finite element methods for non-linear continua · 2019. 5. 7. · an infinitesimal material area...
TRANSCRIPT
ME280B
Finite Element Methods for Non-linearContinua
Panos Papadopoulos
Department of Mechanical Engineering
University of California, Berkeley
2019 edition
Copyright c©2019 by Panos Papadopoulos
Introduction
This is a set of notes written as part of teaching ME280B, a graduate course on non-linear
finite elements in the Department of Mechanical Engineering at the University of California,
Berkeley.
Berkeley, California P. P.
January 2019
i
Contents
1 Review of Continuum Mechanics 1
1.1 Kinematics of Deformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Balance Laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.1 Conservation of Mass . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2.2 Balance of Linear Momentum . . . . . . . . . . . . . . . . . . . . . . 11
1.2.3 Balance of Angular Momentum . . . . . . . . . . . . . . . . . . . . . 13
1.2.4 Balance of Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3 Invariance under Superposed Rigid-body Motions . . . . . . . . . . . . . . . 16
1.4 Closure of the Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2 Consistent Linearization 19
2.1 Sources of Non-linearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1.1 Geometric non-linearity . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1.2 Material non-linearity . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.1.3 Non-linearity in the balance laws . . . . . . . . . . . . . . . . . . . . 20
2.2 Gateaux and Frechet Differentials . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3 Consistent Linearization in Continuum Mechanics . . . . . . . . . . . . . . . 25
3 Incremental Formulations 36
3.1 Strong and Weak Form of the Balance Laws . . . . . . . . . . . . . . . . . . 36
3.2 Lagrangian and Eulerian Methods . . . . . . . . . . . . . . . . . . . . . . . . 43
3.3 Semi-discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.4 Total and Updated Lagrangian Methods . . . . . . . . . . . . . . . . . . . . 48
3.5 Corotational Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.6 Arbitrary Lagrangian-Eulerian Methods . . . . . . . . . . . . . . . . . . . . 62
3.7 Eulerian Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
ii
4 Solution of Non-linear Field Equations 85
4.1 Generalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.2 The Newton-Raphson Method and its Variants . . . . . . . . . . . . . . . . . 90
4.3 The Newton-Raphson Method in Non-linear Finite Elements . . . . . . . . . 96
4.4 Continuation Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4.4.1 Force control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
4.4.2 Displacement control . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
4.4.3 Arc-length control . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.5 Computational Treatment of Constraints . . . . . . . . . . . . . . . . . . . . 126
5 Constitutive Modeling of Deformable Continua 138
5.1 Incompressible Newtonian Viscous Fluid . . . . . . . . . . . . . . . . . . . . 138
5.2 Nonlinear Elasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
iii
List of Figures
1.1 A body B and its subset S . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Reference and current configurations of a body B at times t0 and t, respectively. 2
1.3 Mapping of an infinitesimal material line element dX from the reference to
the current configuration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Mapping of an infinitesimal material surface element dA to its image da in
the current configuration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.5 Interpretation of the right polar decomposition. . . . . . . . . . . . . . . . . . 6
1.6 Angular momentum of an infinitesimal volume element. . . . . . . . . . . . . 14
1.7 Configurations associated with motions χ and χ+ differing by a superposed
rigid-body motion χ+. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.1 Non-linear mapping between Banach spaces . . . . . . . . . . . . . . . . . . . 21
2.2 Configurations of the body for consistent linearization . . . . . . . . . . . . . 25
3.1 Reference and current configuration . . . . . . . . . . . . . . . . . . . . . . . 37
3.2 Partition of boundary ∂R of R . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.3 Tangent vector at a point in the current configuration . . . . . . . . . . . . . 39
3.4 A surface S traversing the configuration R at time t . . . . . . . . . . . . . . 43
3.5 Finite element mesh schematic . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.6 Lagrangian finite element surface . . . . . . . . . . . . . . . . . . . . . . . . 45
3.7 Eulerian finite element mesh with a moving solid or fluid . . . . . . . . . . . 46
3.8 Time discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.9 Configurations at times t0, tn and tn+1 . . . . . . . . . . . . . . . . . . . . . 48
3.10 Partition of element boundary ∂Ωe0 . . . . . . . . . . . . . . . . . . . . . . . 52
3.11 Configurations for updated Lagrangian formulation . . . . . . . . . . . . . . . 57
3.12 Fixed vs. corotational coordinate system . . . . . . . . . . . . . . . . . . . . 61
3.13 General corotational coordinate system . . . . . . . . . . . . . . . . . . . . . 61
iv
3.14 Configurations of ALE method . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.15 Velocity constraint on the boundary of an ALE mesh . . . . . . . . . . . . . 64
3.16 Configurations of ALE method with coincident body and mesh boundaries . . 65
3.17 Mesh region PM in the current configuration . . . . . . . . . . . . . . . . . . 65
3.18 Neumann boundary conditions in ALE methods (R and M are shown apart
from each other for clarity) . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.19 Operator-split approach for ALE implementation (Rn+1, resulting from La-
grangian step andMn+1, resulting from Eulerian remapping are shown apart
from each other for clarity) . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.20 Projection method from one mesh (denoted “-”) to another (denoted “+”) . . 76
3.21 Volumetric and shear distortion indicators for a patch of 4-node quadrilaterals 77
3.22 Smooth map between equilateral triangles in the natural space (ξ, η) and gen-
eral triangles in the physical space (x, y) . . . . . . . . . . . . . . . . . . . . 78
3.23 Smooth map between squares in the natural space (ξ, η) and general quadrilat-
erals in the physical space (x, y) . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.24 Smooth map from a square in the natural space (ξ, η) and a general quadrilat-
eral domain in the physical space (x, y) . . . . . . . . . . . . . . . . . . . . . 80
3.25 Finite difference stencil for linear triangles with unit spacing in the natural
domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.26 Finite difference stencil for bilinear quadrilaterals with unit spacing in the
natural domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.1 Ball of radius δ centered at x0 . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.2 Convex and non-convex sets . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.3 Contraction mapping in S . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.4 One-dimensional geometric interpretation of the Newton-Raphson method . . 90
4.5 Failure of the Newton-Raphson method due to singular J(x(k)) . . . . . . . . 91
4.6 Failure of the Newton-Raphson method due to the large distance of x(k) from
the solution x∗ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.7 One-dimensional geometric interpretation of the modified Newton-Raphson
method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.8 Balls B(x∗; ρ) and B(x∗; δ) around the solution x∗ . . . . . . . . . . . . . . . 94
4.9 Functions with different radii of attractions for the Newton-Raphson method 95
4.10 Convergence under the Newton-Kantorovich conditions . . . . . . . . . . . . 96
v
4.11 Pressure-type follower traction . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.12 Body force in a gravitational field . . . . . . . . . . . . . . . . . . . . . . . . 98
4.13 The Kelvin-Voigt solid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.14 Schematic of line search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.15 Schematic of quasi-Newton method in one-dimension . . . . . . . . . . . . . 109
4.16 Limit and bifurcation points in (u, λ)-space . . . . . . . . . . . . . . . . . . . 114
4.17 Snapping of a deep arch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
4.18 Solution in neighborhood of point (u(s), λ(s)) . . . . . . . . . . . . . . . . . 115
4.19 Relation between the loading vector c and the null eigenvector ψ of KT . . . 119
4.20 Force control in (u, λ)-space . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
4.21 Range of validity of force control in (u, λ)-space . . . . . . . . . . . . . . . . 121
4.22 Displacement control in (ul, λ)-space . . . . . . . . . . . . . . . . . . . . . . 122
4.23 Range of validity of displacement control in (ul, λ)-space . . . . . . . . . . . 122
4.24 Arc-length in (u, τλ)-space . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.25 Spherical arc-length in (u, τλ)-space . . . . . . . . . . . . . . . . . . . . . . . 123
4.26 Normal-plane arc-length in (u, τλ)-space . . . . . . . . . . . . . . . . . . . . 124
4.27 Contact of deformable solid with rigid foundation . . . . . . . . . . . . . . . 127
4.28 Penalty functional and the enforcement of the constraint c = 0 . . . . . . . . 134
vi
Chapter 1
Review of Continuum Mechanics
1.1 Kinematics of Deformation
Define a body B as a collection of material points and identify a typical material point
with P , as in Figure 1.1.
S
B
P
Figure 1.1. A body B and its subset S .
At some time t, define the configuration of B as the region in the three-dimensional Eu-
clidean point space E3 occupied by B, and denote this configuration R and its boundary ∂R.The material point P in the configuration of B at time t is identified with a vector x in
three-dimensional Euclidean vector space E3 drawn from any fixed origin o, see Figure 1.2.
Let x be resolved with respect to a fixed orthonormal basis ei as
x = xiei ; xi = x · ei , (1.1)
where “·” denotes dot-product between two vectors and the Einsteinian summation conven-
tion in enforced on repeated indices. The configuration of B at time t will be referred to as
the current or present configuration.
1
2 Review of Continuum Mechanics
RR0
x
X
O
oEAei
Figure 1.2. Reference and current configurations of a body B at times t0 and t, respectively.
Let the same body B occupy at some other time t = t0 (often set, without loss of
generality, to t0 = 0) the region R0 in E3 with boundary ∂R0. The material point P at time
t0 is identified with another vector X in E3 drawn, in general, from another fixed origin O,
see Figure 1.2. Let X be now resolved with respect to a fixed orthonormal basis EA as
X = XAEA ; XA = X · EA . (1.2)
The configuration of B at time t0 will be referred to as the reference configuration.
The motion χ : E3 × R→ E3 is a mapping of points X in the reference configuration to
points in the current configuration x at given time t, that is,
x = χ (X, t) ; xi = χi (XA, t) . (1.3)
The velocity and acceleration fields are respectively defined at time t as
v = χ (X, t) =∂χ (X, t)
∂t; vi = χi =
∂χi (XA, t)
∂t(1.4)
and
a = χ (X, t) =∂2χ (X, t)
∂t2; ai = χi =
∂2χi (XA, t)
∂t2. (1.5)
where ˙( ) denotes the material time derivative, that is, the time derivative of a quantity for
a fixed material point.
The fundamental measure of relative deformation is the deformation gradient tensor F
which may be viewed as a linear mapping of an infinitesimal material line element dX of the
reference configuration to an infinitesimal line element dx in the current configuration, that
is,
dx = FdX ; dxi = FiAdXA , (1.6)
ME280B May 7, 2019
Kinematics of Deformation 3
x
dx
X
dX
RR0
Figure 1.3. Mapping of an infinitesimal material line element dX from the reference to the
current configuration.
see Figure 1.3. Assuming that there is a motion χ, where
χ : (X, t)→ x = χ(X, t) , (1.7)
one readily concludes from (1.6) that
F =∂χ
∂X. (1.8)
The deformation gradient is a two-point tensor, in the sense that
F = FiAei ⊗ EA , (1.9)
where “⊗” denotes tensor product of two vectors. As seen from (1.8), the deformation
gradient F has one “leg” (the first) in the current configuration and the other “leg” in the
reference configuration.
An infinitesimal material volume element dV in the reference configuration is likewise
mapped to an infinitesimal volume element dv in the current configuration, such that
dv = JdV , (1.10)
where J = detF is the Jacobian of the motion. Appealing to the inverse function theorem of
calculus, it can be established that a continuously differentiable (in X) motion χ is invertible
at a given time t if, and only if,
detF = det∂χ
∂X6= 0 , (1.11)
for all X ∈ R0. Upon assuming, on physical grounds, that the sign of a given material
volume should be always preserved, it can be concluded from (1.10) that invertibility of χ
necessitates that J > 0.
May 7, 2019 ME280B
4 Review of Continuum Mechanics
An infinitesimal material area element dA in the reference configuration is mapped to
an infinitesimal area element da in the current configuration according to Nanson’s formula,
which states that
n da = JF−TN dA , (1.12)
where N (resp. n) is a unit vector normal to the surface dA (resp. da), see Figure 1.4.
xX
R0 R
dX1
dX2
dx1
dx2
N n
Figure 1.4. Mapping of an infinitesimal material surface element dA to its image da in the
current configuration.
Other useful measures of deformation are the right and left Cauchy-Green deformation
tensors C and B defined, respectively, as
C = FTF ; CAB = FiAFiB (1.13)
and
B = FFT ; Bij = FiAFjA . (1.14)
It is clear from the preceding component representations that C is a referential tensor,
while B is a spatial tensor. Both tensors are symmetric, which means that their components
form can be represented in the form of symmetric matrices.
Let M and m be unit vectors in the direction of dX and dx, respectively, such that
dX = MdS , dx = mds , (1.15)
and define the ratio
λ =ds
dS(1.16)
as the stretch of the material line element dX. It can be easily established with the aid
of (1.8), (1.15), and (1.16) that
λm = FM (1.17)
ME280B May 7, 2019
Kinematics of Deformation 5
and, upon also using (1.13), that
λ2 = M ·CM ; λ2 = MACABMB . (1.18)
Likewise, (1.17) can be alternatively expressed as
1
λM = F−1m , (1.19)
hence, with the aid of (1.14), it follows that
1
λ2= m ·B−1m ;
1
λ2= miB
−1ij mj . (1.20)
Two important measures of strain can be obtained by resolving the difference ds2 − dS2
as
ds2 − dS2 = dx · dx− dX · dX= (FdX) · (FdX)− dX · dX= dX · (CdX)− dX · dX = dX · (C− I) dX
= dX · 2EdX (1.21)
or, alternatively,
ds2 − dS2 = dx · dx− dX · dX= dx · dx−
(F−1dx
)·(F−1dx
)
= dx · dx− dx ·B−1dx = dx ·(i−B−1
)dx
= dx · 2edx , (1.22)
where I and i are the referential and spatial identity tensors, respectively. These measures
are the (relative) Lagrangian strain tensor E, given with the aid of (1.21) by
E =1
2(C− I) ; EAB =
1
2(CAB − δAB) . (1.23)
and the (relative) Eulerian (or Almansi) strain tensor e, defined from (1.22) as
e =1
2
(i−B−1
); eij =
1
2
(δij − B−1
ij
). (1.24)
The polar decomposition theorem states that F, being non-singular by virtue of J =
detF > 0, can be uniquely resolved into
F = RU = VR ; FiA = RiBUBA = VijRjA . (1.25)
May 7, 2019 ME280B
6 Review of Continuum Mechanics
Clearly, R is a two-point tensor, U is a referential tensor, and V is a spatial tensor.
More specifically, R is a proper-orthogonal (or indexrotation—seealsotensor!proper orthog-
onalrotation) tensor, that is, one which satisfies RTR = I, RRT = i, and detR = +1. In
addition, U and V are symmetric positive-definite tensors referred to as the right and left
stretch tensors, respectively. In light of these definitions, equations (1.25)1,2 are correspond-
ingly referred to as the right and left polar decompositions of F.
Starting from (1.17) it can be seen that
dx = R (UdX) , (1.26)
which implies that dX is first stretched and rotated byU, and is subsequently further rotated
by R, see Figure 1.5. A corresponding interpretation may be readily formulated for the left
dxdX dX′
U R
Figure 1.5. Interpretation of the right polar decomposition.
polar decomposition.
Along its principal directions, U effects stretch without any rotation, which implies that,
if dX lies along such a principal direction, then
UdX = λdX (1.27)
or, upon recalling (1.15)1,
UM = λM . (1.28)
With reference to the principal stretches λI , I = 1, 2, 3 and the associated principal direc-
tions MI , I = 1, 2, 31 of the preceding linear eigenvalue problem, one may use the spectral
representation theorem to write
U =3∑
I=1
λIMI ⊗MI , (1.29)
1The set of eigenvectors MI may be taken to be orthonormal without loss of generality.
ME280B May 7, 2019
Kinematics of Deformation 7
from which it follows that
F = RU =3∑
I=1
λI (RMI)⊗MI (1.30)
and also, given the definition of C and the orthogonality of R, that
C = FTF = (RU)T (RU) = U2 =3∑
I=1
λ2IMI ⊗MI . (1.31)
Generalized referential measures of deformation can be introduced in the form
Cm =3∑
I=1
λ2mI MI ⊗MI m 6= 0 (1.32)
logC =3∑
I=1
log λ2IMI ⊗MI (Hencky strain) (1.33)
with associated measures of generalized referential strain
E(m) =
1m
(Cm/2 − I
)if m 6= 0
12lnC if m = 0
. (1.34)
Note that m = 2 in (1.34) recovers the Lagrangian strain tensor.
It is readily concluded from (1.25) and (1.29) that
V =3∑
I=1
λI (RMI)⊗ (RMI) (1.35)
and also, given the definition of B and the orthogonality of R, that
B = FFT = (VR) (VR)T = V2 =3∑
I=1
λ2I (RMI)⊗ (RMI) . (1.36)
There exist several ways to extract the polar factors of F. In particular, focusing attention
to the right polar decomposition, one may:
(a) Extract the stretch U directly from F in closed form2, and subsequently obtain R as
R = FU−1 . (1.37)
2A. Hoger and D.E. Carlson, “Determination of the stretch and rotation in the polar decomposition of
the deformation gradient”, Q. Appl. Math., 42:113–117, (1984).
May 7, 2019 ME280B
8 Review of Continuum Mechanics
(b) Extract the rotation R directly from F in closed form3, and subsequently obtain U as
U = RTF . (1.38)
(c) Use a numerical method to solve the matrix equation
[CAB] = [FiAFiB] = [UACUCB] (1.39)
for the components [UAB] of U to within the required tolerance, then evaluate R as
in (a). A computationally efficient algorithm for solving this matrix equation employs
the Newton-Raphson method, according to which
U(i+1) =1
2
(U(i) +CU(i)−1
), (1.40)
with initial guess
U(0) = I (or mI, m > 0) . (1.41)
It can be shown that with this choice of U(0), all U(i) are positive-definite and also
U(i) converges to U at a quadratic rate4
(d) Solve the linear eigenvalue problem (1.31) for the eigenpairs (λ2I ,MI), form U us-
ing (1.29), and then calculate R as R = FU−1.
Lastly, attention is focused on mixed space-time derivatives of the motion, which give
rise to measures of deformation rate. Invoking the chain rule, write
F =d
dt
∂χ
X=
∂
∂X
dχ
dt=
∂v
∂X=
∂v
∂x
∂x
∂X= LF . (1.42)
In (1.42), L is the spatial velocity gradient, defined as
L =∂v
∂x; Lij =
∂vi∂xj
(1.43)
The tensor L may be uniquely decomposed into a symmetric tensor D, termed the rate-
of-deformation tensor, and a skew-symmetric tensor W, termed the vorticity or indexspin
tensor—seealsovorticity tensorspin tensor, such that
L = D+W . (1.44)
3P. Papadopoulos and J. Lu, “On the direct determination of the rotation tensor from the deformation
gradient”, Math. Mech. Solids., 2:17–26, (1997).4N.J. Higham, “Newton’s method for the matrix square root”, Math. Comp., 46:537–549, (1986).
ME280B May 7, 2019
Balance Laws 9
The physical meaning of D can be established by observing that (1.17), (1.42), and (1.44)
imply that
˙lnλ =λ
λ= m ·Dm . (1.45)
Regarding W, first note that owing to its skew-symmetry, it can be put into a one-to-one
correspondence with a vector w of E3, such that for any vector z in E3,
Wz = w × z . (1.46)
The vector w is called the axial vector of W. Then, considering a unit vector m along a
principal direction of D, it can be established with the aid of (1.17), (1.42), (1.44), and (1.45)
that
m = Wm = w ×m . (1.47)
This shows that the axial vector of W can be thought of as the angular velocity of a line
element positioned along a principal direction of D.
Using (1.42), as well as the earlier definitions of the deformation and strain tensors, it
can be readily concluded that
C =˙
FTF = 2FTDF ; CAB = 2FiADijFjB , (1.48)
E = FTDF ; EAB = FiADijFjB , (1.49)
and also
B =˙
FFT = LB+BLT ; Bij = LikBkj + BikLjk , (1.50)
e =1
2
(B−1L+ LTB−1
); eij =
1
2
(B−1
ik Lkj + LkiB−1kj
). (1.51)
It is also easy to show that
dJ
dt= J div v = J trL . (1.52)
1.2 Balance Laws
The underlying physical principal of continuum mechanics are the conservation of mass,
the balance of linear and angular momentum, and the balance of energy. There are briefly
reviewed in this section.
May 7, 2019 ME280B
10 Review of Continuum Mechanics
Preliminary to the introduction of the balance laws, three important results from calculus
and real analysis are stated without proof. The first is the divergence theorem, according
to which, given a continuously differentiable real-valued function φ = φ(x, t) : P × R → R
defined at time t on some subset P of E3 with boundary ∂P , then∫
P
gradφ dv =
∫
∂P
φn da , (1.53)
where gradφ is the gradient of the function φ, defined as
gradφ =∂φ
∂xiei . (1.54)
Versions of this theorem may be easily deduced from the above for vector and tensor func-
tions. The second result is known as the Reynolds’ transport theorem, and states that if φ is
differentiable in both of its variables, then
d
dt
∫
P
φ dv =
∫
P
(φ+ φ div v) dv , (1.55)
where
φ =∂φ
∂t+∂φ
∂x· v . (1.56)
Using 1.56 and the divergence theorem, it is easy to show that (1.55) may be recast as
d
dt
∫
P
φ dv =
∫
P
[∂φ
∂t+ div(φv)
]dv =
∫
P
∂φ
∂tdv +
∫
∂P
φv · n da . (1.57)
Lastly, the localization theorem states that if φ is continuous in x and∫
P
φ dv = 0 , (1.58)
for all subsets P of R at a given time t, then φ = 0 everywhere in R at time t.
Extensive use of these three important results will be made henceforth.
1.2.1 Conservation of Mass
Start by axiomatically admitting the existence of a measure of mass m(S ) for every subset
S of the body B. Upon assuming certain regularity conditions for this measure, it can be
established that there exists a mass density ρ(x, t) such that∫
S
dm =
∫
P
ρ dv (1.59)
ME280B May 7, 2019
Balance Laws 11
for the region P ⊂ R that occupies the part S at time t.
Likewise, there exists a mass density function ρ0(X) at time t0, such that
∫
S
dm =
∫
P0
ρ0 dV (1.60)
for the region P0 ⊂ R0 that occupies the part S at time t0. The preceding two statements
reflect the conservation of mass between the configurations occupied by the body at times t
and t0. Equating the right-hand sides of these statements and recalling (1.10) leads to the
integral statement ∫
P0
ρ0 dV =
∫
P
ρ dv =
∫
P0
ρJ dV (1.61)
or ∫
P0
(ρ0 − ρJ) dV = 0 . (1.62)
Invoking the localization theorem, this implies that
ρ0 = ρJ . (1.63)
This is a local statement of mass conservation between the two configurations.
Alternatively, conservation of mass may be stated in integral form as
d
dt
∫
P
ρ dv = 0 . (1.64)
Applying Reynolds’ transport theorem in the form of (1.55) implies that
∫
P
(ρ+ ρ div v) dv = 0 . (1.65)
Appealing to the localization theorem, this leads to a local statement of mass conservation
in the form
ρ+ ρ div v = 0 . (1.66)
1.2.2 Balance of Linear Momentum
The linear momentum of the material which occupies the region P at time t is∫Pρv dv.
Admit the existence of two types of external forces, both of which quantify the interaction
of P with its surrounding environment. Specifically, consider: (a) body forces per unit mass
b = b(X, t) acting on the material particles the region S which occupies P at time t and (b)
contact forces per unit area t = t(x, t;n) acting on the boundary surface ∂P and depending
May 7, 2019 ME280B
12 Review of Continuum Mechanics
on the orientation n of the surface. The integral statement of linear momentum balance for
the continuum occupying P at time t takes the form
d
dt
∫
P
ρv dv =
∫
P
ρb dv +
∫
∂P
t da . (1.67)
Using linear momentum balance in the preceding primitive form it can be shown that
t(x, t;n) = −t(x, t;−n) . (1.68)
This result is known as Cauchy’s lemma, and asserts that contact forces acting at a point x
on opposite sides of the same surface are equal and opposite.
Applying balance of linear momentum to a right-angled tetrahedral region (the Cauchy
tetrahedron) in conjunction with Cauchy’s lemma, it can be shown that there exists a ten-
sor T, appropriately termed the Cauchy stress tensor, such that
t = Tn ; ti = Tijnj . (1.69)
Upon using mass conservation, (1.68), and the divergence theorem, it can be concluded
from the linear momentum balance statement (1.67) that
d
dt
∫
P
ρv dv =
∫
P
ρa dv
=
∫
P
ρb dv +
∫
P
Tn da
=
∫
P
ρb dv +
∫
P
divT dv . (1.70)
Upon invoking the localization theorem, this results in
divT+ ρb = ρa ; Tij,j + ρbi = ρai . (1.71)
The preceding statement of linear momentum balance can be also resolved on the geometry
of the reference configuration, where, upon again using mass conservation, Cauchy’s lemma
ME280B May 7, 2019
Balance Laws 13
and the divergence theorem, in conjunction with Nanson’s formula (1.12), one finds that
d
dt
∫
P
ρv dv =d
dt
∫
P0
ρ0v dV
=
∫
P
ρb dv +
∫
∂P
Tn da
=
∫
P0
ρ0b dV +
∫
∂P0
TJF−TN dA
=
∫
P0
ρ0b dV +
∫
∂P0
PN dA
=
∫
P0
ρ0b dV +
∫
∂P0
DivP dV , (1.72)
where P, defined as
P = JTF−T ; PiA = JTijF−1Aj , (1.73)
is the first Piola-Kirchhoff stress tensor. Upon again invoking the localization theorem, it
follows that the local statement of linear momentum balance may be expressed in referential
form as
DivP+ ρ0b = ρ0a ; PiA,A + ρ0bi = ρ0ai . (1.74)
Recalling (1.76), one may define the traction p resolved on the geometry of the reference
configuration as
p = PN ; pi = PiANA . (1.75)
where the (differential) force df across a surface is given by
df = tda = pdA . (1.76)
1.2.3 Balance of Angular Momentum
The angular momentum of the material which occupies the region P at time t is∫Px×ρv dv,
see Figure 1.6.
The integral statement of angular momentum balance for the continuum occupying P at
time t isd
dt
∫
P
x× ρv dv =
∫
P
x× ρb dv +∫
∂P
x× t da . (1.77)
Upon invoking mass conservation and linear momentum balance, as well as using the diver-
gence and localization theorems, it follows from the above that
T = TT ; Tij = Tji , (1.78)
May 7, 2019 ME280B
14 Review of Continuum Mechanics
O
x
v
P
Figure 1.6. Angular momentum of an infinitesimal volume element.
namely angular momentum balance requires that T be symmetric. An equivalent local
statement of angular momentum balance may be formulated in terms of the first Piola-
Kirchhoff stress. Indeed, upon using (1.73), this takes the form
PFT = FPT ; PiAFjA = FiAPjA . (1.79)
1.2.4 Balance of Energy
First, consider the mechanical energy balance theorem, which is a direct consequence of the
preceding three balance laws. According to it,
d
dt
∫
P
1
2ρv · v dv +
∫
P
T ·D dv =
∫
P
ρb · v dv +∫
∂P
t · v da , (1.80)
where
K (P) =
∫
P
1
2ρv · v dv (1.81)
is the kinetic energy of the continuum in P ,
S (P) =
∫
P
T ·D dv (1.82)
is the stress power , and
R(P) =
∫
P
ρb · v dv +∫
∂P
t · v da (1.83)
is the rate at which the external forces do work in P . In words, the theorem asserts that the
rate of change of the kinetic energy and the stress power of the material in P are equal to
the rate at which work is done by the external forces acting on R.Note that the stress power can be expressed as
∫
P
T ·D dv =
∫
P0
P · F dV =
∫
P0
S · E dV , (1.84)
ME280B May 7, 2019
Invariance under Superposed Rigid-body Motions 15
where S is the second Piola-Kirchhoff stress tensor. Given (1.42) and (1.73), it can be easily
concluded from (1.84) that
S = F−1P = JF−1TF−T ; SAB = F−1Ai PiB = JF−1
Ai TijF−1Bj (1.85)
or
T =1
JPFT =
1
JFSFT ; Tij =
1
JPiAFiA =
1
JFiASABFjB . (1.86)
It is immediately concluded from the above that S is a symmetric tensor.
Next, admit the existence of a scalar heat supply r = r(x, t) per unit mass, which quan-
tifies the rate at which heat is supplied (or absorbed) by the body. Also, introduce a scalar
heat flux h = h(x, t;n) per unit area across a surface ∂P with outward unit normal n. In
addition, assume that there exists a scalar function ε = ε(x, t) per unit mass, called the
internal energy which quantifies all forms of energy stored in the body other than kinetic
energy.
Now the principle of energy balance is postulated in the form
d
dt
∫
P
[1
2ρv · v + ρε
]dv =
∫
P
ρb · v dv +∫
∂P
t · v da+∫
P
ρr dv −∫
∂P
h da . (1.87)
This is a statement of the first law of thermodynamics. Using a standard procedure akin to
the one used to deduce (1.68), it can be established that
h(x, t;n) = −h(x, t;−n) . (1.88)
Likewise, using (1.88), the Cauchy tetrahedron argument can be repeated for the balance of
energy to show that there exists a heat flux vector q, such that
h = q · n . (1.89)
Now, using the divergence and localization theorems, along with the previously deduced
local forms of mass, linear momentum, and angular momentum balance, a local statement
of energy balance can be obtained from (1.87) as
ρε = ρr − div q+T ·D ; ρε = ρr − qi,i + TijDij . (1.90)
An equivalent referential local statement of energy balance may be readily deduced from (1.90)
as
ρ0ǫ = ρ0r −Div q0 +P · F ; ρ0ε = ρ0r − q0A,A + PiAFiA , (1.91)
where q0 = JF−1q.
May 7, 2019 ME280B
16 1.3. INVARIANCE UNDER SUPERPOSED RIGID-BODY MOTIONS
1.3 Invariance under Superposed Rigid-body Motions
Consider two motions χ and χ+ which map points of the same reference configuration to
points of two current configuration which at time t differ from each other by a superposed
rigid motion χ+, as in Figure 1.7.
x
x+
X
y
y+
Y
R0
R
R+
χ
χ+
χ+
Figure 1.7. Configurations associated with motions χ and χ+ differing by a superposed rigid-
body motion χ+.
It can be shown using a standard procedure that
x+ = χ+(x, t) = Q(t)x+ c(t) (1.92)
where Q is a proper orthogonal tensor and c a vector, both functions of time t only.
It follows easily from chain rule that
F+ =∂χ+
∂X=
∂χ+
∂x
∂x
∂X= QF , (1.93)
This can be used to establish the relations
R+ = QR , U+ = U , V+ = QVQT (1.94)
and also
C+ = C , E+ = E , B+ = QBQT , e+ = QeQT . (1.95)
In addition, it is easy to show that
n+ = Qn , m+ = Qm , ds+ = ds , da+ = da , dv+ = dv . (1.96)
ME280B May 7, 2019
Closure of the Theory 17
Also, taking the material time derivative of (1.92), it follows that
v+ = x+ = Qx+Qv + c
= ΩQx+Qv + c(QQT = Ω
)
= Ω(x+ − c) +Qv + c = ω × (x+ − c) +Qv + c (1.97)
Using the above, it can be readily establish that
L+ = QLQT +Ω , D+ = QDQT , W+ = QWQT +Ω . (1.98)
A physically plausible assumption is made about the transformation of the stress vector t
under a superposed rigid-body motion. Indeed, recalling that
t = Tn, (1.99)
namely that t is linear in n, assume that
t+ = Qt (1.100)
which immediately implies that
T+ = QTQT . (1.101)
From the above, it follows easily that
P+ = QP , S+ = S . (1.102)
Invariance under superposed rigid-body motions is a physically motivated postulate in
continuum mechanics, according to which the stress response remains unaltered when a rigid
motion is superposed onto the actual motion of interest. Following this postulate, assume
that
T = T(F, F, · · · ,GradF, · · · , ρ, · · ·
)(1.103)
then invariance under superposed rigid-body motion implies
T+ = T(F+, F+, · · · ,GradF+, · · · , ρ+, · · ·
), (1.104)
namely T remains unaltered by the superposed rigid motion.
Regarding invariance, it is also postulated that
r+ = r , ε+ = ε , q+ = Qq . (1.105)
The last relation implies
h+ = h . (1.106)
May 7, 2019 ME280B
18 1.4. CLOSURE OF THE THEORY
1.4 Closure of the Theory
It is instructive to summarize the governing equations of continuum mechanics in the context
of a thermomechanical theory. There are a total of 8 equations stemming from the balance
laws, that is,
1. 1 equation from balance of mass, see (1.63) or (1.66),
2. 3 equations from balance of linear momentum, see (1.71) or (1.74),
3. 3 equations from balance of angular momentum, see (1.78) or (1.79)
4. 1 equation from balance of energy, see (1.90) or (1.91).
The unknowns of the problem are the position x (or velocity v), the mass density ρ, the
stress T, the heat flux q, and the temperature θ, which is a total of 3+1+9+3+1=17.
Closure is provided in the theory by constitutive laws for the stress (6 unknowns) and
heat flux (3 unknowns).
ME280B May 7, 2019
Chapter 2
Consistent Linearization
2.1 Sources of Non-linearity
The consistent linearization of the kinematic and kinetic variables, as well as of the balance
laws themselves is of importance in two ways. First, in deriving “linearized” theories con-
sistent with general non-linear theories of continuum mechanics. For instance, consistent
linearization may be used to derive the classical theory of linear elasticity from non-linear
elasticity or a theory of infinitesimal deformations superposed on pre-existing finite deforma-
tions (“small-on-large” theories). The latter has an important application spectral analysis
of non-linear systems at a given time.
A second application of consistent linearization, which is especially relevant to finite ele-
ment analysis, is in obtaining the solution to boundary- and initial- value problem of contin-
uum mechanics using iterative methods methods that rely on instantaneous approximation
of the non-linear system by a linear counterpart.
There are three distinct sources of non-linearity when solving problems in continuum
mechanics. Each of them is briefly discussed here.
2.1.1 Geometric non-linearity
The measures of deformation may be non-linear functions of x (or v) and its gradients. For
instance, consider the (relative) Lagrangian strain tensor E defined in (1.23), and recall that
its components EAB may be expressed in terms of the components xi of the position vector
x as
EAB =1
2
(∂xi∂XA
∂xi∂XB
− δAB
). (2.1)
19
20 Consistent Linearization
This is clearly a non-linear function of xi through its derivative with respect to the referential
position X. This form of non-linearity is referred to as geometric.
Geometric non-linearity can be present even when the strains are “infinitesimal” (that
is, when they can be accurately approximated as linear in x). This could be the case when
the rotations are large.
2.1.2 Material non-linearity
The constitutive equation of the stress or heat flux may be a non-linear function of its
arguments (which are themselves generally functions of the unknowns of the problem). This
form of non-linearity is referred to as material or physical.
Material non-linearity would exist in a general isotropic non-linear elastic solid, for which
the constitutive law is of the form
T = α0i+ α1B+ α2B2 , (2.2)
where α0, α1, and α2 are functions of the scalar invariants of B. Likewise, material non-
linearity exists in a general Reiner-Rivlin fluid, for which
T = β0i+ β1D+ β2D2 , (2.3)
where β1, β2, and β3 are functions of the scalar invariants of D.
2.1.3 Non-linearity in the balance laws
The equation of linear momentum balance can exhibit non-linearity due to the convective
term in the acceleration term. Indeed, recalling (1.71) and also taking into account that
v = v(x, t), it follows that
divT+ ρb = ρv = ρ
(∂v
∂t+∂v
∂x· v)
(2.4)
The second term on the right-hand side of (2.4) is clearly non-linear in v.
Note that this form of non-linearity vanishes when linear momentum balance is written
in the referential form of (1.74), since, in this case v = v(X, t) and
DivP+ ρ0b = ρ0v = ρ0∂v
∂t. (2.5)
ME280B May 7, 2019
Gateaux and Frechet Differentials 21
2.2 Gateaux and Frechet Differentials
In this section, the Gateaux and Frechet differentials are introduced in the context of a non-
linear operator F mapping elements from a Banach space U to another Banach space V ,
as in Figure 2.1.1 As a special case, one may U = V = Rn, where R
n is equipped with the
U
U0V
F
Figure 2.1. Non-linear mapping between Banach spaces
usual Euclidean norm.2
Let F : U0 ⊂ U → V be a non-linear mapping, where U0 is an open set, and take
x ∈ U0, v ∈ U0 and ω ∈ R. The Gateaux differential DGF (x, v) of the operator F at x in
the direction v is formally defined as
DGF (x, v) = limω→0
F (x+ ωv)− F (x)ω
. (2.6)
Equivalently, the Gateaux differential may be defined as
DGF (x, v) =
[d
dωF (x+ ωv)
]
ω=0
, (2.7)
since[d
dωF (x+ ωv)
]
ω=0
=
[lim
∆ω→0
F (x+ ω +∆ωv)− F (x+ ωv)
∆ω
]
ω=0
= lim∆ω→0
[F (x+ ωv +∆ωv)− F (x+ ωv)
∆ω
]
ω=0
= lim∆ω→0
F (x+∆ωv)− F (x)∆ω
= DGF (x, v) ,
(2.8)
where continuity of F is invoked for fixed x and v.
1The treatment in this section follows closely the presentation in M.M. Vainberg, Variational Methods
for the Study of Non-linear Operators(English translation from Russian), Hold-Day, San Francisco, 1964.2It is easy to show that all finite-dimensional normed vector spaces are complete.
May 7, 2019 ME280B
22 Consistent Linearization
The Gateaux differential DGF (x, v) is not necessarily linear in v. However, DGF (x, v) is
necessarily homogeneous in v, that is, for any α ∈ R, α 6= 0,
DGF (x, αv) = limω→0
F (x+ αωv)− F (x)ω
= α limαω→0
F (x+ αωv)− F (x)αω
= αDGF (x, v) .
(2.9)
It can be shown that if:
(a) DGF (x, v) exists in some open neighborhood U0 of x0 and is continuous in x at x = x0,
(b) DGF (x0, v) is continuous in v at v = 0,
then the Gateaux differential DGF (x0, v) is actually linear in v and one may write
DGF (x0, v) = [DGF (x0)] (v) , (2.10)
where the function DGF (x0) is called the Gateaux derivative of F at x = x0. The Gateaux
derivative is a linear operator on the Banach space U , since for a given x = x0 it maps
v ∈ U to [DGF (x0)] (v) ∈ V , that is,
DGF (x0) : v → [DGF (x0)] (v) . (2.11)
Example 2.2.1: Gateaux differential and derivative of a function in R
Let U = V = R and define F according to F (x) = x2. Then, use the definition (2.6) to write
DGF (x, v) = limω→0
F (x+ ωv)− F (x)ω
= limω→0
(x+ ωv)2 − x2ω
= limω→0
[2xv + ωv2
]= 2xv .
Clearly, in this case the Gateaux differential is linear in v, therefore the Gateaux derivative DGF (x) isa linear operator defined as
DGF (x) : v ∈ R → [DGF (x)] (v) = 2xv , (2.12)
for any given x ∈ R.
Note that the differential is not necessarily “infinitesimal” in magnitude. For a given x,
this depends on the magnitude of v.
Let F be again a non-linear mapping from the Banach space U to the Banach space V .
Also, assume that at a given point x ∈ U ,
F (x+ v) = F (x) +DFF (x, v) +O(x, v) , (2.13)
ME280B May 7, 2019
Gateaux and Frechet Differentials 23
for all v ∈ U , where DFF (x, v) is a linear mapping in v ∈ U and
lim‖v‖U→0
‖O(x, v)‖V‖v‖U
= 0 . (2.14)
Then, DFF (x, v) is called the Frechet differential of F at x and O(x, v) is the remainder
of F at x. The Frechet derivative DFF (x0) at a given point x = x0 is defined as the linear
operator on U which takes any v ∈ U and maps it to [DFF (x0)] (v) = DFF (x0, v), that is,
DFF (x0) : u ∈ U → [DFF (x0)] (v) . (2.15)
In view of (2.13) and the properties of its constituent parts, one may define the linear part
of F at x as
L[F ; v]x = F (x) +DFF (x, v) . (2.16)
A corollary of the above definition and the property of the remainder in (2.14) is that if
v 6= 0 is fixed and α > 0, then
lim‖αv‖U→0
‖O(x, αv)‖V‖αv‖U
= 0 (2.17)
implies
limα→0
‖O(x, αv)‖Vα
= 0 . (2.18)
Note that DFF (x, v) is unique as defined in (2.13). To argue this, let, by contradiction,
F (x+ v) = F (x) +DFF (x, v) +O(x, v) = F (x) +D′FF (x, v) +O′(x, v) . (2.19)
Then, clearly,
DFF (x, v)−D′FF (x, v) = O′(x, v)−O(x, v) . (2.20)
If one now replaces in the above v with αv, α 6= 0, it follows that
α [DFF (x, v)−D′FF (x, v)] = O′(x, αv)−O(x, αv) (2.21)
or
DFF (x, v)−D′FF (x, v) =
O′(x, αv)−O(x, αv)α
(2.22)
Now, taking norms of both sides and then considering the limit α→ 0 and recalling (2.18),
it follows that D′FF (x, v) = DFF (x, v).
May 7, 2019 ME280B
24 Consistent Linearization
Example 2.2.2: Frechet differential and derivative of a function in R
Take, again, F (x) = x2 and write
F (x+ v) = (x+ v)2 = x2 + 2xv + v2 = F (x) +DFF (x, v) +O(x, v) ,
where, clearly,
DFF (x, v) = 2xv ,
which is linear in v, and also
O(x, v) = v2 ,
such that
lim|v|→0
|v|2|v| = 0 .
Thus, the Frechet differential of F at x is given in this case by
DFF (x, v) = 2xv
and the Frechet derivative DFF (x) is
DFF (x) : v ∈ R → [DFF (x)] (v) = 2xv .
From the definition of the Gateaux and Frechet derivatives, it follows that if the Frechet
derivative of F exists at a point x = x0, then so does the Gateaux derivative and, furthermore,
the two derivatives coincide. Indeed, if the Frechet derivative exists at x0, then, by definition,
one may write
F (x0 + ωv) = F (x0) + [DFF (x0)] (ωv) +O(x, ωv) , (2.23)
therefore,
limω→0
F (x0 + ωv)− F (x0)ω
= [DGF (x0)] (v) + limω→0
O(x, ωv)
ω= [DGF (x0)] (v) , (2.24)
in view of the property (2.18) of the remainder. The inverse of the above assertion does not
hold, namely the Gateaux derivative is weaker than the Frechet derivative. However, it can
be shown that if DGF (x) exists and is continuous at x, then DFF (x) exists and coincides
with DGF (x).
Note that continuity of an operator such as DGF (x) at a point x = x0 is defined using
the natural norm induced with the mapping F , namely DGF (x) is continuous at x = x0 if
for every ε > 0 there is a δ = δ(x0, ε) such that
‖DGF (x)−DGF (x0)‖UV < ε , (2.25)
ME280B May 7, 2019
Consistent Linearization in Continuum Mechanics 25
whenever
‖x− x0‖U < δ . (2.26)
Here, the natural norm is defined as
‖DGF (x)‖UV = supy∈U
‖DG(x, y)‖V‖y‖U
, (2.27)
where “sup” denotes least upper bound.
2.3 Consistent Linearization in Continuum Mechanics
In this section, linearization will be considered for quantities defined in the current configu-
ration R with respect to another configuration of a body associated with a region R.To this end, start with the region R0 occupied by the body in the reference configuration,
and define a motion χ, such that
χ : (X, t) → x = χ(X, t) = χt(X) , (2.28)
so that at time t the body occupies the region R. In addition, define another motion χ as
in (1.7), that is,
χ : (X, t) → x = χ(X, t) = χt(X) , (2.29)
where the body ends up in the region R at time t, see Figure 2.2. Both motions are assumed
EA
e1
R
R0
R
x
x
ξ
χt
χtχδ
t
Figure 2.2. Configurations of the body for consistent linearization
invertible at any fixed t, that is both mappings χt and χt are invertible.
May 7, 2019 ME280B
26 Consistent Linearization
One may now write
x = χt(X) = χt(χ−1t (x)) = χδ
t (x) (2.30)
where χδt is a motion superposed on R, see, again, Figure 2.2. Symbolically, one may write
χt = χδt χt , (2.31)
where “” denotes functional composition, hence
(χδ
t χt
)(X) = χδ
t (χt(X)) = χδt (x) = x = χ−1
t (X) (2.32)
Next, define a vector field ξ = ξ(X, t), such that
ξ = x− x = χ(X, t)− χ(X, t) ; ξ = ξiei . (2.33)
The vector ξ is called the tangent vector field on R.The consistent linearization of kinematic and kinetic fields, as well as balance laws with re-
spect to the configuration associated with the region R is achieved by application of Gateaux
differentiation to each field (or balance law) atR in the direction of the tangent vector field ξ.
In what follows, all involved functions will be assumed as smooth as necessary to make their
Gateaux and Frechet derivatives coincide. For this reason, the common notation DF (x, ξ)
and DF (x) will be employed henceforth for differentials and derivatives, respectively.
Note that the tangent vector field ξ is not the usual displacement field u, given by
u = x−X = χ(X, t)−X . (2.34)
However, in the special case
χ(X, t) = X , (2.35)
for which R ≡ R0, the displacement and tangent vectors coincide. In this case, all lineariza-
tions are taken with respect to the reference configuration3
Start with the deformation gradient F and note that, by (1.8) and (2.7),
DF(x, ξ) =
[d
dwF(x+ wξ)
]
w=0
=
[d
dw
∂(x+ wξ)
∂X
]
w=0
=∂ξ
∂X, (2.36)
3See J. Casey, “On infinitesimal deformation measures”, J. Elast., 28:257-269, (1992), for derivations of
linearized kinematic quantities relative to the reference configuration.
ME280B May 7, 2019
Consistent Linearization in Continuum Mechanics 27
where ξ = ξ(X, t). This implies that
L [F; ξ]x = F+∂ξ
∂X= F , F =
∂x
∂X, (2.37)
that is, F is itself linear in ξ.
Observe that
ξ = ξ(X, t) = ξ(χ−1
t (x), t)
= ξ(x, t) , (2.38)
therefore, by the chain rule,∂ξ
∂X=
∂ξ
∂x
∂x
∂X=
∂ξ
∂xF (2.39)
and the linear part of F can be alternatively written as
L [F; ξ]x = F+∂ξ
∂xF . (2.40)
Consider next the linearization of the Lagrangian strain tensor E defined in (1.23). Here,
employing again the definition (2.7), it follows that
DE(x, ξ) =
[d
dw
1
2
(FT (x+ wξ)F(x+ wξ)− I
)]
w=0
=1
2
[d
dwFT (x+ wξ)
]
w=0
F+ FT[d
dwF(x+ wξ)
]
w=0
=1
2
[(∂ξ
∂X
)T
F+ FT ∂ξ
∂X
]. (2.41)
Alternatively, in view of (2.39),
DE(x, ξ) =1
2
[F
T(∂ξ
∂x
)T
F+ FT ∂ξ
∂xF
]= F
T(∂ξ
∂x
)S
F , (2.42)
where (∂ξ
∂x
)S
=1
2
[∂ξ
∂x+
(∂ξ
∂x
)T]. (2.43)
Therefore,
L [E; ξ]x = E+ FT(∂ξ
∂x
)S
F (2.44)
and the linear part of E remains symmetric, as expected. Likewise, given (1.13) and (2.42),
it is immediately clear that
L [C; ξ]x = C+ 2FT(∂ξ
∂x
)S
F . (2.45)
May 7, 2019 ME280B
28 Consistent Linearization
Similarly, it is simple to conclude from (1.14) and (2.40) that
L [B; ξ]x = B+∂ξ
∂xB+B
(∂ξ
∂x
)T
. (2.46)
In the special case when the linearization is performed with respect to the reference
configuration, where
x = X , ξ = u , F = I , (2.47)
it is easy to see that, given (2.39),
∂ξ
∂X=
∂ξ
∂xF =
∂ξ
∂x. (2.48)
Hence, once may conclude from (2.42) that
DE(x, ξ) = FT(∂ξ
∂x
)S
F =
(∂ξ
∂x
)S
=
(∂ξ
∂X
)S
(2.49)
and
L [E; ξ]X = E+1
2
∂ξ∂X
+
(∂ξ
∂X
)T = ε(ξ) , (2.50)
where
ε(ξ) =1
2(ξA,B + ξB,A)EA ⊗ EB (2.51)
and since, in view of (2.47)3, E = 0. Given (2.47)2 and (2.50), one may also conclude that
L [E; ξ]X = L [E;u]X = ε(u) =1
2(uA,B + uB,A)EA ⊗ EB , (2.52)
which is the usual measure of strain at infinitesimal deformations. In addition, it follows
from (2.45) and (2.45) that
L [C; ξ]X = L [C;u]X = 2ε(u) . (2.53)
and
L [B; ξ]X = L [B;u]X = 2ε(u) . (2.54)
ME280B May 7, 2019
Consistent Linearization in Continuum Mechanics 29
Next, consider the linearization of J = detF with respect to the configurationR. Startingonce more with the definition (2.7),
DJ(x, ξ) =
[d
dwdetF(x+ wξ)
]
w=0
=
[d
dwdet
(F+ w
∂ξ
∂X
)]
w=0
=
[d
dwdet(F+ wG
)]
w=0
; G =∂ξ
∂X
=
[d
dwdet
wF
[F
−1G−
(− 1
w
)I
]]
w=0
=
[d
dw
w3 detF
[−(− 1
w
)3
+
(1
w
)2
IF
−1G−(− 1
w
)II
F−1
G+ III
F−1
G
]]
w=0
=
[d
dw
detF
[1 + wI
F−1
G+ w2II
F−1
G+ w3III
F−1
G
]]
w=0
= detF IF
−1G
(2.55)
where it is recalled that for any second-order tensor, say T,
det(T− λI) = −λ3 + λ2IT − λIIT + IIIT . (2.56)
In the above, IT, IIT, and IIIT are the principal invariants of T, defined as
IT = trT (2.57)
IIT =1
2
[(trT)2 − trT2
](2.58)
IIIT = detT =1
6
[(trT)3 − 3 (trT)
(trT2
)+ 2 trT3
]. (2.59)
Equation (2.55) may be expressed as
[d
dwdetF(x+ wξ)
]
w=0
= J tr
(F
−1 ∂ξ
∂X
)(2.60)
where J = detF, and the linear part of detF relative to the configuration R takes the form
L [detF; ξ]x = J + J tr
(F
−1 ∂ξ
∂X
)= J
[1 + tr
(F
−1 ∂ξ
∂X
)]. (2.61)
May 7, 2019 ME280B
30 Consistent Linearization
Alternatively, taking into account (2.39), one may obtain an Eulerian form of the above
linearization by rewriting (2.60) as
DJ(x, ξ) = J tr
(F
−1 ∂ξ
∂X
)= J tr
[F
−1(∂ξ
∂xF
)]
= J tr
[FF
−1 ∂ξ
∂x
]; (tr [AB] = tr [BA])
= J tr
(∂ξ
∂x
)= J div ξ , (2.62)
where div ξ = ∂ξi∂xi
4, therefore
L [J ; ξ]x = J + J div ξ = J(1 + div ξ) . (2.63)
For the special case of linearization with respect to the reference configuration, equations (2.47)
and (2.51) imply that
L [J ; ξ]X = 1 + Div ξ = 1 + tr ε(ξ) = 1 + tr ε(u) . (2.64)
Also, consider the linearization ofda
dA, namely the ratio of the surface element areas of the
current over the reference configuration. To this end, start with Nanson’s formula (1.12),
from where it follows thatda
dAn = JF−TN = F∗N , (2.65)
where
F∗ = JF−T (2.66)
is the adjugate of F. Upon taking the dot-product of each side of (2.65) with itself, one finds
that
(da
dA
)2
= (F∗N) · (F∗N)
= N · (F∗TF∗N) = N · (C∗N) , (2.67)
where
C∗ = F∗TF∗ = J2C−1 . (2.68)
4In strict terms, one should use the notation divξ for this term, but the extra overbar is neglected here.
ME280B May 7, 2019
Consistent Linearization in Continuum Mechanics 31
It now follows from the usual definition (2.7) that
D
[(da
dA
)2](x, ξ) =
[d
dwN ·C∗(x+ wξ)N
]
w=0
, (2.69)
where[d
dwC∗(x+ wξ)
]
w=0
=
[d
dw
J2(x+ wξ)C−1(x+ wξ)
]
w=0
= 2J
[d
dwJ(x+ wξ)
]
w=0
C−1
+ J2[d
dwC−1(x+ wξ)
]
w=0
. (2.70)
Since CC−1 = I,[d
dw
C(x+ wξ)C−1(x+ wξ)
]
w=0
= DC(x, ξ)C−1
+CDC−1(x, ξ) = 0 , (2.71)
therefore
DC−1(x, ξ) =
[d
dwC−1(x+ wξ)
]
w=0
= −C−1DC(x, ξ)C
−1(2.72)
or, in light of (2.45),
DC−1(x, ξ) = −F−1F
−T
[2F
T(∂ξ
∂x
)S
F
]F
−1F
−T= −2F−1
(∂ξ
∂x
)S
F−T
. (2.73)
Return to (2.70) and substituting in it (2.62) and (2.45), it follows that
[d
dwC∗(x+ wξ)
]
w=0
= 2J[J div ξ
]C
−1+ J
2
[−2F−1
(∂ξ
∂x
)S
F−T
]
= 2J2
[div ξC
−1 − F−1(∂ξ
∂x
)S
F−T
]. (2.74)
This means that (2.69) takes the form
D
[(da
dA
)2](x, ξ) = 2
(da
dA
)D
[da
dA
](x, ξ)
= N · 2J2
[div ξC
−1 − F−1(∂ξ
∂x
)S
F−T
]N , (2.75)
hence
D
[da
dA
](x, ξ) =
(dA
da
)N · J2
[div ξC
−1 − F−1(∂ξ
∂x
)S
F−T
]N (2.76)
May 7, 2019 ME280B
32 Consistent Linearization
and, accordingly,
L[da
dA; ξ
]
x
=
(da
dA
)+D
[da
dA
](x, ξ) . (2.77)
When the linearization is performed with respect to the reference configuration, it follows
from (2.47) and (2.51) that (2.76) reduces to
D
[da
dA
](X, ξ) = N ·
[div ξI−
(∂ξ
∂x
)S]N
= N ·[Div ξI−
(∂ξ
∂X
)S]N
= N · [tr ε(ξ)I− ε(ξ)]N= tr ε(u)−N · ε(u)N (2.78)
and
L[da
dA; ξ
]
X
= 1 + tr ε(u)−N · ε(u)N . (2.79)
The balance laws are also subject to linearization. For instance, consider the balance of
mass in the form (1.63) and linearize it relative to the configuration R, that is, write
L [ρ0; ξ]x = L [ρJ ; ξ]x ⇒ (2.80)
which leads, with the aid of (2.62) to
ρ0 = ρJ +D[ρJ ](x, ξ)
= ρJ +Dρ(x, ξ)J + ρDJ(x, ξ)
= ρJ +Dρ(x, ξ)J + ρJ div ξ . (2.81)
Assuming that conservation of mass holds in R (which is tantamount to letting ρ be such
that ρ0 = ρJ), it follows from the above that
Dρ(x, ξ) = −ρ div ξ , (2.82)
so that
L [ρ; ξ]x = ρ+ (−ρ div ξ) = ρ(1− div ξ) . (2.83)
In view of (2.82), the special case of linearization with respect to R0 readily yields
L [ρ; ξ]X = ρ0(1− div ξ) = ρ0(1−Div ξ) = ρ0(1− tr ε(ξ)) = ρ0(1− tr ε(u)) , (2.84)
ME280B May 7, 2019
Consistent Linearization in Continuum Mechanics 33
which implies that conservation of mass may be expressed as
ρ = ρ0(1− tr ε(u)) (2.85)
Note that ρ is not equal to ρ0 when linearized with respect to R0.
Lastly, consider the balance of linear momentum (1.71) in spatial form and take its linear
part with respect to R, which is
L [divT; ξ]x + L [ρb; ξ]x = L [ρu; ξ]x (2.86)
The differential of each of the three terms in the preceding equation is considered separately.
First, write the acceleration term as
D [ρu] (x, ξ) =
[d
dwρ(x+ wξ)u(x+ wξ)
]
w=0
= −ρ div ξu+ ρξ (2.87)
with the use of (2.82) and the fact that u(x) = x. Next, the body force term becomes
D [ρb] (x, ξ) =
[d
dwρ(x+ wξ)b(x+ wξ)
]
w=0
= −ρ div ξb+ ρDb(x, ξ) (2.88)
where, again, use is made of (2.82) and also of the definition
L [b; ξ]x = b+Db(x, ξ) . (2.89)
For the stress-divergence term, it can be shown that
D [divT] (x, ξ) =
[d
dwdivT(x+ wξ)
]
w=0
= div
[d
dwT(x+ wξ)
]
w=0
− T ij,kξk,jei
= div [DT(x, ξ)]− gradT · gradT ξ . (2.90)
Note that a constitutive equation for T is required in order to evaluate the first term on the
right-hand side of (2.90)3. To derive (2.90), it is best to resort to component representation,
as follows:
D [divT] (x, ξ) =
[d
dwdivT(x+ wξ)
]
w=0
=
[d
dw
∂Tij(x+ wξ)
∂(xj + wξj)
]
w=0
ei
=
[d
dw
∂Tij(x+ wξ)
∂xk
∂xk∂(xj + wξj)
]
w=0
ei
=
[d
dw
∂Tij(x+ wξ)
∂xk
]
w=0
∂xk∂xj
ei +∂T ij
∂xk
[d
dw
∂xk
∂(xj + wξj)
]
w=0
ei .
(2.91)
May 7, 2019 ME280B
34 Consistent Linearization
Note that it is not possible to switch the order of the differential operators ddw
and ∂∂x
in (2.91)2 as the latter is explicitly dependent on ω.5 To evaluate the second term on the
right-hand side of (2.91)4, observe that
[d
dw
∂(xk + wξk)
∂xl
∂xl∂(xj + wξj)
]
w=0
=
[d
dwδkj
]
w=0
= 0 , (2.92)
therefore,
[d
dw
∂(xk + wξk)
∂xl
]
w=0
δlj + δkl
[d
dw
∂xl
∂(xj + wξj)
]
w=0
= 0 , (2.93)
hence [d
dw
∂xk
∂(xj + wξj)
]
w=0
= −ξk,j . (2.94)
Equation (2.90) follows upon substituting (2.94) into (2.91)4.
Return to the linearized statement of linear momentum balance (2.86) and taking into
account (2.87), (2.88) and (2.90), it follows that
divT+ div [DT(x, ξ)]− gradT · gradT ξ + ρb+−ρ div ξb+ ρDb(x, ξ)
= ρu+−ρ div ξu+ ρξ
. (2.95)
Assuming that linear momentum balance holds in R, the preceding equation reduces to
div [DT(x, ξ)]− gradT · gradT ξ + ρ div ξ(u− b) + ρDb(x, ξ) = ρξ (2.96)
or, equivalently,
div [DT(x, ξ)]− gradT · gradT ξ + divT div ξ + ρDb(x, ξ) = ρξ . (2.97)
Assuming that the refefence configuration R0 is initial stress- and initial acceleration-
free, it is immediately obvious from linear momentum balance that the body force vanishes
as well in this configuration. As a result,
L [T; ξ]X = DT(X, ξ) = σ , (2.98)
L [u; ξ]X = Du(X, ξ) = a , (2.99)
5This point appears to have been overlooked in T.J.R. Hughes and K.S. Pister, “Consistent linearization
in mechanics of solids and structures”, Comp. Struct., 8:391–397, (1978).
ME280B May 7, 2019
Consistent Linearization in Continuum Mechanics 35
and
L [b; ξ]X = Db(x, ξ) = b . (2.100)
It follows from (2.97) and the above three equations that the linearized statement of linear
momentum balance with respect to R0 reduces to
Divσ + ρ0b = ρ0a . (2.101)
May 7, 2019 ME280B
Chapter 3
Incremental Formulations
A typical procedure for the solution of continuummechanics problems using the finite element
method starts with the derivation of a weak (e.g., Galerkin) form of the governing (balance)
equations from the original strong form. In general, the weak form comprises a set of
integro-differential equations in space and time. Next, these equations are considered in a
given time domain (real or implied by the application of the external loads). The solution of
these equations is now performed by partitioning the time into time increments and deducing
the solution from the beginning of the first time increment (where this solution is assumed
to be given) to the end of the first time increment, and so on up to the end of the last
increment. This process gives rise to the term incremental formulation for the solution of
the continuum mechanics equations.
Each incremental solution requires that dependent variables of the problem be interpo-
lated in space. This reduces the original integro-differential equations in space-time to merely
differential equations in time. Upon observing that the incrementation in time induces a
natural discretization, it follows that the preceding differential equations in time are further
reduced to either implicit (generally, non-linear) algebraic equations or merely explicit equa-
tions. The solution to the former requires some iterative approach (e.g., Newton-Raphson
method or its variants), while the latter does not.
3.1 Strong and Weak Form of the Balance Laws
Consider a motion χ : R0 × R 7→ R ⊂ E3 of a body, as in (1.3), which is due to externally
applied loads. Note that the preceding definition of this motion is predicated on the existence
36
Strong and Weak Form of the Balance Laws 37
of a reference configuration R0. In this case, recall that the displacement field u is defined
as in (2.34), see also Figure 3.1. If a reference configuration is not available, then one needs
to formulate the strong and weak forms of the initial/boundary-value problem without using
(explicitly or implicitly) such a configuration.
RR0
Xx
u
EAe1
χ
Figure 3.1. Reference and current configuration
Let the boundary ∂R of R be smooth enough for a unique outward unit normal n defined
everywhere on ∂R, as in Figure 3.2. Assume further that the boundary ∂R is decomposed
into parts Γu and Γq such that at any time t
∂R = Γu ∪ Γq . (3.1)
These two parts are defined as
n
R∂R
ΓqiΓui
Figure 3.2. Partition of boundary ∂R of R
Γu = ∪3i=1Γui
, Γq = ∪3j=1Γqj , (3.2)
such that
ui = ui(x, t) on Γui(Dirichlet boundary conditions in solids) (3.3)
May 7, 2019 ME280B
38 Incremental Formulations
or
vi = vi(x, t) on Γui(Dirichlet boundary conditions in fluids) (3.4)
and
tj = tj(x, t) on Γqj (Neumann boundary conditions) (3.5)
where u (resp., v) and t are the prescribed displacements (resp., velocities) and tractions on
∂R and
Γui∩ Γqi = ∅ . (3.6)
Note that the components in (3.3-3.5) do not need to be those resolved on the basis vectors
ei associated with the current configuration.
The strong form of the general initial/boundary-value problem in continuum mechanics
can be stated as follows: given the external forces and prescribed displacements (resp.,
velocities) b, t, u (resp., v) defined as above, the initial velocity and density fields v0 and
ρ0, and a constitutive law for the Cauchy stress tensor T, such that T = TT , determine
the displacement field u such that det∂(X+ u)
∂X> 0 (or the velocity field v) and the mass
density ρ > 0, such that, for all t, linear momentum holds in the form (1.71) or (1.74) and
mass balance holds in the form (1.66) or (1.63), subject to the boundary conditions (3.3)
(or (3.4)) and (3.5), and the initial conditions
u(X, 0) = 0 , u(X, 0) = v0 in R0 (3.7)
for solids and
v(x, 0) = v0 in R (3.8)
for fluids.
At every point x of the current configuration, define a tangent vector ξ = ξ(x, t), that
is any vector which originates at x, as in Figure 3.3. Define the space of admissible tangent
vector field W as
W =ξ : R× R −→ E3 , ξi(x, t) = 0 on Γui
, (3.9)
that is, such that ξ vanishes identically where a displacement or velocity boundary condition
is enforced.
Now define the weighted residual form
R =
∫
R×I
ξ · (ρu− ρb− divT) dvdt+
∫
Γq×I
ξ · (t− t) dadt , (3.10)
ME280B May 7, 2019
Strong and Weak Form of the Balance Laws 39
R
ξ
Figure 3.3. Tangent vector at a point in the current configuration
where I = (t0, T ) is the time domain of the analysis. Note that the second integral on the
right-hand side of (3.10) is understood as being a short-hand for∑
i
∫Γq,i
ξi(ti− ti) dadt, thatis, it includes the contributions of all the components of the prescribed traction in all of their
directions.
The weak form of the general initial/boundary-value problem in continuum mechanics
reads as follows: given the external forces b and t, a constitutive law for T, such that
TT = T, find the displacement field u ∈ U (or the velocity field v ∈ V), such that
R = 0 (3.11)
for all ξ ∈ W . Here,
U =
u : R0 × R→ E3 | det
(I+
∂u
∂X
)> 0 , ui(X, t) = ui on Γui
,
u(X, 0) = 0 , u(X, 0) = v0 (3.12)
for solids and
V =v : R× R→ E3 | vi(x, t) = vi on Γui
, v(x, 0) = v0
(3.13)
for fluids. In the solids problem, mass conservation is enforced trivially, in the sense that,
given ρ0 > 0 at time t0, one may find ρ as a function of ρ0 and the deformation, according
to (1.63). Therefore, there is no need for mass conservation to be enforced through the
weak form. On the contrary, in fluid mechanics, one needs to additionally enforce mass
conservation according to (1.66). To achieve this, the weighted residual R of (3.10) must be
amended by an additional term as
R =
∫
R×I
ξ ·(ρu−ρb−divT) dvdt+
∫
Γq×I
ξ ·(t−t) dadt+∫
R×I
σ(ρ+ρ div v) dvdt , (3.14)
May 7, 2019 ME280B
40 Incremental Formulations
such that a solution (v, p) ∈ V × Z+ is sought by enforcing (3.11) for all (ξ, σ) ∈ W × Z,where
Z = σ : R× R→ R , Z+ = σ : R× R→ R | σ > 0 . (3.15)
Additional assumptions on smoothness are required for the functions in U , V ,W ,Z in order
to guarantee the existence of the integrals which comprise R in (3.14). These requirements
will be discussed when introducing finite element counterparts of these spaces.
It is worth observing here that the balance of linear momentum (as well as the mass
balance in fluids) and the traction boundary conditions are enforced through the weighted
residual form. On the other hand, the displacement (or velocity) boundary conditions and
initial conditions are satisfied from the outset by the choice of admissible fields U or V . In
addition, balance of angular momentum is satisfied by the constitutive law for stress.
The above strong and weak form are equivalent (that is, derivable from each other) only
under the assumption of continuity of all integrands in (3.10) or (3.14). This is a fairly
severe assumption, which, when violated, leads to solutions obtained from the weak form
(termed weak solutions) that may differ from the strong solution where/when discontinuities
are encountered.
The complete proof of the preceding result may be found elsewhere. Here, only a sketch
of the proof is provided, as follows: first, it is clear by inspection that the strong form implies
the weak form. for the opposite, note that (3.10) (or (3.14)) applies for any ξ ∈ W (and,
σ ∈ Z, if applicable). Now fix a time t and choose ξ such that ξ = 0 on Γq (and also σ = 0
on R, if applicable). Next, argue, by contradiction, that if there is a point x ∈ R where
ρa−ρb−divT 6= 0, then one may choose ξ to vanish everywhere and every time except in a
neighborhood Nx of x in which ξ · (ρa−ρb−divT) > 0, which is always possible due to the
continuity of the preceding scalar quantity. It follows that∫R×I
ξ ·(ρa−ρb−divT) dvdt > 0,
which contradicts the original assumption, thus ρa− ρb− divT = 0 everywhere in R at all
times t. Then, the same argument may be repeated on Γq (as well as on R for the continuity
equation in (3.14), if applicable).
Proceed with separating the space from the time integrals in the weighted residual
form (3.10), and consider first the spatial integrals (that is, “freeze” time). In particular, set
∫
R
ξ · (ρu− ρb− divT) dv +
∫
Γq
ξ · (t− t) da = 0 (3.16)
and note that, upon using integration by parts, the divergence theorem, and the property
ME280B May 7, 2019
Strong and Weak Form of the Balance Laws 41
of ξ specified in (3.9)∫
R
ξ · divT dv =
∫
R
ξiTij,j dv =
∫
R
[(ξiTij),j − ξi,jTij ] dv
=
∫
∂R
ξiTijnj da−∫
R
ξi,jTij dv
=
∫
Γu∪Γq
ξiTijnj da−∫
R
ξi,jTij dv
=
∫
Γq
ξitida−∫
R
ξi,jTij dv
=
∫
Γq
ξ · t da−∫
R
∂ξ
∂x·T dv . (3.17)
Substituting the preceding expression to the weighted-residual form (3.16), it follows that∫
R
ξ · ρu dv +∫
R
∂ξ
∂x·T dv =
∫
R
ξ · ρb dv +∫
Γq
ξ · t da (3.18)
The full space-time integral form can be easily reconstructed from the above. Observing
the symmetry of T, the second term on the left-hand side of (3.18) can be also written as∫R
(∂ξ∂x
)s ·T dv The above expression is computationally convenient, as it involves the term
(∂ξ
∂x
)s
=1
2
[∂ξ
∂x+
(∂ξ
∂x
)T]
(3.19)
which resembles the infinitesimal strain tensor. This observation is important when attempt-
ing to extend finite element formulations originally developed for infinitesimal deformations
to finite deformations.
If ξ is identified as the variation δu1 of u, one obtains a statement of virtual work,
according to which the virtual work done by the internal forces (inertial forces and stresses)
is equal to the virtual work done by the external forces (body forces and surface tractions)
over the whole body occupying the configuration R. A corresponding statement of virtual
power is obtained from (3.18) when ξ is substituted with a virtual velocity. It is instructive
to contrast these statements to the mechanical energy balance equation for the whole body,
see (1.80).
One may assume at the outset that the traction boundary conditions (3.5) are satisfied
at the outset and do not need to be included in the weighted residual form (3.10). While
1Recall that, by definition, the variation of u vanishes where u is specified, which confirms that the space
of admissible variations of u is precisely W.
May 7, 2019 ME280B
42 Incremental Formulations
this goes against the premise of finite element modeling (where one may one directly impose
boundary conditions on the dependent variables that enter the finite element approximation),
it is instructive to consider this option, where the weighted-residual statement is reduced∫
R
ξ · (ρa− ρb− divT) dv = 0 (3.20)
for all ξ ∈ W . Repeating the preceding procedure leads to:∫
R
ξ · ρa dv +∫
R
∂ξ
∂x·T dv =
∫
R
ξ · ρb dv +∫
Γq
ξ · t da . (3.21)
Upon now enforcing (3.5), one immediately recovers the derived weighted-residual statement
in (3.18).
A completely equivalent weighted-residual form can be derived based on the referential
statement of linear momentum balance (1.74). In this case, one observes that the tangent
vector ξ may be represented as
ξ = ξ(X, t) = ξ(x, t) . (3.22)
Next, “freeze” again time and start from the weighted residual statement
R0 =
∫
R0
ξ · (ρ0a− ρ0b−DivP) dV +
∫
Γq0
ξ · (p− p) dA = 0 , (3.23)
for all ξ ∈ W , where p = PN is the traction vector on the boundary of the body in
the current configuration, but resolved using the geometry of the reference configuration.
Repeating the same procedure as with the spatial weak form, it is easy to show that∫
R0
ξ · ρ0a dV +
∫
R0
∂ξ
∂X·P dV =
∫
R0
ξ · ρ0b dV +
∫
Γq0
ξ · p dA . (3.24)
Recalling (2.36), it is concluded that the integrand of the second term on the left-hand side
of (3.24) satisfies∂ξ
∂X·P = DF(X, ξ) ·P . (3.25)
Furthermore,
∂ξ
∂X·P =
∂ξ
∂X· (FS) =
(FT ∂ξ
∂X
)· S
=1
2
[FT ∂ξ
∂X+
(∂ξ
∂X
)T
F
]· S = DE(x, ξ) · S , (3.26)
where use is made of (1.85), (2.41), and the symmetry of S.
ME280B May 7, 2019
Lagrangian and Eulerian Methods 43
3.2 Lagrangian and Eulerian Methods
Preliminary to the presentation of any specific finite element methods, it is essential to
formally distinguish between Lagrangian and Eulerian computations. To this end, consider
a body occupying the configuration R at time and define a surface S(t), as in Figure 3.4,
which may be represented by the parametric equation
xs = xs(ξ, η, t) , (3.27)
where ξ, η are surface parameters. These parameters may be eliminated to obtain an alter-
R
S
v
vs
Figure 3.4. A surface S traversing the configuration R at time t
native representation of the surface as
f(xs, t) = 0 . (3.28)
Taking the time derivative of the above equation, while following any given point on the
surface, it follows that∂f
∂t+∂f
∂x· vs = 0 , (3.29)
where vs is the velocity of the surface, see Figure (3.4), which can be readily obtained
from (3.27). Alternatively, consider a material point which occupies a point x ∈ S at time t.
The material time derivative of f describes the rate of change of f for the given material
point. This equals
f =∂f
∂t+∂f
∂x· v , (3.30)
where v = χ is the velocity of the particle, see, again, Figure 3.4. Generally, at the fixed
time t, vs 6= v. Hence, subtracting (3.29) from (3.30), it follows that
f =∂f
∂x· (v − vs) (3.31)
May 7, 2019 ME280B
44 Incremental Formulations
and, since the unit normal n to the surface may be written as
n =∂f∂x∣∣∂f∂x
∣∣ , (3.32)
it is concluded that the flux of mass through the surface S at time t is given by
∫
S
ρ(v − vs) · n da =
∫
S
ρf
∣∣∣∣∂f
∂x
∣∣∣∣−1
da . (3.33)
Upon invoking the localization theorem, it follows from the above that the necessary and
sufficient condition for the flux of mass through S (or any part of it) to vanishes is f = 0, in
which case
(v − vs) · n = 0 . (3.34)
This mandates that the surface S and the material particles on it travel with the same
normal velocity to S. In this case, the surface S is called a material surface. The necessary
and sufficient condition
f = 0 (3.35)
is referred to as Lagrange’s criterion of materiality.
Assume that at any given time t the region R is spatially discretized according to
R .= ∪eΩ
e, where Ωe is the domain of the typical finite element e, such that all bound-
R
Ωe
Figure 3.5. Finite element mesh schematic
ary surfaces (or edges) of each Ωe are material. Lagrangian finite element computations are
making use of finite elements with domains whose boundaries are material.
To understand how it can be guaranteed that the boundary of a finite element domain are
material, note that in Lagrangian finite element computations the motion of the element is
completely parametrized by the motion of its nodal points. Hence, the finite element system
ME280B May 7, 2019
Lagrangian and Eulerian Methods 45
is finite-dimensional, in the sense that the motion of the discretized body is fully determined
at any given time t by a finite set of parameters.
Consider a typical surface S of a finite element domain Ωe, in the context of classical
isoparametric interpolations. For a typical point X on S in the reference configuration, one
may write
X =∑
i
Ni(ξ, η)Xi (3.36)
where (ξ, η) are the natural coordinates (which coincide, here, with the surface coordinates),
S ξ
η
X
Figure 3.6. Lagrangian finite element surface
Ni are the isoparametric shape functions, and Xi the position vector of node i on S, as inFigure 3.6. The point X is mapped in the current configuration to x = χ(X, t), such that
x =∑
i
Ni(ξ, η)xi
=∑
i
Ni(ξ, η)χ(Xi, t)
=∑
i
Ni(ξ, η) [Xi + ui]
=∑
i
Ni(ξ, η)Xi +∑
i
Ni(ξ, η)ui = X+ u . (3.37)
The surface S is clearly material, since it consists of the same material points at all times
with respect to the motion χ of the nodal points i and the isoparametric interpolation.
Put differently, all xi are material points by definition, while any x on S is also material by
construction of the motion in terms of the nodal motions and the isoparametric interpolation.
Alternatively, one may consider an Eulerian finite element formulation for which the
mesh is stationary, that is vs = 0 on the boundary ∂Ωe of any element element e. In this
case, each element becomes a control volume. In view of (3.34) it is clear that the element
May 7, 2019 ME280B
46 Incremental Formulations
region Ωe is not material. Thus, mass flux takes place across ∂Ωe, which necessitates that
mass balance be enforced through the equation
∫
Ωe
σ(ρ+ ρ div v) dv = 0 , (3.38)
for all admissible σ, as also argued earlier (see (3.14)).
Lagrangian and Eulerian methods have both advantages and disadvantages from an algo-
rithmic standpoint. For instance, when Eulerian methods are used for solids or free-surface
fluids, then
(a) A geometric detection algorithm is required to identify the “full” elements at each
time t, see Figure 3.7.
(b) A special treatment is generally needed for “partial” element at each time t, see again
Figure 3.7.
(c) The application of Dirichlet conditions on the physical boundary is not necessarily
trivial.
full element
partial element
Figure 3.7. Eulerian finite element mesh with a moving solid or fluid
Moreover, the tracking of material-based history variables requires a special procedure, since
such variables cannot be directly stored at nodal or integration points.
Lagrangian methods suffer from excessive mesh distortion, which may cause loss of
uniqueness in the isoparametric mapping, as well as from inability to directly handle loss of
materiality, as is the case, for instance, when modeling crack propagation or fragmentation.
ME280B May 7, 2019
Total and Updated Lagrangian Methods 47
3.3 Semi-discretization
As done earlier, it is customary (and practical) to separate the time from the space integrals
in (3.14), which often referred to as semi-discretization. Assuming that the solution must be
determined in the time domain (t0, T ], perform a time discretization by introducing a finite
sequence of distinct times t1, t2, ..., tN , as in Figure 3.8. Here, ∆tn = tn+1− tn is the discrete
t0 t1 t2 tn tn+1 tN−1 tN = T
∆tn
Figure 3.8. Time discretization
time step at time tn and
∆t =
N∑
n=1
∆tn−1
N=
T − t0N
(3.39)
is the average step size, with the understanding that ∆tn = O(∆t), n = 0, 1, . . . , N − 1.
The basic idea of incremental methods is to satisfy the equations of motion only at
discrete time instances t1, t2, ..., tN , where at each such tn+1, the solution to these equations
is known for all ti, i = 1, 2, . . . , n. In the limit as N → ∞, incremental methods recover a
solution in a dense subset of (t0, T ].
Incremental methods make sense on two grounds. First, they recover the history of a
deformation, which is oftentimes, in itself, useful to the analyst. Second, they are necessary
when modeling continua made of path-dependent material or which are subject to path-
dependent loading. In the latter case, iterative methods are essentially used to resolve the
path-dependency.
A weighted-residual interpretation of semi-discretization can be obtained by starting
from (3.10) (similarly, for (3.14)) and writing ξ(x, t) = θ(t)ξ(x), such that
R =
∫ T
t0
θ
[∫
R
ξ · (ρa− ρb− divT) dv +
∫
Γq
ξ · (t− t) da
]dt = 0 , (3.40)
where the time-dependent weighting function θ(t) is defined as θ(t) =∑N
i=1 θiδ(t− ti). Thisdiscretization leads naturally to the requirement that linear momentum balance and traction
boundary conditions must be enforced at each of the times t1, t2, ..., tN .
May 7, 2019 ME280B
48 3.4. TOTAL AND UPDATED LAGRANGIAN METHODS
3.4 Total and Updated Lagrangian Methods
Consider an incremental Lagrangian method and assume that the balance laws have been
already satisfied (in a weak form) and the motion χ has been determined for times t1, t2, ..., tn,
see Figure 3.9. The objective is now to determine the position of all material particles
(therefore, the configuration of the body) at t = tn+1, such that
R0
Rn
Rn+1
Figure 3.9. Configurations at times t0, tn and tn+1
∫
Rn+1
ξn+1 · ρn+1an+1 dv +
∫
Rn+1
(∂ξ
∂x
)s
n+1
·Tn+1 dv
=
∫
Rn+1
ξn+1 · ρn+1bn+1 dv +
∫
Γqn+1
ξn+1 · tn+1 da , (3.41)
for all ξn+1 ∈ W . The preceding equation may be resolved iteratively for the displacement
field un+1 ∈ U (or, equivalently, the relative displacement field un+1 − un), assuming the
data (that is, b and t) is known in (tn, tn+1].
Alternatively, one may pull-back (3.41) to the reference configuration and write
∫
R0
ξn+1 · ρ0an+1 dV +
∫
R0
∂ξn+1
∂X·Pn+1 dV
=
∫
R0
ξn+1 · ρ0bn+1 dV +
∫
Γq0
ξn+1 · pn+1 dA , (3.42)
for all ξn+1 ∈ W . Note that, since, in general, Γq depends on time, by Γq0 in (3.42),
one denotes the pull-back of Γqn+1 to the reference configuration, rather than the traction
boundary of the reference configuration itself.
ME280B May 7, 2019
Total and Updated Lagrangian Methods 49
Starting with the referential weak form (3.42), concentrate on a typical Lagrangian finite
element e with domain Ωe0 in the reference configuration. For this element, write at t = tn+1
u(X, tn+1)∣∣Ωe
0= ue
n+1(X).=
NEN∑
i=1
N ei u
ei n+1 ,
an+1(X, tn+1)∣∣Ωe
0= ae
n+1(X).=
NEN∑
i=1
N ei a
ei n+1 , (3.43)
ξn+1(X, tn+1)∣∣Ωe
0= ξen+1(X)
.=
NEN∑
i=1
N ei ξ
ei n+1 ,
where the element is assumed isoparametric with NEN nodes and interpolation functions
N ei (X), i = 1, 2 . . . , NEN. Special cases, such as those involving hierarchical, mixed or incom-
patible interpolations can be easily handled by extending the above definitions.
For ease of computer implementation, one may write the preceding interpolations in
matrix form as
uen+1
.=
NEN∑
i=1
N ei u
ei n+1
=
[N e
113︸ ︷︷ ︸(3×3)
N e213 · · · N e
NEN13
]
︸ ︷︷ ︸3×(3×NEN)
ue1
ue2...
ueNEN
n+1︸ ︷︷ ︸(3×NEN)×1
= [Ne][uen+1] , (3.44)
where [uen+1] is a vector that contains all nodal displacements, while 13 denotes the 3 × 3
identity matrix. Similarly, one may write
aen+1
.= [Ne][ae
n+1] , ξen+1.= [Ne][ξ
e
n+1] . (3.45)
Also, to resolve in convenient matrix form the stress Pn+1 and the derivative∂ξen+1
∂Xin (3.42),
May 7, 2019 ME280B
50 Incremental Formulations
write their 9 components as column vectors according to
〈Pn+1〉 =
P11
P22
P33
P12
P23
P31
P13
P21
P32
n+1
,
⟨∂ξen+1
∂X
⟩=
ξe1,1
ξe2,2
ξe3,3
ξe1,2
ξe2,3
ξe3,1
ξe1,3
ξe2,1
ξe3,2
n+1
. (3.46)
Thus, one may write
⟨∂ξen+1
∂X
⟩
︸ ︷︷ ︸(9×1)
= [Be]︸︷︷︸9×(3×NEN)
[ξe
n+1]︸ ︷︷ ︸(3×NEN)×1
, (3.47)
where
[Be]︸︷︷︸9×(3×NEN)
=[[Be
1]︸︷︷︸(9×3)
[Be2] · · · [Be
NEN]]
(3.48)
and [Bei ] is such that
< Bei ξ
e
in+1> =
N ei,1ξ
ei1
N ei,2ξ
ei2
N ei,3ξ
ei3
N ei,2ξ
ei1
N ei,3ξ
ei2
N ei,1ξ
ei3
N ei,3ξ
ei1
N ei,1ξ
ei2
N ei,2ξ
ei3
n+1
(no summation on i) , (3.49)
ME280B May 7, 2019
Total and Updated Lagrangian Methods 51
consequently
[Bei ]︸︷︷︸
9×3
=
N ei,1
N ei,2
N ei,3
N ei,2
N ei,3
N ei,1
N ei,3
N ei,1
N ei,2
, (3.50)
where N ei,A =
∂Nei
∂XA.
Likewise, to express the spatial statement (3.41) in matrix form, one need to express the
(symmetric) stress Tn+1 and the derivative(
∂ξen+1
∂x
)sas 6-dimensional vectors according to
〈Tn+1〉 =
T11
T22
T33
T12
T23
T31
n+1
,
⟨(∂ξe
∂x
)s
n+1
⟩=
ξe1,1
ξe2,2
ξe3,3
ξe1,2 + ξe2,1
ξe2,3 + ξe3,2
ξe3,1 + ξe1,3
n+1
. (3.51)
It follows that
⟨(∂ξe
∂x
)s
n+1
⟩
︸ ︷︷ ︸6×1
= [Ben+1]︸ ︷︷ ︸
6×(3×NEN)
[ξe
n+1]︸ ︷︷ ︸(3×NEN)×1
=
[[Be
1,n+1]︸ ︷︷ ︸6×3
[Be2,n+1] · · · [Be
NEN,n+1]]
ξe1
ξe2...
ξeNEN
n+1
, (3.52)
May 7, 2019 ME280B
52 Incremental Formulations
where
[Be
i,n+1
]︸ ︷︷ ︸
6×3
=
N ei,1
N ei,2
N ei,3
N ei,2 N e
i,1
N ei,3 N e
i,2
N ei,3 N e
i,1
n+1
(3.53)
and N ei,j =
∂Nei
∂xj, where now it is understood that the element interpolation functions are
functions of the current position of the element points, that is, N ei = N e
i (x).
Substituting the interpolations (3.44), (3.45) and (3.47), and the vectorial representation
of stress from (3.46)1 to the weak form (restricted to the domain Ωe0 of element e), one finds
that
[ξe
n+1]T
(∫
Ωe0
[Ne]Tρ0[Ne] dV
)[ae
n+1] + [ξe
n+1]T
∫
Ωe0
[Be]T < Pn+1 > dV
= [ξe
n+1]T
∫
Ωe0
[Ne]Tρ0[bn+1] dV + [ξe
n+1]T
∫
∂Ωe0∩Γq0
[Ne]T [pn+1] dA
+ [ξe
n+1]T
∫
∂Ωe0\Γq0
[Ne]Tpn+1 dA , (3.54)
where ∂Ωe0 ∩ Γq0 is the part of the element boundary which (potentially) intersects with the
Neumann boundary Γq0 and ∂Ωe0 \Γq0 is the rest of the element boundary, as in Figure 3.10.
∂Ωe0 ∩ Γq0
∂Ωe0 \ Γq0
Figure 3.10. Partition of element boundary ∂Ωe0
The preceding statement may be compactly rewritten as
[ξe
n+1]T[[Me][ae
n+1] + [Re(uen+1)]− [Fe
n+1]]
= [ξe
n+1]T
∫
∂Ωe0\Γq0
[Ne]T [pn+1] dA (3.55)
ME280B May 7, 2019
Total and Updated Lagrangian Methods 53
where
[Me]︸︷︷︸(3×NEN)×(3×NEN)
=
∫
Ωe0
[Ne]Tρ0[Ne] dV (3.56)
is the element mass matrix, which is clearly symmetric,
[Ren+1]︸ ︷︷ ︸
(3×NEN)×1
=
∫
Ωe0
[Be]T < Pn+1 > dV (3.57)
is the element stress-divergence vector, and
[Fen+1]︸ ︷︷ ︸
(3×NEN)×1
=
∫
Ωe0
[Ne]Tρ0[bn+1] dV +
∫
∂Ωe0∩Γq0
[Ne]T [pn+1] dA (3.58)
is the element external force vector.
Since ξn+1 is an arbitrary tangent vector on the current configuration, it follows that
[Me][aen+1] + [Re(ue
n+1)] = [Fen+1] +
∫
∂Ωe0\Γq0
[Ne]T [pn+1] dA . (3.59)
Under the action of the assembly operator Ae, the above equation leads to
Ae
[[Me][ae
n+1] + [Re(uen+1)]
]= A
e
[[Fe
n+1] +
∫
∂Ωe0\Γq0
[Ne]T [pn+1] dA
]. (3.60)
which, in turn, yields
[M][an+1] + [R(un+1)] = [Fn+1] , (3.61)
where
[M] = Ae[Me] (3.62)
is the global mass matrix, which is symmetric,
[Rn+1] = Ae[Re
n+1] (3.63)
is the global stress-divergence vector, and
[Fn+1] = Ae[Fe
n+1] (3.64)
is the global external force vector. In (3.61), the termAe
∫∂Ωe
0\Γq0[Ne]T [pn+1] dA, which quan-
tifies the cumulative interelement jump in tractions is neglected, as is customary in finite
May 7, 2019 ME280B
54 Incremental Formulations
element approximations. This term is frequently used to quantify the error in the finite
element analysis a posteriori. In addition, the global displacement and acceleration vectors
are defined as
[un+1] =
u1n+1
u2n+1
· · ·uNUMNP
n+1
, [an+1] =
a1n+1
a2n+1
· · ·aNUMNPn+1
, (3.65)
where NUMNP is the total number of nodes in the mesh. Equations (3.61) constitute the semi-
discrete form of linear momentum balance, that is, the approximate equations of motion
obtained after spatial discretization. This is a (generally) non-linear system of second-order
ordinary differential equations in time, with unknowns the nodal displacements un+1. Under
quasi-static conditions (that is, when the inertia term can be neglected), the above system
of equations becomes purely algebraic. Also, Dirichlet boundary conditions are readily ac-
counted for by merely fixing the relevant displacement degrees-of-freedom on nodes which
lie on Γu and, if desired, calculating the corresponding forces upon solution of the overall
problem as reactions.
Note that the stress-divergence vector in (3.57) may well depend not only on un+1, but
also on vn+1. This would be the case when the stress response is rate-dependent. Also, the
external force vector in (3.58) may be explicitly dependent on the motion, as would be the
case, for instance, when the externally imposed traction is a follower load (e.g., a pressure
load). In this case, Fn+1 = Fn+1(un+1), so the external force is itself a function of the
solution.
If one chooses to start with the spatial statement of linear momentum balance in (3.41),
an analogous procedure would ensue that would yield the element arrays
[Me]︸︷︷︸(3×NEN)×(3×NEN)
=
∫
Ωe
[Ne]Tρ[Ne] dv (3.66)
[Ren+1]︸ ︷︷ ︸
(3×NEN)×1
=
∫
Ωe
[Ben+1]
T < Tn+1 > dv (3.67)
and
[Fen+1]︸ ︷︷ ︸
(3×NEN)×1
=
∫
Ωe
[Ne]Tρ[bn+1] dv +
∫
∂Ωe∩Γqn+1
[Ne]T [tn+1] da . (3.68)
The system of differential equations in (3.61) may be integrated in time using one of
several standard integration methods. The most commonly used such time integrator in
ME280B May 7, 2019
Total and Updated Lagrangian Methods 55
solid mechanics is the Newmark method,2 which is based on a time series expansion of
(un+1, vn+1), such as one may write in the time interval (tn, tn+1]
un+1 = un + vn∆tn +1
2[(1− 2β)an + 2βan+1] ∆t
2n
vn+1 = vn + [(1− γ)an + γan+1] ∆tn
, (3.69)
where the parameters β and γ are chosen such that
0 < γ ≤ 1 , 0 < β ≤ 1
2, (3.70)
and the brackets around each of the matrix terms is dropped for brevity. It is easy to show
that the special case β = 14, γ = 1
2corresponds to the trapezoidal rule. Equation (3.69)1 may
be solved for the acceleration an+1 as
an+1 =1
β∆t2n(un+1 − un − vn∆tn)−
1− 2β
2βan . (3.71)
Substituting the above to (3.61) yields
1
β∆t2nMun+1 +R(un+1) = Fn+1 +M
(un + vn∆tn)
1
β∆t2n+
1− 2β
2βan
︸ ︷︷ ︸known quantities
. (3.72)
The preceding system of non-linear algebraic equations for un+1 can be solved using a stan-
dard iterative method (e.g., the Newton-Raphson method or one of its variants). After un+1
is determined, an+1 and, then, vn+1 are computed from (3.71) and (3.69)2, respectively.
An alternative solution sequence would entail substituting (3.69)1 into (3.61), solving for
the acceleration an+1, then substituting it back into (3.69)1,2 to determine un+1 and vn+1.
In either approach, minor and straightforward modifications suffice to accommodate the
case where Fn+1 depends explicitly on un+1 or the case where Rn+1 depends explicitly on
both un+1 and vn+1.
The above time integration scheme is implicit, in the sense that determining the state at
time tn+1 requires the solution of a system of algebraic equations. A weighted-residual formal-
ization in time can be easily formulated results in deriving the Newmark equation (3.69).3
2N.M. Newmark, “A Method of Computation in Structural Dynamics”, J. Engr. Mech. Div. ASCE,
85:67-94, (1959).3O.C. Zienkiewicz, “A New Look at the Newmark, Houbolt and other Stepping Formulas. A Weighted
Residual Approach”, Earthquake Engin. Struct. Dyn., 5:413-418, (1977).
May 7, 2019 ME280B
56 Incremental Formulations
An explicit integration scheme may be obtained from the Newmark formulae by setting
β = 0. This leads to the equations
un+1 = un + vn∆tn +1
2an∆t
2n
vn+1 = vn + [(1− γ)an + γan+1] ∆tn
. (3.73)
It is again easy to show that in the special case γ = 0.5 one recovers the classical centered-
difference method. The general explicit Newmark integrator is implemented as follows: start-
ing from the semi-discrete form (3.61), one may substitute un+1 from (3.73)1 to get
Man+1 = Fn+1 −R(un + vn∆tn +1
2an∆t
2n) . (3.74)
If M is rendered diagonal,4 then an+1 can be determined without solving equations. Sub-
sequently, the velocity vector vn+1 is computed from (3.73)2. Note that the displacement
vector un+1 is updated through (3.73)1 without using an+1.
As with the implicit Newmark scheme, it is easy to accomodate the case where Fn+1
depends on un+1. On the other hand, if Rn+1 depends on un+1, then the explicit nature
of the time-stepping may not be preserved unless one resorts to a time-shifting modifica-
tion, in which the term Rn+1(un+1, vn+1) is replaced, at an error that depends on ∆tn, by
Rn+1(un+1, vn) or Rn+1(un+1, vn + an∆tn).
Explicit time integration is computationally inexpensive, as there is no need to compute
gradients and solve coupled algebraic equations. However, explicit time integration can only
be conditionally stable, that is, the step-size must be restricted (sometimes severely) to yield
convergent solutions.
The preceding formulation is often referred to total Lagrangian. This means that the mo-
tion and deformation is always measured relative to the original fixed reference configuration
regardless of its magnitude. Alternatively, one may choose to measure the motion and de-
formation at time tn+1 relative to any previously determined (and, therefore, de facto fixed)
configuration at time t = tk, where k = 1, 2, · · · , n. The so-called updated Lagrangian formu-
lation is specifically obtained by measuring the motion and deformation at t = tn+1 relative
to the configuration at time t = tn.5 To this end, consider again a motion χ : R0×R→ E3,
as in Figure 3.11.
4See Appendix H of O.C. Zienkiewicz, R.L. Taylor, and J.Z. Zhu, The Finite Element Method: Its
Basis & Fundamentals, 7th edition, Butterworth-Heinemann, Boston, (2013), for full details on the theory
and practice of mass matrix diagonalization.5K.-J. Bathe, E. Ramm, and E.L. Wilson, “Finite element formulations for large deformation dynamic
analysis”, Int. J. Num. Meth. Engr., 9:353–386, (1975).
ME280B May 7, 2019
Total and Updated Lagrangian Methods 57
X xn
xn+1
unn+1
χn
R0
Rn
Rn+1
Figure 3.11. Configurations for updated Lagrangian formulation
Note that, by definition,
xn = χ(X, tn) (3.75)
and
xn+1 = χ(X, tn+1) , (3.76)
so that a mapping χn : Rn × R→ E3 is defined, such that for any time t
x = χ(X, t) = χ(χ−1tn (xn), t) = χn(xn, t) . (3.77)
Clearly, χn is the relative motion with respect to the (generally deformed) configuration Rn.
It follows that the deformation gradient Fnn+1 at time t = tn+1 relative to the configurationRn
is defined as
Fnn+1 =
∂χn(xn, tn+1)
∂xn
=∂xn+1
∂xn
=∂(xn + un
n+1)
∂xn
= I+∂un
n+1
∂xn
,
where unn+1 = xn+1− xn is the displacement at tn+1 relative to tn, see Figure 3.11. Invoking
the chain rule and recalling (3.77),
∂χ(X, tn+1)
∂X=
∂χn(xn, tn+1)
∂xn
∂χ(X, tn)
∂X(3.78)
or
Fn+1 = Fnn+1Fn . (3.79)
It follows from the above that, since detF > 0 for all (X, t), the same applies to Fnn+1 for all
(xn, t).
May 7, 2019 ME280B
58 Incremental Formulations
Starting from the spatial statement of the weak form in (3.41), one may “pull-back” to
the configuration Rn to find that
∫
Rn
ξnn+1 · ρnann+1 dVn +
∫
Rn
∂ξnn+1
∂xn
·Pnn+1 dVn
=
∫
Rn
ξnn+1 · ρnbnn+1 dVn +
∫
Γqn
ξnn+1 · pnn+1 dAn , (3.80)
where
ξ = ξ(x, t) = ξ(χn t(xn), t) = ξn(xn, t) , (3.81)
hence
ξnn+1 = ξn(xn, tn+1) . (3.82)
Also, for the acceleration and body forces
a = a(x, t) = a (χn(xn, t), t) = an(xn, t) (3.83)
b = b(x, t) = b (χn(xn, t), t) = bn(xn, t) , (3.84)
therefore
ann+1 = an(xn, tn+1) , bn
n+1 = bn(xn, tn+1) . (3.85)
In addition,
ρn = ρn+1Jnn+1 , (3.86)
where Jnn+1 = detFn
n+1, which confirms that ρn in (3.86) is precisely the mass density of the
body at t = tn. Likewise,
dVn =1
Jnn+1
dvn+1 (3.87)
and
nn+1 dan+1 = Jnn+1
(Fn
n+1
)−Tnn dAn , (3.88)
where nn and nn+1 are unit normals to the same material surface at times tn and tn+1,
respectively.
Analogous definitions apply for the traction and stress. In particular, the traction vec-
tor pnn+1 is defined according to
tn+1 dan+1 = pnn+1 dAn , (3.89)
ME280B May 7, 2019
Total and Updated Lagrangian Methods 59
while Pnn+1 is the first Piola-Kirchhoff stress tensor obtained by resolving tractions on the
geometry of the configuration Rn, that is,
Pnn+1 = Jn
n+1Tn+1
(Fn
n+1
)−T. (3.90)
In view of the above and upon recalling (3.44), one may write for a typical Lagrangian
element which at time t = tn occupies the region Ωen
un(xn, tn+1)∣∣Ωe
n= ue n
n+1(xn).=
NEN∑
i=1
N ei u
e ni n+1 = [Ne][ue n
n+1] , (3.91)
and, likewise,
ae nn+1
.= [Ne][ae n
n+1] , ξe nn+1.= [Ne][ξ
e n
n+1] . (3.92)
Furthermore, ⟨∂ξe nn+1
∂xn
⟩
︸ ︷︷ ︸(9×1)
= [Be n]︸ ︷︷ ︸9×(3×NEN)
[ξe n
n+1]︸ ︷︷ ︸(3×NEN)×1
, (3.93)
where
[Be n]︸ ︷︷ ︸9×(3×NEN)
=[[Be n
1 ]︸ ︷︷ ︸(9×3)
[Be n2 ] · · · [Be n
NEN]]
(3.94)
and
[Be ni ]︸ ︷︷ ︸
9×3
=
N ei,1
N ei,2
N ei,3
N ei,2
N ei,3
N ei,1
N ei,3
N ei,1
N ei,2
. (3.95)
Note that here N ei = N e
i (xn) and Nei,j =
∂Nei
∂xnj
=∂Ne
i
∂XAF−1nAj.
Substituting the interpolations in (3.91), (3.92) and (3.93) to the weak form (3.80) applied
to element e yields
[ξe n
n+1]T[[Me][ae n
n+1] + [Re n(ue nn+1)]− [Fe n
n+1]]
= [ξe n
n+1]T
∫
∂Ωen\Γqn
[Ne]T [pnn+1] dAn , (3.96)
May 7, 2019 ME280B
60 Incremental Formulations
where
[Me]︸︷︷︸(3×NEN)×(3×NEN)
=
∫
Ωe0
[Ne]Tρn[Ne] dVn , (3.97)
[Re nn+1]︸ ︷︷ ︸
(3×NEN)×1
=
∫
Ωen
[Be n]T < Pnn+1 > dVn (3.98)
and
[Fe nn+1]︸ ︷︷ ︸
(3×NEN)×1
=
∫
Ωen
[Ne]Tρn[bnn+1] dVn +
∫
∂Ωen∩Γqn
[Ne]T [pnn+1] dAn . (3.99)
Under the action of the assembly operator Ae
over all elements one recovers the semi-
discrete form of the updated Lagrangian formulation as
Mann+1 +Rn(un
n+1) = Fnn+1 , (3.100)
where
[unn+1] =
u1nn+1
u2nn+1
· · ·uNUMNPnn+1
, [ann+1] =
a1nn+1
a2nn+1
· · ·aNUMNPnn+1
. (3.101)
The solution of the above system of non-linear second-order differential equations in time is
achieved as in the total Lagrangian formulation.
The total and updated Lagrangian formulations are completely equivalent, in the sense
that, given the same finite element mesh, they yield the exact same solution. However,
since they involve different unknown vectors ([un+1] for the total Lagrangian and [unn+1] for
the update Lagrangian method), this solution is obtained by solving two different non-linear
algebraic systems. Therefore, it is conceivable that one of the two systems may be more easily
(or, given the limitations of finite arithmetic, more accurately) solvable than the other. This
would form a practical basis for selecting one of the two methods over the other.
3.5 Corotational Methods
When solving problems in solid mechanics, it is sometimes desirable to use coordinate systems
which rotate together with a Lagrangian mesh. This is principally the case when using
structural elements (e.g., bars, beams). By way of motivation, contrast the cases of fixed vs.
corotating coordinate systems for a two-dimensional bar element, as in Figure 3.12.
ME280B May 7, 2019
Corotational Methods 61
R0 R
R
fixed
corotational
Figure 3.12. Fixed vs. corotational coordinate system
Consider the general case shown in Figure 3.13, where one may identify a rotation
field R(x, t), such that the balance laws are expressed at any point x and time t in terms
of the corotating right-hand orthonormal coordinate system e′i. This system is related to
the fixed global coordinate system ei according to
e′i = Rei , (3.102)
for i = 1, 2, 3. Any spatial vector, say v, is resolved relative to ei and e′i as
v = viei = v′ie′i . (3.103)
Rx
ei
e′i
Figure 3.13. General corotational coordinate system
In view of (3.102) and (3.103), one may write
v′i = v · e′i= vjej · e′i = vjej · (Rei) = Rjivj .
(3.104)
May 7, 2019 ME280B
62 Incremental Formulations
In direct notation, equation (3.104) may be stated as
v′ = RTv , (3.105)
where v′ = v′iei is the vector with the corotational components ofv on the global basis (hence,
in general, v′ 6= v).
Similarly, for a spatial tensor, say T, one may write with the aid of (3.102)
T = Tijei ⊗ ej = T ′ije
′i ⊗ e′j
= T ′ij(Rei)⊗ (Rej)
= R(T ′ijei ⊗ ej)R
T ,
(3.106)
which implies that
T′ = RTTR ; T ′ij = RkiTklRlj . (3.107)
Of all terms in the weak form of linear momentum balance (3.41) typically only the
stress-divergence term is resolved using the corotational system (the other terms are resolved
directly in the global system). Ultimately, the stress-divergence term is also rotated back to
the global system before the integration of the semi-discrete equations.
Taking into account (3.105) and (3.106), and using the chain rule, the stress-divergence
term in (3.41) takes the form
∫
R
(∂ξ
∂x
)s
·T dv =
∫
R
(∂ξ
∂ξ′∂ξ′
∂x′
∂x′
∂x
)s
· (RTRT ) dv
=
∫
R
(R∂ξ′
∂x′RT
)s
· (RTRT ) dv
=
∫
R
(∂ξ′
∂x′
)s
·T′ dv .
(3.108)
It is important to emphasize that the coordinate system e′i is not inertial. This means
that if one needs to represent time-dependent effects (such as those of inertia) using a corota-
tional frame, then one needs to explicitly account for the dynamics of the coordinate system.
3.6 Arbitrary Lagrangian-Eulerian Methods
Arbitrary Lagrangian-Eulerian (ALE) methods constitute a creative compromise between
the classical Lagrangian and Eulerian methods. In ALE methods, the mesh undergoes a
ME280B May 7, 2019
Arbitrary Lagrangian-Eulerian Methods 63
motion which is generally different from the motion of the continuum. The mesh motion
is specifically designed to preserve, to the extent possible, the quality of the mesh as the
continuum undergoes significant deformation.
To set up a general ALE method, consider a body B with reference configuration R0 at
time t = 0, and identify this configuration with an initial “mesh configuration”M0 at the
same time, see Figure 3.14. Next, in addition to the usual body motion χ : R0 × R → E3,
X
R
R0 ≡M0
M
x
xM
ei
EA
χ
χM
Figure 3.14. Configurations of ALE method
admit the existence of a “mesh motion” χM :M0 × R→ E3, such that
xM = χM(XM , t) , (3.109)
where, in general, x 6= xM when X = XM , as in Figure 3.14. This mesh motion is sometimes
referred to as “arbitrary”, in the sense that it is generally different from the motion of the
body. The velocity vm of the mesh is now defined as
vM =∂χM(XM , t)
∂t. (3.110)
Two special cases arise naturally: (a) χM = χ, in which case the mesh motion is identical
to body motion and the method is Lagrangian, and (b) χM = i, hence xM = X, in which
case the mesh remains stationary, thereby rendering the method Eulerian.
Attention is focused henceforth to the case χM 6= χ and χM 6= i. Generally, creative
choices for χM can provide serious advantages over pure Lagrangian or Eulerian methods
when the continuum undergoes significant non-homogeneous deformations. The mesh mo-
tion χM is typically chosen to satisfy the following four conditions:
May 7, 2019 ME280B
64 Incremental Formulations
(i) The ALE mesh consists of elements with better aspect ratios and narrower size distri-
bution than the corresponding Lagrangian mesh,
(ii) The element connectivities do not change (otherwise, the procedure yields a new mesh,
and is referred to as remeshing),
(iii) The mesh motion χM is assumed invertible at all times, that is,
JM = det∂χM
∂XM
6= 0 . (3.111)
As argued in Section 1.1, the inverse function theorem enables the unique mapping of a
mesh point xM to its referential counterpart XM and, through it, to a material point x
with which xM coincided at time t = 0. Actually, since JM(XM , 0) = 1 and χM is
assumed smooth, it follows from (3.111) that JM(XM , t) > 0 at all times.
Since the mesh motion is invertible, one may write
vM = vM(XM , t) = vM(χ−1M t(xM), t) = vM(xM , t) , (3.112)
therefore the mesh velocity may be readily expressed in spatial form.
(iv) In certain classes of problems, the body motion and the mesh motion are chosen such
that the normal components of their velocities coincide on the boundary of the body (or
on part of it) at all times. This is generally the case with problems in solid Mechanics
and in free-surface flow problems of fluid mechanics (at least along the free surface).
This restriction is placed to avoid the flow of material outside the mesh, which would
render it unavailable for approximation. At a given time t, let x = xM lie on the
(common) boundary ∂R = ∂M, having outward unit normal n, as in Figure 3.15.
∂R ≡ ∂M
nv
vM
Figure 3.15. Velocity constraint on the boundary of an ALE mesh
ME280B May 7, 2019
Arbitrary Lagrangian-Eulerian Methods 65
The above constraint on the mesh motion implies that
(v − vM) · n = 0 (3.113)
on ∂R ≡ ∂M. Thus, the mesh motion χM is such that it convects with the body
along its boundary. Assuming that the constraint (3.113) is enforced on the whole
(common) boundary of the mesh and body, the kinematic setting for the ALE method
is now depicted in Figure 3.16.
X
R ≡MR0 ≡M0
x
xM
eiEA
χ
χM
Figure 3.16. Configurations of ALE method with coincident body and mesh boundaries
The two main tasks in developing an ALE method are: (a) to formulate and enforce the
balance laws on the ALE finite elements, and (b) to prescribe (in an automated fashion) an
appropriate motion χM .
Consider the first of the above tasks and start with conservation of mass. Let P ⊂ R be
a region occupied by material in the current configuration an let PM ⊂M be a mesh region,
such that P ≡ PM at time t, as in Figure 3.17.
R ≡M
P ≡ PM
n
∂P ≡ ∂PM
Figure 3.17. Mesh region PM in the current configuration
May 7, 2019 ME280B
66 Incremental Formulations
Recall from 1.2.1 that, according to the principle of mass conservation, the rate of change
of the mass of the particles which occupy the material region P at time t is equal to zero.
Starting with the spatial statement of conservation of mass in (1.64), note that
d
dt
∫
P
ρ(x, t)dv =
∫
P
dρ
dtdv
︸ ︷︷ ︸rate of change of mass
for material particles in P
+
∫
P
ρd(dv)
dtdt
︸ ︷︷ ︸rate of change of massdue to change in P
=
∫
P
(ρ+ ρ div v) dv
=
∫
P
(∂ρ
∂t+∂ρ
∂x· v + ρ div v
)dv
=
∫
P
[∂ρ
∂t+ div(ρv)
]dv
=
∫
P
∂ρ
∂tdv
︸ ︷︷ ︸rate of change of mass
in fixed region P
+
∫
∂P
ρv · n da︸ ︷︷ ︸
flux of massthrough boundary ∂P
= 0 , (3.114)
where use is made of (1.10), (1.52), and the divergence theorem. Next, proceed with the
derivation of an ALE statement of mass balance by defining the rate of change of the mass
contained in a mesh region PM at time t as
dMdt
∫
PM
ρM(xM , t)dvM , (3.115)
wheredMdt
denotes the mesh time derivative, that is, the time derivative which keeps the
mesh points fixed.
At this state, it is instructive to consider the different representations of a scalar function,
say f , using material, spatial, or mesh coordinates. Specifically, one may write
f = f(X, t) = fM(XM , t) = f(x, t) = fM(xM , t) . (3.116)
Here, it is clear that fM(xM , t) = f(xM , t) (which means that there is no distinction between
the two spatial representations, hence, in principal, one may drop the subscript “M” from
fM),6 fM(XM , t) 6= f(XM , t) (which is tantamount to observing that at time t the density
of the mesh point which was at XM at time t0 is not equal to that of the material point that
occupied XM at t0), and, alsodMdtfM(XM , t) =
∂fM∂t
. It is also clear from (3.116) that
dMf
dt=
∂fM∂t
=∂fM∂t
+∂fM∂xM
· vM . (3.117)
6It is important to emphasize, in connection to the different representations of a function in (3.116), that
the mesh velocity vM is not equal to the mesh-coordinate representation of the material velocity.
ME280B May 7, 2019
Arbitrary Lagrangian-Eulerian Methods 67
The variables (xM , t) used in ALE methods are often referred to as the mixed variables.
Following the procedure above, one may also write
dMdt
∫
PM
ρM(xM , t) dvM =
∫
PM
dM ρMdt
dvM︸ ︷︷ ︸rate of change of massfor particles in PM
+
∫
PM
ρdM(dvM)
dt︸ ︷︷ ︸rate of change of mass
because of change in PM
=
∫
PM
(dM ρMdt
+ ρ div vM
)dvM
=
∫
PM
(∂ρM∂t
+∂ρM∂xM
· vM + ρ div vM
)dvM
=
∫
PM
[∂ρM∂t
+ div(ρvM)
]dvM
=
∫
PM
∂ρM∂t
dvM +
∫
∂PM
ρvM · n daM . (3.118)
Clearly, since∂ρM∂t
above denotes the rate of change of ρ for fixed xM in the (necessarily)
fixed region PM ≡ P , it follows from (3.118) that∫
PM
∂ρM∂t
dvM =
∫
P
∂ρ
∂tdv = −
∫
∂P
ρv · n da . (3.119)
Therefore, it is concluded from (3.118) and (3.119) that
dMdt
∫
PM
ρM(xM , t) dvM =
∫
∂PM
ρ(vM − v) · n daM , (3.120)
which states that the rate of change of the mass contained in a mesh region equals to the
rate at which mass flows through the boundary of the region. Using the divergence theorem,
one may alternatively express (3.120) as
dMdt
∫
PM
ρM(xM , t) dvM =
∫
PM
div [ρ(vM − v)] dvM . (3.121)
Upon combining (3.118)2 and (3.121), it follows that∫
PM
(dM ρMdt
+ ρ div vM
)dvM =
∫
PM
ρ div(vM − v) dvM +
∫
PM
∂ρM∂xM
· (vM − v) dvM ,
(3.122)
hence, appealing to the localization theorem,
dM ρMdt
+ ρ div v =∂ρM∂xM
· (vM − v) . (3.123)
For consistency, it merits attention to check the two following two cases:
May 7, 2019 ME280B
68 Incremental Formulations
(a) vM = v, where (3.123) reduces to the standard mass balance statementdρ
dt+ρ div v = 0
(since, in this case,dMdt
=d
dt),
(b) vM = 0, where (3.123) reduces to the standard mass balance statement∂ρ
∂t+ div(ρv) = 0
(since, in this case,dMdt
=∂
∂t).
If (vM − v) · n = 0 on ∂M, then setting PM = M it is immediately concluded
from (3.120) that
dMdt
∫
M
ρ(xM , t) dvM = 0 , (3.124)
that is, the total mass of the body is conserved automatically by the ALE method.
The local statements of mass balance (1.66) (in the spatial frame) and (3.123) (in the
mesh frame) jointly imply that
ρ = −ρ div v =dM ρMdt
− ∂ρM∂xm
· (vM − v) . (3.125)
The derivation of linear momentum balance in the mesh frame follows a similar procedure.
First, define the rate of change of linear momentum in a mesh region PM at time t as
dMdt
∫
PM
ρv dvM , (3.126)
and then evaluate the mesh time derivative to conclude that
dMdt
∫
PM
ρvdvM =
∫
PM
dMdt
(ρv) dvM +
∫
PM
ρvdM(dvM)
dt
=
∫
PM
dMdt
(ρv) dvM +
∫
PM
ρv div vM dvM
=
∫
PM
[∂
∂t(ρv) +
∂(ρv)
∂xM
· vM + ρv div vM
]dvM
=
∫
PM
∂
∂t(ρv) dvM +
∫
PM
div(ρv ⊗ vM) dvM
=
∫
PM
∂
∂t(ρv) dvM +
∫
∂PM
ρv(vM · n) daM . (3.127)
ME280B May 7, 2019
Arbitrary Lagrangian-Eulerian Methods 69
Setting PM ≡ P and writing the linear momentum balance statement in Eulerian form as
d
dt
∫
P
ρv dv =
∫
P
ρb dv +
∫
P
t da
=
∫
P
(ρv + ρv div v
)dv
=
∫
P
[∂
∂t(ρv) +
∂(ρv)
∂x· v + ρv div v
]dv
=
∫
P
∂
∂t(ρv) dv +
∫
P
div(ρv ⊗ v) dv
=
∫
P
∂
∂t(ρv) dv +
∫
∂P
ρv(v · n) da . (3.128)
As argued before for mass balance,∫
PM
∂
∂t(ρv) dvM =
∫
P
∂
∂t(ρv) dv =
∫
P
ρb dv +
∫
∂P
t da−∫
∂P
ρv(v · n) da , (3.129)
where use is also made of (3.128). It follows from (3.127), (3.129), and the divergence
theorem that
dMdt
∫
PM
ρv dvM =
∫
PM
ρb dvM +
∫
∂PM
t daM +
∫
∂PM
ρv [(vM − v) · n] daM
=
∫
PM
ρb dvM +
∫
PM
divT dvM +
∫
PM
div [ρv ⊗ (vM − v)] dvM .
(3.130)
Expanding the left- and the right-hand sides of the above leads to
∫
PM
(dM ρMdt
v + ρdM v
dt+ ρv div vM
)dvM =
∫
PM
ρbdvM +
∫
PM
divTdvM
+
∫
PM
v
[∂ρM∂xM
· (vM − v) + ρ div(vM − v)
]dvM +
∫
PM
∂vM
∂xM
ρ(vM − v) dvM (3.131)
or, upon imposing conservation of mass in the form (3.123),∫
PM
ρdM v
dtdv =
∫
PM
ρb dvM +
∫
PM
divT dvM +
∫
PM
∂vM
∂xM
ρ(vM − v) dvM . (3.132)
Hence, the localization theorem, when applied to (3.132) results in the local form of linear
momentum balance
ρdM v
dt= ρb+ divT+
∂v
∂xM
ρ (vM − v) . (3.133)
As with mass conservation, it is instructive to consider the two special cases:
May 7, 2019 ME280B
70 Incremental Formulations
(a) vM = v, where (3.133) reduces to the standard linear momentum balance statement
ρa = ρb+ divT (since, in this case,dMdt
=d
dt),
(b) vM = 0, where (3.133) reduces to ρ
(∂v
∂t+∂v
∂x· v)
= ρb + divT (since, in this case,
dMdt
=∂
∂t).
Angular momentum balance in the mesh frame yields the symmetry of the Cauchy stress,
as usual.
An alternative procedure for the derivation of the balance laws is possible without resort-
ing to any integral statements.7 Indeed, start by recalling (3.117) and use (1.66) to conclude
that
dMρMdt
=∂ρM∂t
+∂ρM∂xM
· vM =∂ρ
∂t+∂ρM∂xM
· vM
= − div(ρv) +∂ρM∂xM
· vM
= −ρ div v − ∂ρ
∂x· v +
∂ρM∂xM
· vM
= −ρ div v − ∂ρM∂xM
· (v − vM) , (3.134)
since ρ(xM , t) = ρM(xM , t) implies that∂ρ
∂t=∂ρM∂t
and also∂ρ
∂x=∂ρM∂x
. Therefore, (3.134)
immediately implies (3.123). Linear momentum balance may be derived analogously.
To derive weak forms of the balance laws (3.123) and (3.133), σ = σM(xM , t) and ξ =
xM(xM , t) be a scalar and vector function at a point xM ofM at time t, respectively. The
weak form of (3.123) amounts to finding ρ ∈ Z+M , such that
∫
M
σ
[dM ρ
dt+ ρ div v − ∂ρ
∂xM
· (vM − v)
]dvM = 0 , (3.135)
for all σ ∈ ZM , where, in analogy to the definitions in (3.15),
ZM = σ :M× R→ R , Z+M = σ :M× R→ R | σ > 0 . (3.136)
7J. Donea, A. Huerta, J.-Ph. Ponthot, and A. Rodriguez-Ferran. Arbitrary Lagrangian-Eulerian methods.
In E. Stein and R. De Borst and T.J.R. Hughes, editors, Encyclopedia of Computational Mechanics, Volume 1:
Fundamentals, chapter 14. John Wiley, New York, 2004.
ME280B May 7, 2019
Arbitrary Lagrangian-Eulerian Methods 71
Likewise, the weighted-residual form of the linear momentum balance statement in (3.133)
including the Neumann boundary conditions at a given time t requires that the displace-
ment u ∈ UM (or velocity v ∈ VM) be such that∫
M
ξ ·[ρdM v
dt− ρb− divT− ∂v
∂xM
ρ(vM − v)
]dvM +
∫
ΓqM
ξ · (t− t) daM = 0 , (3.137)
for all ξ ∈ WM , where
UM =
u :M0 × R→ E3 | det
(I+
∂u
X
)> 0 , ui(X, t) = ui on Γui
,
u(X, 0) = 0 , u(X, 0) = v0 (3.138)
for solids,
VM =v :M× R→ E3 | vi(x, t) = vi on Γui
, v(x, 0) = v0
(3.139)
for fluids, and
WM =ξ :M× R −→ E3 , ξi(x, t) = 0 on Γui
. (3.140)
Note that ΓqM in (3.137) is the part of the exterior boundary ∂M which coincides with
the material boundary Γq at time t, as in Figure 3.18. Using integration by parts and the
R0
R
M
PM
QM
P
PQ
Q
χ
χM
Γq
ΓqM
Figure 3.18. Neumann boundary conditions in ALE methods (R andM are shown apart from
each other for clarity)
divergence theorem, as well as the definition of WM in (3.140), it follows that
∫
M
ξ · divT dvM =
∫
ΓqM
ξ · t daM −∫
M
∂ξ
∂xM
·T dvM . (3.141)
May 7, 2019 ME280B
72 Incremental Formulations
It follows from (3.137) and (3.141) that the weak form of linear momentum balance takes
the form
∫
M
ξ · ρdM v
dtdvM +
∫
M
∂ξ
∂xM
·T dvM
=
∫
M
ξ · ρb dvM +
∫
ΓqM
ξ · t daM +
∫
M
ξ · ρ[∂v
∂xM
(vM − v)
]dvM . (3.142)
There exist two general methodologies for the finite element approximation of (3.135)
and (3.142) in an incremental formulation:
(a) Solve (3.135) and (3.142) simultaneaously at every time step,
(b) Use an operator-split procedure, where a typical ALE time step comprises two sub-
steps:
(b.1) a purely Lagrangian step (where conservation of mass is automatic),
(b.2) if needed, a sub-step in which the mesh convects according to a specified mesh
motion χM (hence, mass balance should be enforced).
In both methodologies, it is important to emphasize that convection of any history vari-
ables (attached, by definition, to material points) requires an additional algorithmic proce-
dure, since the mesh points are not material.
Starting with the first of the above methodologies, interpolate the relevant fields in an
ALE finite element with domain Ωe at time tn+1 as
u(xM , tn+1)∣∣Ωe = ue
n+1(xM).=
NEN∑
i=1
N ei u
ein+1
= [Ne][uen+1] ,
v(xM , tn+1)∣∣Ωe = ve
n+1(xM).=
NEN∑
i=1
N ei v
ein+1
= [Ne][ven+1] ,
dMv
dt(xM , tn+1)
∣∣Ωe = ve
n+1(xM).=
NEN∑
i=1
N ei v
ein+1
= [Ne][ˆven+1] ,
vM(xM , tn+1)∣∣Ωe = ve
M n+1(xM).=
NEN∑
i=1
N ei v
eM in+1
= [Ne][veM n+1] ,
ξ(xM , tn+1)∣∣Ωe = ξen+1(xM)
.=
NEN∑
i=1
N ei ξ
ein+1
= [Ne][ξe
n+1] ,
(3.143)
ME280B May 7, 2019
Arbitrary Lagrangian-Eulerian Methods 73
where N ei = N e
i (xM), i = 1, 2, . . . , NEN. In addition,
ρ(xM , tn+1)∣∣Ωe = ρen+1(xM)
.=
NEM∑
j=1
N ej ρ
ejn+1
= [Ne][ρen+1] ,
dMρ
dt(xM , tn+1)
∣∣Ωe = ρen+1(xM)
.=
NEM∑
j=1
N ej ρ
ejn+1
= [Ne][ˆρen+1] ,
σ(xM , tn+1)∣∣Ωe = σe
n+1(xM).=
NEM∑
j=1
N ej σ
ejn+1
= [Ne][σen+1] ,
(3.144)
where N ej = N e
j (xM), j = 1, 2, . . . NEM, are the element interpolation function for density and
its derivatives. These may, in general, be different from those used for the body motion and
its derivatives.
Substituting the interpolations in (3.143) and (3.144) to (3.135), when the latter is re-
stricted to the domain Ωe of an ALE element, it follows that
[σen+1]
T
∫
Ωe
[Ne]T[Ne][ˆρe
n+1] +([Ne][ρe
n+1]) (
[∆en+1][v
en+1]
)
−([ρe
n+1]T [Λe
n+1]T)[Ne]
([ve
Mn+1]− [ve
n+1])
dvM = 0 , (3.145)
where
[∆en+1] =
[N e
1,1 N e1,2 N e
1,3 · · · N eI,1 N e
I,2 N eI,3 · · · N e
NEN,1 N eNEN,2 N e
NEN,3
](3.146)
and
[Λen+1] =
N e
1,1 · · · N eI,1 · · · N e
NEM,1
N e1,2 · · · N e
I,2 · · · N eNEM,2
N e1,3 · · · N e
I,3 · · · N eNEM,3
. (3.147)
Similarly, substituting the interpolations in (3.143) and (3.144) to (3.142), when the latter
is written for the domain Ωe of an ALE element, it follows that
[ξe
n+1]T
[∫
Ωe
[Ne]T([Ne][ˆve
n+1]) (
[Ne][ρen+1]
)dvM +
∫
Ωe
[Ben+1]
T < Tn+1 > dvM
−∫
Ωe
[Ne]T [bn+1]([Ne][ρe
n+1])dvM −
∫
∂Ωe
[Ne]T [tn+1]daM
−∫
∂Ωe
[Ne]T [Γen+1][N
e][([ve
Mn+1]− [ve
n+1]) (
[Ne][ρen+1]
)dvM
]= 0 . (3.148)
Here,
[Γen+1] =
[[Ne
,1][ven+1] [Ne
,2][ven+1] [Ne
,3][ven+1]
], (3.149)
May 7, 2019 ME280B
74 Incremental Formulations
and all derivatives in [Ben+1] and [Γe
n+1] are taken with respect to xM .
Recalling that σen+1 and ξ
e
n+1 are arbitrary and after assemblying the element equations,
as usual, one recovers a system of coupled non-linear ordinary differential equations in time
with unknowns un+1 (or vn+1) and ρn+1. Note that the mesh motion χM is assumed to be
either prescribed or deduced (an extra task that will be discussed later in this section).
The two sets of equations can be integrated either by an implicit or an explicit method,
along the lines of the corresponding discussion in Section 3.4. In the former case, one may
use, e.g., the Newmark method at the cost of repeated solving of coupled algebraic equations.
In the latter case, one may use, e.g., the centered-difference method, which, for the domain
(tn, tn+1], yields the equations
ρn+1 = ρn + ˆρn∆tn (3.150)
and also
vn+1 = vn +1
2
(ˆvn + ˆvn+1
)∆tn , un+1 = un + vn∆tn +
1
2ˆvn∆t
2n . (3.151)
For the centered-difference method to be free of equation-solving, it is necessary that the
two global associated with the element matrices
[Men+1] =
∫
Ωe
[Ne]T [Ne] dvM , [Men+1] =
∫
Ωe
ρ[Ne]T [Ne] dvM (3.152)
to be rendered diagonal. In addition, it is readily seen from (3.145) and (3.148) that an
explicit velocity estimate ve,0n+1 is needed to estimate the ALE-related flux terms before the
velocity vector is updated using (3.151)1.
Consider next the alternative operator-split algorithm, which consists of a purely La-
grangian step, followed, if needed, by an advective or Eulerian remapping sub-step, where
the nodal/elemental variables are mapped from the previous Lagrangian mesh to the new
ALE mesh, as dictated by the mesh motion. Most often in practice, there are multiple
Lagrangian steps before a single remapping sub-step. Also, the operator-split algorithm fa-
cilitates the automatic determination of the mesh motion based on the distortion encountered
in the Lagrangian mesh between two Eulerian remappings.
The advective remapping sub-step involves the following tasks:
(i) select the mesh regions to be remapped,
(ii) deduce a suitable mesh motion,
(iii) compute the advection of mass and linear momentum,
ME280B May 7, 2019
Arbitrary Lagrangian-Eulerian Methods 75
(iv) compute the advection of history variables.
By way of background, it is instructive to distinguish advective remapping from adaptive
remeshing (or mesh rezoning). In the latter, a new mesh is generated at t = tn+1 not
necessarily with the same connectivities with the old mesh from t = tn, and all variables are
projected from the old mesh. In contrast, Eulerian remapping by the mesh motion χM maps
the mesh from t = tn to its new configuration at t = tn+1, and updates all relevant variables
to account for the transport of material. Hence, Eulerian remapping entails no change in
the element connectivities.
Deferring momentarily the discussion of tasks (i) and (ii) above, concentrate on the last
two tasks of Eulerian remapping. For task (iii), two distinct approaches may be adopted.
In the first approach, the ALE equations are solved in (tn, tn+1] for a fixed motion χ (as
computed from the Lagrangian step) and for prescribed mesh motion χM to obtain updated
values for ρn+1, un+1, vn+1, etc. Indeed, assume that the vectors ρLn+1, uLn+1, v
Ln+1, etc.,
have been computed from the Lagrangian step. Also, assume that the mesh motion in
(tn, tn+1] has been somehow determined. The nodal and elemental quantities will change
Rn ≡Mn
Rn+1
Mn+1
χ = χM
χM (prescribed)
χ (known)
Figure 3.19. Operator-split approach for ALE implementation (Rn+1, resulting from Lagrangian
step and Mn+1, resulting from Eulerian remapping are shown apart from each
other for clarity)
during the remapping sub-step because the nodes are moving independently of the fixed
body, as in Figure 3.19. The new nodal and elemental values can be computed separately for
conservation of mass and balance of linear momentum, using the ALE weak forms introduced
earlier, except that here the density and body motion are already known with reference to
May 7, 2019 ME280B
76 Incremental Formulations
the Lagrangian mesh and need only be calculated on the remapped mesh. In particular, the
discrete counterpart of the mass conservation equation (3.135) becomes
∫
M
σ
[dM ρ
dt+ ρ div vL − ∂ρ
∂xM
· (vM − vL)
]dvM = 0 , (3.153)
and can be integrated in (tn, tn+1] with given vL and vM for ρn+1. Likewise, one may
obtain vn+1 by rewriting the discrete counterpart of (3.142) as
∫
M
ξ · ρLdM v
dtdvM +
∫
M
∂ξ
∂xM
·T dvM
=
∫
M
ξ · ρLb dvM +
∫
ΓqM
ξ · t daM +
∫
M
ξ · ρL[∂v
∂xM
(vM − v)
]dvM , (3.154)
and integrating for vn+1 in (tn, tn+1] with given ρL and vM . Note that the sets of equations
resulting from (3.153) and (3.154) are uncoupled.
Projection methods are addressing the same problem in a purely geometric fashion, which
is also applicable to the advection of history variables in task (iv) above. To introduce these
methods, consider a scalar field f which must be remapped at a given time tn+1. With
“-” mesh “+” mesh
Figure 3.20. Projection method from one mesh (denoted “-”) to another (denoted “+”)
reference to Figure 3.20, let the interpolation of f on the initial mesh (denoted here as the
“-” mesh) be
f− =∑
I
ϕ−I f
−I , (3.155)
where ϕ−I are global interporation functions. After remapping, the same function is interpo-
lated on the new mesh (denoted here as the “+” mesh) as
f+ =∑
I
ϕ+I f
+I , (3.156)
ME280B May 7, 2019
Arbitrary Lagrangian-Eulerian Methods 77
where ϕ+I are the global interpolation functions on the new mesh. Assuming that f− is
given, the goal of a projection method is to determine the nodal values f+I , such that the
error between f− and f+ be minimized in some way. This is generally accomplished by
choosing f+I such that ∫
M
ψ(f+ − f−) dvM = 0 . (3.157)
Here, ψ =∑
I αIψI , where ψI are linearly independent weighting functions and αI are
arbitrary parameters. The projection condition (3.157) leads to the system of linear algebraic
equations
∫
M
ψI(f+ − f−) dvM =
∫
M
ψI
[∑
J
ϕ+J f
+J −
∑
J
ϕ−J f
−J
]dvM = 0 . (3.158)
for all I. This system can be solved for f+J assuming non-singularity of the matrix with
components
∫
M
ψIϕ+J dvM . A typical choice for this projection is ψI = ϕ+
I , which leads to a
classical least-squares problem.
Attention is now turned to tasks (i) and (ii), which pertain to the definition of a suitable
mesh motion. Preliminary to any developments, it is essential to stress that such for such
a motion to be practical, it has to be defined in an automated manner as the ALE solution
evolves. Start by considering a popular heuristic procedure employed for meshes with 4-
node quadrilateral elements.8 The idea here is to maintain a mesh which resembles as
much as possible a uniform mesh, see Figure 3.21. With reference to Figure 3.21, consider a
ideal mesh patch actual mesh patch
A1A2
A3A4
θ1 θ2
θ3θ4
Figure 3.21. Volumetric and shear distortion indicators for a patch of 4-node quadrilaterals
8D.J. Benson. An efficient, accurate, simple ALE method for nonlinear finite element programs. Comp.
Meth. Appl. Mech. Engr., 72:305–350, (1989).
May 7, 2019 ME280B
78 Incremental Formulations
typical node in a mesh of 4-node quadrilaterals, and define a scalar measure Rv of volumetric
distortion around this node as
Rv =minI=1−4
(AI)
maxI=1−4
(AI), (3.159)
where, generally, 0 < Rv ≤ 1 and Rv = 1 for the uniform mesh. A scalar measure Rs is
shear distortion is likewise defined as
Rs = minI=1−4
(sin θi) , (3.160)
where, generally, 0 < Rs ≤ 1 and Rs = 1 for a uniform mesh. Note that Rs detects both
acute and obtuse angles. Subsequently, define an empirical non-negative function f(Rv, Rs),
such that f(1, 1) = 0 (optimal value), and a node I is “flagged” to move when f > f , where
f > 0 is a user-specified constant.
The last outstanding task is to decuce a mesh motion. The simplest possible scheme
would place node I in the coordinate-wise arithmetic mean of the locations of its immediately
neighboring nodes. A more sophisticated (and better performing) scheme is based on the so-
called equipotential relaxation originally propounded by A.M. Winslow for triangular meshes
natural domain (ξ, η) physical domain (x, y)
ξ = const ξ = const
η = const η = const
Figure 3.22. Smooth map between equilateral triangles in the natural space (ξ, η) and general
triangles in the physical space (x, y)
used in hydrocodes.9 Winslow argued that, since any triagular element domain can be
mapped uniquely into a regular equilateral triangle (common for all elements of a mesh) in
the space of natural coordinates (ξ, η), a high-quality mesh can be created by the intersections
of equally-spaced lines of “potential” ξ = constant, η = constant, together with a third set
9A.M. Winslow. Numerical solution of the quasilinear Poisson equation in a nonuniform triangle mesh.
J. Comp. Physics, 2:149–172, (1967).
ME280B May 7, 2019
Arbitrary Lagrangian-Eulerian Methods 79
of lines drawn through the intersection points, as in Figure 3.22. The Laplace-Poisson
equation (being an archetypical example of an elliptic partial differential equation) produces
a smooth set of equipotential lines, provided the boundary conditions and force are smooth.
Therefore, generating a high-quality mesh (or, improving the quality of an existing mesh) is
tantamount to solving the Laplace-Poisson equation for the natural coordinates ξ and η in
the physical domain and defining element edges at equipotential lines of ξ and η. The same
logic applies for quadrilateral meshes mapped from square elements in the natural domain,
see Figure 3.23. Equipotential relaxation has been shown to produce excellent meshes for
natural domain (ξ, η) physical domain (x, y)
ξ = const ξ = const
η = constη = const
Figure 3.23. Smooth map between squares in the natural space (ξ, η) and general quadrilaterals
in the physical space (x, y)
relatively complex two-dimensional domains.10
To derive the equations used for equipotential relaxation, start by considering the general
two-dimensional mapping from the natural to the physical domain in the form
x = x(ξ, η) , y = y(ξ, η) . (3.161)
Since the mapping is assumed to be invertible, one may also write
ξ = ξ(x, y) , η = η(x, y) (3.162)
where the individual functions x, y, ξ, η are to be determined, see Figure 3.24. The two
Laplace-Poisson problems are now defined by
∇2ξ = 0 , ∇2η = 0 , (3.163)
10J.F. Thompson, F.C. Thames, and C.W. Mastin. Automatic numerical generation of body-fitted curvi-
linear coordinate system for fields containing any number of arbitrary two-dimensional bodies, J. Comput.
Phys., 15:299-319 (1974).
May 7, 2019 ME280B
80 Incremental Formulations
where ∇2 = ∂2
∂x2 + ∂2
∂y2. At this stage, it is crucial to observe that the boundary conditions
for the equations in (3.163) are given in the physical space (put differently, it is the domain
in the physical space that is known and that needs meshing). Therefore, since boundary
conditions are enforced on the dependent variable of a differential equation, it is necessary
that (3.163) be expressed inversely, that is, to interchange the dependent and independent
variables. This is possible since x, y are invertible.
x
y
x, y
ξ, ηξ
η
1 2
34
1′ 2′
3′4′
Figure 3.24. Smooth map from a square in the natural space (ξ, η) and a general quadrilateral
domain in the physical space (x, y)
To this end, use the chain rule to write for any differentiable function f = f(x, y),
∂f
∂ξ=
∂f
∂x
∂x
∂ξ+∂f
∂y
∂y
∂ξ∂f
∂η=
∂f
∂x
∂x
∂η+∂f
∂y
∂y
∂η
(3.164)
or, in matrix form, [f,ξ
f,η
]=
[x,ξ y,ξ
x,η y,η
][f,x
f,y
]. (3.165)
Therefore, [f,x
f,y
]=
1
j
[y,η −y,ξ−x,η x,ξ
][f,ξ
f,η
], (3.166)
where j = x,ξy,η − x,ηy,ξ and j 6= 0 owing to the invertibility of x, y. Setting f = ξ and
f = η, it follows from the above equation that
ξ,x =1
jy,η , ξ,y = −1
jx,η (3.167)
and
η,x =1
jy,ξ , η,y = −1
jx,ξ . (3.168)
ME280B May 7, 2019
Arbitrary Lagrangian-Eulerian Methods 81
Repeating this process to calculate expressions for ξ,xx, ξ,yy, η,xx, and η,yy, and substituting
these in (3.163) results in the two coupled non-linear partial differential equations
αx,ξξ − 2βx,ξη + γx,ηη = 0 , αy,ξξ − 2βy,ξη + γy,ηη = 0 , (3.169)
where
α = x2,η + y2,η , β = x,ξx,η + y,ξy,η , γ = x2,ξ + y2,ξ (3.170)
and the overbars have been dropped for brevity. With reference to Figure 3.24, the boundary
conditions take the form
x = x(ξ1, η) on 1′ − 4′ , y = y(ξ1, η) on 1′ − 4′ ,
x = x(ξ2, η) on 2′ − 3′ , y = y(ξ2, η) on 2′ − 3′ ,
x = x(ξ, η1) on 1′ − 2′ , y = y(ξ, η1) on 1′ − 2′ ,
x = x(ξ, η2) on 3′ − 4′ , y = y(ξ, η2) on 3′ − 4′ ,
(3.171)
where x, y are assumed to be known on the boundary of the physical domain. The solutions
of the equations (3.169) provide directly the placements of points (x, y) at the intersection
of equipotential curves ξ = constant and η = constant.
Note that the differential equations (3.169) are more complex than the original Laplacians
in (3.163). However, the domain and boundary conditions of (3.169) are considerably simpler
than those of the original Laplacians.
In practice, equations (3.169) are solved approximately by a finite difference method using
an iterative technique. For example, consider the case of 3-node triangles with equally-spaced
equipotential lines of unit spacing in the natural domain, as in Figure 3.25. Here, it is easy to
ξր, η =const.
ξ =const., ηր
0
12
3
4
5
6
8
Figure 3.25. Finite difference stencil for linear triangles with unit spacing in the natural domain
derive the placement (x0, y0) of the center node in terms of the placement of the neighboring
May 7, 2019 ME280B
82 Incremental Formulations
nodes according to
x,ξ|0.=
1
6[(x1 + 2x6 + x5)− (x2 + 2x3 + x4)] ,
x,η|0.=
1
6[(x2 + 2x1 + x6)− (x3 + 2x4 + x5)] ,
x,ξξ|0.= x6 − 2x0 + x3 ,
x,ηη|0.= x1 − 2x0 + x4 ,
x,ξη|0.=
1
2[(x1 + x6 + x3 + x4)− (x2 + x5 + 2x0)] .
(3.172)
with corresponding equations for y,ξ, y,η, y,ξξ, y,ηη, and y,ξη. To deduce, say, (3.172)1, use
the divergence theorem to write for the shaded region Ω7 of Figure 3.25
∫
Ω7
x,ξ dv2 =
∫
∂Ω7
xnξ da2 . (3.173)
Recalling that all triangles are equilateral with unit side (hence, the area of each triangle is
equal to 1/√3), the above equation implies that, by finite difference approximation,
61√3x,ξ
∣∣∣∣0
.=
x5 + x62
2√3+x6 + x1
2
2√3+ 0− x2 + x3
2
2√3− x3 + x4
2
2√3+ 0 , (3.174)
which reduces to (3.172)1. An even simpler formula for x,ξ|0 may be obtained by considering
the divergence theorem only along the ξ-axis, in which case
x,ξ|0.=
1
2(x6 − x3) . (3.175)
Similar derivations apply to the remaining formulae in (3.172).
ξր, η =const.
ξ =const., ηր
0
123
4
5 6 7
8
Figure 3.26. Finite difference stencil for bilinear quadrilaterals with unit spacing in the natural
domain
ME280B May 7, 2019
Arbitrary Lagrangian-Eulerian Methods 83
For the case of bilinear quadrilateral elements, as in Figure 3.26, use of the divergence
theorem for the shared region ΩBox yields
∫
Ω2
x,ξ dv2 =
∫
∂Ω
xnξ da2 , (3.176)
or
4 x,ξ|0.=
x7 + x82
+x8 + x1
2+ 0 + 0− x3 + x4
2− x4 + x5
2− 0− 0 , (3.177)
hence,
x,ξ|0.=
1
8[(x7 + 2x8 + x1)− (x3 + 2x4 + x5)] , (3.178)
with analogously derived formulae for the remaining derivatives. Again, an even simpler
approximation results from one-dimensional finite differencing, so that
x,ξ|0.=
1
2(x8 − x4)
x,η|0.=
1
2(x2 − x6) ,
x,ξξ|0.= x8 − 2x0 + x4 ,
x,ηη|0.= x2 − 2x0 + x6 ,
x,ξη|0.=
1
4(x1 + x5 − x7 − x3) ,
(3.179)
with corresponding formulae for the derivatives of y.
Upon substituting either (3.172) or (3.179) into the equations in (3.169), one obtains a
non-linear algebraic system of two equations for (x0, y0). In practical computations, Poisson
smoothing is performed using a Gauss-Seidel-like iterative procedure, where the placement
of the nodes is updated successively from one node to the next until all nodes are exhausted.
Several rounds of Poisson smoothing are typically performed until convergence of the node
placements in an appropriate measure.
A variational interpretation of Poisson smoothing is possible and provides useful in-
sights.11 Indeed, it can be easily shown that the solution of (3.163) is attained by the
minimization of the functional
Is =
∫
Ω
(|∇ξ|2 + |∇n|2
)dv . (3.180)
11J.U. Brackbill and J.S. Saltzman. Adaptive zoning for singular problems in two dimensions. J. Comp.
Physics, 46:342–368, (1982).
May 7, 2019 ME280B
84 Incremental Formulations
In this sense, the functional Is measures the variation in mesh spacing along equipotential
curves. Within the variational setting, it is now possible to introduce additional functionals
Io and Iv, as
Io =
∫
Ω
(∇ξ · ∇η)2 dv , Iv =
∫
Ω
wjdV , (3.181)
where w is a given positive function. These measure orthogonality and (weighted) volume
variation, respectively. The minimization of a combined functional of the form I = Is +
c1Io+ c2Iv, where c1, c2 ≥ 0, may be used to simultaneously control mesh quality in all three
measures. In addition, mesh grading (e.g., when one needs to refine the mesh in a specified
area) may be introduced by solving
∇2ξ = f1 , ∇2η = f2 , (3.182)
where the sources f1, f2 are appropriately chosen to effect mesh refinemenet, as needed.
3.7 Eulerian Methods
In computational mechanics, Eulerian finite element methods are used almost exclusively
in applications of fluid dynamics. The weak forms of mass and linear momentum balance
can be readily derived from their ALE counterparts (3.135) and (3.142) by setting the mesh
velocity vM to zero and reinterpreting all mesh time derivatives as partial derivatives with
respect to time at a fixed point in space. This leads to the statements
∫
R
σ
(∂ρ
∂t+∂ρ
∂x· v + ρ div v
)dv = 0 , (3.183)
and
∫
R
ξ · ρ(∂v
∂t+∂v
∂xv
)dv +
∫
R
∂ξ
∂x· T dv =
∫
R
ξ · ρb dv +
∫
Γq
ξ · t da , (3.184)
respectively, with ρ = ρ(x, t) and v = v(x, t), and with admissible spaces and weights defined
accordingly.
The interpolation of velocity and density fields follows (3.143) and (3.144), whereN ei = N e
i (x)
are the element interpolation functions.
A specific application of Eulerian finite elements for incompressible Newtonian fluids is
presented in Section 5.1.
ME280B May 7, 2019
Chapter 4
Solution of Non-linear Field Equations
4.1 Generalities
Recall the semi-discrete form of linear momentum balance (3.61) obtained from a Lagrangian
finite element method or a corresponding form resulting from ALE or Eulerian methods.
With the use of an implicit time integrator, the above system of non-linear ordinary dif-
ferential equations in time gives rise to a system of non-linear algebraic equations for un+1
(typically, in solids) or vn+1 (typically, in fluids). In ALE or Eulerian methods, an additional
system of equations is obtained by imposing conservation of mass and ρn+1 is added to the
list of unknowns to be determined.
By way of background, consider a system of non-linear algebraic equations expressed in
column vector form as
[f(x)] =
f1(x)
f2(x)...
fN(x)
= [0] , (4.1)
where x is a vector in RN (not a position vector!) and f : Ω ⊂ R
N → RN is a non-linear
mapping, in the sense that given any real α, β and any points x1, x2 in Ω, then
f(αx1 + βx2) 6= αf(x1) + βg(x2) . (4.2)
In the interest of simplicity, no brackets are included henceforth to denote general vector
equations, therefore (4.1) would be written simply as
f(x) = 0 . (4.3)
85
86 Solution of Non-linear Field Equations
Next, consider a continuous mapping g : Ω ⊂ RN → R
N , such that x → g(x). A point
x∗ ∈ Ω is called a fixed point of g if
x∗ = g(x∗) . (4.4)
The solution of the system of non-linear algebraic equations in (4.3) can be trivially
reduced to a problem of finding the fixed points of a function, since (4.3) is equivalent to
x = x− cf(x) = g(x) , (4.5)
for any constant c 6= 0. Here, recall that the function g is continuous at x0 ∈ Ω if, given
any ε > 0, there exists a δ = δ(ε) > 0 such that
‖g(x)− g(x0)‖ < ε , (4.6)
for all x ∈ B(x0, δ) = x ∈ Ω | ‖x − x0‖ < δ, where, geometrically, B(x0, δ) is an N -
dimensional ball of radius δ centered at x0, as in Figure 4.3.
x0
x
δ
Figure 4.1. Ball of radius δ centered at x0
Let g be differentiable and its domain Ω be a convex subset of RN , which means that for
any x, y in Ω, the point (1− ω)x+ ωy is also in Ω if 0 < ω < 1, as in Figure 4.2. One may
x y
convex non-convex
Figure 4.2. Convex and non-convex sets
now invoke the mean-value theorem in RN to write for any points x, y in Ω,
‖g(x)− g(y)‖ ≤ supz∈[x,y]
‖J(z)‖‖x− y‖ , (4.7)
ME280B May 7, 2019
Generalities 87
where z is on the line segment connecting x and y, that is, z = x+ω(y−x), 0 ≤ ω ≤ 1, and
Jij(x) =∂gi(x)
∂xj(4.8)
is the Jacobian matrix of g.
Note that any vector norm can be used in the mean-value theorem, such as, for example,
the classical Euclidean 2-norm, defined as
‖x‖2 =(x21 + x22 + · · ·+ x2N
) 12 . (4.9)
Also, recall that all norms in RN are equivalent, in the sense that, given any two such norms
‖ · ‖a, ‖ · ‖b, there exist positive scalar c1 and c2, such that for any x ∈ Ω,
c1‖x‖a ≤ ‖x‖b ≤ c2‖x‖a . (4.10)
Lastly, the natural matrix norm associated with a given vector norm is defined for any given
matrix A (viewed as a linear mapping of a vector x ∈ RN to another vector Ax ∈ R
N) as
‖A‖ = sup‖x‖6=0
‖Ax‖‖x‖ = sup
‖x‖=1
‖Ax‖ . (4.11)
It is noteworthy that the classical equality form of the mean-value theorem in R, that is,
|g(x)− g(y)| = |g′(z)||x− y| , (4.12)
for at least one point z ∈ [x,y], does not generalize to RN .
Returning to (4.7), it is observed that if there is a positive constant λ, such that ‖J(x)‖ ≤λ < 1 for all x ∈ Ω, then clearly
‖g(x)− g(y)‖ ≤ λ‖x− y‖ < < ‖x− y‖ . (4.13)
Under the above condition, the operator g is termed contractive (or a contraction), in the
sense that its application to x − y strictly reduces the distance between the two points
measured by the vector norm. The following theorem provides a useful connection between
contractions and the existence of a solution to (4.3).
Theorem 4.1. (Contraction mapping) Let g map a closed Ω ⊂ RN onto itself, and be
contractive on Ω with constant λ. It follows that:
(a) there exists a unique fixed point x∗ in Ω, and
May 7, 2019 ME280B
88 Solution of Non-linear Field Equations
(b) the fixed point x∗ can be determined by a fixed-point iteration, that is from
x(k+1) = g(x(k)) , k = 0, 1, 2, · · · , (4.14)
starting from any point x(0) in Ω.
Proof. Start with part (b) and note that, in view of (4.13)1,
‖x(k+1) − x(k)‖ = ‖g(x(k))− g(x(k−1))‖ ≤ λ‖x(k) − x(k−1)‖· · ·≤ λk‖x(1) − x(0)‖ = λk‖g(x(0))− x(0)‖ .
(4.15)
Then, for any positive integers m,n, with m > n, (4.15) implies that
‖x(m) − x(n)‖ = ‖x(m) − x(m−1) + x(m−1) − · · · − x(n+1) + x(n+1) − x(n)‖≤ ‖x(m) − x(m−1)‖+ ‖x(m−1) − x(m−2)‖+ · · ·+ ‖x(n+1) − x(n)‖≤ (λm−1 + λm−2 + · · ·+ λn)‖x(1) − x(0)‖= λn(λm−n−1 + λm−n−1 + · · ·+ 1)‖x(1) − x(0)‖
= λn1− λm−n
1− λ ‖x(1) − x(0)‖ ≤ λn
1− λ‖x(1) − x(0)‖ . (4.16)
Hence,
limm,n→∞
‖x(m) − x(n)‖ ≤ limn→∞
λn
1− λ‖x(1) − x(0)‖ = 0 . (4.17)
Thus, x(n) is a Cauchy sequence and since Rn is complete and Ω is closed, it follows that
limn→∞ x(n) = x∗ ∈ Ω.
Returning to part (a), write for the point x∗ obtained above
‖x∗ − g(x∗)‖ = ‖x∗ − g(x(n)) + g(x(n))− g(x∗)‖≤ ‖x∗ − g(x(n))‖+ ‖g(x(n))− g(x∗)‖≤ ‖x∗ − x(n+1)‖+ λ‖x(n) − x∗‖ . (4.18)
Taking the limit as n goes to infinity of both sides of (4.18),
limn→∞
‖x∗ − g(x∗)‖ ≤ limn→∞
‖x∗ − x(n+1)‖+ λ limn→∞
‖x(n) − x∗‖ = 0 , (4.19)
which implies that x∗ is a fixed point.
ME280B May 7, 2019
Newton’s Method and its Variants 89
To show that x∗ is unique, assume, by contradiction, that there exists a second fixed
point y∗ ∈ Ω, such that y∗ 6= x∗. Then,
‖x∗ − y∗‖ = ‖g(x∗)− g(y∗)‖ ≤ λ‖x∗ − y∗‖ < ‖x∗ − y∗‖ (4.20)
which is a contradiction, therefore x∗ = y∗.
If there exists no guarantee that g maps the closed set Ω onto itself, the following theorem
is useful.
Theorem 4.2. Let g : Ω→ RN be contractive on Ω, and assume that x(0) ∈ Ω is such that
S =
x | ‖x− g(x(0))‖ ≤ λ
1− λ‖g(x(0))− x(0)‖
⊂ Ω , (4.21)
where λ is the contraction constant. Then, the fixed-point iteration x(k+1) = g(x(k)) converges
to the unique fixed point x∗ ∈ S.
Proof. For any x ∈ S,
‖g(x)− g(x(0))‖ ≤ λ‖x− x(0)‖ = λ‖x− g(x(0)) + g(x(0))− x(0)‖≤ λ
[‖x− g(x(0))‖+ ‖g(x(0))− x(0)‖
]
≤[
λ2
1− λ + λ
]‖g(x(0))− x(0)‖ =
λ
1− λ‖g(x(0))− x(0)‖ , (4.22)
therefore g(x) ∈ S, thus g maps the closed set S onto itself. Thus, the contraction mapping
x g(x)
ΩΩ
SS
Figure 4.3. Contraction mapping in S
theorem is applicable and yields the desired result.
The preceding analysis can be carried out in the broader framework of Banach spaces
(that is, complete normed linear spaces). However, this is not necessary here since the sub-
sequent development pertains strictly to finite-dimensional spaces (which are, by necessity,
complete).
May 7, 2019 ME280B
90 4.2. THE NEWTON-RAPHSON METHOD AND ITS VARIANTS
4.2 The Newton-Raphson Method and its Variants
The Newton-Raphson method is one of the most widely used iterative methods for solving
non-linear algebraic equations. Its origins may be traced at least as far back as the first
century AD work of Heron of Alexandria, who was clearly aware of the iterative formula
x(k+1) = 12
(x(n) + a
x(k)
)for estimating the square root of a positive real a starting from any
positive guess x(0). It is now widely accepted that the classical Newton-Raphson formula
is due to Raphson, although Newton had earlier employed a related, albeit less general,
formula.1
Recalling (2.13), the idea of the Newton-Raphson method is to approximate f(x) with
its linear part about a point x0 ∈ RN , that is,
f(x) = f(x0) + J(x0)(x− x0)︸ ︷︷ ︸L[f ;x−x0]x0
+O(‖x− x0‖2
)︸ ︷︷ ︸
ignore
= 0 . (4.23)
In the above equation, J(x0) is the (Frechet or Gateaux) derivative of f at x, with components
Jij(x0) =∂fi(x0)
∂xj, (4.24)
and, similarly, J(x0)(x − x0) is the (Frechet or Gateaux) differential at x0 in the direction
x− x0. Then, one may use the iterative formula
x(k+1) = x(k) −[J(x(k))
]−1f(x(k)) (4.25)
to determine, if possible, the solution x∗ = limk→∞ x(k), starting with x(0) as the initial
guess, and assuming that det [J(x(k))] 6= 0 for all iterates k. Equation (4.25) is the classical
Newton-Raphson formula, see Figure 4.4 for a geometric interpretation.
x
x∗
x(k)x(k+1)
f
f(x)
f(x(k))
f ′(x(k)).=
f(x(k))− 0
x(k) − x(k+1)
Figure 4.4. One-dimensional geometric interpretation of the Newton-Raphson method
1For a detailed historical account, see F. Cajori, Historical Note on the Newton-Raphson Method of
Approximation, Amer. Math. Monthly, 18:29–32, (1911).
ME280B May 7, 2019
Newton’s Method and its Variants 91
The three major drawbacks of the Newton-Raphson method are that (a) the deriva-
tive J(x) must be computed explicitly, which is occasionally difficult to do analytically
(although either automatic or symbolic differentiation software may be used to free the pro-
grammer from this task), (b) the condition det [J(x(k))] 6= 0 for all k = 0, 1, 2, · · · is essentialand its violation (or near-violation) results in failure of the method, see Figure 4.5, and (c)
xx∗
x(k)
f
f(x)
Figure 4.5. Failure of the Newton-Raphson method due to singular J(x(k))
convergence is not guaranteed unless the initial guess x(0) is sufficiently close to the solution
(a point that will be expounded upon later in this section), see Figure 4.6.
x xx∗ x∗x(k) x(k)x(k+1)x(k+1) x(k+2)
f f
f(x) f(x)
flip-flop divergence
Figure 4.6. Failure of the Newton-Raphson method due to the large distance of x(k) from the
solution x∗
The Newton-Raphson method can be interpreted as a fixed-point iteration (4.7), where
g(x) = x− [J(x)]−1 f(x) . (4.26)
As an alternative to the Newton-Raphson method, one may define the so-called modified
Newton method, where
g(x) = x−[J(x(0))
]−1f(x) , (4.27)
so that now
x(k+1) = x(k) −[J(x(0))
]−1f(x(k)) , k = 0, 1, 2, · · · (4.28)
May 7, 2019 ME280B
92 Solution of Non-linear Field Equations
x
x∗
x(0)x(1)x(2)
f
f(x)
f(x(0))
Figure 4.7. One-dimensional geometric interpretation of the modified Newton-Raphson method
and J(x(0)) needs to be factorized only once, see Figure 4.7.
The generalized Newton method, where
g(x) = x− [A(x)]−1 f(x) , (4.29)
where A(x) : Ω 7→ RN is an invertible matrix which is typically an approximation to J(x).
The generalized Newton method encompasses a wide class of iterative methods for the so-
lution of (4.3), including the Newton-Raphson, the modified Newton method, as well as all
approximations to Newton-Raphson resulting from numerical differentiation of f .
The preceding contraction mapping theorem may be used to study the convergence of
the Newton-Raphson method. Indeed, in light of (4.26), if the resulting g(x) maps some
closed region Ω to itself and is contractive, then the Newton-Raphson method convergences
to the unique solution x∗ ∈ Ω starting from any point x(0) ∈ Ω. A sharper, and, therefore,
more interesting result, is as follows:2
Theorem 4.3. Let a non-linear operator f : Ω ⊂ RN → R
N be continuously differentiable,
and assume that there exists a point x∗ ∈ Ω, such that f(x∗) = 0 and detJ(x∗) 6= 0.
Then, there is a ρ > 0 such that for all initial guesses x(0) ∈ B(x∗; ρ), the Newton-Raphson
iteration (4.25) is well-defined and converges to x∗.
Proof. Since detJ(x∗) 6= 0, one may find an M > 0, such that
‖ [J(x∗)]−1 ‖ < M . (4.30)
In addition, since J(x) is continuous, then, by the inverse function theorem, so is [J(x∗)]−1
in a neighborhood of x∗. Then, there is a δ > 0, such that
‖ [J(x)]−1 − [J(x∗)]−1 ‖ ≤ M − ‖ [J(x∗)]−1 ‖︸ ︷︷ ︸>0
, (4.31)
2See, e.g., Section 8.1.10 in J.M. Ortega, Numerical Analysis: a Second Course, SIAM, Philadelphia,
1990.
ME280B May 7, 2019
Newton’s Method and its Variants 93
for all x ∈ B(x∗; δ). This, in turn, implies that
‖ [J(x)]−1 ‖ ≤ ‖ [J(x∗)]−1 ‖+M − ‖ [J(x∗)]−1 ‖ = M , (4.32)
that is, [J(x)]−1 is bounded in the ball B(x∗; δ) = x ∈ Ω | ‖x− x∗‖ ≤ δ of radius δ centeredat x∗. Then, for any x(k) ∈ B(x∗; δ) it is concluded that
x(k+1) − x∗ = x(k) −[J(x(k))
]−1f(x(k))− x∗
=[J(x(k))
]−1 J(x(k))[x(k) − x∗]− f(x(k))
. (4.33)
Since J(x) is continuous in B(x∗; δ), one may invoke the fundamental theorem of calculus3
to conclude that
f(x(k)) = f(x∗) +
[∫ 1
0
J(x∗ + ω(x(k) − x∗)
)dω
](x(k) − x∗)
=
[∫ 1
0
J(x∗ + ω(x(k) − x∗)
)dω
](x(k) − x∗) . (4.34)
Substituting the above equation in (4.33) yields
x(k+1) − x∗ =[J(x(k))
]−1∫ 1
0
[J(x(k))− J
(x∗ + ω(x(k) − x∗)
)]dω(x(k) − x∗) . (4.35)
Taking norms of both sides of (4.35) and using standard norm properties, it follows that
‖x(k+1) − x∗‖ ≤ ‖[J(x(k))
]−1 ‖[∫ 1
0
‖J(x(k))− J(x∗ + ω(x(k) − x∗)
)‖dω
]‖x(k) − x∗‖ .
(4.36)
Now, since J(x) is continuous in B(x∗; δ), there exists a ρ, 0 < ρ ≤ δ, such that, for
any x(k) ∈ B(x∗; ρ) and all ω, 0 ≤ ω ≤ 1,
‖J(x(k))− J(x∗ + ω(x(k) − x∗)
)‖ ≤ 1
2M, (4.37)
as in Figure 4.8. Indeed, taking advantage of the aforementioned continuity of J(x), one
may always find a ρ, 0 < ρ ≤ δ, such that
‖J(x(k))− J(x∗ + ω(x(k) − x∗)
)‖ = ‖J(x(k))− J(x∗) + J(x∗)− J
(x∗ + ω(x(k) − x∗)
)‖
≤ ‖J(x(k))− J(x∗)‖+ ‖J(x∗)− J(x∗ + ω(x(k) − x∗)
)‖
≤ 1
4M+
1
4M=
1
2M. (4.38)
3The one-dimensional version of the fundamental theorem of calculus, f(b) = f(a) +∫a
b
df
dxdx readily
generalizes in RN to f(b) = f(a) +
∫ 1
0df(a+ω(b−a))d(a+ω(b−a))
d(a+ω(b−a))dω
= f(a) +[∫ 1
0J (a+ ω(b− a)) dω
](b− a).
May 7, 2019 ME280B
94 Solution of Non-linear Field Equations
x∗
x(k)ρ
δ
Figure 4.8. Balls B(x∗; ρ) and B(x∗; δ) around the solution x∗
Taking advantage of (4.30) and (4.37), it follows from (4.36) that
‖x(k+1) − x∗‖ ≤ M1
2M‖x(k) − x∗‖ =
1
2‖x(k) − x∗‖ . (4.39)
Note that clearly x(k+1) ∈ B(x∗; ρ) and a standard argument of induction leads to
limk→∞‖x(k+1) − x∗‖ = 0 , (4.40)
therefore limk→∞ x(k+1) = x∗.
An additional condition may be included with the original assumptions of the preceding
theorem, whereby it is assumed that there exists a constant λ > 0 such that
‖J(x)− J(y)‖ ≤ λ‖x− y‖ , (4.41)
for all x,y ∈ B(x∗; ρ). Under this condition, J is called Lipschitz continuous in B(x∗; ρ)
with Lipschitz constant λ. Under the additional assumption of Lipschitz continuity for J,
one may start again (4.36) and use (4.30) to conclude that
‖x(k+1) − x∗‖ ≤ ‖[J(x(k))
]−1 ‖∫ 1
0
‖J(x(k))− J(x∗ + ω(x(k) − x∗)
)‖dω
‖x(k) − x∗‖
≤ M
∫ 1
0
λ(1− ω)‖x(k) − x∗‖dω‖x(k) − x∗‖
= Mλ
[ω − ω2
2
]1
0
‖x(k) − x∗‖2 =1
2λM‖x(k) − x∗‖2 . (4.42)
This implies that the Newton-Raphson method converges quadratically in the sense that
‖x(k+1) − x∗‖‖x(k) − x∗‖2 ≤ 1
2λM = constant . (4.43)
ME280B May 7, 2019
Newton’s Method and its Variants 95
The scalar ρ > 0 is referred to as the radius of attraction for the Newton-Raphson
method. Clearly, convergence to x∗ is not guaranteed when starting from x(0) /∈ B(x∗; ρ).
This, in effect, limits the step-size when using the Newton-Raphson method. The radius
of attraction depends on the spatial variation of J around the solution x∗, as depicted in
Figure 4.9. Note that, contrary to statements in some texts, quadratic convergence does
fxfx
x xx∗ x∗
f f
“large” ρ “small” ρ
Figure 4.9. Functions with different radii of attractions for the Newton-Raphson method
not require differentiability of J(x),x ∈ B(x∗; ρ). Indeed, Lipschitz continuity is a stronger
condition than continuity but weaker than differentiability.
The above theorem is local in nature, in the sense that it assumes that the solution x∗
exists, and subsequently asserts that there exists a neighborhood B(x∗; ρ) around x∗, such
that starting with any point x(0) in it, the Newton-Raphson sequence will converge to x∗.
This theorem does not allow for any computable error estimate for the quantity ‖x(k)−x∗‖.A semi-local result (that is, one for in which the existence of a solution x∗ is not assumed
at the outset) is the celebrated Newton-Kantorovich theorem:4
Theorem 4.4. Let a non-linear operator f : RN → RN be differentiable in a convex set
Ω ⊂ RN and its derivative J be Lipschitz continuous with constant λ in Ω. Suppose that
there is a point x(0) ∈ Ω, such that
‖[J(x(0))
]−1 ‖ ≤M , ‖[J(x(0))
]−1f(x(0))‖ ≤ η (4.44)
and
α = ληM ≤ 1
2. (4.45)
Also, assume that B(x(0); ρ∗) ⊂ Ω where
ρ∗ =1
λM
[1− (1− 2α)
12
], (4.46)
4See p. 421 in J.M. Ortega and W. Rheinboldt, Iterative Solution of Nonlinear Equations in Several
Variables, Academic Press, New York, 1970.
May 7, 2019 ME280B
96 Solution of Non-linear Field Equations
see Figure 4.10.
x(0)
ρ∗x∗
Ω
Figure 4.10. Convergence under the Newton-Kantorovich conditions
Then, the Newton-Raphson iteration is well-defined and converges to the unique solu-
tion x∗ of f(x) = 0 in B(x(0); ρ∗). In addition, an error estimate is obtained in the form
‖x∗ − x(k)‖ ≤ (2α)2k
2kλM. (4.47)
Even sharper error estimates can be obtained for ‖x∗ − x(k)‖ by further assuming that
the second derivative of f exists everywhere and is bounded.5
4.3 The Newton-Raphson Method in Non-linear Finite
Elements
The Newton-Raphson method enjoys wide use in non-linear computational mechanics. To
demonstrate its application, consider first the system of ordinary differential equations (3.61)
emanating from the spatial discretization of a Lagrangian finite element method and the cor-
responding algebraic system (3.72) resulting from the application of discrete time integrator.
The latter may be put in the form
f(un+1) =1
β∆t2nMun+1 +R(un+1)−M
(un + vn∆tn)
1
β∆t2n+
1− 2β
2βan
− Fn+1 = 0
(4.48)
To find un+1 using the Newton-Raphson method, start with the initial guess u(0)n+1 = un,
that is, the displacement field of the converged solution at t = tn, and, recalling (4.23), write
0 = f(u(k+1)n+1 )
.= f(u
(k)n+1) +Df(u
(k)n+1)
(u(k+1)n+1 − u
(k)n+1
), (4.49)
5See p. 708 in L.V. Kantorovich and G.P.Akilov, Functional Analysis, Pergamon Press, New York, 1982.
ME280B May 7, 2019
The Newton-Raphson Method in Non-linear Finite Elements 97
where ∆u(k)n+1 = u
(k+1)n+1 − u
(k)n+1 is the (yet unknown) increment of the displacement vector
from iteration (k) to iteration (k + 1). Then,
u(k+1)n+1 = u
(k)n+1 −
[Df(u
(k)n+1)
]−1
f(u(k)n+1) , k = 0, 1, 2, ... . (4.50)
Clearly, the term Df(u(k)n+1)
(u(k+1)n+1 − u
(k)n+1
)is the differential (Frechet or Gateaux) of the
vector f at u(k)n+1 in the direction ∆u
(k)n+1. In order for the Newton-Raphson method to be
employed, the derivative Df(u(k)n+1) must be determined explicitly. This can be done at the
level of the full post-assembly algebraic system, or more conveniently, at the element level
using the original weak forms of the balance laws. To explore the latter option, recall the
discrete linear momentum balance equations (3.59) derived from (3.42), as restricted to the
element e with domain Ωe0, that is,
∫
Ωe0
ξn+1 · ρ0an+1 dV +
∫
Ωe0
∂ξn+1
∂X·Pn+1 dV
=
∫
Ωe0
ξn+1 · ρ0bn+1 dV +
∫
∂Ωe0∩Γq0
ξn+1 · pn+1 dA+
∫
∂Ωe0\Γq0
ξn+1 · pn+1 dA . (4.51)
Next, proceed with obtaining the differential in the direction ∆un+1 of each of the terms in
the above weak form, except for the last one which will not appear in the global balance
statements. For the acceleration term, one may write
D
[∫
Ωe0
ξn+1 · ρ0an+1 dV
](un+1,∆un+1) =
[d
dω
∫
Ωe0
ξn+1 · ρ0 ¨(un+1 + ω∆un+1) dV
]
ω=0
=
∫
Ωe0
ξn+1 · ρ0∆un+1 dV . (4.52)
Assuming that the body forces are known and independent of u, and also that the prescribed
tractions are deformation-independent (that is, they are dead loads), then clearly
D
[∫
Ωe0
ξn+1 · ρ0bn+1 dV +
∫
∂Ωe0∩Γq0
ξn+1 · pn+1 dA
](un+1,∆un+1) = 0 . (4.53)
An additional important assumption made above is that Γq0 is itself independent of u, that
is, the characterization of traction boundaries does not depend on the solution.
Alternatively, it is possible to have deformation-dependent tractions. For example, a fol-
lower load of the form (tn+1 = −pnn+1) may be imposed on Γq0 , so that, upon recalling (1.12)
and (1.75),
pn+1 = −pJn+1F−Tn+1N . (4.54)
May 7, 2019 ME280B
98 Solution of Non-linear Field Equations
nn+1
nn+1
Nn+1
dadA
Figure 4.11. Pressure-type follower traction
where p is a constant pressure, as in Figure 4.11. These traction loads are clearly subject
to non-trivial linearization.
Similarly, it is possible to model body forces which are derived from a potential, such as
gravitational forces which are changing with position, as in Figure 4.12. These also lead to
b
b
Figure 4.12. Body force in a gravitational field
a non-trivial linearization.
Generally, the principal concern in formulating the Newton-Raphson iteration in finite
element methods is to obtain the differential of the stress-divergence term. For the case of
solids, write
D
[∫
Ωe0
∂ξn+1
∂X·Pn+1 dV
](un+1,∆un+1) =
[d
dω
∫
Ωe0
∂ξn+1
∂X·P (un+1 + ω∆un+1) dV
]
ω=0
=
∫
Ωe0
∂ξn+1
∂X·[d
dωP (un+1 + ω∆un+1)
]
ω=0
dV
=
∫
Ωe0
∂ξn+1
∂X·DP (un+1,∆un+1) dV . (4.55)
If the stress depends only on the deformation gradient (as in the case of elastic materials
discussed in Chapter 5), then
DP (un+1,∆un+1) = DP (Fn+1,∆Fn+1) =∂Pn+1
∂Fn+1
∆Fn+1 , (4.56)
ME280B May 7, 2019
The Newton-Raphson Method in Non-linear Finite Elements 99
where
∆Fn+1 =∂∆un+1
∂X. (4.57)
Thus
D
[∫
Ωe0
∂ξn+1
∂X·Pn+1 dV
](un+1,∆un+1) =
∫
Ωe0
∂ξn+1
∂X· ∂Pn+1
∂Fn+1
∂∆un+1
∂XdV . (4.58)
For materials which exhibit rate-dependence, e.g., when P = P(F, F), the differential
DP (un+1,∆un+1) does not capture the rate-dependence and needs to be augmented to
DP(un+1;vn+1,∆un+1; ∆vn+1) . (4.59)
Upon time discretization, this augmented differential is not necessary because the rate quan-
tities can be expressed in terms of the primary variable u through the use of a discrete time
integrator.
Example 4.3.1: By way of example, consider the simple one-dimensional Kelvin-Voigt solid (spring-dashpot in parallel), as in Figure 4.13. Here,
σ = Eε+ ηε , (4.60)
where ε is the infinitesimal strain, σ is the uniaxial stress, E is the Young’s modulus, and η is theviscosity coefficient. Clearly, at t = tn+1,
σn+1 = Eεn+1 + ηεn+1 , (4.61)
therefore
Dσ(εn+1; εn+1,∆εn+1; ∆εn+1) = E∆εn+1 + η∆εn+1 . (4.62)
Upon integrating in time using the backward Euler method, it follows that
σ σ
E
η
Figure 4.13. The Kelvin-Voigt solid
εn+1.= εn +∆tnεn+1 , (4.63)
May 7, 2019 ME280B
100 Solution of Non-linear Field Equations
therefore (4.61) becomes
σn+1.= Eεn+1 + η
εn+1 − εn∆tn
. (4.64)
This, in turn, implies that the differential of the stress is
Dσ(εn+1,∆εn+1) =
(E +
η
∆tn
)∆εn+1 , (4.65)
hence the algorithmic tangent modulus equals E + η∆tn
.
It is important to recognize that in determining the differential of P in an algorith-
mic framework, one needs to first account for all approximations to rate-type quantities on
which P depends. In solid mechanics problems, this leads to a functional expression of the
form
Pn+1 = P(Fn+1, · · · |n, · · · ) , (4.66)
where the notation · · · |n is used to denote generic functional dependency on quantities at
t = tn. Sssuming sufficient smoothness of P, this leads to an algorithmically consistent
differentiation of P as
DP(Fn+1,∆Fn+1) =
(∂P
∂F
)
n+1
∆Fn+1 . (4.67)
The quantity
(∂P
∂F
)
n+1
is referred to as the algorithmic tangent modulus at t = tn+1.
Using the above results, one may obtain the linear part of the weak form (4.51) at u(k)n+1
in the direction ∆(k)un+1, as
∫
Ωe0
ξn+1 · ρ0a(k)n+1 dV +
∫
Ωe0
∂ξn+1
∂X·P(k)
n+1 dV −∫
Ωe0
ξn+1 · ρ0b dV
−∫
∂Ωe0∩Γq0
ξn+1 · pn+1 dA−∫
∂Ωe0\Γq0
ξn+1 · p(k)n+1 dA
+
∫
Ωe0
ξn+1 · ρ0∆u(k)n+1 dV +
∫
Ωe0
∂ξn+1
∂X·(∂P
∂F
)(k)
n+1
∂∆u(k)n+1
∂XdV = 0 , (4.68)
where use is made of (4.52), (4.53), and (4.67). Recalling the implicit Newmark for-
mula (3.71), one may write for element e
a(k)n+1 =
1
β∆t2n
(u(k)n+1 − un − vn∆tn
)− 1− 2β
2βan , (4.69)
ME280B May 7, 2019
The Newton-Raphson Method in Non-linear Finite Elements 101
therefore
∆u(k)n+1 =
1
β∆t2n∆u
(k)n+1 . (4.70)
Substituting (4.70) into (4.69) and introducing the standard element interpolations (3.43)
results in
[ξe
n+1]T[Me][a
e (k)n+1 ] + [Re(u
e (k)n+1)]− [Fe
n+1]− [Fe int (k)n+1 ]
+ [ξe
n+1]T
∫
Ωe0
[Ne]Tρ0
β∆t2n[Ne] dV +
∫
Ωe0
[Be]T
[∂P
∂F
](k)
n+1
[Be] dV
[∆ue (k)
n+1 ] = 0 , (4.71)
where[∂P∂F
](k)n+1
is a 9 × 9 matrix containing the components∂PiA
∂FjB
, consistently with the
convention adopted in (3.46). The quantity inside the large brackets in (4.71) is the element
tangent stiffness matrix Ke (k)n+1 , that is,
Ke (k)n+1 =
∫
Ωe0
[Ne]Tρ0
β∆t2n[Ne] dV
︸ ︷︷ ︸inertial stiffness
+
∫
Ωe0
[Be]T
[∂P
∂F
](k)
n+1
[Be] dV
︸ ︷︷ ︸material stiffness
. (4.72)
Likewise, the quantity inside the small brackets in (4.71) is the residual vector, which quan-
tifies the imbalance of linear momentum at the (k)-th iteration of the Newton-Raphson
method. Upon assemblying (4.71), the global Newton-Raphson equations take the form(Ma
(k)n+1 +R(u
(k)n+1)− Fn+1
)+K
(k)n+1∆u
(k)n+1 = 0 , k = 0, 1, · · · , (4.73)
where K(k)n+1 =A
eK
e (k)n+1 is the global tangent stiffness matrix.
A similar analysis may be carried out starting from the Eulerian form of linear momentum
balance (3.41) restricted to an element e with domain Ωe, such that
∫
Ωe
ξn+1 · ρn+1an+1 dv +
∫
Ωe
∂ξn+1
∂xn+1
·Tn+1 dv =∫
Ωe
ξn+1 · ρn+1bn+1 dv +
∫
∂Ωe∩Γq
ξn+1 · tn+1 da+
∫
∂Ωe\Γq
ξn+1 · tn+1 da . (4.74)
By appealing to mass conservation, it is immediately seen that
D
[∫
Ωe
ξn+1 · ρn+1an+1 dv
](un+1,∆un+1) =
∫
Ωe
ξn+1 · ρn+1Dan+1(un+1,∆un+1) dv
=
∫
Ωe
ξn+1 · ρn+1∆un+1 dv . (4.75)
May 7, 2019 ME280B
102 Solution of Non-linear Field Equations
A similar procedure can be employed for the term involving the body forces.
Neglecting non-linearities due to body force or surface traction, proceed next with the
differentiation of the stress-divergence term in (4.74). First, note that, upon using (1.85)
and the chain rule, one may conclude for a solid mechanics problem that
∂ξ
∂x·T =
(∂ξ
∂XF−1
)·(1
JFSFT
)=
1
J
∂ξ
∂X· (FS) . (4.76)
Taking into account the preceding relation, as well as (2.36), the differentiation of the stress-
divergence term may be obtained as
D
[∫
Ωe
∂ξn+1
∂xn+1
·Tn+1 dv
](un+1,∆un+1) (4.77)
= D
[∫
Ωe0
1
Jn+1
∂ξn+1
∂X· (Fn+1Sn+1) Jn+1 dV
](un+1,∆un+1)
= D
[∫
Ωe0
∂ξn+1
∂X· (Fn+1Sn+1) dV
](un+1,∆un+1)
=
∫
Ωe0
∂ξn+1
∂X·[DF (un+1,∆un+1)Sn+1
]dV +
∫
Ωe0
∂ξn+1
∂X·[Fn+1DS (un+1,∆un+1)
]dV
=
∫
Ωe0
∂ξn+1
∂X·(∂∆un+1
∂XSn+1
)dV +
∫
Ωe0
∂ξn+1
∂X·[Fn+1DS (un+1,∆un+1)
]dV . (4.78)
Assuming, in analogy to earlier discussion on P, that
Sn+1 = S(En+1, · · · |n, · · · ) , (4.79)
one may write
DS (En+1,∆En+1) =
(∂S
∂E
)
n+1
∆En+1 . (4.80)
It is now possible to push-forward to the current configuration the first term on the right-
hand side of (4.77) to find that∫
Ωe0
∂ξn+1
∂X·(∂∆un+1
∂XSn+1
)dV =
∫
Ωe0
(∂ξn+1
∂xn+1
Fn+1
)·[∂∆un+1
∂xn+1
Fn+1
(Jn+1F
−1n+1Tn+1F
−Tn+1
)]dV
=
∫
Ωe
∂ξn+1
∂xn+1
·(∂∆un+1
∂xn+1
Tn+1
)dv , (4.81)
where use is again made of (1.85) and the chain rule. This term gives rise to the geometric
stiffness, namely the part of the tangent stiffness matrix which is due to the change of the
geometry in the current configuration for fixed stress Tn+1.
ME280B May 7, 2019
The Newton-Raphson Method in Non-linear Finite Elements 103
Repeat the above procedure for the second term in the differentiation to conclude that
∫
Ωe0
∂ξn+1
∂X·[Fn+1DS (un+1,∆un+1)
]dV
=
∫
Ωe0
∂ξn+1
∂X·Fn+1
[(∂S
∂E
)
n+1
1
2
(∆FT
n+1Fn+1 + FTn+1∆Fn+1
)]
dV
=
∫
Ωe0
(FT
n+1
∂ξn+1
∂X
)·[(
∂S
∂E
)
n+1
1
2
[(∂∆un+1
∂X
)T
Fn+1 + FTn+1
(∂∆un+1
∂X
)]]dV ,
(4.82)
where use is made of (2.36), (2.41), and the chain rule. Pushing the right-hand side of (4.82)
forward to the current configuration leads to
∫
Ωe0
∂ξn+1
∂X·[Fn+1DS (un+1,∆un+1)
]dV
=
∫
Ωe
1
Jn+1
(FT
n+1
∂ξn+1
∂xn+1
Fn+1
)·[(
∂S
∂E
)
n+1
FTn+1
(∂∆un+1
∂xn+1
)Fn+1
]dv , (4.83)
with the aid of (1.10), the chain rule, and exploiting the symmetry of the fourth-order tensor
∂S
∂Ein its third and fourth legs. This term gives rise to the material stiffness, namely the
part of the tangent stiffness matrix which is due to the non-linearity of the constitutive law
for stress at the fixed geometry of the body in the current configuration. In component
form, one may express the integral in (4.83) as 1J(FiAξi,jFjB)
(SAB
ECDFkC∆uk,lFlD
), where the
subscript “n+ 1” is omitted for brevity.
Now obtain again the linear part of the weak form of linear momentum balance (4.74) at
u(k)n+1 in the direction ∆u
(k)n+1 as
∫
Ωe
ξn+1 · ρ(k)n+1a(k)n+1 dv +
∫
Ωe
∂ξn+1
∂x(k)n+1
·T(k)n+1 dv −
∫
Ωe
ξn+1 · ρ(k)n+1bn+1 dv
−∫
∂Ωe∩Γq
ξn+1 · tn+1 da−∫
∂Ωe\Γq
ξn+1 · tn+1 da
+
∫
Ωe
ξn+1 · ρ(k)n+1∆u(k)n+1 dv +
∫
Ωe
∂ξn+1
∂x(k)n+1
·(∂∆u
(k)n+1
∂x(k)n+1
T(k)n+1
)dv
+
∫
Ωe
1
J(k)n+1
(F
(k)Tn+1
∂ξn+1
∂x(k)n+1
F(k)n+1
)·
(∂S
∂E
)(k)
n+1
F(k)Tn+1
(∂∆u
(k)n+1
∂x(k)n+1
)F
(k)n+1
dv = 0 , (4.84)
May 7, 2019 ME280B
104 Solution of Non-linear Field Equations
where use is made of (4.81) and (4.83). Admitting the standard element interpolations
and recalling the implicit Newmark formula (3.71), as done earlier in this section for the
referential description, it follows that
[ξe
n+1]T[Me][a
e (k)n+1 ] + [Re(u
e (k)n+1)]− [F
e (k)n+1 ]− [F
e int (k)n+1 ]
+ [ξe
n+1]T
∫
Ωe
[Ne]Tρ(k)n+1
β∆t2n[Ne]dv
︸ ︷︷ ︸inertial stiffness
+
∫
Ωe
[Be (k)n+1 ]
T [Ae (k)n+1 ][B
e (k)n+1 ]dv
︸ ︷︷ ︸geometric stiffness
+
∫
Ωe
[Be (k)n+1 ]
T [c(k)n+1][B
e (k)n+1 ]dv
︸ ︷︷ ︸material stiffness
[∆ue (k)n+1 ] = 0 , (4.85)
where [Ae (k)n+1 ] is a 9× 9 array, such that
∂ξ(k)n+1
∂x(k)n+1
· ∂∆u(k)n+1
∂x(k)n+1
T(k)n+1 = [ξ
e
n+1]T [B
e (k)n+1 ]
T [Ae (k)n+1 ][B
e (k)n+1 ][∆u
e (k)n+1 ] (4.86)
in Ωe, while, in light of (4.83), [c(k)n+1] is a 9×9 array whose components in full tensorial form
are
c(k)ijkl =
1
J (k)F
(k)iA F
(k)jB
∂S(k)AB
∂ECD
F(k)kC F
(k)lD , (4.87)
where the subscript “n + 1” is omitted. The symmetry of T is sufficient to deduce that∂ξ
∂x· ∂∆u
∂xT =
∂∆u
∂x· ∂ξ∂x
T, which implies that the matrix [Ae (k)n+1 ] is necessarily symmetric
provided the element interpolation functions for u and ξ are identical. Indeed, it is easy to
show using the ordering convention in (3.46) that
[Ae] =
T11 0 0 T21 0 0 T31 0 0
0 T22 0 0 T32 0 0 T12 0
0 0 T33 0 0 T13 0 0 T23
T12 0 0 T22 0 0 T32 0 0
0 T23 0 0 T33 0 0 T13 0
0 0 T31 0 0 T11 0 0 T21
T13 0 0 T23 0 0 T33 0 0
0 T21 0 0 T31 0 0 T11 0
0 0 T32 0 0 T12 0 0 T22
. (4.88)
ME280B May 7, 2019
The Newton-Raphson Method in Non-linear Finite Elements 105
This, in turn, implies that the geometric stiffness is always symmetric. The material stiffness
is symmetric only when the matrix [c] is symmetric, which, according to (4.87) is contingent
on the symmetry of∂S
(k)AB
∂ECD
in both the first two and the last two pairs of legs. Further
dimensional reduction to the arrays used to represent the geometric and material stiffness
may be introduced by exploiting the fact that∂ξ
∂x·T =
(∂ξ
∂x
)s
·T, owing to the symmetry
of T.
Upon assembly of the element-wise equations stemming from (4.85) one readily recovers
a global system analogous to the one in (4.73).
Equation (4.74) is the starting point for the Newton-Raphson method in fluids. In con-
trast with solids, the inertial term is more complex, since
D
[∫
Ωe
ξn+1 · ρn+1an+1 dv
](vn+1,∆vn+1)
=
∫
Ωe
ξn+1 · ρn+1Dan+1(vn+1,∆vn+1) dv
=
∫
Ωe
ξn+1 · ρn+1∆vn+1 dv
=
∫
Ωe
ξn+1 · ρn+1
(∂∆vn+1
∂t+∂∆vn+1
∂xn+1
vn+1 +∂vn+1
∂xn+1
∆vn+1
)dv .
(4.89)
On the other hand, the term emanating from stress-divergence is much simplified compared
to solids and is given by
D
[∫
Ωe
∂ξn+1
∂xn+1
·Tn+1 dv
](vn+1,∆vn+1) =
∫
Ωe
∂ξn+1
∂xn+1
·DT(vn+1,∆vn+1) dv . (4.90)
This is because the Eulerian nature of the analysis implies that spatial derivatives are not
subject to linearization due to changes in the velocity unless the argument of the derivative
depends itself on the velocity. Of course, the weak form of linear momentum balance must
be supplemented by an equation that enforces mass balance in a weak form.
Revisit the Newton-Raphson method in the form of (4.73) and write it succinctly as
f(u(k)n+1) +K(u
(k)n+1)∆u
(k)n+1 = 0 , k = 0, 1, 2, · · · , (4.91)
with initial guess u(0)n+1 = un. The algorithmic steps for the Newton-Raphson method are
outlined in Algorithm 1 below.
May 7, 2019 ME280B
106 Solution of Non-linear Field Equations
Algorithm 1: Newton-Raphson method
Data: ∆tn, state at time tn, maxit, tol
Set k = 0, u(0)n+1 = un;
while k ≤ maxit
Form residual f(u(k)n+1);
if ‖f(u(k)n+1)‖ < tol then
return u(k)n+1 and rest of state at time tn+1
else
Form and factorize tangent stiffness K(u(k)n+1);
Solve (4.91) for ∆u(k)n+1 and set u
(k+1)n+1 = u
(k)n+1 +∆u
(k)n+1;
Set k ← k + 1;
end
end
The Newton-Raphson iterations are terminated based on a pre-selected stopping criterion.
Such possible criteria include
(a) The residual norm criterion, ‖f(u(k)n+1)‖ < tol, which is shown in Algorithm 1,
(b) The “absolute” error criterion, ‖∆u(k)n+1‖ < tol,
(c) The “relative” error criterion,‖∆u
(k)n+1‖
‖u(k)n+1‖
< tol,
(d) The “energy” norm criterion E(k)n+1 = ∆u
(k)Tn+1 f(u
(k)n+1) < tol. In essence, E(k) is the
work done by the unbalanced forces f(u(k)n+1
)6= 0 going over the displacement incre-
ment ∆u(k)n+1.
The tolerance tol is typically set to be a multiple of the machine epsilon eps, say 10m∗epsfor some positive integer m.
The modified Newton method of (4.28) leads to
f(u(k)n+1) +K(u
(0)n+1)∆u
(k)n+1 = 0 , k = 0, 1, 2, · · · , (4.92)
with initial guess u(0)n+1 = un. This is implemented very similarly to the original Newton-
Raphson method, as shown in Algorithm 2.
ME280B May 7, 2019
The Newton-Raphson Method in Non-linear Finite Elements 107
Algorithm 2: Modified Newton method
Data: ∆tn, state at time tn, maxit, tol
Set k = 0, u(0)n+1 = un;
Form and factorize tangent stiffness K(u(0)n+1);
while k ≤ maxit
Form residual f(u(k)n+1);
if ‖f(u(k)n+1)‖ < tol then
return u(k)n+1 and rest of state at time tn+1
else
Solve (4.92) for ∆u(k)n+1 and set u
(k+1)n+1 = u
(k)n+1 +∆u
(k)n+1;
Set k ← k + 1;
end
end
In highly non-linear problems, it may be necessary to perform a line-search to either
accelerate convergence or to prevent divergence of the Newton-Raphson or modified Newton
iteration. Here, one obtains the increment ∆u(k)n+1 by solving (4.91) or (4.92), as usual.
However, the displacement vector update is defined as
u(k+1)n+1 = u
(k)n+1 + ηk∆u
(k)n+1 , (4.93)
where 0 < ηk ≤ 1 is a scalar parameter. The value of ηk is determined so that the scalar
function
g(η) = ∆u(k)Tn+1 f(u
(k)n+1 + η∆u
(k)n+1) (4.94)
satisfies g(ηk) = 0, that is, given u(k)n+1 and ∆u
(k)n+1, the residual f(u
(k+1)n+1 ) is orthogonal to
∆u(k)n+1, as in Figure 4.14.
u(k)n+1
u(k)n+1 +∆u
(k)n+1
u(k)n+1 + ηk∆u
(k)n+1
f(u(k)n+1)
f(u(k)n+1 +∆u
(k)n+1)
f(u(k)n+1 + ηk∆u
(k)n+1)
Figure 4.14. Schematic of line search
May 7, 2019 ME280B
108 Solution of Non-linear Field Equations
The existence of such an ηk is not guaranteed for any given f , u(k)n+1, and ∆u
(k)n+1. In
practice, one typically seeks to find a ηk, such that
∣∣∣∆u(k)Tn+1 f(u
(k)n+1 + η∆u
(k)n+1)
∣∣∣ ≤ s∣∣∣∆u
(k)Tn+1 f(u
(k)n+1)
∣∣∣ , (4.95)
where s is a user-specified positive parameter less than unity (e.g., s = 0.8). Clearly, line
search involves (relatively inexpensive) residual evaluations for different values of η, until the
above condition is met. A coarse one-dimensional exhaustive search is effective in searching
for an ηk that satisfies (4.95), as shown in Algorithm 3 below.
Algorithm 3: Line search with Newton-Raphson method
Data: u(k)n+1, s, m
Form K(u(k)n+1) and f(u
(k)n+1) and solve (4.91) for ∆u
(k)n+1;
Form f(u(k)n+1 +∆u
(k)n+1);
if∣∣∣∆u
(k)Tn+1 f(u
(k)n+1 +∆u
(k)n+1)
∣∣∣ ≤ s∣∣∣∆u
(k)Tn+1 f(u
(k)n+1)
∣∣∣ thenreturn u
(k+1)n+1 = u
(k)n+1 +∆u
(k)n+1;
else
Set i = 0;
while i ≤ m
Set ηk = (m− i+ 1)/m;
Form f(u(k)n+1 + ηk∆u
(k)n+1);
if∣∣∣∆u
(k)Tn+1 f(u
(k)n+1 + ηk∆u
(k)n+1)
∣∣∣ ≤ s∣∣∣∆u
(k)Tn+1 f(u
(k)n+1)
∣∣∣ thenreturn u
(k+1)n+1 = u
(k)n+1 + ηk∆u
(k)n+1;
else
Set i← i+ 1;
end
end
end
Quasi-Newton methods are iterative schemes with convergence properties generally some-
where between those of the Newton-Raphson and the modified Newton method. The general
idea is to obtain an approximation K(k)
n+1 to the tangent stiffness matrix K(u(k)n+1), which sat-
isfies the condition
K(k)
n+1(u(k)n+1 − u
(k−1)n+1 ) = f(u
(k)n+1)− f(u
(k−1)n+1 ) , (4.96)
ME280B May 7, 2019
The Newton-Raphson Method in Non-linear Finite Elements 109
which is referred to as the quasi-Newton or secant property. Subsequently, one uses the
approximate stiffness to efficiently solve
K(k)
n+1∆u(k)n+1 = −f(u(k)
n+1) (4.97)
for ∆u(k)n+1 and update the displacement vector as
u(k+1)n+1 = u
(k)n+1 +∆u
(k)n+1 . (4.98)
Note that the quasi-Newton property (4.96) shows that the approximate stiffness K(k)
n+1 is
obtained by backward interpolation from the state at the k-th to the (k − 1)-st iterate, see
also Figure 4.15.
f
uu(k)u(k−1)
f(u
(k) )
f(u
(k−1) )
(dfdu
)(k) .=
f(u(k))−f(u(k−1))u(k)−u(k−1)
Figure 4.15. Schematic of quasi-Newton method in one-dimension
There exist many different ways (each corresponding to a different quasi-Newton method)
for constructing K(k)
n+1. Ideally, one expects of a quasi-Newton method that:
(a) The incremental displacement ∆u(k)n+1 can be obtained efficiently, that is, without hav-
ing to factorize K(k)
n+1. This is possible if K(k)
n+1 is obtained from K(k−1)
n+1 as a low-rank
update. In such a case, one may use the Sherman-Morrison-Woodbury formula, ac-
cording to which, given an invertible n× n matrix K and two n×m matrices G and
H with m < n, then
(K+GHT
)−1= K
−1 −K−1G(Im +HTK
−1G)−1
HTK−1
, (4.99)
where Im stands for the m × m identity matrix. Thus, the cost of finding ∆u(k)n+1 is
very small when, say, m = 1 or 2.
(b) The stiffness matrix update preserves properties such as symmetry and positive-definiteness,
if these are present in the initial stiffness K(0)
n+1.
May 7, 2019 ME280B
110 Solution of Non-linear Field Equations
Proceeding constructively, assume that K(0)
n+1 is symmetric and one wants to develop a rank-
one quasi-Newton method that preserves symmetry. Then, by necessity,
K(k)
= K(k−1)
+ zzT , (4.100)
where z is a vector to be determined and the subscript n + 1 is ignored henceforth for the
sake of brevity. Since the quasi-Newton property (4.96) needs to be enforced, it follows that
K(k−1)
∆u(k−1) + zzT∆u(k−1) = f(u(k))− f(u(k−1)) , (4.101)
which implies that
z =1
zT∆u(k−1)
[f(u(k))− f(u(k−1))−K
(k−1)∆u(k−1)
](4.102)
and also
∆u(k−1)TK(k−1)
∆u(k−1) +(zT∆u(k−1)
)2=[fT (u(k))− fT (u(k−1))
]∆u(k−1) . (4.103)
The last two equations lead to
K(k)
= K(k−1)
+1
[f(u(k))− f(u(k−1))]T∆u(k−1) −∆u(k−1)TK
(k−1)∆u(k−1)
[f(u(k))− f(u(k−1))−K
(k−1)∆u(k−1)
] [f(u(k))− f(u(k−1))−K
(k−1)∆u(k−1)
]T. (4.104)
Regrettably, this quasi-Newton method does not necessarily preserve definiteness and, in
fact, can become unbounded due to the term in the denominator of the right-hand side.
Unlike the preceding symmetric case, where satisfaction of the quasi-Newton property
uniquely determines the rank-one update, more assumptions are needed in the unsymmetric
rank-one case
K(k)
= K(k−1)
+ z1zT2 , (4.105)
where z1 and z2 are vectors, and z1 6= z2.
Broyden6 further assumed that,
K(k)q(k−1) = K
(k−1)q(k−1) ; q(k−1)T∆u(k−1) = 0 , (4.106)
6C.G. Broyden, A class of methods for solving nonlinear simultaneous equations, Mathematics of Com-
putation, 19:577-593, (1965).
ME280B May 7, 2019
The Newton-Raphson Method in Non-linear Finite Elements 111
which asserts that the change in f predicted by K(k)
along a direction orthogonal to ∆u(k−1)
is the same with the change predicted by K(k−1)
(stated differently, K(k)
and K(k−1)
are
identical in the direction normal to ∆u(k−1)). It follows from (4.105) that
K(k)q(k−1) = K
(k−1)q(k−1) + z1z
T2 q
(k−1) , (4.107)
so that, in view of (4.106),
z2 = ∆u(k−1) , (4.108)
to within a scalar multiplier. Furthermore, the quasi-Newton property (4.96) necessitates
that [K
(k−1)+ z1∆u(k−1)T
]∆u(k−1) = f(u(k))− f(u(k−1)) , (4.109)
which, in turn, implies that
z1 =1
∆u(k−1)T∆u(k−1)
[f(u(k))− f(u(k−1))−K
(k−1)∆u(k−1)
]. (4.110)
A quasi-Newton method may be implemented as shown in Algorithm 4.
Algorithm 4: Quasi-Newton method
Data: ∆tn, state at time tn, maxit, tol
Set u(0)n+1 = un;
Form K(u(0)n+1) and f(u
(0)n+1) and solve (4.91) for ∆u
(0)n+1;
Set u(1)n+1 = u
(0)n+1 +∆u
(0)n+1;
while k ≤ maxit
Form residual f(u(k)n+1);
if ‖f(u(k)n+1)‖ < tol then
return u(k)n+1 and rest of state at time tn+1
else
Form and factorize K(k)
n+1;
Solve (4.97) for ∆u(k)n+1 and set u
(k+1)n+1 = u
(k)n+1 +∆u
(k)n+1;
Set k ← k + 1;
end
end
The BFGS method (named after Broyden-Fletcher-Goldfard-Shanno) is a quasi-Newton
method intended for problems that lead to symmetric tangent stiffness matrix. The method
makes use of two rank-one updates, such that
K(k)
= K(k−1)
+ z1zT1 + z2z
T2 , (4.111)
May 7, 2019 ME280B
112 Solution of Non-linear Field Equations
where z1 and z2 are vectors with z1 6= z2. The BFGS method is employed as follows:
(i) Assuming K(k−1)
(symmetric) and u(k−1) assumed known, determine the increment
∆u(k−1) from the standard quasi-Newton formula (4.97) as
K(k−1)
∆u(k−1) + f(u(k−1)) = 0 . (4.112)
(ii) Perform a line search along u(k−1), that is determine ηk−1, 0 < ηk−1 ≤ 1, such that
u(k) = u(k−1) + ηk−1∆u(k−1) . (4.113)
(iii) Update the stiffness matrix consistently with a quasi-Newton property (4.96), where,
in particular,
K(k)
= K(k−1)
+
[f(u(k))− f(u(k−1))
] [f(u(k))− f(u(k−1))
]T
ηk−1 [f(u(k))− f(u(k−1))]T∆u(k−1)
−
[K
(k−1)∆u(k−1)
] [K
(k−1)∆u(k−1)
]T
∆u(k−1)TK(k−1)
∆u(k−1). (4.114)
The update formula (4.114) is derived from (4.111) by imposing the quasi-Newton
property (4.96) and also setting z1 to be proportional to the change in the force im-
balance f(u(k))− f(u(k−1)). It can be easily shown that if K(k−1)
is positive-definite,
then K(k)
may or may not be positive-definite but is at least positive semi-definite.
The inverse of the stiffness matrix in (4.114) may be obtained in closed form by two
successive applications of the formula in (4.99). Special case is typically needed to
ensure that the terms in (4.114) are computed in a manner than minimizes errors
associated with algebraic operations involving small numbers.
The BFGS method is implemented as shown in Algorithm 5.
ME280B May 7, 2019
Continuation Methods 113
Algorithm 5: BFGS method
Data: ∆tn, state at time tn, maxit, tol, s
Set u(0)n+1 = un;
Form K(u(0)n+1) and f(u
(0)n+1) and solve (4.91) for ∆u
(0)n+1;
Perform line search according to Algorithm 3 for given s and set
u(1)n+1 = u
(0)n+1 + η0∆u
(0)n+1;
while k ≤ maxit
Form residual f(u(k)n+1);
if ‖f(u(k)n+1)‖ < tol then
return u(k)n+1 and rest of state at time tn+1
else
Form K(k)
n+1 as in (4.114) and factorize;
Solve (4.97) for ∆u(k)n+1;
Perform line search and set u(k+1)n+1 = u
(k)n+1 + ηk∆u
(k)n+1;
Set k ← k + 1;
end
end
4.4 Continuation Methods
Recall the discrete equations of motion for solids, in the form of equation (3.72), and consider
a special case, in which inertial effects may be neglected (rendering the problem quasi-static)
and the external loading is proportional with proportionality factor λ = λ(s). Here, s may
be thought of an “implicit” measure of time and λ is taken to be a smooth function of s.
Under these conditions, the N -dimensional discrete equilibrium equations (3.72) reduce to
f(un+1) = R(un+1)− Fn+1 = f(un+1)− λn+1c ,
where c is a constant force vector in RN and λn+1 effects an alternative parametrization of
loading. Therefore, one may express the equilibrium equation at any load level as
f (u, λ) = f (u)− λc = 0 . (4.115)
Now identify two distinct types of stability points (u, λ) ∈ RN ×R, that is limit points (also
referred to as turning points) and bifurcation points. A one-dimensional representation of
May 7, 2019 ME280B
114 Solution of Non-linear Field Equations
LB
I
II
u u
λλ
limit point bifurcation point
Figure 4.16. Limit and bifurcation points in (u, λ)-space
these points is included in Figure 4.16. The distinction between limit and bifurcation points
was originally suggested by Poincare7.
Some typical examples of stability problems from solid mechanics are shown below.
Example 4.4.1: Snapping of a deep archThis problem involves an arch made of an elastic material. The arch is loaded by a downward-pointingforce at its midpoint. As the load increases, the arch resists the load in compression until it snapsthrough, as shown in Figure 4.17. Subsequently, the arch resists the load in tension. This is an exampleof a structure whose response exhibits two limit loads (one in the pre- and the other in the post-snap-through phase).
Figure 4.17. Snapping of a deep arch
Example 4.4.2: Buckling of a cantilever columnThis is the classical buckling problem of a column made of an elastic material and loaded by a compressiveforce. It is well-known that upon reaching a critical value, the force induces buckling of the column,in which case multiple equilibrium states become possible. This is an example of a bifurcation of theoriginal equilibrium state into a set of possible such states.
7H. Poincare, Sur l’ equilibre d’ une mass fluide animee d’ un mouvement de rotation, Acta Mathematica,
7:259–380, (1885).
ME280B May 7, 2019
Continuation Methods 115
An equilibrium path (or solution path) is defined as a set of functions u(s), λ(s), such that
f (u, λ) = 0 . (4.116)
Let f : RN × R → R
N be continuously differentiable, and also such that for a given s
associated with a solution (u(s), λ(s))
det
[∂f
∂u
(u(s), λ(s)
)]6= 0 . (4.117)
Then, by the implicit function theorem, one may conclude that there exists an open neigh-
borhood U ⊂ RN around u(s), an open neighborhood V ⊂ R around λ(s), and a unique
continuously differentiable function u : V → U , such that
f (u(λ), λ) = 0 , (4.118)
see Figure 4.18.
u
s u(s), λ(s)
Figure 4.18. Solution in neighborhood of point (u(s), λ(s))
Upon differentiating the original equation with respect to s, it follows that
∂f
∂u
du
ds+∂f
∂λ
dλ
ds= 0 . (4.119)
A solution point (u, λ) is termed a stability point if:
(a)
det
[∂f
∂u
(u, λ
)]= 0 , (4.120)
(b)
ker
[∂f
∂u
(u, λ
)]=
n∑
i=1
αiφi , n < N (4.121)
where n < N , αi ∈ R and φi ∈ RN , i = 1, 2, . . . , n, that is,
[∂f
∂u
(u, λ
)]φi = 0 , i = 1, · · · , n , (4.122)
May 7, 2019 ME280B
116 Solution of Non-linear Field Equations
(c)
range
[∂f
∂u
(u, λ
)]=y ∈ R
N | φTi y = 0 , i = 1, · · · , n
. (4.123)
Note that in the context of a finite element analysis,∂f
∂u(u, λ) = K(u, λ). Also, condition (a)
is implied by condition (b), if it is stated explicitly that φi 6= 0, that is, if there exists a
non-trivial null eigenvector of∂f
∂u(u, λ). In addition, condition (c) is implied by condition (b)
when∂f
∂uis symmetric. Indeed, in this case note that, for any x ∈ R
N ,
[∂f
∂u
(u, λ
)]x = y⇒ φ
Ti
[∂f
∂u
(u, λ
)]x = xT
[∂f
∂u
(u, λ
)]Tφi
= xT
[∂f
∂u
(u, λ
)]φi = φ
Ti y = 0 , (4.124)
where use is made of (4.121) and (4.123).
Now consider (4.119) at (u, λ) = (u(s), λ(s)), in the form
∂f
∂u
(u, λ
) [duds
]
s=s
+∂f
∂λ
(u, λ
) [dλds
]
s=s
= 0 (4.125)
and dot this equation with a non-trivial null eigenvector φ of∂f
∂u(u, λ) to conclude that
φT ∂f
∂u
(u, λ
) [duds
]
s=s
+ φT ∂f
∂λ
(u, λ
) [dλds
]
s=s
= 0 . (4.126)
However, owing to condition (c), the first term on the right-hand side of (4.126) vanishes,
therefore
φT ∂f
∂λ
(u, λ
) [dλds
]
s=s
= 0 . (4.127)
One may now distinguish between two cases in connection with (4.127):
(i) φT ∂f
∂λ
(u, λ
)= 0 (or, equivalently, φ
Tc = 0)
In this case, (u, λ) is a bifurcation point. Note that here it is typically true that[dλ
ds
]
s=s
6= 0.
(ii) φT ∂f
∂λ
(u, λ
)6= 0 (or, equivalently, φ
Tc 6= 0)
In this case,(u, λ
)is a limit point. Note that for (4.127) to hold it is now necessary
that
[dλ
ds
]
s=s
= 0.
ME280B May 7, 2019
Continuation Methods 117
In extraordinary circumstances, it is possible that φTc = 0 and
[dλ
ds
]
s=s
= 0. Such a
stability point is simultaneously a bifurcation and limit point.
In the computational investigation of stability, the key tasks are:
(a) To detect stability points and classify them as limit or bifurcation points,
(b) To develop methods which allow the analyst to follow specific non-linear solution paths
past stability points. These are referred to as continuation methods.
(c) To develop methods which effect switching from one solution path to another in bifur-
cation problems. These are referred to as branch-switching methods.
Here, attention is focused on continuation methods. The classical approach is to define an
extended system of the form
f(u)− λc = 0 ,
g(u, λ) = 0 ,(4.128)
where g : RN×R→ R is a given continuously differentiable function. The second of the above
equations is a global constraint equation which completely characterizes the continuation
method.
The extended system may be solved using the classical Newton-Raphson method. Here,
a typical iteration is
[f(u(k)
)− λ(k)c
]+K
(u(k)
)∆u(k) −∆λ(k)c = 0
g(u(k), λ(k)
)+
[∂g
∂u
(u(k), λ(k)
)]T∆u(k) +
[∂g
∂λ
(u(k), λ(k)
)]∆λ(k) = 0
(4.129)
or, put in matrix form,
K(u(k)
)−c[
∂g
∂u
(u(k), λ(k)
)]T ∂g
∂λ
(u(k), λ(k)
)
[∆u(k)
∆λ(k)
]= −
[f(u(k)
)− λ(k)c
g(u(k), λ(k)
)], (4.130)
from where one derives the update relations
u(k+1) = u(k) +∆u(k) , λ(k+1) = λ(k) +∆λ(k) . (4.131)
A crucial observation here is that the tangent stiffness of the extended system of non-linear
equations (4.128) may be non-singular, even when K(u(k)
)is singular. At the same time,
May 7, 2019 ME280B
118 Solution of Non-linear Field Equations
this tangent stiffness matrix of the extended system is neither symmetric (even whenK(u(k)
)
is) nor banded.
If both K(u(k)
)and the stiffness matric of the extended system are non-singular, then
one may solve (4.130)1 for ∆u(k) according to
∆u(k) = −K−1(u(k)
) [f(u(k)
)−(λ(k) +∆λ(k)
)c]
(4.132)
and then substitute the above to (4.130)2 to find that
− ∂g
∂u
(u(k), λ(k)
)K−1
(u(k)
) [f(u(k)
)−(λ(k) +∆λ(k)
)c]
+∂g
∂λ
(u(k), λ(k)
)∆λ(k) = −g
(u(k), λ(k)
). (4.133)
Solving (4.133) for ∆λ(k) leads to
∆λ(k) = −[∂g
∂λ
(u(k), λ(k)
)+∂g
∂u
(u(k), λ(k)
)K−1
(u(k)
)c
]−1
g(u(k), λ(k)
)− ∂g
∂u
(u(k), λ(k)
)K−1
[f(u(k)
)− λ(k)c
], (4.134)
with ∆u(k) subsequently computed by substituting ∆λ(k) from (4.134) in (4.132). This
solution procedure is referred to a bordering algorithm and the quantity inside the square
bracket in (4.134) is known as the Schur complement of K(u(k)
)in the extended tangent
stiffness matrix in (4.130).
It can be shown8 that the necessary and sufficient conditions for the tangent stiff-
ness matrix of the extended system to be non-singular when K(u(k)
)is singular with
ker[K(u(k)
)]= φ 6= 0, that is, when there is a single non-trivial null eigenvector, are
c /∈ range[K(u(k)
)],
∂g
∂u
(u(k), λ(k)
)/∈ range
[KT
(u(k)
)]. (4.135)
Indeed, in this case
K(u(k)
)x 6= c , (4.136)
for any x ∈ RN , therefore if ψ is the null eigenvector of KT , that is,
KT(u(k)
)ψ = 0 , (4.137)
8See Section 4.1.1 in H.B. Keller, Numerical Methods in Bifurcation Problems, Springer-Verlag, Berlin,
1987, and also D.W. Decker and H.B Keller, Multiple limit point bifurcation, J. Math. Anal. Appl., 75:417–
430, (1980).
ME280B May 7, 2019
Continuation Methods 119
it follows, in view of the arbitrariness of x, that
0 = ψTK(u(k)
)x 6= ψTc (4.138)
or
ψTc 6= 0 , (4.139)
as in Figure 4.19.
c
Kx
ψ
Figure 4.19. Relation between the loading vector c and the null eigenvector ψ of KT
Note here that if K is symmetric (hence, the null eigenvectors φ of K and ψ of KT
coincide), then it follows from the above and the defining property of bifurcations that at a
bifurcation point(u, λ
), there is a φ 6= 0, such that
K(u, λ
)φ = 0 , φTc 6= 0 . (4.140)
Therefore, if K is symmetric, then the extended system is singular at any bifurcation point.
Returning to the original condition (4.135)2, note that
KT(u(k)
)y 6= ∂g
∂u
(u(k), λ(k)
), (4.141)
for any y ∈ RN . Therefore, given the null eigenvector φ of K
(u(k)
), it follows as earlier
from the arbitrariness of y that
0 = φTKT(u(k)
)y 6= φT ∂g
∂u
(u(k), λ(k)
)(4.142)
or
φT ∂g
∂u
(u(k), λ(k)
)6= 0 . (4.143)
Now, return to the extended system (4.130) and pre-multiply the first set of equations
with ψT to get
ψTK(u(k)
)∆u(k) −ψTc∆λ(k) = −ψT
[f(u(k)
)− λ(k)c
], (4.144)
May 7, 2019 ME280B
120 Solution of Non-linear Field Equations
so that, in light of (4.137) and (4.139), one may deduce that
∆λ(k) =ψT[f(u(k)
)− λ(k)c
]
ψTc. (4.145)
This implies that (4.130)2 takes the form
K(u(k)
)∆u(k) = −
[f(u(k)
)− λ(k)c
]+ψT[f(u(k)
)− λ(k)c
]
ψTcc . (4.146)
Since K(u(k)
)is singular, one may stipulate that the (unique) solution ∆u(k) of the above
system is of the form
∆u(k) = ∆u(k)p + z(k)φ , (4.147)
where ∆u(k)p is a particular solution obtained by fixing one of the degrees of freedom and
solving for the rest, and z(k) is a scalar to be determined. Substituting (4.147) to (4.130)2,
one gets
[∂g
∂u
(u(k), λ(k)
)]T (∆u(k)
p + z(k)φ)+∂g
∂λ
(u(k), λ(k)
)∆λ(k) = −g
(u(k), λ(k)
), (4.148)
from where z(k) is determined in light of (4.143). Thus, the exact solution(∆u(k),∆λ(k)
)
is obtained as a function of φ, ψ and ∆u(k)p . These arrays can be computed without a
substantially higher effort than required for the original bordering algorithm.
Finally, when ker[K(u(k), λ(k)
)]=
n∑
i=1
αiφi, n > 1, that is, when the tangent stiffness
matrix of the original system possesses more than one non-trivial null eigenvectors, then it
can be shown that the tangent stiffness matrix of the extended system is always singular.
Next, several choices for the constraint function g(u, λ) are reviewed in detail.
4.4.1 Force control
Here, the constraint function takes the form
g(u, λ) = λ− λ , (4.149)
where λ is a specified positive scalar, see Figure 4.20.
In this case, the extended system becomes[K(u(k)
)−c
0T 1
][∆u(k)
∆λ(i)
]= −
[f(u(k)
)− λ(k)c
λ(k) − λ
](4.150)
ME280B May 7, 2019
Continuation Methods 121
λ
λ
u
Figure 4.20. Force control in (u, λ)-space
It follows from the above that
K(u(k)
)∆u(k) = −
[f(u(k)
)− λc
]. (4.151)
Clearly, force control fails when K(u(i))becomes singular. This is consistent with (4.135)2,
whose violation implies that 0T ∈ range[KT
(u(i))], see Figure 4.21.
LB
I
II
u u
λλ
limit point bifurcation point
Figure 4.21. Range of validity of force control in (u, λ)-space
Force control is easily implemented by updating λ = λ(s) at the end of each time step.
4.4.2 Displacement control
Here, the constraint function is defined as
g(u, λ) = eTl u− ul , (4.152)
where ul is a specified scalar and el is an N -dimensional column vector given by
[el] = [0 . . . 0 . . . 1︸︷︷︸l−th entry
. . . 0]T , (4.153)
see Figure 4.22. Essentially, the motion of the body is externally constrained by (4.152)
May 7, 2019 ME280B
122 Solution of Non-linear Field Equations
λ
ul ul
Figure 4.22. Displacement control in (ul, λ)-space
so that the extended system be non-singular and the equilibrium path is traced for the
prescribed displacement ul.
Under displacement control the extended system (4.130) becomes[K(u(k)
)−c
eTl 0
][∆u(k)
∆λ(k)
]= −
[f(u(k)
)− λ(k)c
eTl u(k) − ul
](4.154)
It follows from (4.135)2 that displacement control fails if there is a vector x ∈ RN , such
that K(u(k)
)x = el, in which case displacement control along el is unable to remove the
singularity of the extended system, see Figure 4.23.
ul
λ
Figure 4.23. Range of validity of displacement control in (ul, λ)-space
The implementation of displacement control in finite element codes is trivial, since it
merely amounts to imposing displacement boundary conditions, as routinely done for Dirich-
let boundary conditions.
More generally, one may impose a set of M constraints of the form
gm(u, λ) = eTm(u)− um , m = 1, 2, . . . ,M , (4.155)
where M < N . This would lead to an extended system with N +M equations and M load
vectors λm along directions cm.
ME280B May 7, 2019
Continuation Methods 123
4.4.3 Arc-length control
The idea of arc-length methods is impose a constraint to the arc-length s of the equilibrium
path, defined as
s =
∫ds =
∫ √duTdu+ τ 2dλ2 , (4.156)
where τ is a scaling parameter that effects consistency of units in (4.156), see Figure 4.24.
ds
udu
τλ
τdλ
Figure 4.24. Arc-length in (u, τλ)-space
Arc-length methods fix an s = s0 at a given solution point (un, τλn), and seek to determine
a new equilibrium point under this constraint. Various arc-length methods are possible by
means of different approximations to the actual arc-length of the solution curve (since the
latter is not known in advance).
In the spherical arc-length method around the solution point (un, τλn), the constraint
function takes the form
g(u, λ) = (u− un)T (u− un) + τ 2(λ− λn)2 − s20 , (4.157)
see Figure 4.25.
uun
τλ
τλns0
Figure 4.25. Spherical arc-length in (u, τλ)-space
May 7, 2019 ME280B
124 Solution of Non-linear Field Equations
A simplified version of the spherical arc-length method is obtained by linearizing the
constraint function along the tangent [tn] to the solution path at (un, τλn), where
[tn] =
[unt − un
τ(λnt − λn) ,
](4.158)
with magnitude s0, hence,
(unt − un)T (unt − un) + τ 2(λnt − λn)2 = s20 , (4.159)
see Figure 4.26. If follows that the constraint function at (un, τλn) becomes
tn
uun unt
τλ
τλn
τλnt
Figure 4.26. Normal-plane arc-length in (u, τλ)-space
g(u, λ) = (unt − un)T (u− un) + τ 2(λnt − λn)(λ− λn)− s20 . (4.160)
In view of (4.159), one may also write
g(u, λ) = (unt − un)T (u− un) + τ 2(λnt − λn)(λ− λn)
− (unt − un)T (unt − un) + τ 2(λnt − λn)(λnt − λn)
= (unt − un)T (u− unt) + τ 2(λnt − λn)(λ− λnt) , (4.161)
so that g(u, λ) = 0 implies that [tn] is orthogonal to
[u− unt
τ(λ− λnt)
], see Figure 4.26. Owing
to this property, the preceding method is referred to as the normal-plane arc-length method.
Given (4.160), the extended system now takes the form
[K(u(k)
)−c
(unt − un)T τ 2(λnt − λn)
][∆u(k)
∆λ(k)
]= −
[f(u(k)
)− λ(k)c
g(u(k), λ(k))
]. (4.162)
ME280B May 7, 2019
Continuation Methods 125
Its main advantage over its spherical arc-length counterpart is that all extra terms in the
extended stiffness matrix are constant.
The implementation of normal-plane arc-length in connection with the Newton-Raphson
method is summarized in Algorithm 6 below, following the so-called Riks-Wempner method9
Algorithm 6: Riks-Wempner method
Data: state at time tn, s0
Solve K(un)∆u∗n = ∆λ∗nc
for any ∆λ∗n > 0 (ascending part of curve) or ∆λ∗n < 0 (descending part of curve);
Use (4.158) to write [t∗n] =
[∆u∗
n
τ∆λ∗n
];
Compute [tn] = α∗[t∗n], where α∗ enforces (4.159);
Solve (4.162) using the Newton-Raphson method;
Note that the preceding algorithm requires that K(un) be non-singular and also that
the extended system in (4.159) be non-singular during the Newton-Raphson iterations. As
already argued, the latter is possible even when tracing bifurcation points, as long as the
continuation method does not “hit” the singularity at one the iterations.
A simple brute-force computational method for detecting and classifying stability points
with the aid of arc-length continuation is described in Algorithm 8 below.
9E. Riks, The application of Newton’s method to the problem of elastic stability. ASME J. Appl. Mech.,
39:1060–1065, (1972).
May 7, 2019 ME280B
126 4.5. COMPUTATIONAL TREATMENT OF CONSTRAINTS
Algorithm 7: A method for detecting and classifying stability points
Data: state at time tn, s0, tol, tolk, tolp, maxit
Form K(un);
Perform an LU decomposition on K(un) and store the signs of the diagonal terms in U;
while k ≤ maxit
Use arc-length with s = s0 to advance solution to (un+1, λn+1);
Perform an LU decomposition on K(un+1) and store the signs of the diagonal terms
in U;
if no change in the signs of the diagonal terms of U from n to n+ 1 then
return (un+1, λn+1) and advance time (no stability points detected);
else if change to exactly one diagonal term of U from n to n+ 1 thenUse bisection to bracket stability point until u such that |detK(u)| .= tolk
Use subspace iteration to find the null eigenvector φ of K(u)
if∣∣∣φT
c∣∣∣ ≥ tolp then
(u, λ) is a limit point
else
(u, λ) is a bifurcation point
end
elseSet k ← k + 1, s0 ← s0/2 and repeat
end
end
Note, in connection to the preceding algorithm, that the LU decomposition of a square
matrix such as K is unique as long as the matrix is non-singular.
There exist alternative procedures for the determination of stability points, which are
not discussed here.
4.5 Computational Treatment of Constraints
The motion of a continuum is often subject to constraints. Typical examples of such con-
straints are noted below.
Example 4.5.1: IncompressibilityCertain rubber-like solid materials are practically incompressible, that is their deformation is always
ME280B May 7, 2019
Computational Treatment of Constraints 127
volume-preserving. In light of (1.10), this implies that their motion is subject to the constraint
detF = 1 , (4.163)
or, equivalently in terms of the principal stretches in (1.30),
λ1λ2λ3 = 1 . (4.164)
Many fluids (especially, liquids) are also practically incompressible. In this case, taking into consid-eration (1.52), it is readily concluded that an incompressible fluid is subject to the constraint
trL = divv = 0 . (4.165)
Example 4.5.2: Inextensibility along a given directionSome solids are reinforced by inextensible rod-like structures along a given direction, say M, in thereference configuration. In this case, no stretching is permitted along M, which leads to the constraint
λ2M = M ·CM = 1 . (4.166)
Example 4.5.3: Impenetrability of matterAccording to this fundamental postulate, two material points may not occupy the same position in spaceat the same time. In the context of a deformable solid coming to contact with a rigid foundation, as inFigure 4.27, the gap (penetration) function g, defined as
g = (xf − x) · n (4.167)
must satisfy the constraint conditiong ≥ 0 . (4.168)
nx
xf
rigid foundation
Figure 4.27. Contact of deformable solid with rigid foundation
Constraints are classified in various ways. Here, attention is placed on the distinction
between:
May 7, 2019 ME280B
128 Solution of Non-linear Field Equations
(a) equality (bilateral) and inequality (unilateral) constraints
Equality (resp. inequality) constraints are mathematically expressed by equalities
(resp. inequalities). For instance, incompressibility is an equality constraint, while
impenetrability is an inequality constraint.
(b) internal and external constraints
Internal constraints are expressed in terms of measures of deformation (or rate of
deformation), while external constraints cannot. For instance, in view of (4.163),
(4.165) and (4.168), incompressibility is an internal constraint, while impenetrability
is an external constraint.
(c) rheonomous and scleronomous constraints
Rheonomous constraints depend explicitly on time, while scleronomous constraints do
not. For instance, all time-dependent (resp. time-independent) Dirichlet boundary
conditions may be thought of as rheonomous (resp. scleronomous) constraints.
Consider a bilateral scleronomous constraint written for a solid in the form
c (u(X, t)) = 0 (4.169)
and assumed to apply on a part C0 of either the domain R0 or the external boundary ∂R0.
While the constraint is expressed in (4.169) in terms of the displacement field, it may be
thought of as either internal or external.
When the motion of the solid is subject to (4.169), the space of constraint admissible
displacements becomes
Uc =u : R0 × R→ R
3 | u (X, t) = u (X, t) on Γu , c(u) = 0 on C0 × R.
(4.170)
In the context of finite element approximations, one may choose to construct subsets Uchof Uc, and satisfy (weakly) the balance laws over Uch. Such a so-called primal method may
be impractical if the constraint condition is difficult to enforce by the choice of admissible
displacements.
An alternative, so called dual method retains the original space of admissible functions Uand introduces a new Lagrange multiplier field. To this end, recall that the constraint (4.169)
is imposed on kinematic variables and rewrite the linear momentum balance equations (3.42)
ME280B May 7, 2019
Computational Treatment of Constraints 129
as∫
R0
ξ ·ρ0a dV +
∫
R0
∂ξ
∂X·P dV =
∫
R0
ξ ·ρ0b dV +
∫
Γq0
ξ ·p dA−∫
C0
Dc(u, ξ)·Λ dV , (4.171)
where Λ = Λ(X, t) is a Lagrange multiplier field on C0×R. Note that the term in (4.171)
associated with the constraint could be a surface rather than a volume integral. Also note
that, since c(u) = 0 for all admissible displacement fields, then
Dc(u, ξ) = 0 , (4.172)
and the Lagrange multiplier field Λ which enforces the constraint is workless, in the sense
that ∫
C0
Dc(u, ξ) ·Λ dV = 0 . (4.173)
The preceding observation does not generally hold for rheonomous constraints.
Example 4.5.4: Incompressibility in dual methodsTaking into account (2.60) and (4.163), the constraint integral term in (4.173) becomes
∫
R0
Dc(u, ξ) ·Λ dV =
∫
R0
J∂ξ
∂X· F−TΛ dV . (4.174)
One may now combine this term with the stress-divergence term in (4.171) to conclude that∫
R0
∂ξ
∂X·P dV +
∫
R0
J∂ξ
∂X· F−TΛ dV =
∫
R0
∂ξ
∂X· (P+ ΛF−T ) dV (4.175)
=
∫
R
∂ξ
∂x· (T+ Λi) dv , (4.176)
where use is made of (1.73) and the chain rule. The preceding relation shows that the scalar Lagrangemultiplier field Λ may be interpreted as a pressure. Ideally, the Cauchy stress T should be decomposed toa pressure-independent part (which remains constitutively specified) and a pressure part, which coincideswith the Lagrange multiplier Λ.
A similar analysis applies to incompressibility in fluids, where, according to (4.165),∫
RDc(v, ξ) ·Λ dv =
∫
Rtr
(∂ξ
∂x
)Λ dv =
∫
R
∂ξ
∂x· Λi dv . (4.177)
Example 4.5.5: Tied sliding in dual methodsConsider a special case of the impenetrability constraint in which a given part C is a deformable body’sboundary is in persistent contact with a rigid foundation, but is allowed to slide on it. This is the case oftied sliding. Now, the constraint condition (4.168) is imposed as an equality. Hence, in light of (4.167),the constraint integral term in (4.173) becomes
∫
C0
Dc(u, ξ) ·Λ dA =
∫
C0
ξ · Λn dA , (4.178)
May 7, 2019 ME280B
130 Solution of Non-linear Field Equations
therefore the scalar Lagrange multiplier field may be interpreted as the magnitude of the normal traction(pressure) necessary to resist penetration or separation of the two bodies.
The extended equations of motion now read
∫
R0
ξ · ρ0a dV +
∫
R0
∂ξ
∂X·P dV
−∫
R0
ξ · ρ0b dV −∫
Γq0
ξ · p dA+
∫
C0
Dc(u, ξ) ·Λ dV = 0 , (4.179)
∫
C0
θ · c(u)dV = 0 , (4.180)
where θ = θ(X, t) is an arbitrary weight field on C0 × R.
In discrete form, one may write
∫
C0
Dc(u, ξ) ·Λ dV.= ξ
Td(u)Λ ,
∫
C0
θ · c(u) dV .= θ
Tg(u) ,
(4.181)
where u, ξ ∈ RN and where Λ, θ ∈ R
M with N > M , consequently d : RN 7→ RN ×R
M and
g : RN 7→ RM . Taking into account (4.181), one may write
∫
C0
θ ·Dc(u, ξ) dV.= θ
TDg(u)ξ = ξ
TDTg(u)θ
= ξTd(u)θ
T,
(4.182)
therefore
d(u) = DTg(u) . (4.183)
Given (4.179), (4.180), and (4.181), the discrete equations of motion become
ξT[f(u) + d(u)Λ
]= 0 ,
θTg(u) = 0 ,
(4.184)
or, after accounting for the arbitrariness of ξ and θ,
f(u) + d(u)Λ = 0 ,
g(u) = 0 .(4.185)
ME280B May 7, 2019
Computational Treatment of Constraints 131
Note that in the dual method, the original N equations for the unconstrained problem with u
as unknowns are replaced by N +M equations for the constrained problem with u and Λ
as unknowns.
To use the Newton-Raphson method in (4.185), write
f(u(k)
)+K
(u(k)
)∆u(k) + d
(u(k)
)Λ
(k)+DTd
(u(k)
)Λ
(k)∆u(k) + d
(u(k)
)∆Λ
(k)= 0
g(u(k)
)+Dg
(u(k)
)∆u(k) = 0
(4.186)
where, as usual,
f(u(k) +∆u(k)
) .= f
(u(k)
)+K
(u(k)
)∆u(k) ,
D
[∫
C0
Dc(u, ξ) ·Λ(k) dV
] (u(k),∆u(k)
) .= ξ
TDTd
(u(k)
)Λ
(k)∆u(k)
(4.187)
and also, upon recalling (4.183),
D
[∫
C0
θ · c(u)dV] (
u(k),∆u(k)) .
= θTDg(u(k))Λ
(k)∆u(k)
.= ∆u(k) T
d(u(k)
)θ = θ
TdT(u(k)
)∆u(k) . (4.188)
The linearized system of equations (4.186) may be put in bordered matrix form as[K(u(k)
)+DTd
(u(k)
)Λ
(k)d(u(k)
)
dT(u(k)
)0
][∆u(k)
∆Λ(k)
]= −
[f(u(k)
)+ d
(u(k)
)Λ
(k)
g(u(k)
)]. (4.189)
The extended tangent stiffness matrix in (4.189)above is symmetric provided bothK(u(k)
)
and DTd(u(k)
)Λ
(k)are symmetric. However, this stiffness matrix is not banded. In fact,
given its structure, a special equation solver is required to take care of zero diagonals, as the
classical Gauss elimination procedure without pivoting would generally fail.
Dual methods enable the use of the original (unconstrained) space of admissible dis-
placements U and its finite element subspaces Uh, which is convenient. The price for this
convenience is in the need to solve an extended system of equations, including the Lagrange
multipliers Λ ∈ L, where L is the space of admissible multiplier fields. Also, the choice of
discrete admissible fields Uh and Lh is subject to substantial restrictions and one needs to
exercise caution in order to ensure solvability of (4.189).
In the special case of linear elastic response subject to a constraint which is also linear
in u (or F), the matrix form of the extended system becomes[K d
dT 0
][u
Λ
]= −
[f (0)
g (0)
], (4.190)
May 7, 2019 ME280B
132 Solution of Non-linear Field Equations
where K and d are constant matrices. This would be the case for the problem of incom-
pressible linear elasticity.
A corresponding analysis applies to a constrained problem in fluid dynamics.
The treatment of constraints by Lagrange multipliers in continuum mechanics is often
motivated by the formalization of constrained non-linear optimization problems of the general
form: find the extremum of a functional G(u), subject to constraint condition c(u) = 0.
In solid mechanics, under severe restrictions, there exist variational theorems, according to
which the equations of motion are derived as extrema of a functional G(u) to zero. For
example, in linear elastostatics, one may define the total potential energy functional G(u) of
a body occupying the region R0 as
G(u) =1
2
∫
R0
σ(u) · ε(u) dV −∫
R0
u · ρ0b dV −∫
Γq0
u · p dV . (4.191)
Here σ and ε are the stress and strain of the infinitesimal theory. It is easy to confirm
that the equations of motion are recovered by setting DG(u, ξ) = 0, where, under mild
assumptions, G attains an infimum (which reduces to a minimum for the discrete case) at
the solution u∗, that is G(u∗) = infU G(u), where U is the space of admissible displacements.
A corresponding variational principle may be also established for the problem of finite elastic
deformations, see Section 5.2.
In this restricted context, the Lagrange multiplier method of (4.179) and (4.180) can
be interpreted as a constrained extremization method. To this end, define the Lagrange
multiplier functional GL as
GL(u,Λ) = G(u) +
∫
C0
Λ · c(u) dV . (4.192)
Then, the preceding extended system is recovered by setting
DGL(u, ξ;Λ,θ) = 0 . (4.193)
Indeed, this saddle-point problem may be equivalently expressed as
DGL(u, ξ;Λ,θ) = DGL(u, ξ)︸ ︷︷ ︸Λ fixed
+DGL(Λ,θ)︸ ︷︷ ︸u fixed
= 0 , (4.194)
which implies that
DGL(u, ξ) = 0 , DGL(Λ,θ) = 0 . (4.195)
ME280B May 7, 2019
Computational Treatment of Constraints 133
It is then immediately clear that (4.195)1,2 coincide with (4.179) and (4.180). Various ap-
proximations of the above Lagrange multiplier method are possible to bypass the need for
the direct solution of the extended system of N + M equations. Three such regulariza-
tions discussed here are the classical penalty, the perturbed Lagrangian, and the augmented
Lagrangian methods
In the classical penalty method10, one considers the functional GP defined as
GP (u, ε) = G(u) +1
2
∫
C0
εc(u) · c(u) dV , (4.196)
The perturbed Lagrangian method is associated with the functional GPL given as
GPL(u,Λ; ε) = G(u) +
∫
C0
Λ · c(u) dV 1
2ε
∫
C0
Λ ·Λ dV , (4.197)
where, again, ε > 0 is a penalty parameter. Lastly, the augmented Lagrangian method is
defined in terms of the functional GAL written as
GAL(u,Λ; ε) = G(u) +
∫
C0
Λ · c(u) dV +1
2
∫
C0
εc(u) · c(u) dV . (4.198)
In all three functionals, ε > 0 is a penalty parameters.
First, examine the classical penalty method and assume that G(u) attains an infimum
at the solution u∗ of the equations of motion, such that c(u) = 0. Then, it can be proved
conditional upon sufficient smoothness of the involved functions that if u∗ε is such that
DGP (u∗ε, ξ) = 0 (4.199)
and GP (u, ε) attains an infimum at u∗ε, then
limε→∞
u∗ε = u∗ . (4.200)
While the formal proof of this result is omitted here,11 a simple intuitive argument can
be made by observing that as ε → ∞ the penalty term would become unbounded unless
c(u) = 0, which is necessary to attain an infimum of GP .
Upon setting the differential of GP in the direction ξ to zero, one finds that
DG(u, ξ) +
∫
C0
ε [Dc(u, ξ)] · c(u) dV = 0 . (4.201)
10See R. Courant, Variational methods for the solution of problems of equilibrium and vibration, Bulletin
of the American Mathematical Society, 49:1-23, (1943).11See p. 366 in D.G. Luenberger, Linear and Nonlinear Programming, 2nd edition, Addison-Wesley,
Reading, (1984).
May 7, 2019 ME280B
134 Solution of Non-linear Field Equations
In discrete form, this translates to
ξT[f (u) + εd (u)g (u)] = 0 , (4.202)
where use is made of (4.181). A typical step of the Newton-Raphson method would then be
ξT[f(u(k)
)+K
(u(k)
)∆u(k) + εd
(u(k)
)g(u(k)
)
εDTd
(u(k)
)g(u(k)
)+ d
(u(k)
)Dg
(u(k)
)∆u(k)
]= 0 (4.203)
or, upon recalling (4.183), it follows that
[K(u(k)
)+ ε
DTd
(u(k)
)+ g
(u(k)
)d(u(k)
)dT(u(k)
)]∆u(k)
= −f(u(k)
)− εd
(u(k)
)g(u(k)
). (4.204)
Clearly, the classical penalty method requires the solution of only N algebraic equations
and u∗ε satisfies the constraint only approximately, see Figure 4.28. The extended stiffness
uu∗0 u∗1 u∗
c(u∗) = 0
GP
ε = 0
ε1 > 0
ε 7→ ∞
Figure 4.28. Penalty functional and the enforcement of the constraint c = 0
matrix in (4.204) will be symmetric if K itself is symmetric. This is immediately concluded
from the identity
D
D
[∫1
2εc(u) · c(u)dV
](u, ξ)
(u,∆u)
= D
D
[∫1
2εc(u) · c(u)dV
](u,∆u)
(u, ξ) , (4.205)
ME280B May 7, 2019
Computational Treatment of Constraints 135
as the order of differentiation in the directions ξ and ∆u does not matter, conditional upon
sufficient smoothness of the integral in u. At the same time, the form of the extended
stiffness matrix in (4.204) reveals that ill-conditioning is bound to occur for large values of ε.
Next, consider the perturbed Lagrangian method. Here, extremization of the functional
in (4.197) leads to the set of equations
DG(u, ξ) +
∫
C0
[Dc(u, ξ)] ·Λ dV = 0 (4.206)
and ∫
C0
[c(u)− 1
εΛ
]· θ dV = 0 . (4.207)
The second of the above equations is strongly satisfied when
Λ = εc(u) (4.208)
inR0, in which case the first equation becomes reduces to the equation (4.201) of the classical
penalty method. Alternatively, one may choose to enforce (4.207) weakly, resulting in
θT[g(u)− 1
εΛ
]= 0 , (4.209)
which lead to the discrete equations
Λ = εg(u) . (4.210)
Depending on the interpolation of the multiplier field, it may be possible to efficiently de-
termine Λ at the element level. In such a case, the perturbed Lagrangian method would be
an attractive alternative to the classical penalty method.
Finally, consider the augmented Lagrangian method, which may be thought of as a combi-
nation of the Lagrange multiplier and the classical penalty method. Here, the extremization
of the functional in (4.198) leads to the equations
DG(u, ξ) +
∫
C0
[Dc(u, ξ)] ·Λ dV +
∫
C0
ε [Dc(u, ξ)] · c(u) dV = 0 (4.211)
and ∫
C0
c(u, ξ) · θ dV = 0 . (4.212)
May 7, 2019 ME280B
136 Solution of Non-linear Field Equations
A common algorithmic implementation of the augmented Lagrangian method is the so-
called Uzawa algorithm12 as follows:
Algorithm 8: Uzawa’s algorithm for updated Lagrangian methods
Data: state at time tn, ε, ζ, maxits, tol
Set k = 0, Λ0 = 0 and u0 = un;
while k ≤ maxit and |c(uk)| ≥ tol
Solve discrete counterpart of (4.211) for ∆uk;
Set uk+1 = uk +∆uk;
Set Λk+1 = Λk + ζc(uk);
Set k ← k + 1end
The above algorithm can be proven to converge to the desired solution given any ζ, such
that 0 < ζ < 2ε, for a class of minimization problems with linear constraints.13
The augmented Lagrangian functional with the Uzawa algorithm requires the (repeated)
solution of only N equations. However, in contrast to the classical penalty method, ill-
conditioning of the tangent stiffness matrix is altogether avoided, since the penalty parameter
ε is kept fixed at relatively small. In addition, the constraint equation is satisfied exactly
despite the fixed value of the penalty parameter.
While the classical penalty method and its two variants are mathematically justified only
for under the severe restriction of the existence of a variational problem for the momentum
balance equations, they are used heuristically and with great success on a wide array of
problems in computational mechanics.
A critical question in designing a robust finite element approximation of a continuum-
mechanical problem in the presence of constraints is the choice of admissible displacement (or
velocity) fields and of Lagrange multplier fields. To address this question, one may consider,
for simplicity, a linearized counterpart of the extended system (4.189), in the form[K d
dT 0
][u
Λ
]= −
[f
g
], (4.213)
where nowK, d, f , and g are constant arrays. It can be shown14 that necessary and sufficient
12H. Uzawa, Iterative methods for concave programming. In K.J. Arrow,L. Hurwicz and H. Uzawa editors,
Studies in linear and nonlinear programming, Stanford University Press, Palo Alto, (1958).13See p. 48 of R. Glowinski and P. Le Tallec, Augmented Lagrangian and Operator Splitting
Methods in Non-linear Mechanics, SIAM, Philadelphia, 1989.14See Section 3 in F. Brezzi and K.-J. Bathe, A discourse on the stability conditions for mixed finite
ME280B May 7, 2019
Computational Treatment of Constraints 137
conditions for the above system to be solvable are that: (i) N ≥ M and (ii) if v ∈ kerd
and wTKv = 0 for all w ∈ kerd, then v = 0. These conditions ensure that the solution
to (4.213) exists, but do not guarantee that it will be a convergent solution, say, in the sense
of h-refinement. Regardless, condition (i) is an easy one to check when contemplating a dual
method for the solution of a constrained problem.
Example 4.5.6: Constraint counts for incompressibilityConsider a square block meshed using n×n 4-node elements (n > 1) and assume that it is used to modelan incompressible continuum (solid or fluid). Let the displacements or velocities be fully prescribed onthe boundary (except for a single point). In this case, the number of uknowns displacement or velocitydegrees of freedom is
N = 2(n− 1)2 − 2 . (4.214)
Also, assume that the Lagrange multiplier field is piecewise constant in each element and free ofspecification (except for one point). In this case, the number of uknown Lagrange multiplier degrees offreedom is
M = n2 − 1 . (4.215)
Clearly, in this case N > M so the preceding solvability condition (i) is met. Moreover, it is immediatelyclear from (4.214) and (4.216) that, under uniform h-refinement,
limn→∞
N
M= 2 . (4.216)
This ratio is considered optimal, as it equals the ratio of linear momentum equations to constraintequations in the continuum problem.
element formulations, Computer Methods in Applied Mechanics and Engineering, 82:27–57, (1990).
May 7, 2019 ME280B
Chapter 5
Constitutive Modeling of Deformable
Continua
5.1 Incompressible Newtonian Viscous Fluid
In this problem, the balance of linear momentum is solved together with the condition of
incompressibility, that is,
divT+ ρb = ρ
(∂v
∂t+∂v
∂xv
)in R ,
div v = 0 in R .
(5.1)
These equations are subject to the boundary conditions
vi = vi on Γui
Tijnj = ti on Γqi
(5.2)
where, Γui∩ Γqi = ∅, Γu =
⋃3i=1 Γui
, Γq =⋃3
i=1 Γqi , and ∂R = Γu ∪ Γq, as well as the initial
condition
v(x, 0) = v0(x) , (5.3)
where div v0 = 0. Note that if Γq = ∅, then the boundary velocity v should satisfy∫
∂R
v · n da =
∫
R
div v dv = 0 . (5.4)
The Cauchy stress T is written as
T = −pi+ 2µD , (5.5)
138
Incompressible Newtonian Viscous Fluid 139
where, as argued in Section 4.5, p is the Lagrange multiplier which enforced the incompress-
ibility condition (5.1)2 and µ is the viscosity coefficient (assumed here to be constant). Given
the preceding constitutive law, one may write
divT = div(−pi+ 2µD) = − grad p+ 2µ divD = − grad p+ 2µ div
(∂v
∂x
)s
. (5.6)
Starting from the weak form of linear momentum balance in (3.184), rewrite the stress-
divergence term as
∫
R
(∂ξ
∂x
)s
·[−pi+ 2µ
(∂v
∂x
)s]dv =
∫
R
[−p div ξ +
(∂ξ
∂x
)s
· 2µ(∂v
∂x
)s]dv . (5.7)
Therefore, the weak form of linear momentum balance (3.184) takes the form
∫
R
ξ · ρ(∂v
∂t+∂v
∂xv
)dv +
∫
R
[−p div ξ +
(∂ξ
∂x
)s
· 2µ(∂v
∂x
)s]dv
=
∫
R
ξ · ρb dv +∫
Γq
ξ · t da . (5.8)
In addition, the weak form of the incompressibility condition may be expressed as
∫
R
σ(div v)dv = 0 . (5.9)
Before proceeding with the numerical approximation of the preceding weak forms, recall
the Helmholtz-Hodge decomposition, according to which any vector field v : R → E3 can be
uniquely decomposed into a solenoidal and an irrotational part, that is,
v = vso + vir , (5.10)
where vso satisfies
div vsol = 0 in R ,
vsol · n = 0 on R ,(5.11)
while for vir there exists a scalar function φ : R → R, such that
virr = gradφ , (5.12)
hence
curlvirr = curl gradφ = 0 . (5.13)
May 7, 2019 ME280B
140 Constitutive Modeling of Deformable Continua
A proof of this result can be found elsewhere.1
A very efficient and straightforward method for solving the incompressible Navier-Stokes
equation using finite elements is based on a classical work in the context of finite differences.2
In order to appreciate this approach, it is introduced first using the strong forms. To this
end, suppose that the velocity and pressure fields have already been determined to be vn
and pn, respectively, at time tn, respectively, and the same fields are to be determined at
time tn+1 = tn +∆tn. Therefore, it is known that
ρ
(∂vn
∂t+∂vn
∂xvn
)= − grad pn + 2µ div
(∂vn
∂x
)s
+ ρbn ,
div vn = 0 ,
(5.14)
everywhere in Rn. To advance the solution in time to tn+1, first determine a predictor
velocity v∗n+1, such that
ρ
(v∗n+1 − vn
∆tn+∂vn
∂xvn
)= 2µ div
(∂vn
∂x
)s
+ ρbn+1 in R ,
v∗n+1 = vn+1 on Γu ,
t∗n+1 = 0 on Γq .
(5.15)
Note that the predictor velocity ignores completely any forces due to change in the pressure,
as well as any external tractions. As a result, it does not necessarily satisfy the incom-
pressibility condition. Subsequently, a corrector step is applied to account for the pressure,
as
ρvn+1 − v∗
n+1
∆tn= − grad pn+1 . (5.16)
Note that in the above equation, both vn+1 and pn+1 are unknown, so this equation cannot
be solved directly. Equation (5.16) may be readily rewritten as
v∗n+1 = vn+1 +
∆tnρ
grad pn+1 . (5.17)
If vn+1 · n = 0 on ∂R, then (5.17) represents precisely the Helmholtz-Hodge decomposition
in (5.10–5.13). Hence, in this case, vn+1 and pn+1 are obtained by projecting v∗n+1 onto
its solenoidal and irrotational parts. Even in the case of more general boundary conditions
1See, e.g., Section 6.3.1 in P. Papadopoulos, Introduction to Continuum Mechanics, ME185 course notes,
Berkeley, 2017.2A. Chorin. Numerical solution of the Navier-Stokes equations. Math. Comp., 22:745–762, (1968).
ME280B May 7, 2019
Incompressible Newtonian Viscous Fluid 141
for vn+1, one may take the divergence of both sides of (5.17) and observe (5.1)2 to deduce
the Laplace equation
div v∗n+1 =
∆tnρ
div(grad pn+1) in Rn+1 , (5.18)
for pn+1, subject to boundary conditions
∆tnρ
grad pn+1 · n = 0 on Γu ,
[−pn+1i+ 2µ
(∂vn
∂x
)s]n = tn+1 on Γq .
(5.19)
Note that (5.19)1 by taking the normal projection of (5.17) and recalling (5.15)2. Likewise,
(5.19)2 is obtained from (1.76) and (5.5), where the rate-of-deformation tensor is evaluated
at v = vn (or, possibly, v = v∗n+1) to avoid an implicit dependence on (the yet unknown)
vn+1. It follows that (5.19)1,2 are now respectively the Neumann and Dirichlet boundary
conditions for the Laplace equation (5.18). After pn+1 is calculated from (5.18) and (5.19),
the velocity vn+1 is determined from (5.16).
Turning to the Galerkin-based finite element interpretation,3 this procedure translates to
finding the predictor velocity v∗n+1 from
∫
R
ξn+1 · ρ(∂v∗
n+1
∂t+∂vn
∂xvn
)dv =
∫
R
[−(∂ξn+1
∂x
)s
· 2µ(∂vn
∂x
)s
+ ξn+1 · ρbn+1
]dv ,
(5.20)
where∂v∗
n+1
∂t
.=
v∗n+1−vn
∆tn, with v∗
n+1 = vn+1 on Γu and ξn+1 = 0 on Γu. Subsequently, solve
the weak counterpart of (5.16) together with the incompressibility constraint (5.1)2. Upon
using the divergence theorem for the former, the system of equations may be expressed as∫
R
ξn+1 · ρvn+1 − v∗
n+1
∆tndv =
∫
R
div ξn+1 · pn+1 dv +
∫
Γq
ξn+1 · tn+1 dv ,
∫
R
σn+1 div vn+1 dv = 0 ,
(5.21)
with vn+1 · n = vn+1 · n and ξn+1 · n = 0 on Γu. Note that the tangential component of the
velocity does not affect incompressibility.
Neglecting the element representation for brevity, proceed with the global matrix form
of this two-step algorithm. The predictor step is expressed as
1
∆tn[M]
([v∗
n+1]− [vn])+ [A(vn)] + [K][vn] = [Fb
n+1] , (5.22)
3J. Donea, S. Giuliani, H. Laval, and L. Quartapelle, Finite element solution of the unsteady Navier-Stokes
equations by a fractional step method, Comp. Meth. Appl. Mech. Engrg., 30:53–73, (1982).
May 7, 2019 ME280B
142 Constitutive Modeling of Deformable Continua
subject to [v∗n+1] = [ˆvn+1] on Γu. In (5.22), [M] is the global mass matrix (symmetric,
positive-definite), [A(vn)] is the global advection vector (a non-linear function of vn), [K]
is the global diffusion matrix (symmetric, positive-semidefinite), and [Fbn+1] is the global
external force due to prescribed body forces only. This system is now solved for [v∗n+1]. The
solution becomes trivially simple to compute if [M] is rendered diagonal, as earlier. Either
way, one may write
[v∗n+1] = [vn]−∆tn[M]−1 ([A(vn)] + [K][vn]) + ∆tn[M]−1[Fb
n+1] . (5.23)
For the corrector step, write the matrix counterpart of (5.21) as
1
∆tn[M]
([vn+1]− [v∗
n+1])− [C][pn+1] = [Ft
n+1] ,
[C]T [vn+1] = [Gn+1] ,
(5.24)
where [C] is the matrix emanating from the divergence operator and [Ftn+1] is the external
force vector due to prescribed tractions only. Note that the right-hand side of (5.24)2 becomes
non-zero upon accounting for boundary conditions because it needs to account for the discrete
counterpart of the boundary condition vn+1 · n = vn+1 · n on Γu.
The pressure [pn+1] may be determined from the system (5.24) as
[C]T [M]−1[C][pn+1] =1
∆tn
([Gn+1]− [C]T [v∗
n+1])− [C]T [M]−1[Ft
n+1] , (5.25)
where [C]T [M]−1[C] is symmetric and under sufficiently general conditions has the rank of
[pn+1]. Lastly, the velocity [vn+1] is determined as
[vn+1] = [v∗n+1] + ∆tn[M]−1
([Ft
n+1] + [C][pn+1]). (5.26)
Therefore, obtaining the new state vectors [vn+1] and [pn+1] requires the solution of a sym-
metric linear algebraic system for [pn+1] and performing two updates for [v∗n+1] and [vn+1].
5.2 Nonlinear Elasticity
Recall the mechanical energy balance equation (1.80), and admit the existence of a strain
energy function ψ (F) per unit mass, such that∫
P
T ·D dv =
∫
P
ρψ dv . (5.27)
ME280B May 7, 2019
Nonlinear Elasticity 143
Recalling the result on work-conjugacy of stress and strain or deformation measures, one
may also write∫
P
ρψ dv =
∫
P0
ρ0ψ dV =
∫
P0
P · F dV =
∫
P0
S · E dV . (5.28)
It follows from (5.28) that
∫
P0
(ρ0∂ψ
∂F−P
)· F dV = 0 , (5.29)
so that, using a standard argument,
P = ρ0∂ψ
∂F. (5.30)
This constitutive law characterizes a general homogeneous hyperelastic (or Green-elastic)
material.
The strain energy function is subject to invariance requirements under superposed rigid-
body motions, namely
ψ = ψ (F) = ψ(F+)
= ψ (QF) , (5.31)
for all proper orthogonal Q. It can be easily argued that this implies
ψ = ψ (F) = ψ (U) = ψ (C) = ψ (E) . (5.32)
Therefore, one may conclude from (5.28)2,4 that
∫
P0
(2ρ0
∂ψ
∂C− S
)· E dV = 0 , (5.33)
therefore,
S = 2ρ0∂ψ
∂C= ρ0
∂ψ
∂E. (5.34)
If the stress response of the hyperelastic material is further assumed isotropic, then, given
any orthogonal tensor Q,
ψ = ψ (F) = ψ (FQ) = ψ (C) = ψ(QTCQ
). (5.35)
Recalling the standard representation theorem for isotropic scalar functions of a tensor vari-
able, one may conclude that
ψ = ψ (IC, IIC, IIIC) (5.36)
May 7, 2019 ME280B
144 Constitutive Modeling of Deformable Continua
where IC,IIC, IIIC are the three scalar invariants of C. In view of (5.34)1 and (5.36), one
may conclude, with the aid of the chain rule, that
S = 2ρ0
[∂ψ
∂IC
∂IC∂C
+∂ψ
∂IIC
∂IIC∂C
+∂ψ
∂IIIC
∂IIIC∂C
]. (5.37)
The identities
∂IC∂C
=∂
∂C(trC) = I ,
∂IIC∂C
=∂
∂C
[1
2
(trC)2 − trC2
]= ICI−C ,
∂IIIC∂C
= IIICC−1 ,
(5.38)
can be shown to hold. Indeed, the first two may be confirmed by direct calculation, while the
third one follows immediately from (2.55). In view of (5.38), the constitutive equation (5.37)
takes the form
S = 2ρ0
[(∂ψ
∂IC+ IC
∂ψ
∂IIC
)I− ∂ψ
∂IICC+
∂ψ
∂IIICIIICC
−1
]. (5.39)
Example 5.2.1: Kirchhoff-Saint Venant materialIn this model, the classical stress-strain law of elasticity at infinitesimal strains is written in terms of thesecond Piola-Kirchhoff stress and the Lagrangian strain, that is,
S = 2µE+ λ(trE)I , (5.40)
where λ and µ are material constants. This is the Kirchhoff-Saint Venant (or generalized Hookean)law. The preceding law can be expressed equivalently as
S = µ(C− I) + λ1
2(IC − 3)I =
[1
2λ(IC − 3)− µ
]I+ µC . (5.41)
This constitutive law is derivable from a strain-energy function defined as
ρ0ψ(IC, IIC, IIIC) =1
8λ(IC − 3)2 + frac14µ
(I2C − 2IC − 2IIC
). (5.42)
Recalling (1.86), it follows from (5.41) and (1.14) that
T =1
J
[1
2λ(IB − 3)− µ
]B+
1
JµB2 . (5.43)
Clearly, the Kirchhoff-Saint Venant law reduces to the constitutive equations of isotropic linear elasticityfor infinitesimal deformations.
The Kirchhoff-Saint Venant material law is used to describe materials that undergo a moderateamount of elastic deformation.
ME280B May 7, 2019
Nonlinear Elasticity 145
Example 5.2.2: Compressible neo-Hookean materialHere, the strain energy takes the form
ρ0ψ(IC, IIC, IIIC) =µ
2(IC − 3)− µ ln J +
λ
2(J − 1)2 , (5.44)
where λ and µ are material parameters. Taking into account (5.39) that
S = µ(I−C−1) + λJ(J − 1)C−1 . (5.45)
Pushing the preceding equation forward to the current configuration, it follows that
T =1
Jµ(B− i) + λ(J − 1)i . (5.46)
This constitutive law is consistent with linearized isotropic elasticity. Indeed, recalling (2.54) and (2.64),it can be shown that
L [T;u]X = σ
= µDB(X,u) + λDJ(X,u)I = 2µε+ λ(tr ε)I , (5.47)
as expected.The compressible neo-Hookean material law is used to describe materials that may undergo large
elastic deformations, such as rubber.
Under quasi-static conditions and under mild assumptions on the strain energy func-
tion Ψ, there exists a variational theorem for nonlinear elasticity, according to which the
total potential functional G(u), defined as
G(u) =
∫
R0
ρ0Ψ dV −∫
R0
u · ρ0b dV −∫
Γq,0
u · p dA , (5.48)
attains a local minimum at u which satisfies equilibrium.4
Many nonlinearly elastic materials, such as dense rubber, are incompressible or nearly
incompressible. Here, write the constraint of incompressibility as
c(F) = detF− 1 = 0 . (5.49)
To impose incompressibility, it is important to recognize that it is an internal constraint
that affects the form of the strain energy function. In particular, in view of the discussion in
Section 4.5, write the strain energy functional Ψc of the incompressible elastic material as
ρ0Ψc(F, p) = ρ0Ψ(F) + pc(F) , (5.50)
4See Chapter 11 in M.E. Gurtin, Topics in Finite Elasticity, SIAM, Philadelphia, 1981.
May 7, 2019 ME280B
146 Constitutive Modeling of Deformable Continua
where p is a Lagrange multplier, and proceed as usual to defining the stress starting with
the definition ∫
P0
ρ0Ψc dV =
∫
P0
P · F dV . (5.51)
Recalling that c = 0, this implies that∫
P0
[ρ0
(∂Ψ
∂F+
p
ρ0
∂c
∂F
)−P
]· F dV = 0 , (5.52)
hence, in view of (2.55),
P = ρ0∂Ψ
∂F+ pJF−T . (5.53)
Taking into account (1.86) and (5.53), it follows that
S = ρ0F−1∂Ψ
∂F+ pJC−1 (5.54)
and
T = ρ∂Ψ
∂FFT + pi . (5.55)
The last equation demonstrates that the Lagrange multiplier is a pressure term on the Cauchy
stress.
Example 5.2.3: Compressible neo-Hookean materialStart with the strain energy function of the compressible neo-Hookean model in (5.44) and assumeincompressibility, which reduces the function to
ρ0ψ =µ
2(IC − 3) . (5.56)
Taking into account (5.53-5.55), one may write
P =µ
2
∂
∂F
tr(FTF
)+ pJF−T = µF+ pF−T , (5.57)
S = F−1P = µI+ pC−1 , (5.58)
andT = µB+ pi , (5.59)
subject to c = 0. Alternatively, one may start with the stress relations (5.45) and (5.46), and appendthe pressure term. For instance, for the Cauchy stress, one would deduce the expression
T = µ(B− i) + pi = µB+ (p− µ)i , (5.60)
which suggests that the same meaning of the pressure as in (5.59) to within a constant µ.In the nearly incompressible case, one would use the original stain-energy function of (5.44), but
view λ as a penalty parameter by letting it grow toward infinity to satisfy the constraint of incompress-ibility in approximate fashion.
ME280B May 7, 2019
Nonlinear Elasticity 147
From the standpoint of finite element implementation, it is important to observe that
the material stiffness of the hyperelastic material model is symmetric. Indeed, starting from
the term which is associated with the stress-divergence in linear momentum balance, recall
that
D
[∫
Ωe0
∂ξ
∂X· (FS) dV
](u,∆u) =
∫
Ωe0
(∂∆u
∂X
)T∂ξ
∂X· S dV
︸ ︷︷ ︸geometric term
+
∫
Ωe0
∂ξ
∂X· FDS(u,∆u) dV
︸ ︷︷ ︸material term
.
(5.61)
As already argued in Section 4.3, the geometric term yields a symmetric stiffness provided ξ
and ∆u are interpolated by the same functions. For a hyperelastic solid, the material term
also yields a symmetric stiffness. This is because, according to (5.34),
∂S
∂E= ρ0
∂2Ψ
∂E2(5.62)
hence, it is symmetric. Therefore, in light of (2.41),
∫
Ωe0
∂ξ
∂X· FDS(u,∆u) dV =
∫
Ωe0
FT ∂ξ
∂X·DS(u,∆u) dV =
∫
Ωe0
1
2
[FT ∂ξ
∂X+
(∂ξ
∂X
)T
F
]· ∂S∂E
DE(u,∆u) dV =
∫
Ωe0
1
4
[FT ∂ξ
∂X+
(∂ξ
∂X
)T
F
]· ∂S∂E
[FT ∂∆u
∂X+
(∂∆u
∂X
)T
F
]dV ,
(5.63)
which proves the symmetry.
May 7, 2019 ME280B
Index
acceleration, 2adaptive remeshing, 75advective remapping, 74ALE
mixed variables, 67algorithmic tangent modulus, 100arc-length
normal-place, 124spherical, 123
augmented Lagrangian method, 133axial vector, 9
BFGS method, 111bifurcation point, 113, 116bordering algorithm, 118branch-switching method, 117
Cauchy tetrahedron, 12Cauchy’s lemma, 12Cauchy-Green deformation tensor
left, 4right, 4
centered-difference method, 56classical penalty method, 133configuration
current, 1reference, 2
consistent differentiation, 100constraint
bilateral, 128external, 128internal, 128rheonomous, 128scleronomous, 128unilateral, 128
continuation method, 117contraction, 87control volume, 45
dead load, 97
derivativeFrechet, 23Gateaux, 22
differentialFrechet, 23Gateaux, 21
differentiationautomatic, 91symbolic, 91
divergence theorem, 10dual method, 128
energy norm, 106equilibrium path, 115equipotential relaxation, 78error
“absolute”, 106“relative”, 106
Eulerian remapping, see also advective remapping
fixed point, 86fixed-point iteration, 88follower load, 54, 97force vector
externalelement, 53global, 53
gap function, 127generalized Hookean law, see also Kirchhoff-Saint
Venant lawgradient, 10Green-elastic material, see also hyperelastic mate-
rial
heat flux, 15heat flux vector, 15heat supply, 15Helmholtz-Hodge decomposition, 139hyperelastic material, 143
148
INDEX 149
implicit integration, 55integration
explicit, 56internal energy, 15inverse function theorem, 3
Kelvin-Voigt solid, 99kinetic energy, 14Kirchhoff-Saint Venant law, 144
Lagrange multiplier functional, 132Lagrange’s criterion of materiality, 44limit point, 113, 116line search, 107linear part
function, 23Lipschitz constant, 94Lipschitz continuity, 94loading
proportional, 113localization theorem, 10
mass, 10mass density, 10mass matrix
element, 53global, 53
material surface, 44matrix norm
natural, 87mean-value theorem, 86mechanical energy balance theorem, 14mesh rezoning, see also adaptive remeshingmesh time derivative, 66momentum balance
semi-discrete form, 54motion, 3
Jacobian, 3
Nanson’s formula, 4Newmark method, 55Newton method
generalized, 92modified, 91
Newton-Kantorovich theorem, 95non-linearity
geometric, 20material, 20
operator-split procedure, 72
perturbed Lagrangian method, 133
polar decompositionleft, 6right, 6
polar decomposition theorem, 5primal method, 128principal directions, 6
quasi-Newton method, 108quasi-Newton property, 109quasi-static, 54
radius of attraction, 95rate-of-deformation tensor, 8remainder
function, 23remeshing, 64residual norm, 106Reynolds’ transport theorem, 10Riks-Wempner method, 125
saddle-point problem, 132Schur complement, 118secant property, see also quasi-Newton propertysemi-discretization, 47set
convex, 86Sherman-Morrison-Woodbury formula, 109solution path, see also equilibrium pathspectral representation theorem, 6stability point, 113, 115stiffness
geometric, 102material, 103
strain tensorEulerian, 5Hencky, 7Lagrangian, 5
stress power, 14stress tensor
Cauchy, 12first Piola-Kirchhoff, 13second Piola-Kirchhoff, 15
stress-divergence vectorelement, 53global, 53
stretch, 4stretch tensor
left, 6right, 6
strong form, 38strong solution, 40
May 7, 2019 ME280B
150 Index
tangent stiffness matrixglobal, 101
tangent vector, 38tensor
adjugate, 30identityreferential, 5spatial, 5
proper-orthogonal, 6referential, 4spatial, 4symmetric, 4two-point, 3
tied sliding, 129total Lagrangian formulation, 56
total potential energy functional, 132turning point, see also limit point
updated Lagrangian formulation, 56Uzawa algorithm, 136
variation, 41velocity, 2velocity gradient, 8virtual power, 41virtual work, 41volumetric distortion, 78vorticity tensor, 8
weak form, 39weak solution, 40
ME280B May 7, 2019