Beruflich Dokumente
Kultur Dokumente
net/publication/311607691
CITATIONS READS
0 361
1 author:
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Phase-field modeling of damage and failure in solids and structures View project
All content following this page was uploaded by Vinh Phu Nguyen on 13 December 2016.
0.4pt 0.6pt
This page intentionally contains only this sentence.
Finite elements: from basis concepts to
nonlinear analysis
V. P. Nguyen
c Draft date December 8, 2015
This page intentionally contains only this sentence.
Contents
This page intentionally contains only this sentence.
List of Figures
This page intentionally contains only this sentence.
List of boxes
This page intentionally contains only this sentence.
Preface
The aim of this manuscript is to introduce major issues involved in the nonlinear
stress analysis of solid using the finite element method (FEM).
Nonlinear behavior of solid takes two forms, namely, material nonlinearity
and geometry nonlinearity. Nonlinear materials are widely used in practical ap-
plications. The nonlinear elasticity including the hypoelasticity, hyperelasticity
and viscoelasticity are presented in details.
When the deformation of solids reaches a state for which undeformed and
deformed shapes are considerably different, finite deformation occurs. In this
case, geometry nonlinearity must be taken into account.
The underlying block of the finite element method is the theory of continuum
mechanics. It is thefore to
The outline of the writting is as follows. We first start with some mathematical
preliminaries on vector, tensor and matrix analysis. The next section presents the
theory of continuum mechanics. Various material behaviors are introduced in the
next section. The finite element method is first introduced in the context of one
dimensional linear problem.
This writting reflects what the author have self-studied from the following
interesting books:
1. Nonlinear finite elements for continua and structures. Ted Belytschko, Wing
Kam Liu and Brian Moran.
1.1 Vector
Consider the three dimensional Euclidean space R3 with three orthogonal unit
vectors ei (i = 1, 2, 3). Then any vector, say, a can be written as
a = a1 e 1 + a2 e 2 + a3 e 2 = ai e i (1.1.1)
with ai are scalar variables. Unless other stated, the summation convention is used
for repeated index (Einstein summation convention). For computer implementa-
tion, it is useful to express vector a by a matrix:
a1
[a] = a2
(1.1.2)
a3
a · b = (ai ei ) · (bj ej ) = ai bj ei · ej
Since ei · ej = δij , the Kronecker delta, which is defined by
1 if i = j
δij = (1.1.4)
0 otherwise
Hence
a · b = ai bi = a1 b1 + a2 b2 + a3 b3 (1.1.5)
v = vi ei (1.1.6)
Sec. 1.1 Vector 3
v · ej = vi ei · ej = vi δij = vj (1.1.7)
The component of a vector is given by
vj = v · ej (1.1.8)
a × b = ai ei × bj ej = ai bj ei × ej (1.1.10)
where
e1 × e1 = 0 e1 × e2 = e3 e1 × e3 = −e2
e2 × e1 = −e3 e2 × e2 = 0 e2 × e3 = e1 (1.1.11)
e3 × e1 = e2 e3 × e2 = −e1 e3 × e3 = 0
This allows us to write
ej × ek = ijk ei (1.1.12)
where ijk is the permutation symbol or the Levi-Civita symbol, which is defined
by
+1 if (i, j, k) is (1, 2, 3), (2, 3, 1), or (3, 1, 2)
ijk = −1 if (i, j, k) is (3, 2, 1), (1, 3, 2), or (2, 1, 3) (1.1.13)
0 i = j, j = k, or k = i
a × b = aj bk ej × ek = aj bk ijk ei (1.1.14)
Denote c is the cross product of a × b, then we have c = aj bk ijk ei , i.e., the
components of c are ci = aj bk ijk , written explicitly
4 Preliminaries on vector, matrix and tensor Chap. 1
c1 = aj bk 1jk = a2 b3 − a3 b2
c2 = aj bk 2jk = a3 b1 − a1 b3 (1.1.15)
c3 = aj bk 3jk = a1 b2 − a2 b1
Using the concept of determinant, we can write the vector cross product as follows
e1 e2 e3
a × b = det a1 a2 a3 (1.1.16)
b1 b2 b3
[a, b, c] = (a × b) · c (1.1.17)
a1 a2 a3
[a, b, c] = det b1 b2 b3 (1.1.18)
c1 c2 c3
s = vi vi
= Rij vj0 Rik vk0
= vj0 Rij Rik vk0 (1.1.21)
= vj0 δjk vk0
= vj0 vj0
Sec. 1.2 Matrix 5
1.2 Matrix
1.2.1 Introduction
Consider the familiar system of equations
Ax = b (1.2.2)
Diagonal matrix
a11 a12 a13 . . . a1n
a12 a22 a23 . . . a2n
A = a13 a23 a33 . . . a3n (1.2.5)
.. .. .. .. ..
. . . . .
a1n a2n a3n . . . ann
6 Preliminaries on vector, matrix and tensor Chap. 1
∂Aij
= δim δjn (1.2.9)
∂Amn
The trace of a square matrix
∂tr(A)
= δij (1.2.11)
∂Aij
The matrix inner product
Providing two matrices of the same dimension A and B, the matrix inner
product is a scalar obtained by summing terms in which each element of A is
multiplied by the corresponding element in B. That is
Sec. 1.3 Matrix 7
N X
X M
A∗B= Amn Bmn (1.2.12)
n=1 m=1
Inverse of a matrix
The inverse of a square matrix A, denoted by A−1 , is defined by
(AB)C = I (1.2.15)
B11 + B11 2 B12 + B21 B21 + B12 B22 + B22 2
v1 + v1 v2 + v1 v2 + v2
2 2 2 2
(1.2.19)
1
vT (B + BT )v (1.2.20)
2
Positive definite depends only on the symmetric part of the matrix.
8 Preliminaries on vector, matrix and tensor Chap. 1
1.3 Tensor
1.3.1 Introduction
Most of physical quantities that are important in continuum mechanics such as
temperature, force and stress can be expressed as tensors. Temperature is suffi-
ciently specified by a single scalar quantity called zeroth-order tensor. A force,
however, requires a magnitude and a direction, which is an example of a first-
order-tensor. Stress is more complicated since it requires a magnitude and two
directions, namely the direction of the force and the direction of the plane on
which the force acts. Stresses are represented by second-order tensors.
x = x1 e1 + x2 e2 + x3 e3 (1.3.2)
Applying the linear transformation f to the vector x,
yj ej = xi Fji ej (1.3.8)
The components of vector y can therefore be written in terms of Fij and xi ,
yj = xi Fji (1.3.9)
In matrix form,
y = [F]x (1.3.10)
with
F11 F12 F13
[F] = F21 F22 F23 (1.3.11)
F31 F32 F33
The matrix F is called the matrix of components of the tensor associated with
the linear transformation f . It is worthy noting that matrix F refers to the basis ei .
Then, similar to vector, we could expect that the tensor can be expressed in terms
of its component matrix and its basis
F = Fij ei ⊗ ej (1.3.13)
Since F is a tensor, so ei ⊗ ej is. The basis dyad ei ⊗ ej has the component matrix
that contains all zeros except for unity at the position ij. For example,
10 Preliminaries on vector, matrix and tensor Chap. 1
0 1 0
e1 ⊗ e2 = 0 0 0 (1.3.14)
0 0 0
(a ⊗ b) · v = a(b · v) (1.3.15)
with the right-hand size is a vector which can be written explicitly as
a1 b1 a1 b2 a1 b3 v1
a(b · v) = a2 b1 a2 b2 a2 b3
v2 (1.3.16)
a3 b 1 a3 b 2 a3 b 3 v3
Therefore, the matrix component of a dyad a ⊗ b is the matrix given by
[a ⊗ b] = abT (1.3.17)
tr(S) = S : I = I : S (1.3.21a)
(u ⊗ v) : (x ⊗ y) = (u · x)(v · y) (1.3.21b)
S : W = 0 if S = ST ; W = −WT (1.3.21c)
The proof for the last property is as follows:
A · v = Aij vj ei (1.3.31)
Hence, dotting a second-order tensor from the right by a vector results in a vec-
tor whose the ith component is Aij vj , i.e., components of A · v is the matrix
multiplication Av.
(u · A) · v = u · (A · v) (1.3.33)
The right hand side of the above can be written as
u · A = AT · u (1.3.35)
Sec. 1.3 Tensor 13
A = − · a (1.3.40)
Since A depends on a not on the vector x, we could consider the above equation
as a vector to tensor operation denoted by Ω<a> . We therefore can state that, for
any vector a, it is possible to construct a skew-symmetric axial tensor as
Ω<a> = − · a (1.3.41)
whose components are given by
Ω<a>
ij = −ijk ak (1.3.42)
In three dimensions, the component matrix of the axial tensor is
0 −a3 −a2
Ω<a> = a3 0 −a1 (1.3.43)
a2 a1 0
Cross product of two vectors can be replaced by a dot product between the
axial tensor and a vector.
The derivatives of the invariants to the tensor are important and given by
∂I1 ∂I1
= δij , =I
∂Aij ∂A
∂I2 ∂I2
= Akk δij − Aji , = tr(A)I − AT (1.3.45)
∂Aij ∂A
∂I3 ∂I3
= det(A)A−T ij , = det(A)A−T
∂Aij ∂A
∂I3
The last equation, ∂A ij
can be verified using the concept of directional derivative.
If A is a symmetric tensor, then it has three real eigenvalues λi and three
orthogonal eigenvectors ni . That is
Aij = ni · A · nj
= ni · λj nj
(1.3.47)
= λj ni · nj
= λj δij
The above shows that the component matrix of the tensor A is a diagonal matrix,
explicitly written in matrix form as
λ1 0 0
A = 0 λ2 0 (1.3.48)
0 0 λ3
The tensor A now can be expressed by
A = λi ni ⊗ ni (1.3.49)
This expression is particularly usefull to evaluate, for example, the square root of
tensor A,
p
A1/2 = λi ni ⊗ ni (1.3.50)
If the tensor A is, in addition to symmetric, positive definite then we have
ni · A · ni > 0 (1.3.51)
Sec. 1.3 Tensor 15
ni · λi ni > 0
(1.3.52)
λi ||ni ||2 > 0 =⇒ λi > 0
The above means that any symmetric, positive definite second order tensor has
three positive eigenvalues.
In terms of the eigenvalues, the invariants can be expressed as
I1 (A) = Aii = λ1 + λ2 + λ3
1 2
I2 (A) = Aii − Aij Aji = λ1 λ2 + λ2 λ3 + λ3 λ1 (1.3.53)
2
I3 (A) = det A = λ1 λ2 λ3
As will be shown later, the invariants and eigenvalues are very useful in mod-
elling hyperelastic materials.
v · ej = v̂i êi · ej
(1.3.55)
vj = Qji v̂i , Qji = êi · ej
v = Qv̂ (1.3.56)
Following the same derivation, we have
v̂ = QT v (1.3.57)
Premultiplying Eq. (1.3.56) by QT and compare to Eq. (1.3.57), we obtain
QT Qv̂ = v̂ (1.3.58)
16 Preliminaries on vector, matrix and tensor Chap. 1
v̂ = R · v (1.3.61)
I:S=S (1.3.63a)
I˜ : S = ST (1.3.63b)
C = Cijkl ei ⊗ ej ⊗ ek ⊗ el (1.3.64)
E XAMPLE
∂
∇= ei = ∂i ei (1.4.1)
∂xi
where ∂i denotes the partial differentiation with respect to xi . When applied to a
scalar-valued differentiable function f (x), it gives
∂f
∇f = (∂i ei )f = f,i ei = ei (1.4.2)
∂xi
So, it is the vector field with components f,i . To a differentiable vector field u, the
gradient operator can be applied from the left or from the right
∂uj
∇u = (∂i ei )(uj ej ) = ei ej
∂xi
(1.4.3)
∂ui
u∇ = (uj ej )(∂i ei ) = ei ej
∂xj
∂uj
∇u = ei ⊗ ej (1.4.4)
∂xi
The divergence of a vector field u ( continuously differentiable) is defined to be
the scalar-valued function:
∂Aij
∇·A= ei (1.4.6)
∂xj
Laplacian of a scalar ϕ is
∂ 2ϕ ∂ 2ϕ ∂ 2ϕ
∇2 ϕ = ∇ · ∇ϕ = + 2 + 2 (1.4.7)
∂x21 ∂x2 ∂x3
Laplacian of a vector u is
∇2 u = ∇2 u1 e1 + ∇2 u2 e2 + ∇2 u3 e3 (1.4.8)
Derivative of a vector with respect to itself
∂xi
= δij (1.4.9)
∂xj
(A − λI)x = 0 (1.6.2)
The above will have nontrivial solutions if and only if
Axi = λi xi (1.6.4)
1.7.1 Definition
Consider a scalar function f (x) and a vector u, the directional derivative of f (x)
at x in the direction of an increment u is defined by
f (x + u) − f (x)
Df (x)[u] = lim (1.7.1)
→0
where is an artificial parameter which facilitates the computation of the deriva-
tives. We next derive the formula to compute the directional derivative of a given
function in a given direction. To do that, consider a function given by
g() − g(0)
g 0 (0) = lim
→0
f (x + u) − f (x) (1.7.3)
= lim
→0
= Df (x)[u]
where in the third equality, the definition of directional derivative in Eq. (1.7.1)
has been used. We therefore obtain the following equation
d
Df (x)[u] = f (x + u) (1.7.4)
d =0
20 Preliminaries on vector, matrix and tensor Chap. 1
We could derive the relation between directional derivative and partial deriva-
tives by using the chain rule for the LHS of the above,
∂f
Df (x)[u] = ui = ∇f · u (1.7.5)
∂xi
Similarly, we have the counterpart of the above for second-order tensor fields,
∂Fij ∂F
DF(X)[U] = Ukl = :U (1.7.6)
∂Xkl ∂X
DF(x0 )[u] = DF1 (x0 )[u] · F2 (x0 ) + F1 (x0 ) · DF2 (x0 )[u] (1.7.8)
f (x) = 0 (1.8.1)
Estimate solution x0 . Taylor’s series expansion in the neighborhood of x0
gives
1 d2 f
df
f (x) = f (x0 ) + (x − x0 ) + 2
(x − x0 )2 + · · · (1.8.2)
dx x0 2 dx x0
1 d2 f 2
df
f (x0 + u) = f (x0 ) + u + u + ··· (1.8.3)
dx x0 2 dx2 x0
Sec. 1.8 Nonlinear problems and linearization 21
df 0 1 f (x0 )
− f (x0 ) =(x )(x − x0 ) ⇒ x1 = x0 − (1.8.8)
dx df 0
(x )
dx
Continuing until point xk meets the point x∗ which is the solution of the problem.
F(x) = 0 (1.8.9)
1
named after Sir Isaac Newton (1643-1727) and Joseph Raphson
22 Preliminaries on vector, matrix and tensor Chap. 1
1 d2 F
dF
F() = F(0) + + 2 + · · · (1.8.11)
d =0 2 d2 =0
Replacing F() by F(x0 + u), we have
1 d2
d
F(x0 +u) = F(x0 )+ F(x0 +u)+ F(x0 +u)2 +· · · (1.8.12)
d =0 2 d2 =0
F(x0 + u) ∼
= F(x0 ) + DF(x0 )[u] (1.8.15)
The above is a linear equation in u. We just do a linearization for .
Consequently, the Newton-Raphson iterative scheme is defined by
f1 (x1 , x2 , . . . , xn ) = 0
f2 (x1 , x2 , . . . , xn ) = 0
.. (1.8.17)
.=0
fn (x1 , x2 , . . . , xn ) = 0
Sec. 1.8 Nonlinear problems and linearization 23
d
Df (x0 )[u] = f (x0 + u)
d =0
n
X ∂f d
= d (xi + ui )
i=1
∂x i x0i =0 (1.8.18)
n
X ∂f
= ui
i=1
∂x
i x0i
= K(x0 )u
In the second step, the chain rule for differentiation was used. The tangent matrix
K is given by
∂f1 ∂f1 ∂f1
∂x1 ∂x2
··· ∂xn
∂f2 ∂f2
··· ∂f2
∂x1 ∂x2 ∂xn
K= . .. ... .. (1.8.19)
.. . .
∂fn ∂fn ∂fn
∂x1 ∂x2
··· ∂xn
det(S + U) ∼
= det(S) + D det(S)[U] (1.8.21)
d
D det(S)[U] = det(S + U)
d =0
d
= det S(I + S−1 U)
(1.8.22)
d =0
d
= det S det(I + S−1 U)
d =0
In order to proceed, noting that the characteristic equation of a second order tensor
A with eigenvalues λA A A
1 , λ2 and λ3 is written as
2
−1 U −1 U −1 U
det(I + S−1 U) = (1 + λ1S )(1 + λS2 )(1 + λS3 ) (1.8.24)
−1 −1 −1
where λS1 U , λS2 U and λS3 U are the eigenvalues of the tensor S−1 U. With this,
it is easy to compute
d −1 −1 −1
det(I + S−1 U) = λS1 U + λS2 U + λS3 U (1.8.25)
d =0
−1 −1 U −1 U
D det(S)[U] = det S(λS1 U + λS2 + λ3S )
= det Strace(S−1 U) (1.8.26)
−T
= det S(S : U)
Linearization of det(S−1 )
d
−1
D det(S )[U] = det(S + U)−1
d =0
d 1
=
d =0 det(S + U) (1.8.27)
1 d
=− det(S + U)
(det S)2 d =0
= − det(S−1 )(S−T : U)
where in the last step, the result of previous example has been used.
Linearization of inverse of a tensor
(S + U)−1 ∼
= S−1 + D(S−1 )[U] (1.8.28)
d
D(S )[U] = (S + U)−1
−1
(1.8.29)
d =0
The above derivative is far from obvious. So, we must proceed in an indirect
manner by noting that the directional derivative of constant tensors vanishes. We
consider the identity tensor I = S−1 S, we have
3
Tensor A with eigenvalues λi , then eigenvalues of αA are αλi
Sec. 1.8 Nonlinear problems and linearization 25
2.1 Introduction
Continuum mechanics studies the macroscopic behaviour of where inhomogeneities
such as are not taken into account.
The deformation gradient tensor is the most important concept in express-
ing finite deformation. It is used to define many strain measures such as the left
Cauchy-Green deformation tensor, the Green strain tensor and the Almansi strain
tensor. Rotation plays an important role in nonlinear mechanics and the polar
decomposition theorem
With large deformation, the initial configuration and the deformed configu-
ration are quite different, which leads to different stress measures. The familiar
Cauchy stress is the true stress since it is defined with respect to the deformed
configuration. Other stress measures such as the Piola-Kirchhoff stress tensors
are defined with respect to the initial configuration. Material objectivity
2.2.1 Definitions
The position (vector) of a material point (also called particle) in the initial, unde-
formed configuration of a body is denoted X relative to some coordinate basis.
The position of the same material point in the deformed configuration is denoted
28 Continuum mechanics Chap. 2
2.2.2 Motion
The motion (deformation) of a solid is described by a function φ(X, t). A relation
between spatial coordinates and material coordinates can be established as follows
x = φ(X, t) (2.2.1)
Definition 2.2.2. The velocity of a material point X, denoted by v(X, t), is de-
fined as the rate of change of position of this material point, that is
∂φ(X, t)
v(X, t) := (2.2.3)
∂t
Sec. 2.2 The motion and deformation 29
This is the Lagrangian velocity field. The corresponding Eulerian velocity field is
∂φ(X, t)
v(x, t) := ≡ v(φ−1 (x, t), t) (2.2.4)
∂t X=φ−1 (x,t)
∂v(X, t) ∂ 2 φ(X, t)
a(X, t) := = (2.2.5)
∂t ∂t2
Note that this is a Lagrangian field. Its Eulerian counterpart is
∂ 2 φ(X, t)
a(x, t) := 2
≡ a(φ−1 (x, t), t) (2.2.6)
∂t
X=φ−1 (x,t)
Dϕ(X, t) ∂ϕ(X, t)
≡ ϕ̇ = (2.2.7)
Dt ∂t
where the first two equations indicate standard notation for material time
derivative. The superposed dot is also used for ordinary time derivative if
the variable is only function of time.
2. Eulerian description. The considered field is ϕ(x, t). This case is much
more complicated since not only time changes but also does changes the
spatial position x of the considered particle . We must calculate the par-
tial derivative with respect to time, of the material description of ϕ(x, t),
keeping X fixed.
Writing the above in compact form, we obtain the important formula of the mate-
rial time derivative for an Eulerian scalar field:
Dϕ(x, t) ∂ϕ(x, t)
= + ∇ϕ(x, t) · v(x, t) (2.2.10)
Dt ∂t
The term ∂ϕ/∂t is called the spatial time derivative, and the term ϕ,j vj is the
convective term, which is also called the transport term.
For an Eulerian vector field, by considering each component vi (x, t) a Eulerian
scalar field, we have
Dv(x, t) ∂v(x, t)
= + v · ∇v (2.2.12)
Dt ∂t
where ∇v is the left gradient of a vector field which is given by (from the defini-
tion of dotting a second-order tensor from the left by a vector),
∂vi
(∇v)ji = (2.2.13)
∂xj
For memory, the component matrix of the left gradient of a vector, in two dimen-
sions, is given
vx,x vy,x
∇v = (2.2.14)
vx,y vy,y
In the same manner, consider an Eulerian second-order tensor field σ(x, t), its
material time derivative is given by
xp = φ(XP , t)
(2.2.16)
xq = φ(XP + dX, t)
Hence
∂φ ∂x ∂xi
F= = or Fij = (2.2.19)
∂X ∂X ∂Xj
The Eq. (2.2.18) therefore becomes
x=F·X+C (2.2.22)
32 Continuum mechanics Chap. 2
2. φ(X, t) is one-to-one;
dv = εijk dx1i dx2j dx3k = εijk FiA dX1A FjB dX2B FkC dX3C (2.2.26)
where dx1i is the ith component of vector dx1 and in the second equality, Eq.
(2.2.25) was used. Using the well known formula for determinant,
dv = JdV (2.2.29)
Sec. 2.2 The motion and deformation 33
This shows that J must be positive, the deformation gradient tensor F is therefore
invertible. This allows us to write, from Eq.(2.2.20)
dX = F−1 · dx (2.2.30)
The Jacobian is usually used to relate integrals in the reference and current
configurations. That is
Z Z
f (x, t)dΩ = f (φ(X, t), t)JdΩ0 (2.2.31)
Ω Ω0
Another useful formula is the material time derivative of the Jacobian which
is given by
J˙ = Jvi,i (2.2.32)
In order to prove the above, we must use the following result of matrix algebraic,
d dAij C
det A = A (2.2.33)
dt dt ij
−T
where AC is the cofactor matrix of A given by AC
ij = (det A)Aij . Apply the
above equation for J = det F, we have
DJ d ∂xi −T
= JFij
dt dt ∂Xj
∂vi −T ∂vi ∂xk −1
=J Fij = J F
∂Xj ∂xk ∂Xj ji
∂vi ∂vi
=J Fkj Fji−1 = J δki
∂xk ∂xk
∂vi
=J = Jvi,i = Jdiv(v)
∂xi
Definition 2.2.4. If a motion φ is such that there is no volume change, then the
motion is said to be isochoric. In that case
dx = φ∗ dX = F · dX (2.2.34)
34 Continuum mechanics Chap. 2
dX = φ∗ dx = F−1 · dx (2.2.35)
where n is the unit normal vector and da is the area of the parallel generated by
dx1 and dx2 . Similarly, the volume in the reference configuration is
dV = dX1 · n0 dA (2.2.37)
From the relation between two volumes given in Eq. (2.2.29), we have
dx = R · dX (2.2.41)
Hence,
dx · dx = dX · (RT · R) · dX (2.2.42)
Since length is preserved in rigid body rotation, i.e., dx · dx = dX · dX, it yields
RT · R = I (2.2.43)
The above shows that the inverse of R is equal to its transpose,
RT = R−1 (2.2.44)
The rotation tensor R is therefore said to be an orthogonal tensor.
The velocity of a rigid body motion can be obtained by taking the time deriva-
tive of Eq. (2.2.40). So,
ẋ(X, t) = Ṙ · RT · (x − xT ) + ẋT
(2.2.46)
= Ω · (x − xT ) + ẋT
where
Ω = Ṙ · RT (2.2.47)
The tensor Ω is called the angular velocity tensor. It is a skew-symmetric tensor.
To demonstrate the skew-symmetry of angular velocity tensor, taking the material
time derivative of Eq. (2.2.43),
D DI
(R · RT ) = = 0 =⇒ Ṙ · RT + R · ṘT = 0 =⇒ Ω = −ΩT (2.2.48)
Dt Dt
As having been shown in Chapter 1, any skew-symmetric second-order tensor
can be expressed in term of the components of the axial vector. Recall that
vi = Ωik (xk − xT k ) + vT i
(2.2.50)
= εijk ωk (xk − xT k ) + vT i
In tensor notation, the above writes
v = ω × (x − xT ) + vT (2.2.51)
This is the equation for rigid body motion. The first term is due to the rotation
about the point xT and the second term denotes the translational velocity.
1 1
E = (FT · F − I), Eij = (Fki Fkj − δij ) (2.3.7)
2 2
We also can express the Green strain tensor E in terms of displacement as follows
∂xk ∂xk
Fki Fkj =
∂Xi ∂Xj
∂uk ∂Xk ∂uk ∂Xk
= + +
∂Xi ∂Xi ∂Xj ∂Xj
(2.3.8)
∂uk ∂uk
= + δki + δkj
∂Xi ∂Xj
∂uj ∂ui ∂uk ∂uk
= + + + δij
∂Xi ∂Xj ∂Xi ∂Xj
Hence, we obtain
b = F · FT (2.3.13)
Since b operates on the spatial (Eulerian) vectors, it is an Eulerian tensor. In
addition, this is a symmetric tensor. The inverse of the left Cauchy-Green tensor
is often called the Finger tensor 3 .
The change in the scalar product is then given by
1 1
(dx1 · dx2 − dX1 · dX2 ) = dx1 · (I − b−1 ) · dx2 (2.3.14)
2 2
where e is the Almansi strain tensor, given by
1
e = (I − b−1 ) (2.3.15)
2
The squared lengths of the material dX and spatial vectors dx are
1 2
(ds − dS 2 ) = dX · E · dX (2.3.17)
2
Dividing the above by dS 2 gives
2
in C = FT F, F is on the right, where in b = FFT , F is on the left.
3
named after Josef Finger
Sec. 2.3 Strain measures 39
ds2 − dS 2 dX dX
2
= ·E· (2.3.18)
2dS dS dS
Denoting N as the unit material vector in the direction of dX, the above becomes
ds2 − dS 2
=N·E·N (2.3.19)
2dS 2
In the same manner,
ds2 − dS 2
=n·e·n (2.3.20)
2dS 2
Remark. In the term of the push forward and pull back operations, the Almansi
strain tensor is the pushed forward of the Green strain tensor. From the change in
scalar product of two material vectors, we have
Transform the material vectors to spatial vectors, the RHS of the above is written
as
The above holds for any vectors dx1 and dx2 gives
Therefore, the Almansi strain tensor e is the pushed forward of the Green strain
tensor E with the push forward operator given by
Similarly, the Green strain tensor E is the pulled back of the Almansi strain tensor
e with the pull back operator given by
E = φ∗ [e] = FT · e · F (2.3.25)
The two equations Eqs. (2.3.24, 2.3.25) are typical examples of the push forward
and pull back operations for kinematic second order tensors.
40 Continuum mechanics Chap. 2
∂ ∂φ(X, t)
Ḟ =
∂t ∂X
∂v (2.3.27)
=
∂X
∂v ∂x
= · =L·F
∂x ∂X
where in the second equality, we have used the fact that material time derivative
of Lagrangian fields commute with material gradient. Noting that, this fact does
not hold generally for Eulerian fields. From the above, we obtain
L = Ḟ · F−1 (2.3.28)
As previously described, the strain is defined as the change in the scalar prod-
uct of two vectors. Therefore, the strain rate is defined as the rate of change of the
scalar product of two vectors. Recall that
D −T
(dx1 · dx2 ) = 2dx1 · F Ė · F−1} ·dx2
· {z (2.3.33)
Dt |
D
The tensor D is simply the pushed forward Eulerian counterpart of Ė, and is
known as the rate of deformation tensor which is given by
Ė = FT · D · F (2.3.35)
It is interesting that the rate of deformation tensor D is exactly the symmetric
part of the velocity gradient tensor L. That is
1
D = (L + LT ) (2.3.36)
2
This can be verified by substituting Eq. (2.3.27) into Eq. (2.3.31),
1
Ė = FT (L + LT ) F (2.3.37)
2
and comparing the above to Eq. (2.3.35).
With the rate of deformation tensor, we can write the material time derivative
of the Jacobian J as
J˙ = Jtrace(F−T ĖF−1 )
= Jtrace(C−1 Ė)
= JC−1 : Ė (2.3.39)
1
= JC−1 : Ċ
2
This equation will be useful in the treatment of incompressible materials.
42 Continuum mechanics Chap. 2
The velocity gradient tensor can be decomposed into symmetric and skew-
symmetric tensors by
1 1
L = (L + LT ) + (L − LT ) (2.3.40)
2 2
where the symmetric tensor has been defined as the rate of deformation D and the
spin tensor W is the skew-symmetric part of L. With these definitions, we can
write
L=D+W (2.3.41)
with
1
W = (L − LT ) (2.3.42)
2
D
(dx · dx) = 2dx · D · dx (2.3.43)
Dt
1 D
(ds2 ) = n · D · n (2.3.44)
2ds2 Dt
1 D ln(ds)
n·D·n= (ds) = (2.3.45)
ds Dt dt
Remark. Note that the rate of deformation D is the push forward of the material
time derivative of the Green strain tensor E (see Eq.2.3.34) which is the pull back
of the Almansi strain tensor e. We then could write
D ∗
D = φ∗ φ [e] (2.3.46)
Dt
This operation is known as the Lie derivative of a tensor over the mapping φ and
is generally defined by
D ∗
Lφ [g] = φ∗ φ [g] (2.3.47)
Dt
Sec. 2.3 Strain measures 43
F=R·U (2.3.48)
where R is the rotation tensor (hence orthogonal), and U is the symmetric right
stretch tensor.
By substitution Eq. (2.3.48) into the definition equation of the right Cauhcy-
Green tensor C, Eq. 2.3.2, we have
C = (RU)T RU
= UT RT RU (2.3.49)
= UT U = U2
where λ2i and Ni are the eigenvalues and normalized eigenvectors of C. The right
stretch tensor U is now computed by
3
X
U= λi Ni ⊗ Ni or U = QHQT (2.3.51)
i=1
where the diagonal matrix H contains the eigenvalues λi and the matrix Q whose
columns are the eigenvectors of the tensor C.
Having defined U, the rotation tensor R is simply determined by
44 Continuum mechanics Chap. 2
R = F · U−1 (2.3.52)
In terms of the rotation and stretch tensors, the deformation of a material vec-
tor dX is written by
dx = F · dX = R · (U · dX) (2.3.53)
The above shows that the vector dX is first stretched (by U) then rotated to the
current configuration by R. The tensor U is therefore a material or Lagrangian
tensor whereas R is a two point tensor.
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
Example 1 Polar decomposition (1)
Consider a motion given by
1
x1 = 4X1 + (9 − 3X1 − 5X2 − X1 X2 )t
4
1
x2 = 4X2 + (16 + 8X1 )t
4
det(C − λI) = 0
After some manipulations, we obtain the characteristic equation as
(C − λ2i I)Ni = 0
Writing the above out
65 27
− λi Ni1 + Ni2 = 0
16 16
27 1 41
Ni + − λi Ni2 = 0
16 16
||Ni || = 1
−0.5449 0.8385
N1 = N2 =
0.8385 0.5449
The stretch tensor U is then computed by
U = QHQT
where
−0.5449 0.8385 1.2107 0
Q= , H=
0.8385 0.5449 0 2.2714
Finally, the stretch tensor is given by
1.9564 0.4846
U=
0.4846 1.5256
and the rotation tensor
−1 0.3590 −0.9333
R=F·U =
0.9333 0.3590
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
46 Continuum mechanics Chap. 2
F=V·R (2.3.54)
dx = F · dX = V · (R · dX) (2.3.55)
The above shows that the vector dX is first rotated by R then stretched by V. The
tensor V is therefore a spatial or Eulerian tensor.
The Lagrangian stretch tensor U and Eulerian stretch tensor V are related
together by
V = R · U · RT (2.3.56)
In the same manner with the right Cauchy-Green tensor, the left Cauchy-Green
tensor can be written in terms of V and R as follows
b = F · FT = V · R · RT · VT = V2 (2.3.57)
As a consequence, if the principle directions of b are denoted by ni together
with associated eigenvalues λ̄2i , then V is given by
3
X
V= λ̄i ni ⊗ ni (2.3.58)
i=1
3
X 3
X
T
V =R·( λi Ni ⊗ Ni ) · R = λi (RNi ) ⊗ (RNi ) (2.3.59)
i=1 i=1
λi = λ̄i , ni = R · Ni (2.3.60)
The above shows that the left and right Cauchy-Green tensors have the same
eigenvalues. It is also useful in demonstrating the physical meaning of these
eigenvalues as the stretches in the principle directions. Indeed, taking a mate-
rial vector dX1 with length dS1 in the direction of N1 , this vector deformed to the
spatial vector dx1 given by dx1 = F · dX1 . Using the polar decomposition of F,
we can write dx1 as
Sec. 2.3 Strain measures 47
dx1 = R · U · dS1 N1
= dS1 R · λ1 N1 (2.3.61)
= λ1 dS1 n1
The length of the spatial vector dx1 is therefore related to the length of the vector
dX1 by
ds1
ds1 = λ1 dS1 =⇒ λ1 = (2.3.62)
dS1
Then, the eigenvalues λi are called the principle stretches which are exclusively
used in the constitutive equations of hyperelasticity.
dx̃ = Q · dx (2.3.64)
Relating the vector in the current configuration dx to the corresponding in the
initial configuration dX by the deformation gradient F, the above can be written
as
dx̃ = Q · F · dX = F̃ · dX (2.3.65)
with
F̃ = Q · F (2.3.66)
Although the vector dx̃ is different from the vector dx, their magnitudes are
obviously the same. In this sense it can be said that the vector dx is objective
under rigid body rotation.
It can be shown that the Lagrangian strain tensors such as C and E are un-
changed under rigid body motion. For example, we have
C̃ = F̃T · F̃ = FT · QT · Q · F = FT · F = C (2.3.67)
48 Continuum mechanics Chap. 2
In contrast, the Eulerian strain tensors such as the right Cauchy-Green tensor b
and the Almansi tensor e do change under rigid body motion. Indeed,
b̃ = F̃ · F̃T = Q · F · FT · QT = Q · b · QT (2.3.68)
ẽ = Q · e · QT (2.3.69)
2. The nominal stress P (closely related to the first Piola-Kirchhoff stress ten-
sor)
df
t(n) = lim (2.4.1)
ds→0 ds
(e ) (e ) (e )
t(e1 ) = t1 1 e1 + t2 1 e2 + t3 1 e3
(e ) (e ) (e )
t(e2 ) = t1 2 e1 + t2 2 e2 + t3 3 e3 (2.4.3)
(e ) (e ) (e )
t(e3 ) = t1 3 e1 + t2 3 e2 + t3 3 e3
(ei )
where tj is the jth component of the traction vector t(ei ).
(ei )
σij = tj (2.4.4)
Dividing the above by da, noting that dv/da → 0 and the Newton’s first law given
in the Eq. (2.4.2)
3
X dai
t(n) = t(ei ) (2.4.7)
i=1
da
dai /da = ni
3
X
t(n) = t(ei )ni (2.4.8)
i=1
Indicial notation,
(e )
tj (n) = tj i ni = σij ni (2.4.9)
In tensor notation, we can write the above as
4
This was done by Cauchy, and the well known Cauchy tetrahedron
50 Continuum mechanics Chap. 2
t=n·σ (2.4.10)
The above is known as the Cauchy relation.
P = JF−1 · σ (2.4.14)
then
df = n0 · PdΓ0 (2.4.15)
The above means that n0 · P is the traction vector relating the force in current
configuration to the area in the reference configuration. Therefore P is obviously a
stress tensor which is called the nominal stress tensor. It is a Lagrangian-Eulerian
two-point tensor. The transpose of the nominal stress is the first Piola-Kirchhoff
stress tensor.
From Eq. (2.4.14), we have expression of the Cauchy stress in terms of the
nominal stress as
σ = J −1 F · P (2.4.16)
P = S · FT (2.4.20)
Introducing the above into Eq. (2.4.16), we have
σ = J −1 F · S · FT (2.4.21)
Inversely, the second Piola-Kirchhoff stress can be written in term of the Cauchy
stress as
Dwint
ρ =D:σ (2.4.23)
Dt
Since ρJ = ρ0 , the above can be written as
Dwint
ρ0 = D : (Jσ) (2.4.24)
Dt
The above means that Jσ can be a stress measure which is work conjugate with
the rate of deformation with respect to the initial volume. It is named the Kirchhoff
stress τ or also called the weighted Cauchy stress. For isochoric motion (J = 1),
it is identical to the Cauchy stress.
Using Eq.(2.4.16), we get the relation between the Kirchhoff stress and nomi-
nal stress as
52 Continuum mechanics Chap. 2
τ =F·P (2.4.25)
From equations (2.4.21) and (2.4.22), we obtain the relation between the sec-
ond Piola-Kirchhoff stress and the Kirchhoff stress as
τ = F · S · FT (2.4.26)
Remark. Using the push forward and pull back operations, the Kirchhoff stress is
the pushed forward of the second Piola-Kirchhoff stress,
τ = φ∗ [S] = F · S · FT (2.4.28)
Inversely, the second Piola-Kirchhoff stress is the pull back of the Kirchhoff stress,
The above equations are examples of the push forward and pull back operations
on kinetic tensors.
Similarly, the Cauchy stress and the second Piola-Kirchhoff stress are related
as,
In the above, S and σ are related by the so-called Piola transformation which in-
volves a push forward and pull back operation combined with the volume scaling
J.
where p = 1/3σii .
Substituting the above into Eq. (2.4.22), we obtain the similar decomposition
for the second Piola-Kirchhoff stress tensor S as
Sdev : C = 0 (2.4.33)
Taking the double contraction of the first equation in Eq.(2.4.32) with C gives
S = pJC−1 =⇒ pI = J −1 SC (2.4.35)
1
p = J −1 S : C (2.4.36)
3
Introducing the above into the Eq. (2.4.32), we obtain
1
Shyd = (S : C)C−1 (2.4.37)
3
Dσ
σ ∇J = − W · σ − σ · WT (2.4.38)
Dt
54 Continuum mechanics Chap. 2
Truesdell rate
Recall that the second Piola-Kirchhoff stress tensor is independent of any rigid
body rotation. The Truesdell stress rate is thus defined by the Piola transformation
of the time derivative of the second Piola-Kirchhoff stress. That is
∇T −1 −1 D −1 −T
σ = J φ∗ [Ṡ] = J F (JF σF ) FT (2.4.39)
Dt
The above can be easily written as
˙ + FḞ−1 σ + σ Ḟ−T FT
σ ∇T = σ̇ + Jσ (2.4.40)
Using the fact that
Dσ
σ ∇T = + div(v)σ − L · σ − σ · LT (2.4.42)
Dt
The Truesdell stress rate can be interpreted in term of the Lie derivative of the
Kirchhoff stress tensor as
Jσ ∇T = Lφ [τ ] (2.4.43)
In fact, the above defines what is known as the Truesdell rate of the Kirchhoff
stress tensor or the convected rate of the Kirchhoff stress denoted by τ ∇c , which
can be shown from Eq. () to be
τ ∇c = τ̇ − L · τ − τ · LT (2.4.44)
Green-Naghdi rate
Dσ
σ ∇G = − Ω · σ − σ · ΩT (2.4.45)
Dt
Z Z Z
D 1
f (x, t)dΩ = lim f (x, t + ∆t)dΩ − f (x, t)dΩ (2.5.1)
Dt Ω ∆t→0 ∆t Ωt+∆t Ωt
where Ωt is the spatial domain occupied by the material domain at time t and
Ωt+∆t is the spatial domain occupied by the same material domain at time t + ∆t.
Transforming the integrals in the RHS of the above to the reference configuration
gives,
Z Z Z
D 1
f dΩ = lim f (X, t+∆t)J(X, t+∆t)dΩ0 − f (X, t)J(X, t)dΩ0
Dt Ω ∆t→0 ∆t Ω0 Ω0
(2.5.2)
where the function f = f (φ(X, t), t).
Since the integral domain is now constant in time, we can take the limit inside
the integral, the above becomes
Z Z
D ∂
f dΩ = f (X, t)J(X, t) dΩ0 (2.5.3)
Dt Ω Ω0 ∂t
In the above equation, the partial derivative with respect to time is the material
time derivative since the independent spatial variable is the Lagrangian coordinate.
Using the product rule for derivatives on the RHS of the above, we can write
Z Z
D ∂f ∂J
f dΩ = J +f dΩ0 (2.5.4)
Dt Ω Ω0 ∂t ∂t
Recall that J˙ = Jvi,i , we obtain
Z Z
D ∂f ∂vi
f dΩ = J + fJ dΩ0 (2.5.5)
Dt Ω Ω0 ∂t ∂xi
Transform the integrals in the RHS of the above back to the current configuration,
we come up with
Z Z
D Df (x, t) ∂vi
f dΩ = +f dΩ (2.5.6)
Dt Ω Ω Dt ∂xi
56 Continuum mechanics Chap. 2
This equation is one form of the Reynold’s transport theorem. For completeness,
another form of the Reynold’s transport theorem is given here where the proof
will be done later,
Z Z
D Df
ρf dΩ = ρ dΩ (2.5.7)
Dt Ω Ω Dt
This is known as the Reynold’s theorem for a density-weighted integrand.
where ρ is the mass density. The law of conversation of mass states that
Dm(Ω)
=0 (2.5.9)
Dt
Applying the Reynold’s theorem in Eq. (2.5.6) to the above yields
Z
Dρ
+ ρ∇ · v dΩ = 0 (2.5.10)
Ω Dt
Dρ
+ ρ∇ · v or ρ̇ + ρvi,i = 0 (2.5.11)
Dt
The above is the equation of mass conservation, or often called equation of conti-
nuity.
If the density does not change, i.e., the material is incompressible, hence the
material time derivative of the density vanishes, the continuity equation becomes
vi,i = 0 (2.5.12)
which is the well known incompressibility condition.
For Lagrangian description (material domain moves with material),
Z
Dm(Ω)
= 0 =⇒ ρ(X, t)dΩ = constant (2.5.13)
Dt Ω
Hence,
Sec. 2.5 Conservation equations 57
Z Z
ρ(X, t)dΩ = ρ0 (X)dΩ0 (2.5.14)
Ω Ω0
Transforming the RHS of the above to the reference configuration by using dΩ =
JdΩ0 gives
Z
ρ(X, t)J − ρ0 (X) dΩ0 = 0 (2.5.15)
Ω0
Arbitrariness of the material domain Ω0 gives
The second term in the above vanishes since it is exactly the continuity equation,
the above then becomes
Z Z
D Dv
ρvdΩ = ρ dΩ (2.5.22)
Dt Ω Ω Dt
Convert the second term on the right hand side of Eq. (2.5.20) to domain integral
by Cauchy’s relation and Gauss’s theorem,
Z Z Z
tdΓ = n · σdΓ = ∇ · σdΩ (2.5.23)
Γ Γ Ω
Dv
ρ = ∇ · σ + ρb or ρv̇i = σji,j + ρbi (2.5.25)
Dt
The above is the so-called the momentum equation; it is also called the Cauchy’s
first law of motion.
This form of the momentum equation is applicable to both Lagrangian and
Eulerian descriptions. For a Lagrangian description, the independent variables are
the material coordinates X and time t, so the corresponding momentum equation
is
∂v(X, t)
ρ(X, t) = ∇ · σ(φ−1 (x, t), t) + ρ(X, t)b(X, t) (2.5.26)
∂t
Noting that the stress have been written in term of spatial coordinates so that the
spatial derivative can be evaluated. In practice, however, the inverse of φ is never
built and implicit differentiation is employed. The above would not be considered
a true Lagrangian description in classical texts on continuum mechanics due to the
presence of derivative with respect to Eulerian coordinates. However, the above is
indeed a Lagrangian form since all dependent variables are functions of material
coordinates. This form can be called the updated Lagrangian momentum equation
since we will use it to derive the update Lagrangian finite elements.
In an Eulerian description, the material time derivative of the velocity is given
by Eq.(2.2.12), i.e., replacing the material time derivative by partial derivative
with respect to time plus the transport term, and the independent variables are the
spatial coordinates and time. The Eq.(2.5.25) becomes
Sec. 2.5 Conservation equations 59
∂v(x, t)
ρ(x, t) + v(x, t) · ∇v(x, t) = ∇ · σ(x, t) + ρ(x, t)b(x, t) (2.5.27)
∂t
Using the cross product formula, the above equation can be written as (in Carte-
sian components)
Z Z Z
D
εijk ρxj vk dΩ = εijk ρxj bk dΩ + εijk xj tk dΓ (2.5.30)
Dt Ω Ω Γ
Z Z
D D
εijk ρxj vk dΩ = εijk ρ (xj vk )dΩ
Dt Ω Dt
ZΩ (2.5.31)
= εijk ρ(vj vk + xj v̇k )dΩ
Ω
By using the Cauchy’s relation and the divergence theorem, the second term in the
right hand size of Eq. (2.5.30) can be rewritten as
60 Continuum mechanics Chap. 2
Z Z
εijk xj tk dΓ = εijk xj σkp np dΓ
Γ Γ
Z
∂
= εijk
(xj σkp )dΩ
Ω ∂xp
Z (2.5.32)
∂σkp
= εijk δjp σkp + xj dΩ
Ω ∂xp
Z
∂σkp
= εijk σkj + xj dΩ
Ω ∂xp
After substitution of Eqs.(2.5.31) and (2.5.32), the equation (2.5.30) now becomes
Z
∂σkp
εijk ρ(vj vk ) − σkj + xj ρv̇k − − ρbk dΩ = 0 (2.5.33)
Ω ∂xp
Since Ω is arbitrary,
∂σkp
εijk ρ(vj vk ) − σkj + xj ρv̇k − − ρbk =0 (2.5.34)
∂xp
Noting that the term in parentheses is the linear momentum equation, so it van-
ishes; and εijk (vj vk ) vanishes also because it is the cross product of the same
vector, the above equation then simplifies as
So, the conversation of angular momentum leads to the symmetry of the Cauchy
stress tensor.
The internal energy per unit volume is denoted by ρwint where wint is the
internal energy per unit mass. The heat flux per unit area is denoted by vector q
and the heat source per unit volume is denoted by ρs.
The rate of change of the total energy
Z Z Z Z Z
D int 1
ρw + ρv · v dΩ = v · ρbdΩ + v · tdΓ + ρsdΩ − n · qdΓ
Dt Ω 2 Ω Γ Ω Γ
(2.5.43)
Using the Reynold’s theorem for the first term, we have
Dwint 1 D
Z Z
D int 1
ρw + ρv · v dΩ = ρ + ρ (v · v) dΩ
Dt Ω 2 Ω Dt 2 Dt
Z int
(2.5.44)
Dw Dv
= ρ + ρv · dΩ
Ω Dt Dt
62 Continuum mechanics Chap. 2
As usual, we convert the second term of the right hand size in Eq. (2.5.43) which
is surface integral to domain integral by applying the Cauchy’s relation and the
divergence theorem,
Z Z Z
v · tdΓ = vi ti dΓ = vi nj σij dΓ
Γ ZΓ Γ Z
= (vi σji ),j dΩ = (vi,j σji + vi σji,j )dΩ
ZΩ Ω
Dwint
Z
Dv
ρ − D : σ + ∇ · q − ρs + v · ρ − ∇ · σ − ρb dΩ = 0 (2.5.46)
Ω Dt Dt
The last term in the above is exactly the momentum equation, so it vanishes. By
taking the arbitrariness of Ω, the energy equation is obtained
Dwint
ρ = D : σ − ∇ · q + ρs (2.5.47)
Dt
Dwint
ρ =D:σ (2.5.48)
Dt
which is no longer a PDE. Rate of deformation and the Cauchy stress is called
conjugate in energy or work conjugate.
Sec. 2.6 Lagrangian conservation equations 63
Z Z Z
d
ρ0 v(X, t)dΩ0 = ρ0 b(X, t)dΩ0 + t0 (X, t)dΓ0 (2.6.4)
dt Ω0 Ω0 Γ0
On the LHS, the material derivative can be taken inside the integral since the
reference domain is fixed in time, so
Z Z
d ∂v(X, t)
ρ0 v(X, t)dΩ0 = ρ0 dΩ0 (2.6.5)
dt Ω0 Ω0 ∂t
Applying the Cauchy’s relation and Gauss’s theorem in sequences, we have
Z Z Z
t0 (X, t)dΓ0 = n0 · PdΓ0 = ∇0 · PdΩ0 (2.6.6)
Γ0 Γ0 Ω0
∂v ∂Pji
ρ0 = ∇0 · P + ρ0 b or ρ0 v̇i = + ρ 0 bi (2.6.8)
∂t ∂Xi
The above is the linear momentum equation for the total Lagrangian formulation.
Noting the similarity to the updated Lagrangian momentum equation given in
64 Continuum mechanics Chap. 2
Eq.(): the density is replaced by the initial density, the Cauchy stress replaced by
the nominal stress, and certainly derivative with respect to the spatial coordinates
is replaced by derivative with respect to material coordinates.
If we neglect the inertia effect, i.e., for a static problem, we then obtain the
familiar equilibrium equation given by
∇0 · P + ρ0 b = 0 (2.6.9)
σ = σT
(2.6.10)
σ = J −1 F · P
we obtain
J −1 F · P = (J −1 F · P)T (2.6.11)
Hence
F · P = P T · FT (2.6.12)
J −1 F · S · FT = (J −1 F · S · FT )T (2.6.13)
Or
F · S · FT = F · ST · FT (2.6.14)
S = ST (2.6.15)
Z Z Z
D int 1
ρ0 w + ρ0 v · v dΩ0 = v · ρ0 bdΩ0 + v · t0 dΓ0
Dt Ω0 2 Ω0 Γ0
Z Z (2.6.16)
+ ρ0 sdΩ0 − n0 · q
edΓ0
Ω0 Γ0
Since the reference domain is fixed in time, the first term can be rewritten as
Dwint
Z Z
D int 1 Dv
ρ0 w + ρ0 v · v dΩ0 = ρ0 + ρ0 v · dΩ0 (2.6.17)
Dt Ω0 2 Ω Dt Dt
By applying the Cauchy’s relation and the divergence theorem, the second term
of the right hand size in Eq. (2.6.16) can be written
Z Z Z
v · t0 dΓ0 = vj t0j dΓ0 vj n0i Pij dΓ0
=
Γ0 Γ Γ0
Z 0 Z
∂ ∂vj ∂Pij
= (vj Pij )dΩ0 = Pij + vj dΩ0
Ω0 ∂Xi Ω0 ∂Xi ∂Xi
Z (2.6.18)
∂Fji ∂Pij
= Pij + vj dΩ0
Ω ∂t ∂Xi
Z 0 T
∂F
= : P + v · (∇0 · P) dΩ0
Ω0 ∂t
Inserting all of the above into Eq. (2.6.16), and applying the divergence theorem
to the heat flux term of the right hand size in Eq. (2.6.16), we obtain
Dwint ∂FT
Z
Dv
ρ0 − : P + ∇0 · q
e − ρ0 s + v · ρ0 − ∇0 · P − ρ0 b dΩ0 = 0
Ω0 Dt ∂t Dt
(2.6.19)
The last term in the above is exactly the momentum equation, so it vanishes. By
taking the arbitrariness of Ω, we obtain the energy equation
ρ0 ẇint = ḞT : P − ∇0 · q
e + ρ0 s (2.6.20)
If heat energy is neglected, the above becomes
It shows that the nominal stress P is conjugate in power to the material time
derivative of the deformation gradient tensor F.
In what follows, we seek the similar expression for the second Piola-Kirchhoff
stress tensor by simply transform the nominal stress tensor to the second Piola-
Kirchhoff stress tensor. Indeed,
T
= Fkj Ḟji Sik = (FT · Ḟ) : S (sym. of S) (2.6.22)
1
= (FT · Ḟ + ḞT · F) : S
2
where we have used the tensor decomposition into symmetric and skew symmetric
parts for FT and the double contraction of symmetric tensor S with skew symmet-
ric tensor vanishes.
Noting that the term in parentheses in the above is the time derivative of the
Green strain tensor E, so
Constitutive models
A stress-strain curve is a graph derived from measuring load (stress - ?) versus ex-
tension (strain - ?) for a sample of a material. The nature of the curve varies from
material to material. The following diagrams illustrate the stress-strain behaviour
of typical materials in terms of the engineering stress and engineering strain where
the stress and strain are calculated based on the original dimensions of the sample
and not the instantaneous values. In each case the samples are loaded in tension
although in many cases similar behaviour is observed in compression.
Tangent modulus is defined as the slope of a line tangent to the stress-strain
curve at a point of interest. Tangent modulus can have different values depending
on the point at which it is determined. For example, tangent modulus is equal
to the Young’s Modulus when the point of tangency falls within the linear range
of the stress-strain curve. Outside the linear elastic region, at point A shown
for example, tangent modulus is always less than the Young’s modulus. Tangent
modulus is mostly used to describe the stiffness of a material in the plastic range,
and it is denoted by Et.
3.2 Elasticity
A material is said to be elastic if it deforms under stress (e.g., external forces), but
then returns to its original shape when the stress is removed.
68 Constitutive models Chap. 3
1
εx = (σx − νσy − νσz )
E
1
εy = (σy − νσz − νσx ) (3.2.1)
E
1
εz = (σz − νσx − νσy )
E
where E and ν are the Young’s modulus and Poisson’s ratio, respectively.
Solving this system of equations with the stresses as unknowns, we have the
following relation
Eν E
σx = (εx + εy + εz ) + εx (3.2.2)
(1 + ν)(1 − 2ν) 1+ν
The shearing stress σxy is given by
E
µ= (3.2.4)
2(1 + ν)
Define the constants λ by
Eν
λ= (3.2.5)
(1 + ν)(1 − 2ν)
Similarly, we obtain the expressions for the other components of the stress
tensor which allows us to write generally as
where εkk = εx + εy + εz is the so-called dilatation and δij is the Kronecker delta
and λ, µ are the Lamé’s constants. The Young’s modulus and Poisson’s ratio can
be expressed in terms of the Lamé’s constans through equations (from Eq. (3.2.5))
Sec. 3.2 Elasticity 69
µ(3λ + 2µ)
E=
λ+µ
(3.2.7)
λ
ν=
2(λ + µ)
σ = C : ε, C = λI ⊗ I + 2µI (3.2.9)
where I and I are the identity second-order and fourth-order tensors, respectively.
We have just derived the general Hook law for linear elastic materials which
gives the constitutive equation as
We chose the derivation process from simple isotropic materials to the general
case of anisotropic materials. We believe that this is indeed very simple and avoids
complexity of tensors theory.
Since the stress and strain are symmetric, the elastic modulus tensor C must
possess the minor symmetries given by
which further reduces the number of components down to n(n + 1)/2 = 21, for
n = 6.
It is convenient to use the Voigt notation where the symmetry of the stress and
strain tensors are used.
11 → 1 22 → 2 33 → 3
(3.2.13)
23 → 4 13 → 5 12 → 6
70 Constitutive models Chap. 3
σ11
σ22
σ11 σ12 σ13
σ33
σ= σ22 σ23 → {σ} = (3.2.14)
σ23
sym σ33
σ
13
σ12
Then
σ1
C11 C12 C13 C14 C15 C16 ε1
σ2 C22 C23 C24 C25 C26 ε2
σ3 C33 C34 C35 C36 ε3
= (3.2.19)
σ4
C44 C45 C46
ε4
σ SY M C55 C56 ε5
5
σ6 C66 ε6
The above holds true for any anisotropic linear elastic materials.
Return back to the case of isotropic linear elastic materials, using Eq. (3.2.8)
we can compute the components Cijkl and the Voigt mapping allows us to trans-
form them into the matrix form as
Sec. 3.2 Elasticity 71
σ1
λ + 2µ λ λ 0 0 0 ε1
σ2 λ + 2µ λ 0 0 0 ε2
σ3 λ + 2µ 0 0 0 ε3
= (3.2.20)
σ4
µ 0 0 ε4
σ SY M µ 0 ε5
5
σ6 µ ε6
1−ν ν ν 0 0 0
ν
1−ν ν 0 0 0
ν ν 1 − ν 0 0 0
1 − 2ν
E
0
[C] = 0 0 0 0
(1 + ν)(1 − 2ν) 2
0 1 − 2ν
0 0 0 0
2
1 − 2ν
0 0 0 0 0
2
(3.2.23)
and in plane strain
1−ν
ν 0
E ν 1 − ν 0
[C] = (3.2.24)
(1 + ν)(1 − 2ν) 1 − 2ν
0 0
2
whereas in plane stress
72 Constitutive models Chap. 3
1 ν 0
E ν 1 0
[C] = (3.2.25)
1 − ν2 1 − 2ν
0 0
2
∂ψ(C)
S=2 (3.2.29)
∂C
A consequence of the existance of a stored energy function is that the work
done on a hyperelastic mateiral is independent of the deformation path. To illus-
trate the independence of work on deformation path, consider two deformation
state C1 to C2 ,
Sec. 3.2 Elasticity 73
Z C2
1
S : dC = ψ(C2 ) − ψ(C1 ) (3.2.30)
2 C1
∂ 2ψ ∂ 2ψ
Ṡ = 2 : Ċ = 4 : Ė (3.2.32)
∂C∂C ∂C∂C
Ṡ = CSE : Ė (3.2.33)
where
∂ 2ψ ∂ 2ψ
CSE = 4 , Cijkl = 4 (3.2.34)
∂C∂C ∂Cij ∂Ckl
is the material or Lagrangian tangent modulus, which is also called the second
elasticity tensor. This fourth-order tensor has minor symmetries, Cijkl = Cjikl
and Cijkl = Cijlk . In addition, since the stored energy is smooth, the tangent
modulus of hyperelastic materials has the major symmetry Cijkl = Cklij .
In what follows, we derive the equivalent of Eq. (3.2.33) in the Eulerian for-
mulation. We start by the relation of the rate of second Piola-Kirchhoff stress to
the convected rate of the Kirchhoff stress and the relation of Ė and D,
τ ∇c = F · Ṡ · FT , Ė = FT · D · F (3.2.35)
Working in indicial notation, we can write
where in the second step, Eq. (3.2.33) has been used. In tensor notation, the above
becomes
τ ∇c = Cτ : D, τij∇c = Cijkl
τ
Dkl (3.2.37)
where Cτ is the spatial or Eulerian elasticity tensor (the spatial form of the second
elasticity tensor CSE ) which is given by
74 Constitutive models Chap. 3
τ SE
Cijkl = Fip Fjq Fkm Fln Cpqmn (3.2.38)
It is easy to transform the above to the Truesdell rate of the Cauchy stress as
follows
I1 = tr(C) = Cii
1 2 2 1 2
I2 = tr(C) − tr(C ) = Cii − Cij Cji (3.2.41)
2 2
I3 = det(C)
The derivatives of the invariants with respect to the tensor C, which is necessary
in calculation of the stress tensor, are
∂ψ ∂ψ ∂I1 ∂ψ ∂I2 ∂ψ ∂I3
S=2 =2 + +
∂C ∂I1 ∂C ∂I2 ∂C ∂I3 ∂C
(3.2.43)
∂ψ ∂ψ ∂ψ ∂ψ −1
=2 + I1 I−2 C + 2I3 C
∂I1 ∂I2 ∂I2 ∂I3
where Eq. (3.2.42) was used in the second step. The Kirchhoff stress tensor is
given by
1
also invariants of the right Cauchy-Green tensor b
Sec. 3.2 Elasticity 75
τ = F · S · FT
(3.2.44)
∂ψ ∂ψ ∂ψ 2 ∂ψ
=2 + I1 b−2 b + 2I3 I
∂I1 ∂I2 ∂I2 ∂I3
where b is the left Cauchy-Green tensor and the expression of S in Eq. (3.2.43)
was used. The expression for the Cauchy stress is then easily obtained by divide
the above by the Jacobian J.
The last thing left is the analytical expression of the stored energy function ψ.
For completeness, we present here some of the most widely used hyperelasticity
models such as the Neo-Hookean, the Mooney-Rivlin models. For others, reader
is referred to the literature in the subject.
1 1
ψ(C) = λ0 (ln J)2 − µ0 ln J + µ0 (I1 − 3) (3.2.45)
2 2
where λ0 , µ0 are the Lamé constants and J = det(F), I3 = det(C) = J 2 .
From Eq. (3.2.45), we can compute the following quantities
∂ψ 1 ∂ψ ∂ψ 1
= µ0 , = 0, = λ0 ln JJ −2 − µ0 J −2 (3.2.46)
∂I1 2 ∂I2 ∂I3 2
The second Piola-Kirchhoff stress tensor is therefore given by
τ = µ0 (b − I) + λ0 ln JI (3.2.48)
where b is the left Cauchy-Green strain tensor.
The Lagrangian elasticity tensor is given by
∂S
CSE = 2
∂C (3.2.49)
−1 −1 ∂C−1
= λ0 C C − 2(µ0 − λ0 ln J)
∂C
where, for the second equality, Eq. (3.2.47) has been used. In order to compute
the derivative of inverse of a tensor, recall that the directional derivative of the
inverse of a tensor is given by
76 Constitutive models Chap. 3
∂C−1
D(C−1 )[∆C] = : ∆C (3.2.51)
∂C
Consequently,
∂C−1
−1 −1
= −Cik Cjl (3.2.52)
∂C ijkl
SE
Cijkl = λCij−1 Ckl
−1 −1 −1
+ 2µCik Cjl (3.2.53)
From Eq. (3.2.38), the Eulerian elasticity tensor is given by
τ
Cijkl = λδij δkl + 2µδik δjl (3.2.54)
Noting that this tensor is very similar to the one of small strain linear elasticity
except that the shear modulus µ depends on the deformation. In the plane strain
condition, the Voigt form of the Eulerian elasticity tensor is given by
λ + 2µ λ 0
[Cτ ] = λ λ + 2µ 0 (3.2.55)
0 0 µ
Mooney-Rivlin material
2. A yield function f (σ, q) which governs the onset and continuance of plastic
deformation; q are internal variables;
σ̇ = E ε̇e (3.3.2)
σ̇ = E tan ε̇ (3.3.3)
where E tan is the elastic-plastic tangent modulus.
The plastic strain rate is given by a flow rule which is often specified in term
of a plastic flow potential ψ
∂ψ
ε̇p = λ̇ (3.3.4)
∂σ
where λ̇ is the plastic rate parameter. An example of a flow potential is
∂ψ
ψ = σ = |σ| = σsign(σ), = sign(σ) (3.3.5)
∂σ
where σ is the effective stress. Plastic deformation occurs only when the yield
function f is zero:
f = σ − σY (ε) = 0 (3.3.6)
with σY (ε) is the yield strength in uniaxial tension and ε is the effective plastic
strain given by
Z √
ε = ε̇dt, ε̇ = ε̇p ε̇p (3.3.7)
ε̇ = λ̇ (3.3.8)
The flow rule can be rewritten by
∂f
ε̇p = ε̇sign(σ) = ε̇ (3.3.9)
∂σ
78 Constitutive models Chap. 3
dσY (ε)
σ̇ = ε̇ = H ε̇ (3.3.11)
dε
From Eq. 3.3.3,
σ̇ = E tan ε̇
= E tan (ε̇e + ε̇p ) (3.3.12)
1 1
= E tan ( σ̇ + σ̇)
E H
Finally, it yields
1 1 1 tan E2
tan
= + , E =E− (3.3.13)
E E H E+H
Consider a plastic switch parameter β with β = 1 corresponding to plastic loading
and β = 0 corresponding to elastic loading or unloading, we can rewrite the
tangent modulus
E2
E tan = E − β (3.3.14)
E+H
Extension to kinematic hardening
Flow rule
∂ψ
ε̇p = λ̇ , ψ = |σ − α| (3.3.15)
∂σ
Yield function
f = |σ − α| − σY (ε) = 0 (3.3.16)
An evolution equation for the internal variable α is necessary. The simplest form
is the linear kinematic hardening: α̇ = κε̇p .
1
ε̇ = (σ̇ − α̇)sign(σ − α) (3.3.18)
H
Sec. 3.4 Multidimensional plasticity 79
E ε̇sign(σ − α)
ε̇ = (3.3.19)
E+H +κ
E2
E tan = E − β (3.3.20)
E+H +κ
D = De + Dp (3.4.1)
The stress rate is related to the elastic part of the rate of deformation tensor via the
hypoelastic law. So
σ̇ = CσJ e σJ p
el : D = Cel : (D − D ) (3.4.2)
The plastic rate of deformation Dp is given by
Dp = λ̇r(σ, q) (3.4.3)
where λ̇ is and r is the plastic flow direction which is defined as the normal to the
plastic flow potential Ψ, so
∂Ψ
r= (3.4.4)
∂σ
Evolution equation for the internal variables
q̇ = λ̇h(σ, q) (3.4.5)
The yield function
f (σ, q) = 0 (3.4.6)
Consistency condition f˙ = 0
fσ : σ̇ + fq · q̇ = 0 (3.4.7)
where we have adopted the symbols fσ and fq for the partial derivatives of f
with respect to the stress and the internal variables, respectively. From Eqs.
(3.4.2,3.4.3,3.4.5), the above can be written by
80 Constitutive models Chap. 3
n o
σJ
fσ : Cel : (D − λ̇r + fq · λ̇h = 0 (3.4.8)
fσ : C∆J
el : D
λ̇ = (3.4.9)
−fq · h + fσ : C∆J
el : r
fσ : CσJ
el : D
σ̇ = CσJ
el : D− r (3.4.10)
−fq · h + fσ : CσJ
el : r
σ̇ = Cep : D (3.4.11)
where Cep is the so-called elasto-plastic tangent modulus which is given by
(CσJ σJ
el : r) ⊗ (fσ : Cel )
Cep = CσJ
el − (3.4.12)
−fq · h + fσ : CσJ
el : r
1 1
f = JII − σy2 = 0, JII = s : s (3.4.20)
3 2
f˙ = fs : ṡ = s : ṡ = 0 (3.4.21)
Sec. 3.4 Multidimensional plasticity 81
4. Yield function
f (σ, q) = 0 (3.4.16)
σ̇ = Cep : ε̇ (3.4.18)
(C : r) ⊗ (fσ : C)
Cep = C − (3.4.19)
−fq · h + fσ : C : r
82 Constitutive models Chap. 3
∂f
ε̇p = λ̇ = λ̇s (3.4.22)
∂s
From f˙ = 0 and the flow rule ε̇p = λ̇a, we can determine the plastic rate parame-
ter λ̇
aT Del ε̇
λ̇ = (3.4.23)
aT Del a
Having defined λ̇, it is now ready to determine the elasto-plastic tangent constitu-
tive matrix Dep , σ̇ = Dep ε̇
aT Del
Dep = Del − Del a (3.4.25)
aT Del a
Von Mises elasto-plastic material with hardening
Consider the effective (or equivalent) stress σ eq and effective plastic strain εpeq
which are given by
r
3
σ eq = sij sij (3.4.26)
2
r
2 p p
εpeq = ε ε (3.4.27)
3 ij ij
2.
∂ w̄(Ēe )
S̄ = (3.5.2)
∂ Ēe
Sec. 3.6 Viscoelasticity 83
D e p
L = Ḟ · F−1 = (F · F ) · (Fe · Fp )−1
Dt (3.5.3)
= Ḟe · (Fe )−1 + Fe · Ḟp · (Fp )−1 · (Fe )−1
1 1
De = (Le + LeT ), We = (Le − LeT )
2 2 (3.5.5)
1 1
Dp = (Lp + LpT ), Wp = (Lp − LpT )
2 2
Velocity gradient L̄ on Ω̄ is defined as the pull back of L by Fe
L̄ = FeT · L · Fe
eT e e −1 e p p −1 e −1
= F · Ḟ · (F ) + F · Ḟ · (F ) · (F ) · Fe (3.5.6)
= FeT · Ḟe + FeT · Fe · Ḟp · (Fp )−1
3.6 Viscoelasticity
This page intentionally contains only this sentence.
4
4.1 Introduction
The aim of this section is to introduce the underlying features of the finite element
methods (FEM). Since this method has origin in the field of mechanics, many ter-
minologies of the FEM are named using of mechanics such as stiffness matrix,
virtual displacement etc. The standard procedure to obtain a numerical solution
for the partial differential equations by the finite element methods is presented in
details. For this purpose, we usually start with the governing equations of prob-
lem at hand and then build the equivalent weak form using the weighted residual
methods. The finite element approximations of the unknown function are then
built and substituted into the weak form to get the discrete equations which can be
solved.
The problem of linear elasticity for both static and dynamic is presented. The
incompressible elasticity which requires the use of the mixed finite element ap-
proximation, or the penalty formulation with selective reduced integration is also
introduced.
The family of most widely used finite elements in two and three dimensions
are presented and the numerical integration rules such as the Newton-Cotes and
Gauss quadrature is
d2 u
+ x = 0, 0<x<1 (4.2.1)
dx2
u(0) = 1 (4.2.2)
u,x (1) = 1 (4.2.3)
The Eqs. (4.2.2) and (4.2.3) are the so-called boundary conditions and the above
problem is called the boundary value problem which is often abbreviated as BVP.
For this simple BVP, one can easily find the exact solution which is given by
−x3 + 9x + 6
u(x) = (4.2.4)
6
f(x)
1
x
0 1 2
f’(x)
1
x
0 1 2
−1
where the test function v(x) can be any function which is sufficiently well-behaved
to make sense the integration.
If you look at the Eq. (4.2.5), you will notice that the derivative of u(x) is
of second order while there is no derivative of v(x). This will lead to a lack of
symmetry in the formulation which we would like to avoid because of numerical
reasons. To do this, applying the integration by part for the first term as following
Z 1 Z 1
u,xx vdx = u,x v|10 − u,x v,x dx (4.2.6)
0 0
Substituting Eq. (4.2.7) into Eq. (4.2.5), after rearranging terms, gives
Z 1 Z 1
u,x v,x dx = v(1) + xvdx (4.2.8)
0 0
Looking at the weak form given in Eq.(4.2.8), the test and trial functions are only
C 0 functions so that the weak form is integrable. In addition, the trial function
must satisfy the essential boundary condition. Such functions are indicated sym-
bolically by
88 Finite element method Chap. 4
where the highest derivative of both trial and weighting functions are the same,
so we can choose them from the same function space Sh . In addition, the highest
derivative is now reduced to one which makes the choice of these functions easier.
The construction of a finite element approximation begins by partition the
domain in question (here is [0, 1]) into m non-overlaping subdomains called finite
elements. Nodes are then placed at the vertex of these elements. Denote n as the
total number of nodes in the domain. The coordinates of the nodes are denoted
by xI and the element domains are denoted by Ωe . In one dimension, Ωe =
[xI , xI+1 ]. Associated with each node is a shape function NI (x) which is nonzero
only over a small region called nodal support. The support of a nodal shape
function is defined by the union of the elements connected to this node. Fig. 4.2
is an example of finite element discretization of one dimensional domain into five
two node elements. The supports of node 1 and node 4 are also indicated.
n
X
h
u (x) = NI (x)uI (4.2.16)
I=1
where uI are the values of the function u(x) at nodes. They are ususally called
degrees of freedom and are unknowns need be determined of the problem.
The test function v h (x), according to the Bunov-Galerkin method, is approxi-
mated by the same shape functions
n
X
h
v (x) = NI (x)vI (4.2.17)
I=1
By substituting the finite element approximations for the trial and test functions
given in Eqs. (4.2.16), (4.2.17) into the weak form (4.2.15), we obtain
90 Finite element method Chap. 4
Z 1 Z 1
NI,x NJ,x uJ vI dx = NI (1)vI + xNI vI dx (4.2.18)
0 0
Since vI are arbitrary except I = 1, i.e., node on essential boundary condition, we
obtain (n − 1) equations
Z 1 Z 1
NI,x NJ,x dx uI = NI (1) + xNI dx I = 2, . . . , n (4.2.19)
0 0
Let us define
Z 1
KIJ = NI,x NJ,x dx ≡ a(NI , NJ ) (4.2.20)
0
which is called the stiffness matrix and
Z 1
fI = NI (1) + xNI dx ≡ NI (1) + L(NI ) (4.2.21)
0
is the external nodal force. The equation (4.2.19) is then written as
KIJ uJ = fI I = 2, . . . , n (4.2.22)
The above can be written in matrix notation as follows 1
Ku = F (4.2.23)
where the global stiffness matrix is given by
K11 K12 · · · K1n
K21 K22 · · · K2n
K = .. (4.2.24)
.. ... ..
. . .
Kn1 Kn2 · · · Knn
and the external nodal force vector and nodal displacement vector are given by
F 1
u 1
F2
u2
F= .. , u= .. (4.2.25)
.
.
F
u
n n
Solution of Eq.(4.2.23) gives the nodal displacements uI , the function u(x) at any
point is then computed by Eq.(4.2.16).
1
usually the whole matrix is calculated, the essential displacement boundary is then imposed.
Sec. 4.2 One dimensional PDE 91
E XAMPLE
In this example, the domain [0, 1] is discretized by two linear elements. Hence,
there are three nodes with associated shape functions are given by
1 − 2x 0 ≤ x ≤ 21
N1 (x) = 1 (4.2.26a)
0 2
≤x≤1
0 ≤ x ≤ 12
2x
N2 (x) = (4.2.26b)
2(1 − x) 12 ≤ x ≤ 1
0 ≤ x ≤ 21
0
N3 (x) = (4.2.26c)
2x − 1 12 ≤ x ≤ 1
The integration over the domain is taken as the sum of integrals on each element
domain as follows (since the shape functions are elementwise continuous)
Z 1 Z 1/2 Z 1
KIJ = NI,x NJ,x dx = NI,x NJ,x dx + NI,x NJ,x dx (4.2.27)
0 0 1/2
Z 1/2 Z 1
fJ = NJ (1) + xNJ dx + xNJ dx (4.2.28)
0 1/2
Using Eq.(4.2.26), we can compute the stiffness matrix and the force vector.
Therefore, the system of equations reads
2 −2 0 u1 1/24
−2 4 −2 u2 = 1/4 (4.2.29)
0 −2 2 u3 29/24
Before solving the system of equation, the boundary condition u(0) = 0 must be
imposed. Since finite element shape functions satisfy the Kronecker delta prop-
erty, i.e., NI (xJ ) = δIJ , we then have
n
X n
X
h
u (xJ ) = NI (xJ )uI = δIJ uI = uJ (4.2.30)
I=1 I=1
4 −2 u2 1/4 −2
= −1 (4.2.32)
−2 2 u3 29/24 0
Solution of this system gives us u2 = 83/48 and u3 = 56/24, then the nodal
displacement vector is
1
u= 83/48 (4.2.33)
56/24
due to the local support property of finite element shape functions. Before giving
the proof of the positive definiteness of the stiffness matrix, the definition of what
positive definite matrix is given.
2. cT Ac = 0 if c = 0
cT Kc = cI KIJ cJ
= cI a(NI , NJ )cJ
= a(cI NI , cJ NJ ) (4.2.37)
Z 1
h h
= a(u , u ) = (uh,x )2 dx ≥ 0
0
Z 1
T
From the proof of (1), c Ac = 0 means (uh,x )2 dx = 0
0
5. Interpolation function
1. Domain . [ξ1 ξ2 ]
3. Degrees of freedom u1 , u2
5. Interpolation function
1 1
x = (1 − ξ)xI + (1 + ξ)xI+1 (4.2.38)
2 2
Z 1 Z 1/2 Z 1
KIJ = NI,x NJ,x dx = NI,x NJ,x dx + NI,x NJ,x dx
0 0 1/2
Z Z (4.2.39)
= NI,x NJ,x dx + NI,x NJ,x dx
Ω1 Ω2
Z
K11 = N1,x N1,x dx (4.2.40)
Ω1
Z
K12 = N1,x N2,x dx (4.2.41)
Ω1
Z Z
K22 = N2,x N2,x dx + N2,x N2,x dx (4.2.42)
Ω1 Ω2
Z
K23 = N2,x N3,x dx (4.2.43)
Ω2
Z
K33 = N3,x N3,x dx (4.2.44)
Ω2
Z
Kije = e
Ni,x e
Nj,x dx (4.2.45)
Ωe
xe2 − x x − xe1
N1e (x) = , N2e (x) = (4.2.46)
he he
1 1
K11 K12 0
1 1 2 2
K = K21 K22 + K11 K12 (4.2.47)
2 2
0 K21 K22
1 1 −1
Ke = (4.2.48)
he −1 1
Z
e
fJ = NJ (1) + xNJ dx (4.2.49)
Ωe
Sec. 4.3 Galerkin methods 95
Z X N
Ψ L( ΦI (x)uI ) − f (x) dΩ = 0 (4.3.5)
Ω I=1
In the above equations, it was implicitly assumed that integrals are capable of
being evaluated. This places certain restrictions on the families to which functions
Ψ and Φ must belong. In general, if nth order derivatives occur in the operator
L, then the trial and test functions must be Cn−1 (n − 1 continuous derivatives).
Usually, integration by parts is applied in Equation 4.3.5 to lower the order of
derivation, decreasing the order of continuity required for the test and trial spaces.
The form of the partial differential equation is called the weak form associated
with the strong form given in Equation 4.3.1.
In order to obtain the discrete equations, the unknown function u(x) and the
test function Ψ are approximated by
2
R
in the sense of the inner product < u, v >= Ω
uvdΩ
96 Finite element method Chap. 4
N
X
h
u (x) = ΦI (x)uI
I=1
N
(4.3.6)
X
δuh (x) = ΨI (x)δuI
I=1
Collocation method
Assume the xJ to denote the set of points in the computational domain, in the
collocation method, the test functions Ψ are chosen to be Dirac delta distributions
δ(x−xJ ), because of the sifting property of the Dirac delta distributions, the weak
form, Equation 4.3.5, reduces to the strong form, evaluated at all the nodes in the
domain. The discrete equation can be written as
d2 u
−a (x) + cu(x) + f = 0 0 < x < 1 (4.3.8a)
dx2
u(0) = u(1) = 0 (4.3.8b)
with specific parameters for solution are chosen a = 0.01, c = 1 and f = −1. The
domain is divided into an equally spaced set of nodes located at xJ , J = 1, ..., N
where the boundary points are nodes x1 and xN . By imposing the equations given
in Equation 4.3.8 at the N nodes, we obtain the following equations
n n
d2 X
X
−a 2 Φi (xJ )ui +c Φi (xJ )ui +f = 0 J = 2, ..., N −1 (4.3.9a)
dx i=1 i=1
Sec. 4.4 Finite element for linear elasticity 97
N
X
Φi (x1 )ui = 0 (4.3.9b)
i=1
N
X
Φi (xN )ui = 0 (4.3.9c)
i=1
The first equation is rewritten in the familiar form
X N N
X
−a Φi,xx (xj ) + c Φi (xj ) ui + f = 0 j = 2, ..., N − 1 (4.3.10)
i=1 i=1
Ku = f (4.3.11)
where the assembly procedure is performed by looping on separate sets of nodes
(herein, there are interior and essential boundary nodes).
Galerkin methods
The test and trial functions in Galerkin methods are given by
N
X
uh (x) = ΦI (x)uI
I=1
N
(4.3.12)
X
h
δu (x) = ΨI (x)δuI
I=1
If different shape functions are used for the approximation of the test and trial
functions, a Petrov-Galerkin method is obtained, otherwise we have a Bubnov-
Galerkin method. We will assume now that ΨI = ΦI though all derivations apply
also for a Petrov-Galerkin method.
2. Strain measure
1
εij = (ui,j + uj,i ) in Ω (4.4.2)
2
3. Constitutive equation
σij = Cijkl εkl in Ω (4.4.3)
4. Boundary conditions
ui = ui on Γu (4.4.4)
ti = σij nj = ti on Γt (4.4.5)
Γu ∪ Γt = Γ; Γu ∩ Γt = ∅ (4.4.6)
5. Initial conditions
U0 = { vi | vi ∈ C 0 (x), vi = 0 on Γu } (4.4.10)
Taking the fact that vi σij,j = (vi σij ),j − vi,j σij , the above can be written as
Z Z Z Z
(vi σij ),j dΩ − σij vi,j dΩ + vi bi dΩ − vi ρüi dΩ = 0 (4.4.12)
Ω Ω Ω Ω
Using the Gauss’s theorem for the first term of the above gives
Z Z Z
(vi σij ),j dΩ = vi [[σij nj ]]dΓ + vi σij nj dΓ (4.4.13)
Ω Γd Γ
The first integral vanishes due to the continuity condition. For the second integral,
applying the natural condition given in Eq. (4.4.5) and the fact that vi vanishes on
the complement of Γt gives
Z Z
(vi σij ),j dΩ = vi ti dΓ (4.4.14)
Ω Γt
Z Z Z Z
vi ti dΓ − σij vi,j dΩ + vi bi dΩ − vi ρüi dΩ = 0 (4.4.15)
Γt Ω Ω Ω
Decomposing the tensor vi,j into symmetric and skew symmetric tensors, the term
σij vi,j can be written as
100 Finite element method Chap. 4
1 1
σij vi,j = σij (vi,j + vj,i ) + (vi,j − vj,i )
2 2
1
= σij (vi,j + vj,i ) (inner product of sym. and skew sym. tensors vanishes)
2
= σij (u)εij (v)
(4.4.16)
Z Z Z Z
vi ρüi dΩ + σij (u)εij (v)dΩ = vi ti dΓ + vi bi dΩ (4.4.17)
Ω Ω Γt Ω
Z Z Find u ∈ U such
Z that Z
ρv · üdΩ + ε(u) : D : ε(v)dΩ = v · bdΩ + v · t̄dΓ ∀v ∈ U0
Ω Ω Ω Γt
(4.4.19)
In case of static problems, the above weak form is identical to the principle of
minimum of potential energy which states that
Z Z Z
1
ΠTPE = σij εij dΩ + ui ti dΓ + ui bi dΩ (4.4.20)
Ω 2 Γt Ω
Sec. 4.4 Finite element for linear elasticity 101
1 1 1
δ σij εij = δσij εij + σij δεij
2 2 2
1 1 (4.4.21)
= Cijkl δεkl εij + σij δεij
2 2
1 1
= σkl δεkl + σij δεij = σij δεij
2 2
Figure 4.4: A typical two dimensional finite element discretization, showing the
nodal support and a triangle element
In the finite element method, the domain in question is discretized into a set of
elements. The nodes are then inserted at the vertices of these elements. Each node
is associated with a shape function which is nonzero over a small subdomain. Fig-
ure (4.4) shows a finite element discretization using triangle elements, the support
of node is also indicated. Finite-dimensional subspaces Uh ∈ U and Uh0 ∈ U0
The weak form for the discrete problem can be stated as
Z Z Find uh ∈ UZ
h
such that Z
h h
ρv · ü dΩ + h
σ(u ) : ε(v )dΩ = h h
v · bdΩ + vh · t̄dΓ ∀vh ∈ Uh0
Ω Ω Ω Γt
(4.4.22)
The finite element approximation for the displacement field reads
uh = NI (x)uI (t); uT
I = uI vI wI (4.4.24)
The approximation of the acceleration is then given by
The test function is also approximated by the same shape functions (Bubnov-
Galerkin method),
vh = NI (x)vI (4.4.26)
Using the small strain measure given in Eq. (4.4.2), we can write
∂
0 0
ε x ∂x
0 ∂ 0
ε
y ∂y
0 0 ∂ u
εz
=
∂z v (4.4.27)
∂ ∂
εyz 0 ∂z ∂y
w
εxz ∂ ∂
0
∂z ∂x
εxy ∂ ∂
0
∂y ∂x
{ε} = BI uI (4.4.28)
where
∂N
∂x
I
0 0
0 ∂NI
∂y
0
0 ∂NI
0 ∂z
BI = (4.4.29)
∂NI ∂NI
0 ∂z ∂y
∂NI ∂NI
∂z 0 ∂x
∂NI ∂NI
∂y ∂x
0
The B matrix for plane strain problems is given by
∂NI
∂x
0
∂NI
BI = 0 ∂y (4.4.30)
∂NI ∂NI
∂y ∂x
Z Z Z Z
vIT ρNT
I NJ dΩüJ + vIT BT
I DBJ dΩuJ = vIT NI bdΩ + vIT NI t̄dΓ
Ω Ω Ω Γt
(4.4.34)
Taking the arbitrariness of vIT
except on the essential boundary Γu (where the test
function vanishes), we obtain the finite element discrete equation for node I ∈
/ Γu
Z Z Z Z
ρNT
I NJ dΩüJ + T
BI DBJ dΩ uJ = NI bdΩ + NI t̄dΓ
Ω Ω Ω Γt
(4.4.35)
In compact form of matrix notation, the above is written as
Mü + Ku = f (4.4.36)
where
Z
M= ρNT NdΩ (4.4.37)
ZΩ
K= BT DBdΩ (4.4.38)
ZΩ Z
T
f= N bdΩ + NT t̄dΓ (4.4.39)
Ω Γt
Equation (4.4.36) is called the semidiscrete finite element equations since only
the space is discretized not the time. This is a second order ordinary differential
equation in time whose solution will be presented in Section...When the inertia
effect is neglected, the resulting equation Ku = f is called the equilibrium equa-
tion.
giving unity at point k and passing through n points. More terms than complete
polynomial expansion !!!
Linear elements
1 1
N1 = (1 − ξ), N2 = (1 + ξ) (4.5.2)
2 2
Quadratic elements
1 1
N1 = ξ(ξ − 1), N2 = 1 − ξ 2 , N3 = ξ(ξ + 1) (4.5.3)
2 2
1
N1 (ξ, η) = l11 (ξ)l11 (η) = (1 − ξ)(1 − η)
4
1
N2 (ξ, η) = l21 (ξ)l11 (η) = (1 + ξ)(1 − η)
4 (4.5.5)
1
N3 (ξ, η) = l21 (ξ)l21 (η) = (1 + ξ)(1 + η)
4
1
N4 (ξ, η) = l11 (ξ)l21 (η) = (1 − ξ)(1 + η)
4
The above can be written more compactly as
1
NI (ξ, η) = (1 + ξI ξ)(1 + ηI η) (4.5.6)
4
where ξI , ηI are natural coordinate of node I.
Higher order elements in this family are constructed similarity. For example,
the nine node quadrilateral elements are
Sec. 4.5 Isoparametric finite elements 105
1
NI = ξη(1 + ξ0 )(1 + η0 ) I = 1, 4
4
1 1
N5 = (1 − ξ 2 )(η 2 − η), N6 = (ξ 2 + ξ)(1 − η 2 )
2 2 (4.5.7)
1 1
N7 = (1 − ξ 2 )(η 2 + η), N8 = (ξ 2 − ξ)(1 − η 2 )
2 2
2 2
N9 = (1 − ξ )(1 − η )
1
NI (ξ, η) = (1 + ξI ξ)(1 + ηI η)(1 + ζI ζ) (4.5.8)
8
where ξI , ηI are natural coordinate of node I.
i.e., when the nodal values of the approximation are given by p(xI ), then the
function p(x) is exactly reproduced by the approximation NI uI . For example,
if the shape functions are able to reproduce constant, then if uI = 1, then the
approximation must be unity:
X
NI (x)uI = NI (x)1 = 1 (4.5.10)
I
If the shape functions reproduce linear functions, for two dimensions, a+bx+
cy, then if uI = a + bxI + cyI , then the reproducing condition implies that the
approximation is given by a + bx + cy
X
NI (x)uI = NI (x)(a + bxI + cyI )
I
! ! !
X X X
= NI (x) a + NI (x)xI b+ NI (x)yI c (4.5.11)
I I I
| {z } | {z } | {z }
1 x y
= a + bx + cy
106 Finite element method Chap. 4
The above holds for every isoparametric finite elements. Therefore, isoparametric
elements reproduce exactly linear functions. Since rigid body motion is a linear
field, isoparametric elements can represent this special kind of deformation.
∂xi ∂NI
= xiI (t) (4.5.15)
∂ξj ∂ξj
Since the square matrix in Eq. 4.5.13 is the inverse of the matrix Fξ , we can write
T
NI,x T
= NI,ξ F−1
ξ (4.5.16)
The determinant of Fξ denoted by Jξ is the Jacobian of the transformation between
the current configuration and the parent domain.
Four node quadrilateral element
The shape functions are given by
1
NI (ξ) = NI (ξ, η) = (1 + ξI ξ)(1 + ηI η) (4.5.17)
4
nodal coordinates are given by
−1 1 1 −1
ξI = (4.5.18)
−1 −1 1 1
Sec. 4.6 Numerical integration 107
The derivatives of shape functions with respect to the parent element coordinates
are
∂N1 /∂ξ ∂N1 /∂η ξ1 (1 + η1 η) η1 (1 + ξ1 ξ)
T
∂N2 /∂ξ ∂N2 /∂η 1 ξ2 (1 + η2 η) η2 (1 + ξ2 ξ)
NI,ξ =
∂N3 /∂ξ
= (4.5.19)
∂N3 /∂η 4 ξ3 (1 + η3 η) η3 (1 + ξ3 ξ)
∂N4 /∂ξ ∂N4 /∂η ξ4 (1 + η4 η) η4 (1 + ξ4 ξ)
∂x ∂x
4
∂ξ ∂η = 1 ξI (1 + ηI η)xI ηI (1 + ξI ξ)xI
X
Fξ = ∂y ∂y 4 (4.5.20)
ξI (1 + ηI η)yI ηI (1 + ξI ξ)yI
I=1
∂ξ ∂η
εx NI,x 0 N,x 0
u xI u x
εy = 0 NI,y B = 0 N,y (4.5.21)
uyI uy
2εxy NI,y NI,x N,y N,x
Z xn n
X n Z
X xn
I≈ lin−1 (x)f (xi )dx = lin−1 (x)dxf (xi ) (4.6.3)
x1 i=1 i=1 x1
1
I≈ (f (0) + f (1)) (4.6.7)
2
For three points x = 0, x = 1/2 and x = 1, we have three Cotes numbers
1
(x − 1/2)(x − 1)
Z
1
W1 = dx =
0 (0 − 1/2)(0 − 1) 6
1
(x − 0)(x − 1)
Z
2
W2 = dx = (4.6.8)
0 (1/2 − 0)(1/2 − 1) 3
1
(x − 0)(x − 1/2)
Z
1
W3 = dx =
0 (1 − 0)(1 − 1/2) 6
We then get the well known Simpson rule
1 2
I≈ (f (0) + f (1)) + f (1/2) (4.6.9)
6 3
Sec. 4.6 Numerical integration 109
To show this
and
From Eqs.(4.6.14) and (4.6.15), one can see that we must construct the polynom-
inal Pn (x) which is orthogonal to any polynominal of degree less than n and take
the root of equations Pn (xi ) = 0 as the quadrature points. Such polynominals
exist and called the Legendre polynominals.
As example, consider the first three Legendre polynominals given by
110 Finite element method Chap. 4
1
√ 1
√
x − 1/ 3
Z Z
x + 1/ 3
W1 = √ √ dx = 1, W2 = √ √ dx = 1
−1 −1/ 3 − 1/ 3 −1 1/ 3 + 1/ 3
The quadrature points and the associated Cotes numbers (often called weights)
are tabulated in table.
nQ ξi wi p = 2nQ − 1
1 0√ 2 1
2 ±1/ 3 1 3
3 p0 8/9 5
s± 3/5 p
5/9
3 − 2 6/5 1 1
± − p 7
4 7 2 6 6/5
s p
3 + 2 6/5 1 1
± − p
7 2 6 6/5
Z 1 nQ
X
f (ξ)dξ = wQ f (ξQ ) (4.6.17)
−1 Q=1
which can be easily extended to two dimensions by applying the above separately
in each direction
Z 1 Z 1 nQ1 nQ2
X X
f (ξ, η)dξdη = wQ1 wQ2 f (ξQ1 , ηQ2 ) (4.6.18)
−1 −1 Q1 =1 Q2 =1
Z Z X
T
B DBdΩe = BT DBJd = B(ξ Q )T DB(ξ Q )J(ξ Q )wQ (4.6.20)
Ωe Q
4.7 Assembly
In finite element method, the global stiffness matrix and force vector are built
from the contributions of every elements. Therefore, one computes the stiffness
matrix and force vector for each of element and assemble them to construct the
global matrix and vector. This procedure is often called assembly in litterature.
More refinement, the procedure corresponding to matrix is called matrix assembly
while for vectors, it is called the vector assembly or scatter operation. Inversely,
it is necessary to extract the element displacements from the global displacement
vector. This operation is called gather since one collect small vector from the big
one.
Let us denote the elementary mass matrix, stiffness matrix and force vector by
Me , Ke and f e , respectively. They are computed by using the formulas given in
Eqs.(4.4.37), (10.4.2) and (10.4.3) with the integration evaluated on the element
domain Ωe . The element displacement vector is denoted by ue .
The gather operation can be written as
112 Finite element method Chap. 4
ue = Le u; δue = Le δu (4.7.1)
where the connectivity matrix Le is a Boolean matrix (consist of one and zero).
Figure shows an example.
The virtual external energy can be written as the sum of virtual external energy
of all elements as
X
δwext = δweext (4.7.2)
e
Equivalently,
X X
δuT f ext = δuT ext
e fe = δuT LT ext
e fe (4.7.3)
e e
where Eq.(4.7.1) have been used in the second equality. Since the above holds for
any δu, we get the formula for scattering the element external nodal force into the
global force vector
X
f ext = LT ext
e fe (4.7.4)
e
To derive the expression for matrix assembly, we use the fact that the virtual in-
ternal energy can be written as the sum of the elementary virtual internal energy.
That is
X X
vT BT DBu = veT BT
e De Be ue = v T LT T e
e Be De Be L u (4.7.5)
e e
Since the above holds true for any v and u, we come up with
X
BT DB = LT T
e Be De Be L
e
(4.7.6)
e
The above shows that the contribution of elementary stiffness matrix to the global
stiffness matrix is given by
X
K= LT
e Ke Le (4.7.7)
e
Similarly, the mass matrix is also assembled into the global mass matrix by the
same equation
Sec. 4.8 Finite elements for incompressible elasticity 113
X
M= LT
e Me Le (4.7.8)
e
2. Constitutive equation
3. Incompressibility condition
p
uk,k + = 0 in Ω (4.8.3)
λ
For the case incompressibility, the above equations are also the governing
equations of the Stokes flow. Only the physical interpretation of the variables
is different. In Stoke flow, u is the velocity of the fluid and µ is the dynamic
viscosity.
Z Z Z Z p
− σij (u)εij (v)dΩ+ vi ti dΓ+ vi bi dΩ+ qi uk,k + dΩ = 0 (4.8.5)
Ω Γt Ω Ω λ
Introducing the constitutive equation given in Eq.(4.8.2) into the first term of the
above gives
Z Z Z
σij (u)εij (v)dΩ = 2µεij (u)εij (v)dΩ − pvi,i dΩ (4.8.6)
Ω Ω Ω
Finally
Z Z Z
Z Z
p
2µεij (u)εij (v)dΩ− pvi,i dΩ− vi ti dΓ− vi bi dΩ− qi uk,k + dΩ = 0
Ω Ω Γt Ω Ω λ
(4.8.7)
It is more convenient to write the above in tensor notation as
Z Z Z Z Z
p
2µ ε(u) : ε(v)dΩ− p∇·vdΩ− v·t̄dΓ− v·bdΩ− q ∇·u+ dΩ = 0
Ω Ω Γt Ω Ω λ
(4.8.8)
Z Z Z Z
T Te T
B DBdΩ u − (∇ · N) NdΩ p = N bdΩ + NT t̄dΓ
Ω Ω Ω Γt
Z Z (4.8.13a)
T 1 Te
N ∇ · NdΩ u +
e N NdΩ p = 0
e (4.8.13b)
Ω λ Ω
The above can be written in the matrix form as
K B u f
T = (4.8.14)
B M p 0
where
Z
K= BT DBdΩ (4.8.15)
ΩZ
Ku + Bp = f (4.8.18a)
BT u + Mp = 0 (4.8.18b)
At first, we consider the nearly incompressible materials, i.e., M 6= 0. There-
fore, from Eq.(4.8.18b), we can compute the pressure as
116 Finite element method Chap. 4
p = −M−1 BT u (4.8.19)
Introducing the above into Eq.(4.8.18a), we get the equation to solve for the dis-
placement
(K − BM−1 BT )u = f (4.8.20)
After imposition of the boundary conditions, the solution of the above equations
give the displacement fields, then the pressure is computed via Eq.(4.8.19).
Secondly, we consider the case of incompressibile where M = 0. Solving for
u from Eq.(4.8.18a) and substituting it into Eq.(4.8.18b), we obtain the equation
to deterline the pressure p as
−1 −1
(BT K B)p = BT K f (4.8.21)
? HOW TO IMPOSE THE BOUNDARY CONDITIONS ???
P D2 D3
2
ty = − −y ; I= (4.8.22)
2I 4 12
The exact solution of this problem is given in [?]
D2
Py 2
ux = − (6L − 3x)x + (2 + ν) y −
6EI 4
2
(4.8.23)
P 2 D x 2
uy = 3νy (L − x) + (4 + 5ν) + (3L − x)x
6EI 4
2.
Sec. 4.9 The penalty formulation 117
Z Z Z Z
2µ ε(u) : ε(v)dΩ+ γ(∇·v)(∇·u)dΩ− v·t̄dΓ− v·bdΩ = 0 (4.9.2)
Ω Ω Γt Ω
After introducing the finite element approximations into the above, we get the
following discrete equations
Z Z Z Z
T T
B DBdΩ+γ (∇·N) (∇·N)dΩ u = N bdΩ+ NT t̄dΓ (4.9.3)
T
Ω Ω Ω Γt
Z Z Z Z
T T¯ T
B D̄BdΩ + B D̄BdΩ u = N bdΩ + NT t̄dΓ (4.9.4)
Ω Ω Ω Γt
Z Z Z
u u
0= δσ : ε − δσ : ε + σ : δε dΩ + δv · t̄dΓ + δv · bdΩ (4.9.9)
Ω Γt Ω
ε = Bdev u+ (4.9.10)
It is obvious that the additive decomposition of the constitutive matrix D into
two parts is only straightforward for isotropic materials. In the other hand, the
anisotropic and nonlinear materials, the is not symmetric which causes the de-
composition.
From the decomposition of the strain tensor into the deviatoric and dilatational
parts as
1
εij = eij − δij uk,k (4.9.11)
3
The matrix BI for any node I can be additively decomposed into two parts as
BI = Bdev
I + Bdil
I (4.9.12)
with
B1 0 0
0 B2 0
0 0 B3 ∂NI
BI =
0 B3 B2 Bi = (i = 1, 2, 3) (4.9.13)
∂xi
B3 0 B1
B2 B1 0
In order to find the matrix expression for the dilatational part Bdil
I , we transform
it to column matrix using the Voigt rule:
uk,k
uk,k
uk,k 0 0
uk,k
δij uk,k = uk,k 0 → {δij uk,k } = (4.9.14)
0
sym uk,k
0
0
Hence,
Sec. 4.9 The penalty formulation 119
B1 B2 B3
B1 B2 B3
1B1 B2 B3
Bdil
I =
(4.9.15)
30 0 0
0 0 0
0 0 0
Bdil dil
I is replaced by an improved version B̄I where Bi replaced by B i
dil
BI = Bdev
I + BI (4.9.17)
dil
From the explicit expression of Bdev
I and BI , we obtain
B7 B5 B6
B4 B8 B6
B4 B5 B9
BI =
0
(4.9.18)
B3 B2
B3 0 B1
B2 B1 0
where
B 1 − B1 B 2 − B2 B 3 − B3
B4 = , B5 = , B6 =
3 3 3 (4.9.19)
B7 = B4 + B1 , B8 = B5 + B2 , B9 = B6 + B3
The final step is to determine the appropriate B i . Then, the stiffness matrix is
computed by
Z
T
K= B DBdΩ (4.9.20)
Ω
120 Finite element method Chap. 4
Z
Bi dΩe
Ωe
Bi = Z (4.9.21)
dΩe
Ωe
∂Na ∂Na
b1a = , b2a = (4.10.1)
∂x1 ∂x2
Isoparametric expression of the geometry reads
xi = Na xia (4.10.2)
∂Na
δij = xia = bja xia (4.10.3)
∂xj
sa = 1, ha = (−1)a (4.10.4)
where the stress was chosen as Lagrange multiplier due to the work conjugate with
strain. Applying the usual procedure of integration by parts, boundary conditions,
the first term of the above can be written as ( for shorter derivation, see section on
)
Z Z Z
δvi σij,j dΩ = δvi ti dΓ − σij δεuij dΩ (4.11.2)
Ω Γt Ω
Z Z Z
u σ u
δσij (εij − εij ) + σij δεij dΩ − δvi bi dΩ − δvi ti dΓ = 0 (4.11.3)
Ω Ω Γt
The above can be shown as the variation of a functional named the Reissner-
Hellinger functional given by
Z Z Z
1 −1
ΠHR [ui , σij ] = σij εuij − σij Cijkl σkl dΩ − vi bi dΩ − vi ti dΓ (4.11.4)
Ω 2 Ω Γt
A B σ 0
= (4.11.6)
BT 0 u f
where
Z
−1
A = − NT σ D Nσ dΩ
Z Ω
Z Z Z
δvi (σij,j + bi )dΩ − δσij (εuij − εij )dΩ − δεij (σijε − σij )dΩ = 0 (4.11.8)
Ω Ω Ω
Equivalently,
Z Z Z
0= δvi ti dΓ − σij δεuij dΩ + δvi bi dΩ
Γt Ω Ω
Z Z (4.11.9)
− δσij (εuij − εij )dΩ − δεij (σijε − σij )dΩ
Ω Ω
Z Z Z
u u ε
0= δvi ti dΓ + δvi bi dΩ − δσij (εij − εij ) + σij δεij + δεij (σij − σij ) dΩ
Γt Ω Ω
(4.11.10)
It is not so hard to point out that the above is exactly the variation of the three field
variational principle, the Hu-Washizu variational principle, which is given by
Z Z Z
1
ΠHW = σij (εuij ε
− εij ) + εij σij dΩ − vi bi dΩ − vi ti dΓ (4.11.11)
Ω 2 Ω Γt
Sec. 4.12 Enhanced strain formulation 123
Z Z Z
1 T T T T
ΠHW = ε Cε − σ ε + σ Bu dΩ − v bdΩ − vT t̄dΓ (4.11.12)
Ω 2 Ω Γt
εu = Bu (4.11.15)
ε = Nu (4.11.16)
5.1 Introduction
Heat transfer three forms: conduction, convection and radiation. Conduction heat
transfer.
Convection heat transfer
Thermal radiation
2. Fourier law relating the heat flux with the temperature gradient
where q is the heat flux per unit surface area (in W/m2 ) coming out of Ω through
the boundary Γ; n is the unit outward normal vector and Q is the internal heat
source per unit volume (in W/m3 ). C is the heat capacity, ρ is the mass density,
126 Heat transfer Chap. 5
Using the Reynold’s transport theorem to the LHS of the above and the diver-
gence theorem to the heat flux term of the RHS of the above equation gives
Z Z Z
∂T (x, t)
C(x)ρ(x) dΩ = QdΩ − ∇ · qdΩ (5.2.2)
Ω ∂t Ω Ω
∂T (x, t)
C(x)ρ(x) + ∇ · q(x, t) = Q(x, t) (5.2.3)
∂t
∂T (x, t)
C(x)ρ(x) − ∇ · (k(x) · ∇T (x, t)) = Q(x, t) (5.2.5)
∂t
The above is called the transient inhomogeneous heat equation which encompass
many well known PDE equations. If the material is isotropic and the problem is
steady state then the above equation simplifies to
− k∇2 T = Q (5.2.6)
which is the Poisson equation. If there is no internal heat source, i.e., Q vanishes,
we get the Laplace equation
∇2 T = 0 (5.2.7)
T =T on S1 (5.2.9)
The specified heat flux condition is
q·n=q on S2 (5.2.10)
− (k · ∇T ) · n = q on S2 (5.2.11)
Written in indicial notation, the above reads
q · n = h(Ts − Te ) on S3 (5.2.13)
where h is the convection coefficient, Ts is an unknown surface temperature and
Te is a known environment temparature. The radiation condition is given by
2. Boundary conditions
T =T on ΓT (5.2.17)
ΓT ∩ Γq = 0, ΓT ∪ Γq = Γ (5.2.19)
3. Initial condition
T (x, 0) = T0 (x) (5.2.20)
Z
w CρṪ − (kij T,j ),i − s dΩ = 0 (5.2.21)
Ω
The second term of the above can be rewritten as
Z Z Z
w(kij T,j ),i dΩ = (wkij T,j ),i dΩ − kij T,j w,i dΩ (5.2.22)
Ω Ω Ω
Using the divergence theorem and the boundary conditions (noting that w vanishes
on ΓT ), we have
Z Z Z
(wkij T,j ),i dΩ = wkij T,j ni dΓ = − wqdΓ (5.2.23)
Ω Γ Γq
Z Z Z Z
CρwṪ dΩ + kij T,j w,i dΩ = wQdΩ − wqdΓ (5.2.24)
Ω Ω Ω S2
T = NI TI (t) (5.2.25)
Sec. 5.2 Problem statement 129
where NI are the finite element shape functions and TI (t) are the nodal tempera-
tures to be determined. The time derivative of the temperature is then given by
Ṫ = NI ṪI (5.2.26)
Using the Bubnov-Galerkin method, the test function w is approximated by the
same shape function NI ,
w = NI wI (5.2.27)
where wI are arbitrary parameters.
Introducing the above approximations into Eq.(5.2.24) and using the arbitrari-
ness of wI 3 , we obtain the discrete finite element equations
Z Z Z Z
CρNI NJ ṪJ dΩ + kij NJ,j NI,i TJ dΩ = NI QdΩ − NI qdΓ (5.2.28)
Ω Ω Ω Γq
Z
MIJ = ρCNI NJ dΩ (5.2.30)
ZΩ
KIJ = NI,i kij NJ,j dΩ (5.2.31)
ZΩ Z
fJ = NI QdΩ − NI qdΓ (5.2.32)
Ω Γq
The capacity matrix is also non diagonal since finite element shape functions are
not orthogonal. The conductivity matrix is symmetric.
The Eq.(5.2.29), in matrix notation, is written as
MṪ + KT = f (5.2.33)
This is the general semidiscrete equation for transient heat transfer problems. The
problem is linear in the following cases
1. the mass density ρ and the heat capacity C are independent of the tempera-
ture;
3
the imposition of the essential boundary condition is considered later
130 Heat transfer Chap. 5
2. the thermal conductivity of the material k does not depend on the tempera-
ture;
Otherwise, i.e., the physical properties of material are functions of the tem-
perature and/or the boundary conditions on the heat source and heat flux depend
on temperature, the problem is nonlinear which the corresponding semidiscrete
equation given by
6.1 Introduction
This chapter deals with the methods solving the finite element equations. We
start with the simplest one, the equations for linear equilibrium problems. They
are systems of simultaneous linear algebraic equations which can be solved either
by direct methods such as the Gauss elimination method and LDU factorization
method, or by indirect (iterative) methods, the preconditioned conjugate method,
for instance.
The discrete equation of nonlinear equilibrium equations
The equation of transient heat transfer problems is a parabolic equation which
can be solved by the generalized midpoint rule is next introduced.
The chapter ends with the discussion of methods to solve the finite element
equations of the elastodynamics problem. These are second-order ordinary differ-
entiable equations in time.
second one is the preconditioned conjugate gradient method with sparse row-wise
format of matrix storage.
t − tn
u(t) = un + (un+1 − un ) = θ(t)un+1 + (1 − θ(t))un (6.3.1)
∆t
where un = u(tn ); ∆t = tn+1 − tn and θ(t) = (t − tn )/∆t. It is obvious that
un+1 − un
u̇ ≈ (6.3.2)
∆t
Substituting the above equations into , we obtain
un+1 − un
M +K[θ(t)un+1 +(1−θ(t))un ] = θ(t)fn+1 +(1−θ(t))fn (6.3.3)
∆t
where we have made the same linear approximation to the RHS. The above equa-
tion is still continuous since θ(t). By choosing a fixed value for θ in [0, 1], the
Eq.(6.3.8) is known as the generalized midpoint method.
∆t ∆t ∆t
M+ K un+1 = (fn + fn+1 ) + M − K un (6.3.5)
2 2 2
MṪ + KT = f (6.3.7)
where M is the capacity matrix, K is the conductivity matrix, f is the heat supply
vector, T is the temperature vector and Ṫ is the time derivative of T. The problem
to be solved is to find the temperature vector T(t), t ∈ [0, T ], satisfy Eq.(6.3.7)
subjected to the boundary condition T = T̄ on S2 and the initial condition T(0) =
T0 .
The most widely used algorithms used to solve this problem are members of
the generalized trapezoidal family of methods, which consists of the following
equations
Tn+1 = T
e n+1 + θ∆tṪn+1 (6.3.11)
134 Solution of FE equations Chap. 6
where
MṪn+1 + K(T
e n+1 + θ∆tṪn+1 ) = fn+1 (6.3.13)
After rearranging terms, the above becomes
(M + θ∆tK)Ṫn+1 = fn+1 − KT
e n+1 (6.3.14)
The above can be solved to get Ṫn+1 , then Tn+1 is obtained through Eq. (6.3.11).
Obviously, in this implementation, we get the derivative of the temperature first,
then the temperature itself. This is why it is called the v-form of the generalized
midpoint rule.
From Eq.(6.3.11), one can compute Ṫn+1 in terms of Tn+1 and T e n+1 , then sub-
stituting this expression into the first equation of Eq. (6.3.8), we obtain
!
Tn+1 − T e n+1
M + KTn+1 = fn+1 (6.3.15)
θ∆t
The above can be written as
(M + θ∆tK)Tn+1 = θ∆tfn+1 + MT
e n+1 (6.3.16)
Solution of the above gives Tn+1 .
In the case θ = 0, the method is the well-known forward Euler which is explicit
if the mass matrix M is diagonal. In this case, the solution of Eq. (6.3.14) or
Eq.(6.3.16) is trivial, hence no equation solving is required. This is the most
advantage of explicit integration methods.
In case where θ 6= 0, the method is called implicit and at each time step, one
must solve a system of equation to advance the solution. For θ = 1 the method is
the backward Euler. The most useful of the implicit methods encompassed in the
Sec. 6.3 Solution of the heat equations 135
φT
i Mφj = δij (6.3.19)
From the above condition, in what follows, we will show that the eigenvectors are
orthogonal to the K matrix. Pre-multiplying Eq.(6.3.18) with φTj gives
φT T
j Kφi − λi φj Mφi = 0 (6.3.20)
Using Eq.(6.3.19), the above becomes
φT
j Kφi = λi δij (6.3.21)
The orthornormality and orthogonality of eigenvectors are very useful in decou-
pling the Eq.(6.3.7) into uncoupled equations. Since the eigenvectors form a basis
then we can write
α̇i (φT T T
j Mφi ) + αi (φj Kφi ) = φj f (6.3.23)
α̇j + αj λj = fj (6.3.25)
uncoupled equation
136 Solution of FE equations Chap. 6
αin+1 − αin
+ λi (1 − θ)αin + λi θαin+1 = fi (6.3.26)
∆t
1 − λi ∆t(1 − θ) n ∆tfi
αin+1 = αi + (6.3.27)
1 + λi ∆tθ 1 + λi ∆tθ
1 − λi ∆t(1 − θ)
−1< <1 (6.3.28)
1 + λi ∆tθ
It can be shown easily that the right-and inequality is automatically satisfied. The
left-hand inequality requires that λi ∆ < 2/(1 − 2θ) when θ < 1/2, but has no re-
strictions for θ >= 1/2. When there is no restriction on the time step, the method
is called unconditionally stable, otherwise the method is called conditionally sta-
ble. Therefore, the Crank-Nicolson (θ = 1/2) and the backward Euler (θ = 1) are
implicit, unconditionally stable integration schemes; the forward Euler (θ = 0) is
an explicit conditionally stable integration method.
2
∆t < ∆tcrit = (6.3.29)
λmax (1 − θ)
Z
MIJ = ρCNI NJ dΩ (6.3.31)
ZΩ
KIJ = NI,i kij NJ,j dΩ (6.3.32)
ZΩ Z
fJ = NI QdΩ − NI qdΓ (6.3.33)
Ω Γq
(∆t)2
d̃n+1 = dn + ∆tvn + (1 − 2β)an
2 (6.4.3)
ṽn+1 = vn + (1 − γ)∆tan
The above gives the acceleration an+1 , then Eqs.(6.4.4) and (6.4.5) are used to
compute the velocity vn+1 and displacement dn+1 . In this implementation, the
acceleration is calculated first , hence the name a-form implementation.
In the case of β = 0, if the mass and damping matrices are diagonal, the
method is called explicit since no equation solving is necessary to advance the
solution.
The updated acceleration and velocity can be computed using Eqs. (6.4.4) and
(6.4.5)as
138 Solution of FE equations Chap. 6
1
an+1 = (dn+1 − d̃n+1 ) with β > 0
β∆t2
(6.4.7)
1
vn+1 = ṽn+1 + γ∆t (dn+1 − d̃n+1 )
β∆t2
Substituting Eq. (6.4.7) into the first equation of Eq.(6.4.2), we obtain, after some
rearrangements
1 γ 1 γ
M+ C + K dn+1 = fn+1 + M+ C d̃n+1 − Cṽn+1
β∆t2 β∆t β∆t2 β∆t
(6.4.8)
Solution of the above gives us the displacement dn+1 firstly. That is why this
implementation is called the d-form implementation. Then, the acceleration and
velocity are calculated using Eq.(6.4.7).
Since at each time step, the solution of system of equations must be performed,
this implementation is called implicit. This form also works for equilibrium equa-
tions by neglecting the mass and damping matrices. Therefore, in practical fi-
nite element codes, the transient nonlinear problems using the implicit integration
methods and the nonlinear equilibrium problems are implemented in one proce-
dure with a switch α which is unity for transient and zero for equilibrium prob-
lems.
7
Stabilization methods
7.1 Introduction
7.2 One dimensional advection-diffusion equation
Consider the one dimensional advection-diffusion equation which is given by:
find φ(x)
where w(x) is the test function which is a C 0 function satisfying w(0) = w(L) =
0. After integration by parts, the above becomes
Z L Z L
uwφ,x dx + vw,x φ,x dx = 0 (7.2.5)
0 0
140 Stabilization methods Chap. 7
The finite element approximations for the trial and test function are given (Bubnov-
Galerkin method)
u∆x
Pe = (7.2.13)
2v
In terms of the Peclet number, Eq. (7.2.12) becomes
1 + Pe
µ = 1 and µ= (7.2.18)
1 − Pe
Therefore, the nodal solution of the problem is
j
1 + Pe
φj = c1 + c2 (7.2.19)
1 − Pe
where c1 and c2 are determined from the boundary conditions. Comparing this
approximate solution with the exact solution, given in Eq.(7.2.3), we conclude
that
1. If Pe < 1, then [(1 + Pe )/(1 − Pe )]j > 0 for all j. The discrete solution is
therefore similar to the exact solution.
2. If Pe > 1, then the sign of [(1 + Pe )/(1 − Pe )]j depends on j. That means
the spatial instability occurs.
w
e= w
|{z} + γw,x (7.2.20)
|{z}
Galerkin test function discontinuous test function
with γ = α ∆x 2
, and α is chosen to eliminate the oscillations of the numerical
solutions for Pe > 1. Noting that while w ∈ U0 , w
e∈/ U0 1 .
The weak form of the Eq.(7.2.1) is given by
1
U0 is the set of C 0 functions which vanish on essential boundaries.
142 Stabilization methods Chap. 7
Z L
w(uφ
e ,x − vφ,xx )dx = 0 (7.2.21)
0
Z L Ne Z
X
w(uφ,x − vφ,xx )dx + γw,x (uφ,x − vφ,xx )dx = 0 (7.2.22)
0 Ωe
| {z } |e=1 {z }
Galerkin term
upwind Petrov-Galerkin term
Z L Z L Z L Z L
uwφ,x dx + vw,x φ,x dx + γuw,x φ,x dx − γvw,x φ,xx dx = 0
0 0 0 0
(7.2.23)
Consider linear elements whose second derivatives of shape functions vanish, i.e.,
φ,xx = 0, the above simplifies as
Z L Z L
∆x
uwφ,x dx + v + αu w,x φ,x dx = 0 (7.2.24)
0 0 2
∆x
v∗ = v + v with
(7.2.26) v = αu
2
where v can be considered as artificial viscosity which stabilizes the numerical
solution. Let us define the Peclet numbers associated with the real viscosity v, the
artifical viscosity v and the total viscosity v ∗ as
Sec. 7.2 One dimensional advection-diffusion equation 143
1 2v ∗ v+v 1 1
∗
= =2 = + (7.2.28)
Pe u∆x u∆x Pe P e
The finite element solution of Eq. (7.2.24) is (see Eq.(7.2.19))
j
1 + Pe∗
φj = c1 + c2 (7.2.29)
1 − Pe∗
Comparing the above with the exact solution given in Eq. (7.2.3), we have
1 + Pe∗ u
∗
= e v ∆x (7.2.30)
1 − Pe
The RHS of the above can be written in term of Pe as
1 + Pe∗ u u
∗
= e v ∆x = e2 2v ∆x = e2Pe (7.2.31)
1 − Pe
Hence
1 1 1
= + (7.2.33)
tanh(Pe ) Pe P e
The above allows us to compute the artificial viscosity v
2v 1
= coth(Pe ) − (7.2.34)
u∆x Pe
Therefore, one obtains
u∆x 1
v= coth(Pe ) − (7.2.35)
2 Pe
Recall that v = α u∆x
2
, we then come up with the formula determining α:
1
α = coth(Pe ) − (7.2.36)
Pe
144 Stabilization methods Chap. 7
The above is the well known coth-formula. With this definition of the stabiliza-
tion parameter, the nodal finite element solutions will be coincident to the exact
solution in case of linear elements and a regular mesh.
u · ∇φ − v∇2 φ = 0 in Ω (7.3.1)
subject to the boundary conditions
e = w + ku · ∇w
w (7.3.3)
where k is the stabilization parameter given by
h 1 ||u||h
k = |α| , α = coth(Pe ) − , Pe = (7.3.4)
||u|| Pe 2v
and h is the characteristic element length.
The weak form is built by multiplying the strong form (7.3.1) by the test func-
tion w
e and integrating over the domain, so
Z
e · ∇φ − v∇2 φ)dΩ = 0
w(u (7.3.5)
Ω
After introducing the test function given in Eq.(7.3.3), the above becomes
Z
(w + ku · ∇w)(u · ∇φ − v∇2 φ)dΩ = 0 (7.3.6)
Ω
where the first term vanishes due to boundary conditions. Substituting Eq.(7.3.7)
into Eq.(7.3.6), after rearrangement, we obtain the weak formulation:
Z Z
0= wu · ∇φdΩ + v∇w · ∇φdΩ
Ω Ω
| {z }
standard Galerkin terms
Ne Z
X Ne Z
X (7.3.8)
2
+ k(u · ∇w)(u · ∇φ)dΩ − k(u · ∇w)v∇ φdΩ
e=1 Ωe e=1 Ωe
| {z }
upwind streamline Petrov-Galerkin terms
(LIJ + βLstab
IJ + KIJ )φJ = 0 (7.3.11)
where β is a switch between the Galerkin and Petrov-Galerkin formulations.
In what follows, numerical computations are given to illustrate the robustness
of the SUPG. The value of the velocity is u = (1000, 500) and the domain is
a unit square. The values of φ are zeroes on the external boundary except on a
small segment on the left edge, φ equals to unity. The regular mesh of 50 × 50 Q4
elements is used. Fig. (7.1) shows the results for various values of v.
In Fig. (7.2), the same problem is solved with the streamline upwind Petrov-
Galerkin method.
146 Stabilization methods Chap. 7
Figure 7.1: Results for the advection-diffusion equation for various values of the
viscosity obtained with the standard Galerkin formulation.
Sec. 7.3 Multidimensional SUPG 147
Figure 7.2: Results for the advection-diffusion equation for various values of the
viscosity obtained with the Petrov-Galerkin stabilization.
This page intentionally contains only this sentence.
8
Lagrangian finite elements
8.1 Introduction
Finite elements using Lagrangian meshes are commonly classified as total La-
grangian formulation and updated Lagrangian formulation. In both formulations,
the independent variables are the material coordinates X and time. In the total
Lagrangian formulation, the stress and strain are Lagrangian, i.e., they are de-
fined with respect to the reference configuration (for example, the nominal or
second Piola-Kirchhoff stress are employed), the derivatives are computed with
respect to the material coordinates. The corresponding weak form therefore in-
volves integrals over the reference configuration. On the orther hand, the updated
Lagrangian formulation uses the Eulerian strain and stress measures (the Cauchy
stress), the derivatives are computed with respect to the spatial coordinates x. The
corresponding weak form therefore involves integrals over the current (deformed)
configuration.
2. Conservation of momenta
3. Conservation of energy
4. Constitutive equations
Z Z Z
∂
(δvi σij )dΩ = δvi [[nj σij ]]dΓ + δvi nj σij dΓ (8.3.5)
Ω ∂xj Γint Γ
From the traction continuity, the first term on the RHS vanishes. For the second
integrand, using the traction boundary condition and the fact that δvi vanishes on
the complement of Γt , the above becomes
Z Z
∂
(δvi σij )dΩ = δvi ti dΓ (8.3.6)
Ω ∂xj Γt
4. Conservation of energy
5. Constitutive equation
σ ∇ = StσD (D, σ) (8.2.5)
6. Rate of deformation
1 ∂vi ∂vj
D = sym(∇v), or Dij = + (8.2.6)
2 ∂xj ∂xi
7. Boundary conditions
8. Initial conditions
Z Z Z
∂σij ∂δvi
δvi dΩ = δvi ti dΓ − σij dΩ (8.3.7)
Ω ∂xj Γt Ω ∂xj
Introducing Eq. (8.3.7) into Eq. (8.3.3) yields
Z Z Z Z
∂δvi
σij dΩ − δvi ti dΓ − δvi ρbi dΩ + δvi ρv̇i dΩ = 0 (8.3.8)
Ω ∂xj Γt Ω Ω
∂δvi
σij = δvi,j σij
∂xj
= δLij σij (8.4.1)
= (δDij + δWij )σij
= δDij σij = δD : σ
Since δDij σij is the virtual internal power per unit volume, the total virtual inter-
nal power δP int is defined by integrating over the entire domain
Z
int ∂δvi
δP = σij dΩ (8.4.2)
Ω ∂xj
The second and third term in Eq. (8.3.8) are the virtual external power because
they are done by the external forces.
Z Z
ext
δP = δvi ρbi dΩ + δvi ti dΓ (8.4.3)
Ω Γt
The last term in Eq. (8.3.8) is the virtual kinetic (or inertial) power
Z
kin
δP = δvi ρv̇i dΩ (8.4.4)
Ω
Inserting Eqs. (8.4.2-8.4.4) into Eq. (8.3.8), we can write the principle of virtual
power as:
Sec. 8.5 Finite element discretization 153
Z
int
δP = ∇δv : σdΩ
ZΩ Z
ext
δP = ρb · δvdΩ + t̄ · δvdΓ
ZΩ Γt
where NI (X) are the shape functions. It is emphasized that the shape functions
are expressed in terms of the material coordinates although we will use the weak
form defined in the current configuration. This is done because we want the time
dependence
Write the above equation at the time t = 0, we obtain
nN
X
Xi = NI (X)XiI , X = NI (X)XI (8.5.2)
I=1
The velocity is determined by taking the time material derivative of the dis-
placement field,
Z Z Z Z
∂NI
δviI σij dΩ − NI ti dΓ − NI ρbi dΩ + NI ρv̇i dΩ = 0 (8.5.10)
Ω ∂xj Γt Ω Ω
Z Z Z Z
∂NI
σij dΩ − NI ti dΓt − NI ρbi dΩ + NI ρv̇i dΩ = 0 ∀(i, I) ∈
/ Γvi
Ω ∂xj Γt Ω Ω
(8.5.11)
Sec. 8.5 Finite element discretization 155
Z Z
ext
δP = δvi ρbi dΩ + δvi ti dΓ
Ω Γt
Z Z (8.5.14)
= δviI NI ρbi dΩ + NI ti dΓt
Ω Γt
The terms in parentheses can therefore be defined as the external nodal force
Z Z
fiIext = NI ρbi dΩ + NI ti dΓt (8.5.15)
Ω Γt
Z
kin
δP = δvi ρv̇i dΩ
Ω
Z (8.5.16)
= δviI ρNI NJ v̇iJ dΩ
Ω
Z
kin
ρNI NJ dΩ 0
fxI Ω
v̇xJ
= (8.5.18)
Z
kin
fyI
v̇yJ
0 ρNI NJ dΩ
Ω
This is the consistent mass matrix which is constant for Lagrangian meshes. To
show this, the above is transformed to the reference configuration
Z
MIJ = I ρNI NJ JdΩ0 (8.5.21)
Ω0
It is clear that the integrand of the above integral is constant. So, the mass matrix
for Lagrangian meshes is constant and therefore can be only computed at the
beginning of the computation. It is noting that Eq.(8.5.22) can be considered as
the total Lagrangian. We have used a total Lagrangian form for the mass matrix
in an updated Lagrangian formulation.
Figure 8.1: Initial and current configurations and their relation to the parent ele-
ment domain
Using the shape function expressed in terms of the natural coordinates, the
motion is approximated
m
X
xi (ξ, t) = NI (ξ)xiI (t) (8.5.26)
I=1
The above shows that if this map is one-to-one, then the natural coordinates can
be considered surrogate material coordinates in a Lagrangian mesh.
158 Lagrangian finite elements Chap. 8
T
NI,x T
= NI,ξ F−1
ξ (8.5.29)
where the matrix F is given by in three dimensions
x,ξ x,η x,ζ
Fξ = y,ξ y,η y,ζ (8.5.30)
z,ξ z,η z,ζ
whose components are given by (from Eq. (8.5.26)),
∂xi ∂NI
= xiI (t) (8.5.31)
∂ξj ∂ξj
The determinant of Fξ denoted by Jξ is the Jacobian of the transformation between
the current configuration and the parent domain.
The Jacobian of the transformation between the reference configuration and the
parent domain Jξ0 is the determinant of the following matrix
X,ξ X,η X,ζ
F0ξ = Y,ξ Y,η Y,ζ (8.5.32)
Z,ξ Z,η Z,ζ
This will be used in the total Lagrangian finite elements described in the next
chapter.
8.5.7 Implementation
In the finite element computation, all the internal nodal force, external nodal force
and the mass matrix are computed for each of the elements.
For ease of reading, we present the implementation in two dimensions first, the
generalization to three dimensions is straightforward. Using the Voigt notation,
the rate of deformation tensor D and the Cauchy stress tensor σ are considered as
column matrices as
Sec. 8.5 Finite element discretization 159
Dx σ11
{D} = Dy , {σ} = σ22 (8.5.33)
2Dxy σ12
From its definition, Dij = 1/2(vi,j + vj,i ), the rate of deformation can be
written by
vx1
vy1
Dx N1,x 0 ··· Nm,x 0
Dy = 0 N1,y · · · 0 Nm,y .. (8.5.34)
.
2Dxy N1,y N1,x · · · Nm,y Nm,x
vxm
vym
or
{D} = Bv (8.5.35)
where
B = {B1 , B2 , . . . , Bm } (8.5.36)
with the matrix BI in two dimensions given by
NI,x 0
BI = 0 NI,y (8.5.37)
NI,y NI,x
and in three dimensions,
NI,x 0 0
0 NI,y 0
0 0 NI,z
BI =
0
(8.5.38)
NI,z NI,y
NI,z NI,x 0
NI,y NI,x 0
This B matrix is identical to the one used in the linear finite elements. There-
fore, the updated Lagrangian formulation is quite easy to implement since one
just needs the deformed nodal coordinates 1 to evaluate the derivative of shape
functions. Then use the available routines to construct the matrix B.
The virtual internal power is now given by
1
simply the sum of the initial coordinates plus the displacements
160 Lagrangian finite elements Chap. 8
Z Z
int
δP = δD : σdΩ = {δD}T {σ} dΩ
ZΩ Ω
(8.5.39)
T T
= δv B {σ} dΩ
Ω
Therefore, the internal nodal force vector is given by
Z
int
f = BT {σ} dΩ (8.5.40)
Ω
The above is the most important equation in nonlinear finite element program.
The external nodal force vector is simpler and given by
Z Z
ext T
f = ρN bdΩ + NT t̄dΓt (8.5.41)
Ω Γt
4. Conservation of energy
ρ0 ẇint = ḞT : P (8.7.4)
162 Lagrangian finite elements Chap. 8
5. Constitutive equation
6. Strain measure
1 T
E= F ·F−I (8.7.6)
2
7. Boundary conditions
8. Initial conditions
P(X, 0) = P0 (X) (8.7.8a)
u̇(X, 0) = u̇0 (X) (8.7.8b)
Z Z Z
∂
(δui Pji )dΩ0 = δui [[Pji n0j ]]dΓ0 + δui Pji n0j dΓ0 (8.8.5)
Ω0 ∂Xj Γ0int Γ0
The first integral of the RHS vanishes due to the continuity condition. Applying
the traction boundary condition and the fact that δvi vanishes on the complement
of Γ0t gives
Z Z
∂
(δui Pji )dΩ0 = δui t̄0i dΓ0 (8.8.6)
Ω0 ∂X j 0
Γt
Z Z Z Z
∂δui ∂ui ∂xi
dΩ0 = δ dΩ0 = δ dΩ0 = δFij dΩ0 (8.8.7)
Ω0 ∂Xj Ω0 ∂Xj Ω0 ∂Xj Ω0
∂ui ∂ ∂xi
= (xi − Xi ) =
∂Xj ∂Xj ∂Xj
Substituting Eqs. (8.8.6) and (8.8.7) into Eq. 8.8.4 yields
Z Z Z
∂Pji 0
δui dΩ0 = δui t̄i dΓ0 − δFij dΩ0 (8.8.8)
Ω0 ∂Xi Γ0t Ω0
By substitution Eq. (8.8.8) into Eq. (8.8.3), we obtain after some rearrangements
Z Z Z Z
δFij Pji dΩ0 − δui t̄0i dΓ0 − δui ρ0 bi dΩ0 + δui ρ0 üi dΩ0 = 0 (8.8.9)
Ω0 Γ0t Ω0 Ω0
Z
int
δW = (∇0 δu)T : PdΩ0
Ω0
Z Z
δW ext = ρ0 b · δudΩ0 + t̄0 · δudΓ0
Ω0 Γ0t
Z
δW kin = ρ0 δu · üdΩ0
Ω0
∂xi ∂NI
Fij = = xiI (t) (8.10.5)
∂Xj ∂Xj
where in the second equality, Eq. (8.10.1) has been used. The Green strain ten-
sor E can be computed from the deformation gradient F. However, it should be
computed using the gradient of displacement as
1
Eij = (Hij + Hji + Hki Hkj ) (8.10.6)
2
with
∂ui ∂NI
Hij = = uiI (8.10.7)
∂Xj ∂Xj
The variation of F is therefore defined by
∂NI ∂NI
δFij = δxiI = δuiI (8.10.8)
∂Xj ∂Xj
where we have used δxiI = δ(uiI + XiI ) = δuiI . It is convenient to define the
matrix B0 as B0 = [B01 B02 · · · ] with each sub-matrix defined by
∂NI
B0jI = (8.10.9)
∂Xj
Explicitly,
T
B0I = NI,X NI,Y
NI,Z (8.10.10)
With the defined matrix B0 , the internal virtual work is given by
Z
int
δW = δuiI B0jI Pji dΩ0 (8.10.11)
Ω0
The above is the formula to compute the internal nodal force using the full stress
matrix. We can use the Voigt notation (stress vector), we first transform the un-
symmetric nominal stress to the symmetric second Piola-Kirchhoff stress. Using
the transformation P = S · FT , the internal nodal force is written as
Sec. 8.10 Finite element semidiscretization 167
Z
∂NI
fiIint = Sjk Fik dΩ0 (8.10.13)
Ω0 ∂Xj
It is easier to write the above explicitly as follows in two dimensions
int
∂NI∂x ∂NI ∂x ∂NI ∂x ∂NI ∂x
Sx
fxI +
= ∂X ∂X ∂Y ∂Y ∂X ∂Y ∂Y ∂X Sy (8.10.14)
int ∂NI ∂y ∂NI ∂y ∂NI ∂y ∂NI ∂y
fyI +
∂X ∂X ∂Y ∂Y ∂X ∂Y ∂Y ∂X Sxy
NI,X F11 NI,X F21 NI,X F31
NI,Y F12 NI,Y F22 NI,Y F32
0
N I,Z F 13 N I,Z F 23 NI,Z F 33
BI = NI,Y F13 + NI,Z F12 NI,Y F23 + NI,Z F22 NI,Y F33 + NI,Z F32
NI,Y F13 + NI,Z F12 NI,Y F23 + NI,Z F22 NI,Y F33 + NI,Z F32
NI,Y F13 + NI,Z F12 NI,Y F23 + NI,Z F22 NI,Y F33 + NI,Z F32
(8.10.17)
0
It can be shown that the above matrix BI relates the material time derivative
of the Green strain tensor to the nodal velocities as follows
1
Ėij = (Ḟki Fkj + Fki Ḟkj )
2
1 ∂ u̇k ∂xk ∂xk ∂ u̇k
= + (8.10.18)
2 ∂Xi ∂Xj ∂Xi ∂Xj
1 ∂NI ∂xk ∂xk ∂NI
= + u̇kI
2 ∂Xi ∂Xj ∂Xi ∂Xj
Z Z Z Z
fiIext = NI t̄0i dΓ0 + NI ρ0 bi dΩ0 , fIext = 0
NI t̄ · ei dΓ0 + NI ρ0 bdΩ0
Γ0t Ω0 Γ0t Ω0
(8.10.21)
This is identical to the external nodal force of linear finite elements. From the
inertial virtual work
Z
kin
δW = δuiI ρ0 NI NJ dΩ0 üiJ (8.10.22)
Ω0
Z
MIJ = ρ0 NI NJ dΩ0 I
Ω0
Z Z
fIint = BT
0I {S}dΩ0 , (fIint )T = BT
0I PdΩ0 (8.10.28)
Ω0 Ω0
Z Z
fIext = 0
NI t̄ · ei dΓ0 + NI ρ0 bdΩ0
Γ0t Ω0
where Jξ0 is the Jacobian of the mapping between the reference domain and the
parent element domain. The well known Gauss quadrature is then used to evaluate
the integrals over the parent element domain.
Consider an element with m nodes. The integration point has coordinate ξ Q
and weight wQ . The procedure for computation of the internal nodal force for this
element is as follows
dn+1 − dn 1
vn+1/2 = n+1 n
= d − d (8.11.2)
tn+1 − tn ∆tn+1/2
After rearrangement, the above can be written as
vn+1/2 − vn−1/2
an = , vn+1/2 = vn−1/2 + ∆tn an (8.11.4)
tn+1/2 − tn−1/2
Substitution of Eq. (8.11.2) and its counterpart for the previous time step into the
above, the acceleration is expressed directly in terms of displacements as
(∆t)2
d̃n+1 = dn + ∆tvn + (1 − 2β)an
2 (8.11.12)
n+1
ṽ = vn + (1 − γ)∆tan
sD
r= M(dn+1 − d̃n+1 ) + f int (dn+1 , tn+1 ) − f ext (dn+1 , tn+1 ) = 0
β∆t2
(8.11.16)
The above are nonlinear algebraic system of equations in the nodal displacement
dn+1 which can be solved efficiently by the iterative Newton-Raphson method.
After solution of this equation, we can compute the acceleration an+1 by Eq.
(8.11.15). Finally, the velocity vn+1 is determined through Eq. (8.11.14).
where
sD ∂f int ∂f ext
KT = M + − (8.11.19)
β∆t2 ∂u ∂u
This matrix is called the finite element Jacobian of the system of equation or most
often the tangent stiffness matrix. The procedure to derive this matrix is called the
linearization and is subject of the next section.
ui+1 i i
n+1 = un+1 + ∆un+1 (8.11.20)
8.11.4 Linearization
We start with the linearization of the internal nodal force. By comparing the ex-
pression of the total and updated Lagrangian formulations, we choose the former
to do the linearization because in this formula, only the nominal stress P is time
dependent or more precisely depends on the increment of displacement.
Taking the material time derivative of the internal nodal force, we have
Sec. 8.11 Solution procedures 173
Z
∂NI
f˙iIint = Ṗji dΩ0 (8.11.21)
Ω0 ∂Xj
The constitutive equations are usually not expressed in terms of Ṗ because
this stress rate is not objective. We therefore use the material time derivative of
the second Piola-Kirchhoff stress, which is objective. Indeed, by using P = SFT ,
the above equation becomes
Z
˙int ∂NI
fiI = (Ṡjk Fik + Sjk Ḟik )dΩ0 (8.11.22)
Ω0 ∂Xj
This equation shows that the rate of the internal nodal force consists of two parts:
1. Material part.
2. Geometric part.
where {Ṡ} is the rate of the second Piola-Kirchoff stress tensor written in the
Voigt column matrix form. The constitutive equation in rate form is
The above forms can be easily converted to the updated Lagrangian forms by
letting the current configuration be the reference configuration.
Z
Kmat
IJ = BT στ
I [C ]BJ dΩ
Ω
Z (8.11.32)
Kgeo
IJ = I BT
I σBJ dΩ
Ω
τ̇ = Ḟ · P + F · Ṗ (8.11.33)
Solving for Ṗ gives
−1 ∇c
Ṗ = F−1 · (τ ∇c + τ · LT ), or Ṗji = Fjk (τki + τkl Lil ) (8.11.36)
Sec. 8.11 Solution procedures 175
Z
∂NI ∂Xj ∇c
f˙iIint = (τ + τkl Lil )dΩ0
∂Xj ∂xk ki
ZΩ0 (8.11.37)
∂NI ∇c
= (τ + τkl Lil )dΩ0
Ω0 ∂xk ki
Separating the above into the material and geometric parts gives
Z Z
˙mat ∂NI ∇c ˙geo ∂NI
fiI = τki dΩ0 , fiI = τkl Lil dΩ0 (8.11.38)
Ω0 ∂xk Ω0 ∂xk
It is easily to transform the above result to the updated Lagrangian forms using
the fact that
dΩ = JdΩ0 , τ ∇c = Jσ ∇T
Z
f˙iIint = ∇T
NI,k (σki + σkl Lil )dΩ (8.11.39)
Ω
τij∇c = Cijkl
τ
Dkl (8.11.40)
Substituting the above into the first part of Eq.(), we obtain
Z
˙mat ∂NI τ
fiI = Ckijl Djl dΩ0 (8.11.41)
Ω0 ∂xk
Recall that the rate of deformation is the symmetric part of the velocity gradient
L, and the minor symmetry of the tangent moduli. That is,
1 τ τ
Djl = (Lj,l + Ll,j ), Ckijl = Ckilj (8.11.42)
2
We can write
τ 1 τ 1 τ τ
Ckijl Djl = Ckijl Lj,l + Ckilj Ll,j = Ckijl Lj,l (8.11.43)
2 2
The Eq. (8.11.41) therefore becomes
176 Lagrangian finite elements Chap. 8
Z
∂NI τ
f˙iImat = C Lj,l dΩ0
∂xk kijl
ZΩ0
τ
= NI,k Ckijl NJ,l u̇jJ dΩ0 (8.11.44)
Ω0
Z
στ
= NI,k Ckijl NJ,l dΩu̇jJ
Ω
where in the second equality, just the finite element approximation for the velocity
has been used and in the third step, we have transformed to the current configura-
tion.
Z
mat στ
KijIJ = NI,k Ckijl NJ,l dΩ (8.11.45)
Ω
In Voigt notation, the above reads
Z Z
mat
KIJ = T στ
BI [C ]BJ dΩ = J −1 BT τ
I [C ]BJ dΩ (8.11.46)
Ω Ω
which is identical to Eq. ().
9.1 Introduction
As introduced in the chapters on Lagrangian finite elements, the mesh is deformed
with the material which makes the treatment of boundaries and material interfaces
trivially. In addition, and obviously the most salient feature of the Lagrangian for-
mulation is that integration points at particles remain unchanged. This facilitates
the modeling of history-dependent materials. However, since the mesh deformed
with the material, for large deformation problems, secure mesh distortion is in-
evitable and must be recourse to remeshing to get reliable results. Due to these
reasons, the Lagrangian formulation is best suited to solid mechanics problems
and much relies on the ability of elements to distortion.
In the Eulerian formulation, contrast to the Lagrangian formulation, the mesh
is fixed with respect to the motion of the material. It is therefore clearly that large
deformation is no problem since the mesh does never deformed, i.e., their shapes
never change. However, the boundaries and interfaces move with respect to the
mesh which... Hence, the Eulerian formulation is widely used in fluid mechanics.
It is naturally to derive a formulation combining the advantages of the both
methods. The Arbitrary Lagrangian Eulerian formulation (ALE)
x = φ(X, t) (9.2.1)
which is the map between the current configuration Ω and the material configura-
tion Ω0 .
In the ALE formulation, one more domain is defined which is called the refer-
ential domain or the ALE domain Ω̂. The initial positions of particles are denoted
by χ, so
χ = φ(X, 0) (9.2.2)
The coordinates χ are called the ALE coordinates. The purpose of the ALE do-
main Ω̂ is to describe the motion of the mesh independently with the motion of
the material.
The motion of the mesh is given by
x = φ̂(χ, t) (9.2.3)
From Eqs. (9.2.1) and (9.2.3), we can relate the ALE coordinates to the mate-
rial coordinates by a composition of functions
∂v̂(χ, t) ∂ 2 û(χ, t)
â(χ, t) = = = û,tt [χ] (9.2.8)
∂t ∂t2
D ∂f (χ, t) ∂f ∂Ψi
f (χ, t) = + (9.2.9)
Dt ∂t ∂χi ∂t
Defining the referential particle velocity wi as
∂Ψi
wi = (9.2.10)
∂t
Introducing the previously defined variable into Eq. (9.2.9) gives
D ∂f
f (χ, t) = f,t [χ] + wi (9.2.11)
Dt ∂χi
Defining the convective velocity denoted by c as the difference between the mate-
rial and mesh velocities:
ci = vi − v̂i (9.2.14)
From Eq. (9.2.13), we obtain
∂xi
ci = vi − v̂i = wj (9.2.15)
∂χj
This relation between the convective velocity, material velocity and mesh velocity
will be used much in the ALE formulation.
180 Arbitrary Lagrangian Eulerian formulation Chap. 9
D ∂f
f (χ, t) = f,t [χ] + wi
Dt ∂χi
∂f ∂xj
= f,t [χ] + wi (9.2.16)
∂xj ∂χi
∂f
= f,t [χ] + cj
∂xj
D
f (χ, t) = f,t [χ] + c · ∇f (9.2.17)
Dt
ρ̇ + ρ∇ · v = 0 (9.3.1)
In the above, replacing the material time derivative of ρ by the ALE form given in
Eq. (9.2.17) yields
The time derivative of vi is also replaced by the ALE form given in Eq. (9.2.17),
the above becomes
2. Momentum equation
3. Energy equation
6. Initial conditions
182 Arbitrary Lagrangian Eulerian formulation Chap. 9
where δe
vi vanishes on Γvi . Using the results derived in chapter ??, we obtain
Z Z Z
vi,j σij dΩ −
δe vi ti dΓ −
δe δe
vi ρbi dΩ
Ω Γt Ω
Z Z
vi ρvi,t [χ]dΩ + δe
+ δe vi ρcj vi,j dΩ = 0 (9.5.3)
Ω Ω
χ = NI (ξ)χI (9.5.4)
mesh motion is approximated by
In ALE description, the density ρ is also a dependent variable which is ap-
proximated by
Due to the presence of convective terms (), use of the standard Galerkin method
will leads to the so-called spatial instability. To overcome this, we adopt the
Petrov-Galerkin method where the trial and test functions come from different
function spaces. The trial shape functions for the density and material velocity
are denoted by NIρ and NI , respectively whereas the test shape functions for the
density and material velocity are given by N̄Iρ and N̄I .
Z Z Z
N̄Iρ NJρ ρJ,t dΩ + N̄Iρ ci NJ,i
ρ
ρJ dΩ + N̄Iρ vi,i NJρ ρJ dΩ = 0 (9.5.7)
Ω Ω Ω
dρ
Mρ + Lρ ρ + Kρ ρ = 0 (9.5.8)
dt
where Mρ , Lρ and Kρ are capacity, transport and divergence matrices, respec-
tively,
Z
ρ
MIJ = N̄Iρ NJρ dΩ
ZΩ
LρIJ = N̄Iρ ci NJ,i
ρ
dΩ (9.5.9)
ZΩ
ρ
KIJ = N̄Iρ vi,i NJρ dΩ
Ω
Z Z Z
N̄I,j σij dΩ − N̄I ti dΓ − N̄I ρbi dΩ
Ω
Z Γt Z Ω
dv
M + Lv + f int − f ext = 0 (9.5.11)
dt
where M and L are the mass and convective matrices, respectively, whereas f int
and f ext are the internal and external nodal force:
Z
MIJ = I N̄I ρNJ dΩ
ZΩ
LIJ = I N̄I ρcj NJ,j dΩ
Z Ω (9.5.12)
fiIint = N̄I,j σij dΩ
ZΩ Z
ext
fiI = N̄I ti dΓ + N̄I ρbi dΩ
Γt Ω
vi = δvi + δviPG
δe (9.6.1)
where δvi = 0 on Γvi while δviPG not. Therefore, the Petrov-Galerkin weak form
is
Z
(δvi + δviPG )(ρv̇i − σji,j − ρbi )dΩ = 0 (9.6.2)
Ω
Z Z
δvi (ρv̇i − σji,j − ρbi )dΩ + δviPG (ρv̇i − σji,j − ρbi )dΩ = 0 (9.6.3)
Ω Ω
Applying the usual procedure of integration by parts for the first term (see the
derivation of weak form in chapter ??), the above becomes
Sec. 9.6 Petrov-Galerkin formulation for the momentum equation 185
Z Z Z Z
δvi ρv̇i dΩ + δvi,j σji dΩ − δvi ti dΓt − δvi ρbi dΩ
Ω Ω Γt Ω
| {z }
Galerkin terms
Z (9.6.4)
+ δviPG (ρv̇i − σji,j − ρbi )dΩ = 0
|Ω {z }
Petrov-Galerkin stabilization terms
k= (9.6.6)
Z Z
δviI ρkcj NI,j NJ v̇iJ dΩ + ρkcj NI,j ck NJ,k viJ dΩ
Ω Ω
Z Z (9.6.10)
− kcj NI,j σki,k dΩ − ρkcj NI,j bi dΩ = 0
Ω Ω
Taking the arbitrariness of δviI gives the stabilization equation in matrix form as
dv int ext
Mstab + Lstab v + fstab − fstab =0 (9.6.11)
dt
where
Z
Mstab = I[MIJ ]stab = I ρkcj NI,j NJ dΩ
ZΩ
Lstab = I[LIJ ]stab = I ρkcj NI,j ck NJ,k viJ dΩ
Ω
Z (9.6.12)
int
[fiI ]stab = − kcj NI,j σki,k dΩ
Z Ω
[fiIext ]stab = ρkcj NI,j bi dΩ
Ω
Finally, combining two equations (9.6.7) and (9.6.11), we get the discrete
Petrov-Galerkin equations for the linear momentum as
dv
(M + Mstab ) + (L + Lstab )v + f int + fstab
int
− (f ext + fstab
ext
)=0 (9.6.13)
dt
The above can be written as
dv
M + Lv + f int − f ext = 0 (9.6.14)
dt
Z
LIJ = (∇NI )T ρkccT ∇NJ dΩ
stab
(9.6.15)
Ω
Z c1c1 c1c2 c1c3 NJ,x
Lstab
IJ = NI,x NI,y NI,z ρk c2c1 c2c2 c2c3 NJ,y dΩ (9.6.16)
Ω c3c1 c3c2 c3c3 NJ,z
Sec. 9.7 Linearization of the discrete equations 187
dfiIint
Z
∂NI b
= Pki,t[χ] dΩ̂ (9.7.4)
dt Ω ∂χk
10.1 Introduction
The most widely used numerical methods for solving the partial differential equa-
tions are the finite difference method, finite element methods and the finite vol-
ume method. The finite difference method (FDM) are suitable only to problems
with simple geometries and hence there are little practical applications. The fi-
nite element methods (FEM), in contrary, can be applied to problems with com-
plex geometries and complex boundary conditions following a unified procedure
(same steps for many different problems, hence ease the computer implementa-
tion). That is the reason why FEMs are numerical tools of choice for practical
applications ranging from structure mechanics, solid mechanics, fluid mechanics
to electromagnetic, just to name afew. Nowadays, there are many general purpose
finite element softwares such as ANSYS, ABAQUS, LSDYNA, just to name a
few, which are extensively used in daily work of engineers.
Despite of this gross success, limitations do exist with the finite element meth-
ods. By constructing the approximate solution based on the mesh, inaccurate re-
sults would be obtained in case of severely distorted mesh in large deformation
problems. While distorted mesh can be solved by building a new not distorted
mesh (often called remeshing), this remeshing procedure usually leads to degra-
dation of accuracy and results in an excessive computational cost. The Arbitrary
Lagrangian Eulerian (ALE) methods have been used with great success for large
deformation problems, but coming with its complexity and additional treatment
of transport terms (require stabilization techniques).
Another class of problems which can not be efficiently solved by standard
FEMs is the so-called moving boundaries problems. Typical examples are crack
propagation, interface modeling (solid-fuild, fluid-fluid interfaces) etc. Since finite
190 Meshless and Partition of Unity methods Chap. 10
element mesh must be conform to these interfaces (interfaces locating along the
element surfaces), when these interfaces change, a new mesh must be built to
conform with the updated interfaces.
The meshfree methods (also called meshless methods) have been born in ef-
forts of constructing approximations not relying on mesh. According to mesh-
less methods, the approximation is build based on a set of nodes scattered over
the domain under consideration. In MMs, the approximation is built from nodes
only. One of the first meshfree methods is the Smooth Particle Hydrodynamics
(SPH) method by [?] and [?]. It was born to solve problems in astrophysics and,
later on, in fluid dynamics [?, ?, ?]. [?] were the first to employ SPH in solid
mechanics (impact). Since the original SPH version suffered from spurious insta-
bilities and inconsistencies [?, ?, ?], many improvements were incorporated into
SPH [?, ?, ?, ?, ?, ?, ?, ?]. While SPH and their corrected versions were based
on a strong form, other methods were developed in the 1990s, based on a weak
form. Major applications of these methods are in solid mechanics. The element-
free Galerkin (EFG) method [?] was developed in 1994 and was one of the first
meshfree methods based on a global weak form. The Reproducing kernel particle
method (RKPM) [?] was developed one year later. Though the final equations
are very similar to the equations of the EFG method, RKPM has its origin in
wavelets. In contrast to RKPM and the EFG method, that use a so-called intrinsic
basis, other methods were developed that use an extrinsic basis and the partition
of unity concept. This extrinsic basis was initially used to increase the approxi-
mation order similar to a p-refinement as e.g. in the hp-cloud method [?, ?]. [?]
pointed out the similarities of meshfree and finite element methods and developed
a method that he called Partition of Unity Finite Element Method (PUFEM). The
method is very similar to the hp-cloud method and only different in the way the
approximation order is increased. While in PUFEM, Shepard functions were used
exclusively as shape functions, the shape functions in the hp-cloud method were
varied. [?] pointed out in their generalized finite element method (GFEM) that
different partition of unities can be used for the usual approximation and the so-
called enrichment. In the XFEM [?, ?, ?], the extrinsic enrichment was modified
such that it can handle strong discontinuities without remeshing. Another class
of meshfree methods are methods that are based on local weak forms. The most
popular method is the Meshless Local Petrov Galerkin (MLPG) method, [?, ?, ?].
The main difference of the MLPG method to methods such as EFG or RKPM is
that local weak forms are generated on overlapping subdomains rather than using
global weak forms. The integration of the weak form is then carried out in these
local subdomains. [?] introduced the notion "truly" meshfree since no construc-
tion of a background mesh is needed for integration purposes.
It is emphasized that the meshless methods should only be used for specific
problems which can not treated by standard finite elements efficiently. For others,
Sec. 10.3 Reproducing conditions 191
the finite element methods should be the first choice due to less computational
cost, availability of reliable softwares. In conclusion, choose appropriate methods
for your problems since no methods are
or in indicial notation,
X
φI (x)xIα = xα (10.2.2)
I
1
noting that we are interested in second order partial differential equations as frequently en-
countered in solid mechanics.
192 Meshless and Partition of Unity methods Chap. 10
where φI (x) are the meshless shape functions, uI are the particle parameters and
N is number of particles where φI (x) is nonzero. Comparing to finite element
approximation, Eq.(10.3.1) has the following differences:
1. Meshless shape functions do not satisfy the Kronecker delta property, i.e.,
φI (xJ ) 6= δIJ . Hence, uI 6= uh (xI ). Therefore, uI are often called the
fictitious nodal values.
2. The number N , which defines particles contributing to the approximation at
point x, is calculated in the course of computation. It is completely different
to finite element methods, where N is the number of nodes per element
containing point x, which has been defined in advance (when building the
mesh).
The meshless approximating functions φI (x) are constructed from the weight-
ing functions. A weighting function is defined to have compact support, i.e., it is
nonzero only at a subdomain small relative to rest of the domain. The support
is often called the domain of influence of a particle. In practice, the domain of
influence is usually circle in 2D or sphere in 3D. Figure (10.1) shows a typical 2D
meshless model with irregular particles with circular supports.
where w(x − y, s(y)) is the kernel function (also called weighting functions), s is
the
The discrete SPH approximation is obtained by nodal integration
uh (x) ∼
X
= w(x − xI , s(xI ))uI ∆VI (10.3.3)
I∈S
with S contains particles I where the kernel functions w(x−xI , s(xI )) are nonzero,
uI are the function value at particle I and finally, ∆VI , the integration weight, is
the area (2D) or volume (3D) associated with particle I.
For solution of PDE, the derivative of function is needed. Similar to kernel
approximation of a function, Eq. (10.3.2), the kernel approximation of a function
gradient is given by
Z
h
∇u (x) = w(x − y, s(y))∇u(y)dΩy (10.3.4)
Ωy
Z Z
h
∇u (x) = w(x−y, s(y))u(y)ndΓ− ∇w(x−y, s(y))u(y)dΩy (10.3.5)
Ωy
The boundary term is usually omitted in SPH with arguments that either the kernel
function vanishes on the boundary or the function does. The discrete function
gradient is then given by nodal integration
∇uh (x) ∼
X
=− ∇w(x − xI , s(xI ))uI ∆VI (10.3.6)
I
The SPH approximation can be written in form of the familiar finite element ap-
proximation as
X
uh (x) = φI (x)uI (10.3.7)
I
1 function
0.35 approximation
0.9
0.3 0.8
0.7
0.25
0.6
0.2
0.5
0.15 0.4
0.3
0.1
0.2
0.05
0.1
0 0
−3 −2 −1 0 1 2 3 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
(a) ρ/h = 1
1 function
0.35 approximation
0.9
0.3 0.8
0.7
0.25
0.6
0.2
0.5
0.15 0.4
0.3
0.1
0.2
0.05
0.1
0 0
−3 −2 −1 0 1 2 3 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
(b) ρ/h = 2
1 function
0.35 approximation
0.9
0.3 0.8
0.7
0.25
0.6
0.2
0.5
0.15 0.4
0.3
0.1
0.2
0.05
0.1
0 0
−3 −2 −1 0 1 2 3 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
(c) ρ/h = 4
|xI − x|
r= (10.3.10)
dI
where dI is the support size of node I. Some commonly-used weight functions
are given
• Circular support
||xI − x||
w(x − xI ) = w1D (10.3.13)
dI
• Rectangular support
|xI − x| |yI − y|
w(x − xI ) = w1D w (10.3.14)
dxI dyI
p
where ||x|| = x21 + x22 .
where C(x, y) is defined such that the approximation is n-th order consistent.
Z Z
p(y)w(x − y)u(y)dΩy = p(y)pT (y)w(x − y)dΩy a (10.3.19)
Ωy Ωy
Sec. 10.3 Point-based approximations 197
This is a system of equation from which a is solved and substituted into the ap-
proximation uh (x) = pT (x)a, it yields
Z −1 Z
h T T
u (x) = p (x) p(y)p (y)w(x − y)dΩy p(y)w(x − y)u(y)dΩy
Ωy Ωy
(10.3.20)
with the corrected function
Z −1
T T
C(x, y) = p (x) p(y)p (y)w(x − y)dΩy p(y)
Ωy (10.3.21)
= pT (x)[M(x)]−1 p(y)
To evaluate this continuous expression, numerical integration must be em-
ployed. This step leads from the reproducing kernel method to its discrete version,
the reproducing kernel particle method.
Z
h
u (x) = C(x, y)w(x − y)u(y)dΩy
Ωy
N
X
= C(x, xi )w(x − xi )ui ∆Vi (10.3.22)
i=1
N
X
T −1
= p (x)[M(x)] p(xi )w(x − xi )ui ∆Vi
i=1
where the vector p(x) is called the intrinsic basis which is a complete polynomial
of order m:
The weighted square of difference between the function and the approximate
function J(x) is given by
n
X n
X
h 2
J(x) = w(x − xI )[u(xI ) − u (xI )] = w(x − xI )[u(xI ) − pT (xI )a(x)]2
I=1 I=1
(10.3.31)
where n is the number of points around the point x where the weight function
w(xI − x) 6= 0.
The minimum of J(x) in Eq. (10.3.31) with respect to a(x) can be obtained
by setting the derivative of J(x) with respect to a(x) equal to zero. The following
equations result
n
X
w(x − xI )2p1 (xi )[pT (xI )a(x) − uI ] = 0
I=1
n
X
w(x − xI )2p2 (xi )[pT (xI )a(x) − uI ] = 0
I=1 (10.3.32)
..
.
n
X
w(x − xI )2pm (xi )[pT (xI )a(x) − uI ] = 0
I=1
n
X n
X
w(x − xI )p(xI )pT (xI )a(x) = w(x − xI )p(xI )uI (10.3.34)
I=1 I=1
200 Meshless and Partition of Unity methods Chap. 10
Solving a(x) from Eq.(10.3.35) and substituting it into Eq. (10.3.29), the MLS
approximants can be defined as
where uI is value of the function u(x) at particle I, we can immediately write for
the MLS shape functions
Remark At this time, it is worthy to note the difference between finite element
approximation and MLS approximation. In FEMs, the connectivity of an element
is defined in advance while in MLS the connectivity of a particle x i.e., the set of
particles xI where the weight functions w(x − xI ) do not vanish, is determined
during the calculation.
Reproducing conditions of MLS We now show that the MLS approximation can
reproduce any functions in the basis vector p. A function in this basis is of the
form
Sec. 10.3 Point-based approximations 201
w(x − xI )
Φ0I (x) = Pn (10.3.44)
I w(x − xI )
which is the Shepard functions, the lowest order form of MLS shape functions.
It
Pcan0be shown with ease, the Shepard functions form a partition2 2of unity, i.e.,
I ΦI (x) = 1. Advantages of Shepard approximants are of C and can be
computed at relatively low computational cost.
Graphical illustration of MLS shape functions is obviously the best way to
well understand them. For this purpose, consider an interval 0 ≤ 0 ≤ 4 which is
divided into 5 equally spaced nodes. The weight and shape function of nodes is
plotted in Figure 10.3 with bold symbol for ones of the central node. In this exam-
ple, the quartic spline function was used as weight function, the size of domain of
influence of nodes is 2.5, and the linear basis pT = [1 x] was employed. From
this figure, one can easily see that MLS shape functions do not satisfy the Kro-
necker delta property. The derivatives of MLS shape functions are given in Figure
10.4. To get this smooth figure, we computed these derivatives at 150 sampling
points on the interval 0 ≤ 0 ≤ 1. An important property of the first derivatives can
be observed from this figure: the first derivative of node I vanishes at this node.
This makes meshless methods using collocation procedure instable.
Figure (10.5) illustrates the MLS shape functions for various support sizes.
For the case of ρ = h, i.e., the nodal support is the same as in finite elements, the
MLS shape functions coincide linear finite element shape functions. The figure
also shows that there exits an optimal ratio between ρ and h.
As illustration of two dimensional MLS shape functions, considering a 5x5
uniformly distributed nodes given in Figure ??. Here, the circular support of
radius 1.4 and linear basis are used. For weight functions, the quartic spline is
2
since all commonly used weighting functions are C 2 .
202 Meshless and Partition of Unity methods Chap. 10
1 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
(a) (b)
Figure 10.3: Weight and shape function of the central node: (a) Weight function
and (b) Shape function
0.7 15
0.6
10
0.5
5
0.4
0.3 0
0.2
−5
0.1
−10
0
−0.1 −15
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
200
100
−100
−200
−300
−400
−500
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(c) ΦI,xx
Figure 10.4: MLS shape functions and derivatives with quadratic basis
Sec. 10.3 Point-based approximations 203
1
1 function
0.9 approximation
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
(a) ρ/h = 1
1
1 function
0.9 approximation
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
(b) ρ/h = 2
1
1 function
0.9 approximation
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
(c) ρ/h = 4
chosen. The weight function as well as the shape function of the central node are
plotted in Figure 10.6.
1 0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0 0
2 2
1 2 1 2
0 1 0 1
0 0
−1 −1
−1 −1
−2 −2 −2 −2
(a) (b)
Figure 10.6: Two dimensional weighting and MLS shape function of the central
node: (a) Weight function and (b) Shape function
1
0.5
0.5
0
0
−0.5
−0.5
−1
−2 2
−1 −1 1
−2 0 0 2
−1 1
0 1 −1 0
1 −1
2 2 −2 −2
(a) (b)
Figure 10.7: Shape function first derivatives of the central node: (a) Φ,x and (b)
Φ,y
Since the Separd functions Φ0I (x) form a partition of unity, we can construct
an approximation as
X
uh (x) = Φ0I (x)(a0I + a1I x) (10.3.46)
I
which exactly reproduces linear functions. However, with this form of approxima-
tion, at each node I, there are two parameters (in 1D) to satisfy linear reproducing
condition.
X
uh,x (x) = GxI uI
I
X (10.3.47)
uh,y (x) = GyI uI
I
X
α1 (x)Φ0I,x + α2 (x)Φ0I,y xI = 1
I
X (10.3.49)
(α1 (x)Φ0I,x + α2 (x)Φ0I,y )yI = 0
I
Aα = r1 , Aβ = r2 (10.3.50)
where
X Φ0 xI Φ0 xI
I,x I,y
A= (10.3.51)
Φ0I,x yI Φ0I,y yI
I
206 Meshless and Partition of Unity methods Chap. 10
rT rT
1 = 1 0 , 2 = 0 1 (10.3.52)
The Petrov-Galerkin method with Shepard approximant as test function and
derivatives of trial function being the consistent pseudo derivatives (defined to
satisfy the consistency) will pass the patch test for the Poisson equation. To show
this, we start with the strong form of the 2D Poisson equation
u = a1 + a2 x + a3 y (10.3.54)
where ai are arbitrary constants.
The weak form of the above Poisson equation is
Z
u,i v,i dΩ = 0 (10.3.55)
Ω
Substituting the test function (Shepard approximants) and derivatives of trial func-
tion (pseudo derivatives) into the above gives
Z
(GxJ uJ Φ0I,x vI + GyJ uJ Φ0I,y vI )dΩ = 0 (10.3.56)
Ω
KIJ uJ = 0 (10.3.57)
with the stiffness matrix given by
Z
KIJ = (Φ0I,x GxJ + Φ0I,y GyJ )dΩ (10.3.58)
Ω
The method is said to pass the test if the numerical solution at any interior par-
ticle, uI is given by Eq.(10.3.54). Hence, for any interior particle I, the following
must hold (substitution of Eq.(10.3.54) into Eq.(10.3.57))
Z
(Φ0I,x GxJ (a1 + a2 xJ + a3 yJ ) + Φ0I,y GyJ (a1 + a2 xJ + a3 yJ ))dΩ = 0 (10.3.59)
Ω
for arbitrary constants ai . Since the pseudo correctives have been defined to satisfy
the reproducing conditions Eq.(10.2.3), the above will be satisfied if
Sec. 10.4 Some common meshfree methods 207
Z Z
Φ0I,x dΩ = Φ0I,y dΩ = 0 (10.3.60)
Ω Ω
Ku = f (10.4.1)
where
Z
KIJ = BTI DBJ dΩ (10.4.2)
ZΩ Z
fI = φI bdΩ + φI t̄dΓ (10.4.3)
Ω Γt
φI,x 0
BI = 0 φI,y (10.4.4)
φI,y φI,x
where φI (x) are MLS shape functions.
Remark. It is emphasized that the essential boundary conditions are not con-
sidered yet. Except for the use of coupling with finite elements to treat Dirichlet
boundary conditions, Eq. (10.4.1) will be modified to account for essential bound-
ary conditions. Noting that the solution of the discrete equations, uI are not the
nodal displacement. Hence, in post processing, to get uJ = φI (xJ )uI .
ui = ūi on Γu (10.5.2)
According to the Lagrangian multiplier method, we look for the displacement
field and the Lagrangian multiplier field which make the modified functional min-
imum
Z Z Z Z
1
Π̄(u, λ) = σij εij dΩ− ui ti dΓ− ui bi dΩ− λi (ui − ūi )dΓu (10.5.3)
Ω 2 Γt Ω Γu
where λ is the Lagrangian multiplier. Taking the variation of the above gives
Z Z Z
0 = δ Π̄(u, λ) = σij δεij dΩ − δui ti dΓ − δui bi dΩ
Ω Γt Ω
Z Z (10.5.4)
− δλi (ui − ūi )dΓu − λi δui dΓu
Γu Γu
Or
K G u f
= (10.5.7)
GT 0 λ q
with
210 Meshless and Partition of Unity methods Chap. 10
Z
GIK = − φI (x)φLK (x)SdΓu
ZΓu (10.5.8)
qK = − φLK (x)SūdΓu
Γu
where S is a diagonal matrix of size 2 × 2 (in two dimensions), with Sii equal to
one if the displacement is imposed on xi and zero otherwise; the expression of the
external nodal force remains unchanged.
Z Z Z Z
1 α
Π̄(u, α) = σij εij dΩ− ui ti dΓ− ui bi dΩ− (ui −ūi )2 dΓu (10.5.9)
Ω 2 Γt Ω 2 Γu
Z Z Z
0 = δ Π̄(u, α) = σij δεij dΩ − δui ti dΓ − δui bi dΩ
Ω Γt Ω
Z (10.5.10)
−α (ui − ūi )δui dΓu
Γu
After substitution of approximations of test and trial functions into the above, we
get the discrete equations
Ku = f (10.5.11)
with
Z Z
KIJ = BT
I CBJ dΩ −α
ΦI ΦJ SdΓ
Ω Γu
Z Z Z (10.5.12)
fI = ΦI tdΓ + ΦI bdΩ − α ΦI uSdΓ
Γt Ω Γu
Remark. From the previous formula, the stiffness matrix is the difference of two
stiffness matrices, one defined for the interior domain and one for the Dirichlet
boundary.
The advantage of the penalty method is that no additional unknowns are intro-
duced. The choice of penalty number, however, is not easy. In fact, the penalty
number is very large and depends on material properties (see the above equation).
In addition, the more larger the penalty number is, the more ill-conditionning ma-
trix.
u = Dû (10.5.15)
û = D−1 u (10.5.16)
X
−1
uh (x) = φI (x)DIJ uJ (10.5.17)
I
Now we can apply the essential boundary conditions directly. However, this
method requires the inversion of a matrix of size N × N which is, obviously,
not efficient at all and hence is strongly recommended not be used in practice.
This method, however, is useful in deriving the new one that we call the partial
or boundary transformation method. The basis idea is quite the same but now we
want to inverse just a smaller matrix.
Denote the number of interior particles as NΩ and the number of particles on
the essential boundary Γ as NΓ . We could write the approximation as follows
NΩ
X NΓ
X
uh (xJ ) = ΦI (xJ )uˆI Ω + ΦI (xJ )uˆI Γ (10.5.18)
I=1 I=1
212 Meshless and Partition of Unity methods Chap. 10
Ω Ω Γ Γ
D
| {zû } + D
| {zû } = g (10.5.20)
|{z}
(NΓ ×NΩ )(NΩ ×1) (NΓ ×NΓ )(NΓ ×1) (NΓ ×1)
With these modified shape functions, we can directly impose the essential bound-
ary conditions as done for the finite element methods.
where SΓP is the set of nodes on ΓP , i.e., r(x) is the sum of finite element shape
functions of all nodes on the particle boundary.
Sec. 10.6 Integration of the weak form 213
X X
uh (x) = R(x)φI (x)uI + (1 − R(x))NJ (x)uJ , x ∈ ΩB (10.5.26)
I J
where S is the set of particles !!! and VI0 is the volume associated with particles
I.
are then inserted at the centers of the triangles, see Figure (10.8(a)). Second, the
Voronoi tessellation 3 is performed to generate Voroinoi polygons for master as
well as slave particles, see Figure (10.8(b)).
master particles
stress points
(a)
master particles
stress points
(b)
Figure 10.8: Stress point insertion and volume calculation: (a) stress point, (b)
Voronoi cell
Z
fI = ∇φI (x) · σdΩ
Ω
X X (10.6.3)
= VJ0P ∇φI (xPJ ) · σ(xPJ ) + VJ0S ∇φI (xSJ ) · σ(xSJ )
J∈N P J∈N S
3
For each point in a set of coplanar points, you can draw a polygon that encloses all the in-
termediate points that are closer to that point than to any other point in the set. Such a polygon
is called a Voronoi polygon, and the set of all Voronoi polygons for a given point set is called a
Voronoi diagram.
Sec. 10.6 Integration of the weak form 215
where N P (N S ) are sets of master particles (stress points) falling within the do-
main of influence of the master particle I. The stress field at the slave particles
σ(xSJ ) is computed via constitutive equation with kinematical variables such as
strain rate computed from those of neighbouring master particles, for example
X
uSI = φJ (xSI )uPJ (10.6.4)
J
Z
KIJ = BT
I DBJ dΩ (10.6.5)
Ω
(a) (b)
Figure 10.9: Gauss quadrature for Galerkin weak forms: (a) background mesh
and (b) structure cell
• Find nodes within the support of xg . For each of these nodes, compute
• weight function, shape function and shape function derivatives
• Assemble B matrix
• Assemble K matrix
7. Solve for the nodal displacement since meshless approximation does not
satisfy the Kronecker delta property;
LUc(x) = p(x)
Uc(x) = L−1 p(x) (10.7.3)
c(x) = U−1 L−1 p(x)
ΦI,k (x) = cT T
,k (x)p(xI )wI (x) + c (x)p(xI )wI,k (x) (10.7.4)
with
218 Meshless and Partition of Unity methods Chap. 10
11.1 Introduction
The Partition of Unity Method (PUM) [?] allows for the addition of a priori knowl-
edge about the solution of a boundary value problem into the approximation space
of the numerical solution through the addition of enrichment functions that may
be exactly reproduced by the enriched numerical scheme. The extended finite
element method (XFEM) uses a local partition of unity [?, ?], where some of
the enrichment functions are chosen to be discontinuous thereby allowing for the
reproduction of strong discontinuities within an element. Only a portion of the
mesh is enriched. Another similar application of PUM is known as the General-
ized Finite Element Method (GFEM) [?,?]. Very recently, another related method
emerged with the work of Hansbo and Hansbo [?], based on Nitsche’s method, in
which the authors give a convergence proof. Further, in Reference [?], it is shown
that the kinematics of the method described in [?] is in fact equivalent to that
of the extended finite element method. Based on Nitsche’s method, recent work
has been performed in the area of discontinuity modeling within a finite element
framework [?, ?, ?]. A review on computational modeling of cohesive cracks is
given in [?].
Any function may be introduced in the approximation. In particular, singu-
lar functions may be added into the approximation to decrease the required mesh
density close to singularities such as crack tips in linear elastic fracture mechan-
ics (LEFM). Enriched finite element methods have been successfully applied to
numerous solid mechanics problems such as 2D crack growth problems in linear
elastic fracture mechanics with small displacements [?,?,?,?], and large displace-
ment [?, ?]. Extensions to 3D were presented in [?, ?, ?] and further improved
to handle crack initiation and propagation in [?] and multiple cracks [?]. Other
222 Partition of Unity methods Chap. 11
applications include multiple crack growth in brittle materials [?], crack growth
in shells and plates [?], cohesive crack growth [?, ?, ?, ?, ?, ?, ?, ?], bi-material
interface cracks [?], holes and inclusions [?, ?], brittle fracture in Polycrystalline
Microstructures [?], shear bands [?] and, finally, contact problems [?]. Multiscale
work has been performed with micro-macro crack models based on the LATIN
methods (LArge Time INcrement method) in [?]. The XFEM has also been used
to model computational phenomena in areas such as fluids mechanics, phase trans-
formations [?], material science and biofilm growth [?, ?], Chemically-induced
swelling of hydrogels [?], among others. Fluid-structure interaction problems are
also free-boundary problems, where the “free" boundary usually is the structure.
Recent work has been performed in this area by several researchers [?]. Recent de-
velopments of enriched finite element methods in conjunction with discontinuous
Galerkin in time, for dynamics and time dependent problems have been recently
made [?, ?, ?, ?]. XFEM was also recently introduced in spectral elements [?]. A
nice work on interface conditions is given in [?].
For a complete review on recent developments of both XFEM and GFEM, in-
terested readers can refer to the papers of Q.Z. Xiao and B.L. Karihaloo at Cardiff
University [?, ?].
The PUM may be applied in the context of meshfree methods such as the El-
ement Free Galerkin method (EFG) [?, ?, ?, ?], yielding enriched methods where
crack tip fields [?], discontinuous derivatives [?] can be incorporated into the
meshfree approximation. A recent overview of meshfree methods may be found
in [?,?]. One of the drawbacks of the meshfree methods lies in the tricks needed to
handle the essential boundary conditions due to the lack of the so-called Kronecker-
Delta property of meshfree approximation functions. Another drawback of the
meshfree methods is that it is difficult and awkward to convert in-house finite
element codes into a meshfree code. However, meshfree methods are very pow-
erful tools that were successfully used to model fracture in concrete, for instance
[?, ?, ?, ?, ?, ?, ?, ?, ?, ?].
Enriched finite element methods such as the XFEM, contrary to meshfree
methods, can be implemented within a finite element code with relatively small
modifications: variable number of degrees of freedom (dofs) per node; mesh ge-
ometry interaction (a procedure to detect elements intersecting with the geometry
of the discontinuities); enriched stiffness matrices; numerical integration.
X X
uh (x) = φI (x)uI + φJ (x)f (x)aJ (11.2.1)
I∈S J∈S enr
| {z } | {z }
standard part enriched part
where S and S enr are the set of nodes and set of enriched nodes, respectively.
The definition of S enr will be defined later. The shape functions φI (x) are finite
element shape functions (or meshless shape functions for the meshless methods).
The parameters uI and aJ are the regular and enrichment nodal parameters, re-
spectively. The enrichment function f (x) depends on the problem under consid-
eration.
Z
int
δW = (∇0 δu)T : PdΩ0
Ω0 /Γ0c
Z Z Z
ext
δW = ρ0 b · δudΩ0 + t̄0 · δudΓ0 + [[δu]] · tc0
Ω0 Γ0t Γ0c
Z
δW kin = ρ0 δu · üdΩ0
Ω0
where
[[u]] := u+ − u−
X
= NJ (X)[[ψJ (X)]]aJ (t)
J∈Nenr (11.3.3)
X
=2 NJ (X)aJ (t)
J∈Nenr
where in the second equality, the continuity of shape functions have been used. In
the final step, the definition of the step function was used.
According to the Bubnov-Galerkin method, the test function δu is also ap-
proximated as the trial function, hence
X X
δu = NI (X)δuI + NJ (X)ψJ (X)δaJ (11.3.4)
I∈N J∈Nenr
X
[[δu]] = 2 NJ (X)δaJ (11.3.5)
J∈Nenr
Z Z Z
ext
δW = ρ0 b·(NI δuI +NJ ψJ δaJ )dΩ0 + t̄0 ·(NI δuI +NJ ψJ δaJ )dΓ0 + 2NJ δaJ ·tc0
Ω0 Γ0t Γ0c
(11.3.7)
Z
int
δW = (∇0 (NI δuI + NI ψI δaI ))T : PdΩ0
Ω /Γ0
Z 0 c Z (11.3.8)
T T
= (∇0 NI ) δuI : PdΩ0 + (∇0 NI ψI ) δaJ : PdΩ0
Ω0 /Γ0c Ω0 /Γ0c
Z
muu
IJ =I ρ0 NI NJ dΩ0
Ω0
Z
mua
IJ =I ρ0 NI (NJ ψJ )dΩ0 (11.3.12)
Ω0
Z
maa
IJ =I ρ0 (NI ψI )(NJ ψJ )dΩ0
Ω0
where I is the identity matrix with dimension equals to the number of spatial
dimensions. The external nodal force is
u,ext a,ext T
Fext
I = fI fI (11.3.13)
with the external nodal force corresponding to the regular dofs given by
Z Z
u,ext
fI = ρ0 b · NI dΩ0 + t̄0 · NI dΓ0 (11.3.14)
Ω0 Γ0t
and the external nodal force associated with the additional dofs given by
Z Z Z
fIa,ext = ρ0 b · (NI ψI )dΩ0 + t̄0 · (NI ψI )dΓ0 + 2 NJ · tc0 (11.3.15)
Ω0 Γ0t Γ0c
and the internal nodal force associated with additional dofs is given by
Z
a,int
fI = ∇T
0 (NI ψI ) · PdΩ0 (11.3.18)
Ω0 /Γ0c
226 Partition of Unity methods Chap. 11
NT = N1 N2 N3 N1 ψ1
(11.3.19)
int
fx1 fy1 Z N1,X N1,Y
fx2 fy2 N2,X N 2,Y
P11 P12
feint =
fx3
= dΩ0 (11.3.20)
fy3 Ω0 /Γ0c
N3,X N3,Y P21 P22
qx1 qy1 N1,X ψ1 N1,Y ψ1
ext
fx1 fy1 Z 0 0
fx2 fy2 0 0
feext = = tx 0 dΓc (11.3.21)
fx3 fy3 0 0 0 ty
Γc
qx1 qy1 N1 N1