2 ,y +2ν u,x v , y + Gb 2 u, y +v ,x 2 ú dx dy
úû
(7.14)
88
u = a0 + a1x + a2 y + a3 xy (7.15a)
v = b0 + b1x + b2 y + b3 xy (7.15b)
We need to examine only the shear strain field that will become
constrained (i.e. become vanishingly small) while modeling the flexural action
of a thin beam. This will be interpolated as,
and the shear strain energy within the element will be,
( )2 dx dy
t
U e (G ) = ò-l ò Gb 2 a2 + b1 + a3 x + b3 y
l
[(a ]
-t
Equation (7.17) suggests that in the thin beam limit, the following constraints
will emerge:
a2 + b1 → 0 (7.18a)
a3 → 0 (7.18b)
b3 → 0 (7.18c)
We can see that Equation (7.18a) is a constraint that comprises constants from
both u and v functions. This can enforce the true constraint of vanishing shear
strains in a realistic manner. In contrast, Equations (7.18b) and (7.18c) impose
constraints on terms from one field function alone in each case. These are
undesirable constraints. We shall now see how these lead to the stiffening
effect that is called parasitic shear.
will be enforced more rapidly than the constraint b3→0 as the beam becomes
thinner. The spurious energies generated from these two terms will also be in a
similar proportion. In Equation (7.17), one can interpret the shear strain field
to comprise a constant value, which is `field-consistent' and two linear
variations which are `field-inconsistent'. We see now that the variation along
the beam length is the critical term. Therefore, let us consider the use of the
shear strain along y=0 as a measure of the averaged shear strain across a
vertical section. This gives
γ = (a2 + b1 ) + a3 x (7.19)
Thus, along the length of the beam element, γ have a constant term that reflects
the averaged shear strain at the centroid, and a linear oscillating term along
the length, which is related to the spurious constraint in Equation (7.18b). As
in the linear beam element in Section 4.1.6, this oscillation is self-
equilibrating and does not contribute to the force equilibrium over the element.
However, it contributes a finite energy in Equation (7.17) and in the modeling
89
of very slender beams, this spurious energy is so large as to completely
dominate the model of the beam behavior and cause a locking effect. We shall now
use the functional re-constitution technique to determine the magnitude of
locking and the extent of the shear stress oscillations triggered off by the
inconsistency in the shear strain field.
We shall simplify the present example and derive the error estimates from
the functional re-constitution exercise. We shall denote by a3 , the coefficient
determined in a solution of the beam problem by using a field-consistent
representation of the total energy in the beam, and by a3 , the value from a
solution using energies based on inconsistent terms. The field-consistent and
field-inconsistent terms will differ by,
(
a3 a3 = 1 + Gl 2 Et 2 ) (7.20)
(
This factor, e L = 1 + Gl2 Et2) is the additional stiffening parameter that
determines the extent of parasitic shear.
V = V0 + (3M 0 l ) (x l ) (7.21)
The oscillations are proportional to the centroidal moment, and for e L >> 1 , the
oscillations are inversely proportional to the element length, i.e. they become
more severe for smaller lengths of the plane stress element. It is important to
note that a field-inconsistent two node beam element based on linear
interpolations for w and θ (see Equation (7.19)) produces an identical stress
oscillation.
It is well known that the plane stress elements behave reliably in general 2D
applications where no constraints are imposed on the normal or shear strain
fields. The need for field-consistency becomes important only when a strain
field is constrained. Earlier, we saw that in the plane stress modeling of thin
beam flexure, the shear strain field is constrained and that these constraints
are enforced by a penalty multiplier Gl2 Et2 . The larger this term is, the more
severely constrained the shear strain becomes. Of the many strategies that are
available for removing the spurious constraints on the shear strain field, the
use of a variationally correct re-distribution is recommended. This requires the
field-consistent γ to be determined from γ using the orthogonality condition
ò δγ
T
(γ − γ ) dx dy = 0 (7.22)
Thus instead of Equation (7.16), we can show that the shear strain-field
for the 4-node rectangular plane stress element will be,
90
γ = (a2 + b1 ) (7.23)
So far, we have dealt with the finite element modeling of problems which were
simplified to one-dimensional or two-dimensional descriptions. There remains a
large class of problems which need to be addressed directly as three dimensional
states of stress. Solid or three dimensional elements are needed to carry out
the finite element modeling of such cases. A variety of 3D solid elements exist,
e.g. tetrahedral, triangular prism or hexahedral elements. In general cases of
3D stress analysis, no problems are encountered with the use of any of these
elements as long as a sufficiently large number of elements are used and no
constrained media limits are approached. However, under such limits the
tetrahedral and triangular prism elements cannot be easily modified to avoid
these difficulties. It is possible however to improve the hexahedral 8-node and
27-node elements so that they are free of locking and of other ill-effects like
stress oscillations. One problem encountered is that of parasitic shear or shear
locking when solid elements are used to model regions where bending action is
predominant. In fact, in such models, parasitic shear and shear locking merge
indistinguishably into each other. We briefly examine this below.
91
the bi-linear plane stress element above. The problem of designing a useful 8-
node brick element has therefore received much attention and typical attempts to
alleviate this have included techniques such as reduced integration, addition of
bubble functions and assumed strain hybrid formulations. We shall now look at
this from the point of view of consistency of shear strain interpolation.
We can observe here that the shear strain energy is computed from
Cartesian shear strains γ xy , γ xz and γ yz . In the isoparametric formulation,
these are to be computed from terms involving the derivatives of u, v, w with
respect to the natural co-ordinates ξ, η, ζ and the terms from the inverse of
the three dimensional Jacobian matrix. It is obvious that it will be a very
difficult, if not impossible task, to assure the consistency of the
interpolations for the Cartesian shear strains in a distorted hexahedral.
Therefore attention is confined to a regular hexahedron so that consistency
requirements can be easily examined.
92
Similarly, the other shear strain fields can be modified. These strain fields
should now be able to respond to bending modes in all three planes without
locking.
é 1
u=ρ ê +
(1 − 2ν ) r ù
ú (7.27)
ë 4r 2 2 (1 + ν ) û
[(
where ρ = Pba3 G b 3 − a3 )] and r=R/b non-dimensional radius and ν is the Poisson's
ratio.
ε r = u,r εθ = u r εφ = u r
ε = ε r + ε θ + ε φ = u,r + 2u r
(ε )
1
U = ò
2
r + ε θ2 + ε φ2 + 1 2 α ε 2 r 2 dr (7.28)
a
93
and α = 2ν (1 − 2ν ) when G=1 and a = a b is the dimensionless inner radius. The
displacement type finite element representation of this very simple problem runs
into a lot of difficulties when ν→0.5, i.e., when α→∞ in the functional
represented by Equation (7.28).
òε r dr = ò (ru,r +2u )
2 2 2
dr (7.30)
é (r1 + r2 ) ù
(ru,r +2u ) = ê (u2 − u1 ) + (u1 + u2 )ú + [3 2 (u2 − u1 )] ξ (7.31)
ë 2 (r2 − r1 ) û
Consider now the constraints that are enforced when this discretized field
is used to develop the strain energy arising from the volumetric strain as shown
in Equation (7.31) when α→∞.
(r1 + r2 ) (
u2 − u1 ) + (u1 + u2 ) → 0 (7.32a)
2 (r2 − r1 )
(u2 − u1 ) → 0 (7.32b)
94
Condition (7.32a) represents a meaningful discretized version of the
condition of vanishing volumetric strain and is called a true constraint in our
field-consistency terminology. However, condition (7.32b) leads to an undesired
restriction on the u field as it implies u,r → 0 . This is therefore the spurious
constraint that leads to locking and spurious pressure oscillations.
U =1 2G ò ε d D ε d dV + 1 2 λ ò ε n ε n dV
T T
(7.33)
where εn is the volumetric strain. From Equations (7.24) and (7.34) we have,
Equation (7.35) can be constrained to zero only when the coefficients of each of
its terms vanish, giving rise to the constraint conditions,
a2 + b3 + c4 → 0 (7.36a)
b5 + c7 → 0 (7.36b)
a5 + c6 → 0 (7.36c)
a7 + b6 → 0 (7.36d)
a8 → 0 (7.36e)
b8 → 0 (7.36f)
c8 → 0 (7.36g)
Equation (7.36b) to (7.36) are the inconsistent terms that cause locking. To
ensure a field-consistent representation, only the consistent terms should be
retained, giving
ε n = a2 + b3 + c4 (7.37)
95
7.4.4 Concluding remarks
7.5 References
96
Chapter 8
Stress consistency
8.1 Introduction
In initial strain problems, the initial strain variation and the total
strain variations may be of different order. This is a familiar problem in
finite element thermal stress analysis. Here, it is necessary to obtain
kinematically equivalent thermal nodal forces from temperature fields and also
ensure that the discretised thermal strains corresponding to this are consistent
with the discretised description of total strains. Again, one must carefully
identify the conflicting requirements on the order of discretised functions to
be used.
We shall now conduct a simple numerical experiment with a tapered bar element to
show that the force field computed directly from strains in a structural element
of varying sectional rigidities has extraneous oscillations. A linear element
would have sufficed to demonstrate the basic principles involved. However, we
use a quadratic tapered bar element so that the extraneous oscillations, which
97
for a general case can be of cubic form, are not only vividly seen but also need
special care to be reduced to its consistent form. In a linear element, this
exercise becomes very trivial, as sampling at the centroid of the element gives
the correct stress resultant.
We shall first examine how the element stiffness is formulated when the minimum
total potential principle is used. We start with a functional written as,
π = 1 2 ò N T ε dx − W
where,
U =1 2 òN ε dx
T
From this product the terms of the stiffness matrix emerge. Due to the
orthogonal nature of the Legendre polynomials, terms from N which are linked
with the quadratic and cubic Legendre polynomials, N3 and N4 respectively, will
not contribute to the energy and therefore will not provide terms to the
stiffness matrix! It is clear that, the displacements recovered from such a
formulation cannot recognize the presence of the quadratic and cubic terms N3
and N4 in the stress field N as these have not been accounted for when the
stiffness matrix was computed. Hence, in a displacement type finite element
formulation, the stresses recovered from the displacement vector will have
extraneous oscillations if N3 and N4 are not eliminated from the stress field
during stress recovery. We shall designate by BAR3.0, the conventional element
using N for stiffness matrix evaluation and recovery of force resultant.
Next, we must see how the consistent representation of the force field
denoted by N must be made. N should comprise only the terms that will
contribute to the stiffness and strain energy and the simplest way to do this is
to expand the kinematically determined N in terms of Legendre polynomials, and
retain only terms that will meaningfully contribute to the energy in (N ε ) .
T
98
Thus, N must be consistent with ε, i.e. in this case, retain only up to linear
terms:
N = N 1 + N 2ξ (8.1)
We shall perform the computational exercises with the two versions of the
element; note that in both cases, the stiffness matrices and computed
displacements are identical. Figure 8.1 shows a tapered bar clamped at node-1
and subjected to an axial force P at node-3. The taper is defined by the
parameters,
α = (A 3 − A1 ) 2A 2 and β = (A1 − 2A 2 + A 3 ) 2A 2
Finite element results from a computational exercise using the two versions
described above for a bar with cross section tapering linearly from the root to
the tip for which β=0 are obtained. Thus α is given by,
α = (A 3 − A1 ) (A1 + A3 )
Thus α can vary from 0 to –1.0. Fig. 8.2 shows the axial force patterns
obtained from the finite element digital computation for a case with α=-0.9802
(with A3=0.01 A1) It can be noted here that the results are accurate at
ξ = ±1 3.
99
Fig. 8.2 Axial force pattern for linearly tapered bar (α=0.9802 and β=0.0) with
A3=0.01 A1.
π = ò {1 2 (EAε ) }
ε + N T (ε − ε ) dx − W
T
100
Fig. 8.3 Axial force pattern for bar with combined linear and quadratic taper
(α=-4/3 and β=4/9).
ò δu {dN dx − p} dx = 0
T
ò δε
T
{- N + EA ε } dx = 0 (8.2)
and variation with respect to the assumed force field N gives rise to the
condition
ò δN
T
{ε - ε } dx = 0 (8.3)
Now Equations (8.2) and (8.3) are the orthogonality conditions required
for reconstituting the assumed fields for the stress resultant and the strain.
The consistency paradigm suggests that the assumed stress resultant field N
101
should be of the same order as the assumed strain field ε . Then Equation (8.3)
gives the orthogonality condition for strain field-redistribution.
ò δε
T
{N - N } dx = 0 (8.4)
Earlier in this chapter, we saw that stress resultant fields computed from
strain fields in a displacement type finite element description of a domain with
varying sectional rigidities showed extraneous oscillations. This was traced to
the fact that these stress resultant fields were of higher interpolation order
than the strain fields and that the higher degree stress resultant terms did not
participate in the stiffness matrix computations. In this section, we show that
much the same behavior carries over to the problem of thermal stress
computations.
With the introduction of initial strains ε0 due to thermal loading, the stress
to strain relationship has to be written as
102
σ = D (ε - ε0 ) = Dε m (8.5)
The strain terms now need to be carefully identified. {ε} is the total strain
and {εm} is the mechanical or elastic strain. The free expansion of material
produces initial strains
ε0 = α T (8.6)
ò δ {ε } D {ε0 } dV
T
(8.8)
Again, the problem is that the `higher order' components of the thermal (or
initial) stress vector are not sensed by the total strain interpolations in the
integral shown above. In other words, the total strain terms "do work" only on
the consistent part of the thermal stress terms in the energy or virtual work
integral. Thus, a thermal load vector is created which corresponds to a initial
strain (and stress) vector that is `consistent' with the total strain vector.
The finite element displacement and total strain fields which are obtained in
the finite element computation then reflect only this consistent part of the
thermal loading. Therefore, only the consistent part of the thermal stress
should be computed when stress recovery is made from the nodal displacements;
the inclusion of the inconsistent part, as was done earlier, results in thermal
stress oscillations.
103
Fig. 8.4 Linear bar element.
To form the element stiffness and thermal load vector, an integral of the
form
ò δε σ dx
T
(8.10)
where F1 and F2 are the consistently distributed nodal loads arising from the
distributed external loading. By observing the components of the interpolation
fields in Equations (8.9b) to (8.9d) carefully (i.e. constant and linear terms)
and tracing the way they participate in the `work' integral in Equation (8.10),
it is clear that the (T2 − T1 ) 2 term associated with the linear (i.e. ξ) term in
σ (originating from ε0) cannot do work on the constant term in δεT and therefore
vanishes from the thermal load vector; see Equation (8.11). Thus the equilibrium
equations that result from the assembly of the element equilibrium equations
represented by Equation (8.11) will only respond to the consistent part of the
initial strain and will give displacements corresponding to this part only.
104
Fig. 8.5 Clamped bar subject to varying temperature
hand using a simple example below and compare it with the analytical solution.
Figure 8.5 shows a bar of length L=4l clamped at both ends and subjected to a
varying temperature field. We shall consider a case where two conventionally
derived linear elements are used, so that we require the nodal temperatures T1,
T2 and T3 as input. The nodal reactions are F1 and F3 (corresponding to the
clamped conditions u1=u3=0). We have the assembled equations as,
é 1 −1 ù ì 0 ü ìF1 ü ì − (T1 + T2 )ü
EA ê ú ï ï ï ï EAα ï ï
−1 2 −1 ú íu2 ý = í 0 ý + í (T1 − T3 ) ý (8.12)
2l ê ï 0 ï ïF ï 2 ï (T + T )ï
êë −1 1úû î þ î 3þ î 2 3 þ
From this, we can compute the displacements and nodal reactions as,
and these are the correct answers one can expect with such an idealization. If
the stress in the bar is computed from the nodal reactions, one gets a constant
stress σ = −Eα (T1 + 2T2 + T3 ) 4 , in both elements, which is again the correct answer
one can expect for this idealization. It is a very trivial exercise to show
analytically that a problem in which the temperature varies linearly from 0 to T
at both ends will give a constant stress field σ = − EαT 2 , which the above model
recovers exactly.
It is now very clear that a linear oscillation is introduced into each element
and this corresponds to that inconsistent part of the initial strain or stress
interpolation which was not sensed by the total strain term. This offers us a
straightforward definition of what consistency is in this problem - retain only
that part of the stress field that will do work on the strain term in the
105
functional. To see how this part can be derived in a variationally correct
manner, we must proceed to the Hu-Washizu theorem.
We may note now that if a quadratic bar element had been the basis for the
finite element idealization, the total strains would have been interpolated to a
linear order; the initial strain and thermal stress field will now have a
quadratic representation (provided the temperature field has a quadratic
( )
variation) and the inconsistency will now be of the 1 − 3ξ 2 type; thus thermal
stresses derived using a formal theoretical basis will show these quadratic
oscillations which will vanish to give correct answers at the points
corresponding to ξ = ± 1 3 ; i.e. the points corresponding to the 2-point Gauss
integration rule.
We now seek to find a variational basis for the use of the re-constituted
consistent initial strain and stress interpolations in the Hu-Washizu principle.
The minimum total potential principle states the present problem as, find
the minimum of the functional,
Π MTP = ò {σ }
ε m 2 + P dV
T
(8.15)
where σ and εm are as defined in (8.5) to (8.7) and the displacement and strain
fields are interpolated from the nodal displacements using the element shape
functions and their derivatives and P is the potential energy of the prescribed
loads.
Let the continuum linear elastic problem have a discretized solution based
on the minimum total potential principle described by the displacement field u,
strain field ε and stress field σ (we project that the strain field ε is derived
from the displacement field through the strain-displacement gradient operators
106
of the theory of elasticity and that the stress field σ is derived from the
strain field ε through the constitutive laws as shown in (8.2). Let us now
replace the discretized domain by another discretized domain corresponding to
the application of the Hu-Washizu principle and describe the computed state to
be defined by the quantities ε, ε0 and σ , where again, we take that the stress
fields σ are computed using the constitutive relationships, i.e. σ = D (ε − ε0 ) .
It is clear that ε0 is an approximation of the strain field ε0. Note that we
also argue that we can use ε = ε as there is no need to introduce such a
distinction here (in a constrained media elasticity problem it is paramount that
ε be derived as the consistent substitute for ε.)
δΠ HW = 0 (8.16)
where
Π HW = ìíσ T ε m 2 + σ
ò
T
(ε m − ε m ) + P üý dV (8.17)
î þ
In the simpler minimum total potential principle, which is the basis for
the derivation of the displacement type finite element formulation in most
textbooks, only one field (i.e. the displacement field u), is subject to
variation. However, in this more general three field approach, all three fields
are subjected to variation and leads to three sets of equations which can be
grouped and classified as follows:
σ Orthogonality òδσ
T
(ε0 − ε0 ) dV =0 (8.18b)
(Compatibility)
T
æ σ − σ ö dV = 0
ε0 Orthogonality ò δε0 ç
è
÷
ø
(8.18c)
(Equilibrium)
107
Fig. 8.6 (a) Clamped bar under linear temperature variation and its (b) Bar
element model, (c) Plane stress model.
Therefore, we now see how the consistent initial strain field ε0 can be
derived from the inconsistent initial strain field ε0 without violating any
variational norm. We thus have a variationally correct procedure for re-
constituting the consistent initial strain field - for the bar element above,
this can be very trivially done by using an expansion of the strain fields in
terms of the orthogonal Legendre polynomials, as done in the previous section
for the consistent stress-resultants.
108
8.3.6 Numerical examples
In this section we take up the same example of a bar with fixed ends subjected
to linearly varying temperature along its axis and model it with the bar element
and a plane stress element for demonstrating the extraneous stress oscillations
resulting from the lack of consistency in initial strain definition.
Fig.8.7 Consistent and inconsistent thermal stress recovery for a clamped bar
problem.
109
8.3.7 Concluding remarks
110
Chapter 9
Conclusion
9.1 Introduction
Technology and Science go hand in hand now, each fertilizing the other and
providing an explosive synergy. This was not so in the beginning; as there was
Technology much before there was Science. In fact, there was technology before
there were humans and it was technology which enabled the human race to evolve;
science came later. The FEM is also an excellent example of a body of knowledge
that began as Art and Technology; Science was identified more slowly. This
chapter will therefore sum up the ideas that constitute the science of the
finite element method and also point to the future course of the technology.
One would ideally like to find a set of cardinal or first principles that govern
the entire finite element discretisation procedure, ensuring the quality of
accuracy and efficiency of the method. At one time, it was believed that two
rules, namely continuity and completeness, provided a necessary and sufficient
basis for the choice of trial functions to initiate the discretisation process.
Chapter 3 of this book reviewed the understanding on these two aspects. However,
finite element practitioners were to be confronted with a class of problems
where the strict implementation of the continuity condition led to conflicts
where multiple strain fields were required. This was the locking problem which
we examined carefully in chapters 5 to 7. Let us recapitulate our fresh insight
in an epistemological framework below.
Let us again take up the simple problem of a Timoshenko beam element. Two
strain energy components are important; the bending energy which is constituted
from the bending strain as κ = θ ,x and the shear strain energy which is based on
the shear strain γ = θ − w,x . The simplistic understanding that prevailed for a
long time was that the existence of the θ ,x and w,x terms in the energy
expression demanded that the trial functions selected for the θ and w fields
must be C continuous; i.e. continuity of θ and w at nodes (or across edges in a
0
111
producing spurious constraints. The idea of field consistent formulation was to
anticipate these conditions on the constrained strain field. Thus for θ − w,x to
vanish without producing gratuitous constraints, the trial functions used for θ
and w in the shear strain field alone must be complete to exactly the same order
- this is the consistency paradigm. We can also interpret this as a multiple
continuity requirement. Note that although the bending strain requires θ to be
C continuous, in the shear strain component, we require θ not to have C
0 0
-1
continuity, i.e. say a C continuity, is. Thus, such a conflict is resolved by
operating with consistently represented constrained strain fields.
The consistency requirement now demands that when trial functions are chosen for
displacement fields, these fields (such as θ in γ = θ − w,x ) which need to be
constrained must have a reconstituted consistent definition. The minimum total
potential principle, being a single field variational principle, does not
provide a clue as to how this reconstitution of the strain fields can be
performed without violating the variational rules.
The study so far with consistency and correctness showed the primacy of the Hu-
Washizu theorem in explaining how the finite element method worked. It also
became very apparent that if the internal working of the discretisation
procedure is examined more carefully, it turns out that it is strain and stress
fields which are being approximated in a "best-fit" sense and not the
displacement fields as was thought earlier. This brings us to the stress
correspondence paradigm, and the discovery that this paradigm can be axiomatised
from the Hu-Washizu theorem, as shown in chapter 2.
Thus, if and when the first principles are available, then the logical
sequence in which one teaches a subject will be to start with first principles
and derive the various facets of the phenomena from these principles. In
pedagogical practice, we can therefore have a choice from two courses of action:
112
teach in a chronological sense, explaining how the ideas unfolded themselves to
us, or teach in the logical sense, from first principles to observable facts. It
is not clear that one is superior to the other.
CSM has virtually grown out of the finite element method (FEM). Algorithms
for the use of the finite element method in a wide variety of structural and
thermomechanical applications are now incorporated in powerful general purpose
software packages. The use of these packages in the Computer-Aided-
Design/Computer-Aided-Manufacturing cycle forms a key element in new
manufacturing technologies such as the Flexible Manufacturing Systems. These
allow for unprecedented opportunities for increase in productivity and quality
of engineering by automating the use of structural analysis techniques to check
designs quickly for safety, integrity, reliability and economy. Very large
structural calculations can be performed to account for complex geometry,
loading history and material behavior. Such calculations are now routinely
performed in aerospace, automotive, civil engineering, mechanical engineering,
oil and nuclear industries.
113
1. There are a large number of unsolved practical problems of current interest
which still await experimental and/or numerical solutions. Some of these
demand large computational power. Some of the examples described in reference
9.2 are: simulation of response of transportation vehicles to
multidirectional crash impact forces, dynamics of large flexible structures
taking into account joint nonlinearities and nonproportional damping, study
of thermoviscoelatic response of structural components used in advanced
propulsion systems etc. In many structural problems, the fundamental
mechanics concepts are still being studied (e.g. in metal forming, adequate
characterization of finite strain inelasticity is still needed).
3. Emerging and future computer systems are expected to provide enormous power
and potential to solve very large scale structural problems. To realize this
potential fully, it is necessary to develop new formulations, computational
strategies, and numerical algorithms that exploit the capabilities of these
new machines (e.g. parallelism, vectorization, artificial intelligence).
Noor and Atluri [9.2] also expect that high-performance structures will
demand the following technical breakthroughs:
114
h) analysis and design of intelligent structures with active and/or passive
adaptive control of dynamic deformations, e.g. in flight vehicles, large
space structures, earthquake-resistant structures
Special hardware and software requirements must become available to meet the
needs described above. These include:
5. Special and general purpose application software systems that have advanced
modeling and analysis capabilities and are easy to learn and use.
It is expected that CSM will continue to grow in importance. Three areas which
are expected to receive increasing attention are: 1) modeling of complex
structures; 2) predata and postdata processing and 3) integration of analysis
programs into CAD/CAM systems.
115
9.6 Concluding remarks
In this concluding chapter, we have summarized the ideas that should inform the
rational development of robust finite elements and we have also described the
current trends in the use of structural modeling and analysis software in
engineering design. It has been seen that CSM has greatly improved our
capabilities for accurate structural modeling of complex systems. It has the
potential not only to be a tool for research into the basic behavior of
materials and structural systems but also as a means for design of engineering
structures.
9.7 References
116