Thesis for the Degree of Licentiate of Philosophy

A generalized nite element method for linear

thermoelasticity

Anna Persson

Division of Mathematics
Department of Mathematical Sciences
Chalmers University of Technology
and University of Gothenburg
Gothenburg, Sweden, 2016
A generalized nite element method for linear thermoelasticity
Anna Persson


c Anna Persson, 2016.

Department of Mathematical Sciences

Chalmers University of Technology
and University of Gothenburg
SE412 96 Gothenburg
Sweden
Phone: +46 (0)31-772 10 00

Printed in Gothenburg, Sweden, 2016.

 Abstract

In this thesis we develop a generalized nite element method for linear

thermoelasticity problems, modeling displacement and temperature in an elastic
body. We focus on strongly heterogeneous materials, like composites. For
classical nite element methods such problems are known to be numerically
challenging due to the rapid variations in the data.
The method we propose is based on the local orthogonal decomposition
technique introduced in [12]. In short, the idea is to enrich the classical nite
element nodal basis function using information from the diusion coecient.
Locally, these basis functions have better approximation properties than the
nodal basis functions.
The papers included in this thesis rst extends the local orthogonal de-
composition framework to parabolic problems (Paper I) and to linear elasticity
equations (Paper II). Finally, using the theory developed in these papers, we
address the linear thermoelastic system (Paper III).
Keywords: Thermoelasticity, parabolic equations, linear elasticity, multi-
scale, composites, generalized nite element, local orthogonal decomposition, a
priori analysis.

i
 List of included papers

The following papers are included in this thesis:

Paper I. Axel Malqvist and Anna Persson: Multiscale techniques for

parabolic equations, Submitted.

Paper II. Patrick Henning and Anna Persson: A multiscale method

for linear elasticity reducing Poisson locking, Submitted.

Paper III. Axel Malqvist and Anna Persson: A generalized nite el-
ement method for linear thermoelasticity, Submitted.

iii
 Acknowledgments

First and foremost, I would like to thank my supervisor Axel Malqvist for
his guidance throughout my work on this thesis. Thank you for introducing
me to the subject, for sharing your expertise and knowledge, and for teaching
me how the academic world works. I would also like to thank my co-advisors
Patrick Henning and Stig Larsson for many fruitful discussions and your helpful
advice. Especially, thank you Patrick for the last months of collaboration and
for always being so positive.
Furthermore, I am grateful to all my colleagues at the mathematics depart-
ment for creating such a friendly working environment. Many thanks also goes
to my family for always believing in me and giving me encouragement when I
need it.
Last, but not least, thank you Joakim for your love and immense support.

v
 Contents

Abstract i

List of included papers iii

Acknowledgments v

Part 1. Introduction 1

Introduction 3
1. Background 3
2. Classical nite element 6
3. A generalized nite element method 8
4. Summary of papers 15
5. Future work 16
References 16

Part 2. Papers 19

Paper I - Multiscale techniques for parabolic equations 21

Paper II - A multiscale method for linear elasticity reducing Poisson locking 49

Paper III - A generalized nite element method for linear thermoelasticity 73

vii
 Part 1

Introduction
 Introduction

1. Background
In many applications the expansion and contraction of a material exposed to
external forces and temperature changes are of great importance. For instance,
it may be crucial when designing parts for aircrafts or when constructing a
bridge.
In this thesis we study numerical solutions to linear thermoelastic systems,
which consist of partial dierential equations (PDEs) simulating displacement
and temperature changes in materials over time. In particular, we are inter-
ested in applications where the material under consideration is strongly het-
erogeneous, e.g. composites. Composite materials are constructed using two
or more dierent constituents. Typically, the material properties in composites
vary on a very ne scale, as in, for instance, ber reinforced materials. Model-
ing physical behavior in these materials results in equations with highly varying
and oscillating coecients. Such problems, that exhibit a lot of variations in the
data, often on multiple scales, are commonly referred to as multiscale problems.
Classically, numerical solutions to thermoelasticity equations have been ob-
tained using nite element methods (FEMs) based on continuous piecewise poly-
nomials. These methods work well for homogeneous materials, or materials that
are not varying too much in space. However, for strongly heterogeneous mate-
rials the classical FEMs struggle to approximate the solution accurately unless
the mesh width is suciently small. Indeed, the mesh width must be small
enough to resolve all the ne variations in the data. In practice, this leads to
issues with computational cost and available memory.
Todays increasing interest in and usage of composite materials thus pose a
demand for other types of numerical methods. Several such methods have been
proposed over the last two decades, see, for instance, [9, 5, 1, 10]. However,
the analysis of many of these methods require restrictive assumptions on the
material, such as periodicity or separation of scales.
In [12] a generalized nite element method (GFEM), cf. [2], is proposed and
rigorous analysis is provided. Convergence of the method is proven for an ar-
bitrary positive and bounded coecient, that is, no assumptions on periodicity
or separation of scales are needed.

3
 Introduction

The purpose of this thesis is to generalize the method proposed in [12] to

solve linear thermoelasticity equations with highly varying and oscillating coef-
cients. This is done in three steps. In Paper I we extend the method to linear
parabolic problems, in Paper II we consider (stationary) linear elasticity equa-
tions and in Paper III we nally address the thermoelastic system. In all three
papers we prove convergence of optimal order for highly varying coecients and
provide several numerical examples that conrm the analysis.
In the upcoming section we describe the system of equations used to model
the displacement and temperature of an elastic material. In Section 2 the issue
with applying the classical FEM to multiscale problems is described in more
detail. In Section 3 we introduce the GFEM proposed in [12] for elliptic equa-
tions and discuss the main idea behind the extension to linear thermoelasticity.
Finally, in Section 4 we summarize the appended papers and highlight the main
results.

1.1. Linear thermoelasticity. Linear thermoelasticity refers to a cou-

pled system of PDEs describing the displacement and temperature of an elastic
body, see [3, 4]. To introduce the mathematical formulation of this system we
let Rd , d = 1, 2, 3, be a domain describing the initial conguration of an
elastic medium. For a given simulation time T > 0, we let the vector valued
function u : [0, T ] Rd denote the displacement eld and : [0, T ] R
denote the temperature. To dene boundary conditions for u we let uD and
uN be two disjoint parts of the boundary such that uD uN = . On the
part denoted uD we impose Dirichlet boundary conditions corresponding to a
clamped part of the material. On uN , corresponding to the traction boundary,
we impose Neumann boundary conditions. Similarly, we dene D and N to
be the drained and ux part of the boundary for the temperature .
Under the assumption that the displacement gradients are small, the strain
tensor is given by the following linear relation
1
(u) = (u + u ).
2
For isotropic materials the total stress tensor is given by
= 2(u) + ( u)I I,
where I is the d-dimensional identity matrix and is the thermal expansion
coecient. Furthermore, and denotes the Lame coecients satisfying
E E
= , = ,
2(1 + ) (1 + )(1 2)
where denotes Poissons ratio and E denotes Youngs elastic modulus. Pois-
sons ratio is a measure on the materials tendency to shrink (expand) when
stretched (compressed) and Youngs modulus describes the stiness of the ma-
terial. The coecients , , and are all material dependent and thus rapidly
varying in space for strongly heterogeneous (multiscale) materials.
4
 Background

Now, Cauchys equilibrium equations states that

= f,

where f : Rd denotes the external body forces. Furthermore, the temper-

ature in the material can be described by the parabolic equation

+ u = g,

where : Rdd is the heat conductivity parameter and g denotes internal

heat sources. Note that is material dependent and thus rapidly varying. To
summarize, the linear thermoelastic system is given by the following system of
equations
(1.1) (2(u) + uI I) = f, in (0, T ] ,
(1.2) + u = g, in (0, T ] ,
(1.3) u = 0, in (0, T ] uD ,
(1.4) n = 0, in (0, T ] uN .
(1.5) = 0, on (0, T ] D ,
(1.6) n = 0, on (0, T ] N .
(1.7) (0) = 0 , in ,
where we for simplicity assume homogeneous boundary conditions. Note that
the equations (1.1)-(1.2) are coupled.

Remark 1.1. The system (1.1)-(1.7) is formally equivalent to a linear model

for poroelasticity. In this case denotes the uid pressure, the hydraulic
conductivity, and the Biot-Willis coupling-deformation coecient. Hence,
the results in this thesis also apply to the linear poroelastic system.

To dene a FEM (and a GFEM) for (1.1)-(1.7) we dene the corresponding

variational (or weak) formulation. For this purpose we rst need to introduce
some notation and spaces. We use (, ) to denote the inner product in L2 ()
and   the corresponding norm. Let H 1 () := W21 () denote the classical
Sobolev space with norm v2H 1 () = v2 + v2 and let H 1 () denote the
dual space to H 1 . Furthermore, let Lp ([0, T ]; X) denote the Bochner space with
norm
 T 1/p
vLp ([0,T ];X) = vpX dt , 1 p < ,
0
vL ([0,T ];X) = ess sup vX ,
0tT

where X is a Banach space equipped with the norm  X . The dependence

on the interval [0, T ] and the domain is frequently suppressed and we write,
for instance, L2 (L2 ) for L2 ([0, T ]; L2 ()). We also use the double-dot product
5
 Introduction

notation to denote the Frobenius inner product of two matrices A and B


d
A:B= Aij Bij , A, B Rdd .
i,j=1

Now, dene the following spaces

V 1 := {v (H 1 ())d : v = 0 on uD }, V 2 := {v H 1 () : v = 0 on D }.
Multiplying (1.1) with v1 V 1 and (1.2) with v2 V 2 and using Greens
formula together with the boundary conditions (1.3)-(1.6) we arrive at the fol-
lowing variational formulation; nd u(t, ) V 1 and (t, ) V 2 such that, for
a. e. t > 0,
(1.8) ((u) : (v1 )) (, v1 ) = (f, v1 ), v1 V 1 ,
(1.9) (, v2 ) + (, v2 ) + ( u, v2 ) = (g, v2 ), v2 V 2 ,
and the initial value (0, ) = 0 is satised. Here (u) := 2(u) + uI is
the rst part of involving only the displacement u, commonly referred to as
the eective stress tensor.
Two functions u and are weak solutions if (1.8)-(1.9) are satised and
u L2 (V 1 ), u L2 (H 1 ), L2 (V 2 ), and L2 (H 1 ). Existence and
uniqueness of such weak solutions are proved in, e.g., [17, 16], and in [14] within
the framework of linear degenerate evolution equations in Hilbert spaces. In
[14] it is also proved that the system is of parabolic type, meaning that it is
well posed for nonsmooth initial data with regularity estimates depending on
negative powers of t.

2. Classical nite element

In this section we explain more carefully why the classical FEM fails to
approximate the solution to problems with rapidly varying data. To simplify
the discussion we start by considering elliptic equations.

2.1. Elliptic equations. Consider the elliptic equation

Au = f, in ,
u = 0, on ,
with the variational formulation; nd u V , such that
(2.1) a(u, v) = (f, v), v V,
where V = H01 () and a(u, v) := (Au, v). Here A : Rdd the diusion
coecient is assumed to be rapidly oscillating.
To dene a FEM we need a triangulation of the domain. Let {Th }h>0 be a
family triangulations of with the mesh size hK := diam(K), for K Th and
denote the largest diameter in the triangulation by h := maxKTh hK . Now
let Vh V denote the space of continuous piecewise ane functions on the
6
 Classical nite element

triangulation Th . The nite element formulation then reads; nd uh Vh , such

that,
(2.2) a(uh , v) = (f, v), v Vh .
Classical a priori error analysis gives the bound
(2.3) uh uH 1 ChD2 u,
where D2 u denotes the second order (weak) derivatives of u. Not only does this
bound require additional regularity of the solution, the norm D2 u may also
be very large if A is rapidly oscillating. Indeed, if A varies with frequency
1
for some
> 0, then AL = O(
1 ). Estimating D2 u with the problem
data gives
D2 u Cu CA (u) = C (Au) Au
C (Au) + AL u C(1 +
1 )f ,
where we used elliptic regularity in the rst inequality and the bound uH 1
Cf , derived from (2.1), in the last inequality. Furthermore, we can derive
the bound uh H 1 Cf  from (2.2), which gives the following upper bound
of the error; uh uH 1 Cf . Hence, the error bound (2.3) takes the form
 
h
uh uH 1 C min h + , 1 f ,

and convergence does not take place unless h <

. If
is small, the condition
h <
, can be devastating considering computational cost and available memory.
2.2. Linear thermoelasticity. As in the previous section we dene a
family of triangulations {Th }h>0 and we let Vh1 V 1 and Vh2 V 2 denote
nite element spaces consisting of continuous piecewise linear functions on this
triangulation. Furthermore, we let 0 = t0 < t1 < ... < tN = T be a uniform
discretization of the time interval such that tj tj1 = > 0 for j = 1, ..., N .
The classical FEM with a backward (implicit) Euler discretization in time for
(1.8)-(1.9) reads; for n {1, ..., N } nd unh Vh1 and hn Vh2 , such that
(2.4) ((unh ) : (v1 )) (hn , v1 ) = (f n , v1 ), v1 Vh1 ,
(2.5) (t n , v2 ) + (n , v2 ) + ( t un , v2 ) = (g n , v2 ),
h h h v2 Vh2 ,
where t hn := (hn hn1 )/ and similarly for t unh . Here u0h = uh,0 and
h0 = h,0 , where uh,0 Vh1 and h,0 Vh2 denote suitable initial conditions. The
right hand sides are evaluated at time tn , that is, f n := f (tn ) and g n := g(tn ).
A priori analysis for the system (2.4)-(2.5) can be found in [7]. It follows
that the error is bounded by
 n 1/2
unh un H 1 + hn n  + hj j 2H 1 C1 h + C,
j=1

where the constant C1 depends on both u(tn )H 2 and (tn )H 2 . By argu-
ments similar to the ones used for the elliptic equation in Section 2.1, we get that
7
 Introduction

u(tn )H 2 = O(
1 ) and (tn )H 2 = O(
1 ), if the material has variations on
a scale of size
.

3. A generalized nite element method

In [12] a GFEM, often referred to as local orthogonal decomposition, is
proposed and analyzed for elliptic equations of the form (2.1). In Section 3.1
below we describe this method and the main ideas used in the analysis. Finally,
in Section 3.4 we describe how this method can be generalized to dene a GFEM
for linear thermoelasticity, which is the main objective of this thesis.

3.1. Elliptic equations. The method proposed in [12] builds on the ideas
from the variational multiscale method [10, 11], where the solution space is
decomposed to into a coarse and a ne part. The nodal basis functions in the
coarse space is then modied by adding a correction from the ne space.
We begin by assuming that the mesh size h used in the classical FEM in
(2.2) is x and suciently small, that is h <
, such that the error (2.3) is small.
In this case, the solution uh and the space Vh are referred to as the reference
solution and the reference space, respectively. Now dene VH similarly to Vh
but with a larger mesh size H > h, such that VH Vh . Note that the classical
FEM solution uH in the coarse space VH is not a good approximation to u. It
is, however, cheaper to compute than uh since dim(VH ) < dim(Vh ). The aim
is now to dene a new multiscale space Vms with the same dimension as the
coarse space VH , but with better approximations properties.
To dene such a space, we need an interpolation operator IH : Vh VH
with the properties IH IH = IH and for K TH
1
(3.1) HK v IH vL2 (K) + IH vL2 (K) CI vL2 (K ) , v Vh ,

where K := {K TH : K K = }. For a quasi-uniform mesh, the bounds

in (3.1) can be summed to achieve a global bound
(3.2) H 1 v IH v + IH v Cv,
There are many interpolations operators that satisfy these conditions, for in-
stance, the global L2 -projection. In Paper II and Paper III we use an in-
terpolation of the form IH = EH H , where H is the L2 -projection onto
P1 (TH ), the space of functions that are ane on each triangle K TH and
EH : P1 (TH ) VH is an averaging operator. We refer to [13, 6] for further
details and possible choices of IH .
Now let Vf denote the kernel to the operator IH
Vf := ker IH = {v Vh : IH v = 0}.
The space Vh can be decomposed as Vh = VH Vf , meaning that vh Vh can
be decomposed into
(3.3) vh = vH + vf , v H VH , v f Vf .

8
 A generalized nite element method

The kernel Vf is a ne scale (detail) space in the sense that it captures all
features that are not captured by the coarse space VH . Let Rf : Vh Vf denote
the Ritz projection onto Vf , that is,
(3.4) a(Rf v, w) = a(v, w), w Vf , v Vh .
Because of the decomposition (3.3) we have the identity
vh Rf vh = vH + vf + Rf (vH + vf ) = vH Rf vH ,
since vf Vf . Using this we can dene the multiscale space Vms
(3.5) Vms := Vh Rf Vh = VH Rf VH .
Note that Vms is the orthogonal complement to Vf with respect to the inner
product a(, ) and must have the same dimension as VH . Indeed, with N
denoting the inner nodes in TH and z the basis function at node z, a basis for
Vms is given by
{z N : z Rf z }.
Hence, that basis functions are the classical nodal basis functions modied by
corrections Rf z computed in the ne scale space.
Replacing Vh with Vms in (2.2) we can now dene the GFEM; nd ums
Vms , such that,
(3.6) a(ums , v) = (f, v), v Vms .
The following theorem gives an a priori bound for the GFEM and can be found
in [12]. We include the proof here since it is short and highlights the main ideas
used in the analysis.
Theorem 3.1. Let uh be the solution to (2.2) and ums the solution to (3.6).
Then
ums uh H 1 CHf ,
where C does not depend on the derivatives of A.
Proof. Dene e := ums uh and note that e Vf . Hence, IH e = 0.
Furthermore we have due to Galerkin orthogonality a(e, vms ) = 0 for vms Vms .
Using this together with the interpolation bound (3.2) we have
a(e, e) = a(e, uh ) = (f, e) f e = f e IH e CHf e,
and the bound follows by using equivalence of the energy norm induced by a(, )
and the H 1 -norm. 
From Theorem 3.1 we have that the solution given by the GFEM converges
to uh , with optimal order, independently of the derivatives (variations) of A.
We emphasize that the total error is bounded by
ums uH 1 ums uh H 1 + uh uH 1 ,
where the error in the second term is due to the classical FEM and assumed to
be of reasonable size, since h is assumed to be suciently small.
9
 Introduction

Although the a priori analysis seems promising, the GFEM as suggested

above suers from some drawbacks. The problem of nding the corrections
Rf z , which are needed to construct the basis, are posed in the entire ne scale
space Vf which has the same dimension as Vh . Furthermore, the corrections
generally have global support and therefore destroys the sparsity of the resulting
discrete system. Both issues are resolved by performing a localization of the
corrections. The localization is motivated by the observation that the correction
Rf z decay exponentially away from node z.
3.2. Localization. In [12] it is proved that the corrections decay expo-
nentially and a localization procedure is proposed. However, in [8] a dierent
localization technique is proposed which allows for smaller patches to be used.
We describe the procedure in [8] here, which is also the procedure that is used
in the appended papers.
We dene patches of size k in the following way; for K TH
0 (K) := int K,


k (K) := int {K TH : K k1 (K) = } , k = 1, 2, ...,
and let Vf (k (K)) := {v Vf : v(z) = 0 on \ k (K)} be the restriction of Vf
to the patch k (K).
We proceed by noting that Rf in (3.4) can be written as the sum

Rf = RfK ,
KTH

where RfK : Vh Vf and fullls

(3.7) a(RfK v, w) = a(v, w)K , w Vf , v Vh , K TH ,
where we dene
a(v, w)K := (Av, w)L2 (K) , K TH .
The aim is to localize these computations by replacing Vf with Vf (k (K)). De-
K
ne Rf,k : Vh Vf (k (K)) such that
K
a(Rf,k v, w) = a(v, w)K , w Vf (k (K)), v Vh , K TH ,
 K
and set Rf,k := KTH Rf,k . We can now dene the localized multiscale space
(3.8) Vms,k = {vH Rf,k vH : vH VH }.
By replacing Vms with Vms,k in (3.6) a localized GFEM can be dened; nd
ums,k Vms,k such that
(3.9) a(ums,k , v) = (f, v), v Vms,k .
Since the dimension of Vf (k (K)) can be made signicantly smaller than
the dimension of Vf (depending on k), the problem of nding Rf,k z is compu-
tationally cheaper than nding Rf z . Moreover, the resulting discrete system
is sparse. It should also be noted that the computation of Rf,k z for all nodes
z is suitable for parallelization, since they are independent of each other.
10
 A generalized nite element method

The convergence of the method (3.9) depends on the size of the patches. In
[12, 8] the following Theorem is proved.
Theorem 3.2. Let uh be the solution to (2.2) and ums,k the solution to
(3.9). Then there exists (0, 1) such that
ums,k uh H 1 C(H + k d/2 k )f ,
where C does not depend on the derivatives of A.
To achieve linear convergence k should be chosen proportional to log H 1 ,
that is, k = c log H 1 , for some constant c.

3.3. Parabolic equations. A natural rst step in generalizing the GFEM

to linear thermoelasticity is to rst extend it to a time dependent problem of
parabolic type. Recall that the thermoelastic system (1.8)-(1.9) is parabolic
[14]. This is the subject of Paper I.
We consider a parabolic problem on the following weak form; nd u(t) V ,
such that, u(0) = u0 and
(3.10) (u, v) + a(u, v) = (f, v), v V,
where a(u, v) = (Au, v) as in the elliptic equation (2.1). The diusion
coecient A : Rdd is assumed to not depend on time.
The classical FEM for (3.10) with a backward Euler discretization reads;
for n {1, ..., N } nd unh Vh , such that, u0h = uh,0
(3.11) (t unh , v) + a(unh , v) = (f n , v), v Vh ,
with the notation and time discretization as in Section 2.2 and uh,0 a suitable
approximation of u0 . It is well known, see, e.g., [15], that the following error
estimate holds for the parabolic equation
unh u(tn )H 1 C1 h + C,
where C1 is constant depending on, among other terms, u(tn )H 2 and is thus
of size
1 if A varies on scale of size
. Hence, parabolic problems suers from
the same issues as elliptic problems when using classical nite element.
In the error analysis of the classical FEM, the error is usually split into the
two parts
unh u(tn ) = unh Rh u(tn ) + Rh u(tn ) u(tn ) =: n + n ,
where Rh : V Vh is the Ritz projection given by
a(Rh v, w) = a(v, w), w Vh , v V.
The error of the Ritz projection is given by the analysis of the elliptic problem
(3.12) Rh v vH 1 ChD2 v.
This directly gives the error of n . Indeed, n H 1 ChD2 u(tn ), where
D2 u(tn ) C1  Au(tn ) = C1 f n u(tn ) and C1 depend on the
11
 Introduction

derivatives of A. Furthermore, to bound n H 1 we put n into (3.11), which

gives
(t n , v) + a(n , v) = ((Rh I)t u(tn ) + (t u(tn ) u(tn )), v)
=: (t n + , v),
where the error of t n follows from (3.12) and the error of follows from
Taylors formula. In order to bound n in the H 1 -norm we can choose v = t n .
Inspired by this we propose the following GFEM for the parabolic problem,
where the space Vh in (3.11) is simply replaced by the multiscale space Vms
dened in Section 3.1; for n {1, ..., N } nd unms Vms , such that, u0ms = ums,0
(3.13) (t unms , v) + a(unms , v) = (f n , v), v Vms ,
with ums,0 a suitable approximation of uh,0 . Now, because of the choice of the
space Vms we can dene a Ritz projection Rms : Vh Vms by
a(Rms v, w) = a(v, w) = (Ah v, w), v Vms ,
where Ah : Vh Vh is the operator dened by
(Ah v, w) = a(v, w), w Vh .
The error analysis for the elliptic problem in [12] gives the bound
(3.14) Rms v vH 1 CHAh v, v Vh ,
where C is independent of the derivatives of A. The assumption that A does
not depend on time is crucial here. Otherwise, we would have to dene a new
space and compute a new set of basis functions at each time step tn .
As for the elliptic equation we assume that h is suciently small to resolve
the variations in A. This means that the reference solution uh given by (3.11)
approximates u in (3.10) suciently well. In the error analysis we can thus split
unms u(tn )H 1 unms unh H 1 + unh u(tn )H 1 ,
where the second part is bounded by classical FEM error analysis. For the rst
part we can use a similar analysis, but with the new Ritz projection Rms . We
split the error into the parts
unms unh = unms Rms unh + Rms unh unh =: ms
n
+ nms ,
where the error of nms is given by (3.14) and Ah unh = Ph f n t unh with Ph
n n
denoting the L2 -projection onto Vh . For ms we get by plugging ms into (3.13)
n n
(t , v) + a( , v) = (t , v), v Vms .
n
ms ms ms

Naturally, the error bound in this case depends on the regularity of the
(discrete) time derivative of the reference solution. Since the initial data is not
in H 2 we expect, for instance, t unh  to depend on negative powers of tn . This
is possible since the backward Euler scheme preserves the smoothing eect of
parabolic problems. In Paper I this is thoroughly investigated and error bounds
involving negative powers of tn are derived.
12
 A generalized nite element method

To utilize the localization introduced in Section 3.1 we can replace Vms by

Vms,k , dene a new Ritz projection Rms,k : Vh Vms,k , and perform similar
splits of the error.

3.4. Linear thermoelasticity. In the classical nite element error anal-

ysis for linear thermoelasticity, a Ritz projection related to the stationary form
of the problem is used to split the error into two terms. This Ritz projec-
tion is dened by the following Rh (v1 , v2 ) : V 1 V 2 Vh1 Vh2 , such that,
Rh (v1 , v2 ) = (Rh1 (v1 , v2 ), Rh2 v2 ) and for all (v1 , v2 ) V 1 V 2 ,

((v1 Rh1 (v1 , v2 )) : (w1 )) ((v2 Rh2 v2 ), w1 ) = 0, w1 Vh1 ,

((v2 Rh2 v2 ), w2 ) = 0, w2 Vh2 .

with error estimates (see [7, Lemma 2.2])

(3.15) v1 Rh1 (v1 , v2 )H 1 ChD2 v1  + Cv2 Rh2 v2 ,

(3.16) v2 Rh2 v2 H 1 ChD2 v2 .

The error can now be split according to

unh u(tn ) = unh Rh1 (u(tn ), (tn )) + Rh1 (u(tn ), (tn )) u(tn ) =: h,u
n
+ nh,u ,
hn (tn ) = hn Rh2 (tn ) + Rh2 (tn ) (tn ) =: h,
n
+ nh, ,

where the error of nh,u and nh, follows from (3.15)-(3.16). The rst parts h,u
n
n
and h, can be plugged into the equation (2.4)-(2.5) to derive error estimates
for these. Compare to the parabolic problem in Section 3.3. For the details we
refer to [7].
To derive a GFEM for the thermoelasticity problem (1.8)-(1.9) we need
to decompose two dierent spaces; Vh1 and Vh2 . The decomposition of Vh1 is
performed with respect to the bilinear form (() : ()) and the decomposition
of Vh2 with respect to (, ). This is done by mimicking the procedure
1
described in Section 3.1. First dene two interpolations IH : Vh1 VH1 and
IH : Vh VH into the coarse nite element spaces VH Vh and VH2 Vh2 .
2 2 2 1 1

Now, the corresponding kernels are Vf1 := ker IH 1

and Vf2 := ker IH
2
, and we can
dene the Ritz projections onto the these Rf : Vh Vf and Rf2 : Vh2 Vf2
1 1 1

given by

((v1 Rf1 v1 ) : (w1 )) = 0, w1 Vf1 , v1 Vh1

((v2 Rf2 v2 ), w2 ) = 0, w2 Vf2 , v2 Vh2 .

The multiscale spaces are nally dened as

1
Vms := VH1 Rf1 VH1 , 2
Vms := VH2 Rf2 VH2 ,

as in (3.5). With these spaces we can now dene a Ritz projection corresponding
to the stationary system. Dene Rms (v1 , v2 ) : Vh1 Vh2 Vms
1 2
Vms , such that,
13
 Introduction

1 2
Rms (v1 , v2 ) = (Rms (v1 , v2 ), Rms v2 ) and for all (v1 , v2 ) Vh1 Vh2 ,
1 2 1
((v1 Rms (v1 , v2 )) : (w1 )) ((v2 Rms v2 ), w1 ) = 0, w1 Vms ,
2 2
((v2 Rms v2 ), w2 ) = 0, w2 Vms .
1 2
The spaces Vms and Vms are designed to handle multiscale behavior in the
coecients , , and respectively. However, is also material dependent and
can be expected to vary at the same scale. For this reason, we shall add an
extra correction to the solution Rms (v1 , v2 ) inspired by the techniques in [11, 8].
This additional correction is dened as Rf : Vh2 Vf1 , such that,
2
((Rf v2 ) : (w1 )) = (Rms v2 , w1 ), w1 Vf1 ,
1 1
and we dene Rms (v1 , v2 ) = Rms (v1 , v2 ) + Rf v2 . Using the two operators A1 :
Vh Vh Vh and A2 : Vh Vh2 dened by
1 2 1 2

(A1 (v1 , v2 ), w1 ) = ((v1 ) : (w1 )) (v2 , w1 ), w1 Vh1 ,

(A2 v2 , w2 ) = (v2 , w2 ), w2 Vh2 ,
we prove, in Paper III, that the following error bounds hold for any (v1 , v2 )
Vh1 Vh2
1 2
(3.17) v1 Rms (v1 , v2 )H 1 CHA1 (v1 , v2 ) + Cv2 Rms v2 ,
2
(3.18) v2 Rms v2 H 1 CHA2 v2 ,
where C is independent of the variations in , , , and .
The following system denes a GFEM for the time dependent problem (2.4)-
1
(2.5). For n {1, ..., N } nd unms = unms + unf , with unms Vms and unf Vf1 ,
2
and ms Vms , such that
n

1
(3.19) ((unms ) : (v1 )) (ms
n
, v1 ) = (f n , v1 ), v1 Vms ,
2
(3.20) (t ms
n
, v2 ) + (n , v2 ) + ( t un , v2 ) = (g n , v2 ),
ms ms v2 Vms ,
(3.21) n
((uf ) :
(w1 )) (ms
n
, w1 ) = 0, w1 Vf1 ,
where u0ms = ums,0 and ms0
= ms,0 are suitable approximations of uh,0 and
h,0 (see Paper III). Here we have added an additional correction, unf , on unms
inspired by the correction in the stationary setting. Following the classical nite
element analysis one can now split the error according to
1 1
unms unh = unms Rms (unh , hn ) + Rms (unh , hn ) unh =: ms,u
n
+ nms,u ,
2 2
n
ms hn = ms
n
Rms hn + Rms hn hn =: ms,
n
+ nms, ,
where the error of nms,u and nms, are bounded by (3.17)-(3.18). The error of
n n
u,ms and ,ms follows by plugging these into (3.19)-(3.21). However, in this
1
case ms,u  Vms
n
, which needs to be taken into account in the analysis.
1 2
To proceed we need to perform a localization of both spaces Vms and Vms .
1
We use the patches k (K) dened in Section 3.2 to dene localized spaces Vms,k
2
and Vms,k , as in (3.8). To motivate this we need to show that the corrections
Rf x and Rf2 y decay exponentially away from node x and y, where x and y
1

14
 Summary of papers

denotes the classical hat functions in VH1 and VH2 respectively. The correction
Rf2 y is based on the bilinear form (, ) of the same type as in Section 3.1
and the decay thus follows directly from [12, 8]. The correction Rf1 x is based
on the elasticity form (() : ()) and the decay does not follow directly from
the earlier results. This is instead proven in Paper II.
The localized GFEM for (2.4)-(2.5) is now dened as; for n {1, ..., N }
nd
 n,K
1 1
unms,k = unms,k + uf,k , with unms,k Vms,k , un,K
f,k Vf (k (K)),
KTH

2
and n
ms,k Vms,k , such that
1
(3.22) ((unms,k ) : (v1 )) (ms,k
n
, v1 ) = (f n , v1 ), v1 Vms,k ,
(t n , v2 ) + (n , v2 )
ms,k ms,k
2
(3.23) + ( t unms,k , v2 ) = (g n , v2 ), v2 Vms,k ,
(3.24) ((un,K
f,k ) : (w1 )) (ms,k , w1 )K = 0,
n
w1 Vf1 (k (K)).

The main theorem in this thesis is Theorem 3.3 below and is proved in Paper
III under certain conditions on the size of H. Here Cf,g denotes a constant
depending on f and g, see Paper III for details.
Theorem 3.3. Let {unh }N n=1 and {h }n=1 be the solutions to (2.4)-(2.5) and
n N

{unms,k }N
n=1and {ms,k }n=1 the solutions to (3.22)-(3.24). For n {1, ..., N }
n N

we have


unh unms,k H 1 + hn ms,k
n
H 1 C(H + k d/2 k ) Cf,g + tn1/2 h0 H 1 ,
where C and Cf,g are constants independent of the variations in , , , and .

4. Summary of papers
Paper I. In Paper I we propose and analyze the GFEM (3.13) for parabolic
equations with highly varying and oscillating coecients. We prove convergence
of optimal (second) order in the L2 -norm to the reference solution assuming
initial data only in L2 . We do not assume any structural conditions on the
multiscale coecient, such as, periodicity or scale separation. Furthermore, we
show how to extend this method to semilinear parabolic problems, where the
right hand side in (3.10) is replaced by f (u).

Paper II. In Paper II we propose a GFEM for linear elasticity equations

with applications in heterogeneous materials. In particular, we prove expo-
nential decay of the corrections Rf1 z in Section 3.4. Furthermore, we prove
that the GFEM reduces the locking eect that occur for materials with large
Lame parameter when using classical continuous and piecewise linear nite
elements.
15
 Introduction

Paper III. In Paper III we build on the theory developed in Paper I and
Paper II (originating from [12]) to dene a GFEM for linear thermoelasticity
with highly varying coecients describing a heterogeneous material. We prove
linear convergence to the reference solution in the H 1 -norm independent of the
variations in the data, see Theorem 3.3 in Section 3.4.

5. Future work
In Paper I on parabolic equations we assume that the diusion coecient
A(x) is independent of time. A natural extension would be to include time
dependent coecients A(t, x). However, the main idea of the paper, to replace
Vh with the space Vms in (3.11), then fails. We would need to have a new
space Vmsn
for each time tn , since the diusion coecient A(tn , ) takes dierent
values for dierent times tn . This is considerably more expensive than the time
independent case, since we need to compute new corrections at each time step.
It is possible that a more rened strategy could be developed by working with
the parabolic problem in a space-time framework and perform localization in
both time and space.
In applications involving composite materials there may be uncertainties in
the material parameters, such as position or rotation, coming from the assembly
procedure. These uncertainties can, for instance, be modeled by letting the
coecients depend on a random variable . A rst step in extending the GFEM
framework to such problems could be to consider an elliptic problem of the form
A(x, )u(x, ) = f (x, ),
where A(, ) is multiscale in space for a x . This problem suers from the
same problem as the time dependent case, since A(, ) now takes dierent
values for dierent outcomes .
In the analysis of the localization the constant (0, 1), see e.g. Theo-
rem 3.2, depends on the contrast / of A, that is, the ratio between the max-
imal and minimal value obtained by A. Also the constant C in Theorem 3.2
depends on this ratio. However, in available numerical examples, see Paper I
and Paper II, but also, e.g., [12, 8], the size of the patches and the resulting
convergence does not seem to be aected by large contrasts. Thus, the error
bounds derived for the localization could be too crude. This should be further
investigated to derive sharper error bounds for special classes of A.

References
[1] I. Babuska and R. Lipton: Optimal local approximation spaces for gen-
eralized nite element methods with application to multiscale problems,
Multiscale Model. Simul. 9 (2011), no. 1, p. 373406.
[2] I. Babuska and J. E. Osborn: Generalized nite element methods: their
performance and their relation to mixed methods, SIAM J. Numer. Anal.
20 (1983), no. 3, 1983.
16
 References

[3] M. A. Biot: Thermoelasticity and irreversible thermodynamics, J. Appl.

Phys., 27 (1956), p. 240253.
[4] B. A. Boley and J. H. Weiner: Theory of thermal stresses, John Wiley &
Sons, Inc., New York London, 1960.
[5] E, Weinan and Engquist, Bjorn: The heterogeneous multiscale methods,
Commun. Math. Sci. 1 (2003), no. 1, p. 87132.
[6] Ch. Engwer and P. Henning and A. Malqvist and D. Peterseim: Ecient
implementation of the Localized Orthogonal Decomposition method, Sub-
mitted.
[7] A. Ern and S. Meunier: A posteriori error analysis of Euler-Galerkin ap-
proximations to coupled elliptic-parabolic problems, M2AN Math. Model.
Numer. Anal. 43 (2009), no. 2, p. 353 375.
[8] P. Henning and A. Malqvist: Localized orthogonal decomposition tech-
niques for boundary value problems, SIAM J. Sci. Comput. 36 (2014),
no. 4, p. A1609A1634.
[9] T. Y. Hou and X.-H. Wu: A Multiscale Finite Element Method for Elliptic
Problems in Composite Materials and Porous Media, J. Comput. Phys.
134 (1997), no. 1, p. 169189.
[10] T. J. R. Hughes, G. R. Feijoo, L. Mazzei, and J-B. Quincy: The vari-
ational multiscale methoda paradigm for computational mechanics,
Comput. Methods Appl. Mech. Engrg. 166 (1998), no. 1-2, p. 324.
[11] M. G. Larson and A. Malqvist: Adaptive variational multiscale methods
based on a posteriori error estimation: energy norm estimates for elliptic
problems, Comput. Methods Appl. Mech. Engrg. 196 (2007), no. 21-24,
p. 23132324.
[12] A. Malqvist and D. Peterseim: Localization of elliptic multiscale prob-
lems, Math. Comp. 83 (2014), no. 290, p. 25832603.
[13] D. Peterseim: Variational Multiscale Stabilization and the Exponential
Decay of Fine-scale Correctors, Submitted.
[14] R. E. Showalter: Diusion in poro-elastic media, J. Math. Anal. Appl.
251 (2000), no. 1, p.310340.
[15] V. Thomee: Galerkin Finite Element Methods for Parabolic Problems,
Springer Series in Computational Mathematics, Springer-Verlag, Berlin,
Second edition, 2009.
[16] A. Zensek: Finite element methods for coupled thermoelasticity and cou-
pled consolidation of clay, RAIRO Anal. Numer. 18 (1984), no. 2, p.
183205.
[17] A. Zensek: The existence and uniqueness theorem in Biots consolidation
theory, Apl. Mat. 29 (1984), no. 3, p. 194211.

17