Lecture Notes

Nonlinear Finite Element Methods

Julia Mergheim

Institute of Applied Mechanics

Friedrich-Alexander-Universität Erlangen-Nürnberg
Contents

1 Introduction - Nonlinear Phenomena 3

1.1 Geometrical Nonlinearities . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Physical Nonlinearities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Nonlinear Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . 9
1.4 Linear vs. Nonlinear FEM . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2 Basic Equations in Continuum Mechanics 13

2.1 Kinematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Stress Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3 Constitutive Equations: Hyperelasticity . . . . . . . . . . . . . . . . . . . . 18
2.4 Balance Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.5 Weak Form of Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.6 Linearisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.7 Nonlinear Analysis of a Three-Hinged Arch . . . . . . . . . . . . . . . . . . 33

3 Finite Element Discretisation 37

3.1 Isoparametric Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2 Discretisation and Linearisation of the Weak Form . . . . . . . . . . . . . . 46
3.3 Implementation of the FE method . . . . . . . . . . . . . . . . . . . . . . . 51

4 Solution Methods for Time-Independent Problems 53

4.1 Newton-Raphson Method (Load Control) . . . . . . . . . . . . . . . . . . . 53
4.2 Modified Newton-Raphson Method (Load Control) . . . . . . . . . . . . . 55
4.3 Quasi-Newton Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.4 Newton-Raphson Method (Displacement Control) . . . . . . . . . . . . . . 60
4.5 Arc-Length Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

5 Solution Methods for Time-Dependent Problems 67

5.1 Time Discretisation with Finite Differences . . . . . . . . . . . . . . . . . . 67
5.2 Explicit Time Integration Methods . . . . . . . . . . . . . . . . . . . . . . 68
5.3 Implicit Time Integration Methods . . . . . . . . . . . . . . . . . . . . . . 70
5.4 Modifications of the Newmark Method . . . . . . . . . . . . . . . . . . . . 73
5.5 Conserving Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

i
Bibliography

[1] Bathe, K. J. [1995]. Finite Element Procedures. Prentice Hall, Englewood Cliffs,
New Jersey.

[2] Bathe, K. J. [2000]. Finite–Element–Methoden. Springer Verlag, Berlin.

[3] Belytschko, T., W. K. Liu & B. Moran [2000]. Nonlinear Finite Element Analysis
for Continua and Structures. John Wiley & Sons.

[4] Bonet, J. & R. D. Wood [1997]. Nonlinear Continuum Mechanics for Finite Element
Analysis. Cambridge University Press.

[5] Crisfield, M. A. [1996]. Non–linear Finite Element Analysis of Solids and Structures.
John Wiley & Sons.

[6] Holzapfel, G. A. [2000]. Nonlinear Solid Mechanics: A Continuum Approach for

Engineering. John Wiley & Sons.

[7] Hughes, T. J. R. [2000]. The Finite Element Method – Linear Static and Dynamic
Finite Element Analysis. Prentice Hall, Englewood Cliffs, New Jersey.

[8] Wriggers, P. [2001]. Nichtlineare Finite–Element–Methoden. Springer Verlag, Berlin.

[9] Wriggers, P. [2008]. Nonlinear Finite Element Methods. Springer Verlag, Berlin.

[10] Zienkiewicz, O. C. & R. L. Taylor [2000]. The Finite Element Method, Volume I&II .
Butterworth Heinemann, fifth edition.

1
1 Introduction - Nonlinear Phenomena

The modeling of problems in solid mechanics often leads to ordinary or partial differential
equations of nonlinear nature. The finite element (FE) method is a powerful tool to solve
these equations. FE methods were developed over the last 50 years and a number of
all-purpose codes exists to solve problems from different disciplines. However, nonlinear
simulations can have non-unique solutions, bifurcation points, lead to convergence prob-
lems etc. Often no mathematical error analysis is available. Therefore, the application of
nonlinear FE programs requires practical experience (which can not be provided in this
course) and a good understanding of the theory. These lecture notes are mainly based on
the comprehensive book of Wriggers [9]. For further reading on nonlinear finite elements
the textbooks [1, 3, 5, 10] are recommended.

Figure 1.1: nonlinearities: car crash, rubber material, Tacoma bridge collapse

There are different sources of nonlinearities, depending highly on the considered problem.
In solid mechanics one distinguishes mainly between geometrical nonlinearities, material
(or physical) nonlinearities and nonlinearities due to boundary conditions.

• Geometrical nonlinearities occur due to large displacements or rotations or large

strains. In structural elements like cables or shells the rotations might be large,
whereby the strains remain small. In problems like metal forming or the mechanics
of tires finite deformations occur which can imply arbitrarily large strains. Fur-
thermore, geometrical instability problems, like the buckling of frames or shells or
snap-through behavior, are considered.

• Many materials show a nonlinear relation of stresses and strain. This physical
nonlinearity follows for instance from the visco-elastic behavior of polymers or the
elastoplastic behavior of metals, alloys or soils. Furthermore, failure and damage in
materials result in a nonlinear stress-strain response.

3
1 Introduction - Nonlinear Phenomena

• Nonlinear boundary conditions can result from contact between different struc-
tures or structural parts or deformation-dependent loading (e.g. blowing up a bal-
loon).

Nonlinearities can also occur in coupled problems, e.g. thermo-mechanical, chemo-

mechanical, electro-mechanical problems, where nonlinearities in all sets of equations
have to be taken into account. In the following some simple examples, taken from [9],
which can be solved analytically, are introduced to clarify the effects of nonlinearities.

1.1 Geometrical Nonlinearities

It is usually sufficient in structural analysis to consider only small strains and small
displacements, since most of the structures lose its load carrying capacities when large
strains occur. Nevertheless, certain materials like polymers or biomaterials can sustain
large strains and it is important to take these into account. Furthermore, shell or beam
like structures or cables depict large displacements or rotations and a nonlinear theory is
required to capture the exact geometry. The difference between finite strains and small
strains is further discussed in section 2, the following simple examples focus on large
displacements or large rotations and instabilities.

Example 1.1: Large displacements of a rigid beam

l l cos(ϕ)
rigid beam
F

ϕ
c rotational spring

F
undeformed system deformed system

Figure 1.2: example 1.1: system and loading

A rigid beam is simply supported with an additional rotational spring on the left end and
loaded by a force F on the right end as shown in figure 1.2. We consider large rotations
and therefore equilibrium has to be calculated in the deformed configuration, see figure
1.2, right. We obtain a nonlinear relation between force and rotation

F l cos(ϕ) = c ϕ. (1.1.1)

When the rotations are small, the linear theory can be applied and leads to the following
(linear) result

cos(ϕ) ≈ 1 ⇒ F l = c ϕ, (1.1.2)

4
 1.1 Geometrical Nonlinearities

which also results from computing the equilibrium equation in the undeformed config-
uration. The force-rotation relations are qualitatively depicted in the figure 1.3. It is
obvious that the error which is made by using the linear theory increases rapidly for large
rotations. For small rotations of < 5◦ the error is negligible.

Fl
c
2

30 60 ϕ

Figure 1.3: example 1.1: load-rotation relation

Example 1.2: Large displacements of an elastic system

In this example an elastic system, which consists of two linear springs, is examined,
compare figure 1.4. The springs are loaded by the force F . First, the kinematic relations

l F l l l
c c ϕ F
w

l+f
undeformed system deformed system

Figure 1.4: example 1.2: system and loading

are derived in the deformed configuration, see figure 1.4, right

"r #
 w 2 w
[l + f ]2 = w2 + l2 ⇒ f = l 1+ −1 and sin(ϕ) = . (1.1.3)
l l+f

The equilibrium condition has to be satisfied in the deformed configuration. Due to

symmetry, the forces in the springs SF have to be equal.

F = 2 SF sin(ϕ). (1.1.4)

The springs are linear, i.e. they show a linear constitutive relation between force and
elongation f with the spring stiffness c

SF = c f. (1.1.5)

5
1 Introduction - Nonlinear Phenomena

Possible nonlinear characteristics of springs for large elongations are neglected here. The
force-deflection relation results from the combination of the equations (1.1.3)-(1.1.5)
 
1
F = 2 SF sin(ϕ) ⇒ F = 2 c w 1 − q

 (1.1.6)
w 2
1+ l

and is shown in the diagram 1.5. This example can not be formulated in the linear
F
2cl
0, 3

0, 2

0, 1

0, 5 1 w
l

Figure 1.5: example 1.2: load-deflection relation

theory, since it has a horizontal tangent at wl = 0. That can be verified by formulating

the equilibrium equation in the undeformed configuration, which has no possible solution.

Example 1.3: Bifurcation problem

rigid beam
F
ϕ
c rotational spring
l sin(ϕ)
l F
undeformed system deformed system

Figure 1.6: example 1.3: system and loading

The same system as in the first example is considered, but now the beam is loaded by a
compressive force. In this case a stability problem is obtained that has multiple solutions.
The equilibrium equation is again formulated for the deformed system
Fl ϕ
F l sin(ϕ) = cϕ ⇒ = . (1.1.7)
c sin(ϕ)
This equation has multiple solution for the rotation ϕ and it has to be studied if the
equilibrium states are stable or unstable. If ϕ = 0, the equilibrium equation holds for

6
 1.2 Physical Nonlinearities

all values of F . If ϕ 6= 0 and if F l/c > 1 (since |ϕ|/| sin(ϕ)| > 1), there are two more
solutions and in total three possible solutions exist. The point from which the three
different solutions exist, i.e. F = Fcrit = c/l, is known as the bifurcation point and the
related force Fcrit as the critical load.
The theory of stability shows that the trivial solution ϕ = 0 is instable for F > Fcrit .
That means that equilibrium at ϕ = 0 is lost for small pertubations and the system
changes to a new stable equilibrium state. This is usually related with large deformations,
dynamical effects and failure of the system. To avoid instabilities it is often of interest
and sufficient in practical applications to compute the critical load Fcrit of a structure.
However, for certain problems, e.g. in crash simulations, it might also be of interest to
follow the deformation of the structure beyond the bifurcation point. Figure 1.7 shows
the load-displacement relation for the present problem. At the value of the critical load
the equilibrium path bifurcates and three different solutions are possible, whereby the two
non-trivial ones describe a stable equilibrium state.

F
cl
unstable solution
2

1
bifurcation point stable solution

ϕ
−120 −60 0 60 120

Figure 1.7: example 1.3: load-rotation relation

1.2 Physical Nonlinearities

It is sufficient in many simulations to consider only linear elastic material behavior,

for instance for metallic materials under small loads. However, most materials show
a nonlinear stress-strain response. One can distinguish between nonlinear elasticity,
viscoelasticity, elastoplasticity, damage and further combinations of nonlinear material
responses. In figure 1.8 schematic stress-strain responses are depicted. Nonlinear elastic
behavior can be observed for rubber, foams or biomaterials. Viscoelastic behavior is
common for polymers, especially at high temperatures. Most metals and alloys as
well as soils or granular materials behave elastoplastic. Damage is related to emerging

7
1 Introduction - Nonlinear Phenomena

σ σ σ

ǫ ǫ ǫ
nonlinear elastic viscoelastic elastoplastic

Figure 1.8: schematic representation of different constitutive relations

microcracks and can be observed in brittle materials like concrete.

Example 1.4: Elastoplastic bars

σ
E 1 , σy 1 σy
F
E
E 2 , σy 2
1
ǫ
l u

Figure 1.9: example 1.4: system and loading, schematic stress-strain curve for perfect elastoplasticity

This example is geometrically linear and only material nonlinearities are considered. Two
bars with same geometry (cross section A), that are made of different materials are fixed
on the left hand side and loaded via a rigid plate on the right had side by a force F .
Both materials behave perfectly elasto-plastic with different Young’s moduli E and yield
stresses σy . A schematic stress strain relation of perfect elasto-plasticity is shown in figure
1.9, right. The materials have the following parameters. Material 1: E1 = 2 E, σy 1 = 3σy
and material 2: E2 = E, σy 2 = σy . The unknown load-displacement relation has to be
derived step-by-step. The displacements and the strains are equal in both bars
u
u1 = u2 = u ⇒ ǫ 1 = ǫ 2 = . (1.2.1)
l
Equilibrium is obtained when the sum of the normal forces Ni in the bars is equal to the
prescribed force F
F
N 1 + N 2 = F ⇒ σ1 + σ2 = . (1.2.2)
A
For small loading an elastic response is obtained. In the elastic region the stresses are
calculated by the linear constitutive relations

σ 1 = E 1 ǫ1 and σ2 = E2 ǫ2 . (1.2.3)

8
 1.3 Nonlinear Boundary Conditions

Combining kinematics, equilibrium and constitutive equations yields the initially linear
relation of displacement and force
Fl
u= (1.2.4)
[E1 + E2 ]A
This result is used to calculate the stresses for a certain displacement u as
F F
σi = E i = Ei (1.2.5)
[E1 + E2 ]A 3EA
σy l
When the force is increased bar 2 starts yielding at σ2 = σy ⇒ F = 3Aσy ⇒ u = E .
At further increase of the load the stress in bar 2 remains constant and only the stress in
bar 1 increases
 
u F l
N2 = A σy → const. ⇒ E1 A + Aσy = F ⇒ u = − σy (1.2.6)
l A 2E
When σ1 = 3σy bar 1 starts yielding. The displacement u at that point is derived as
3σy l
N1 = 3 A σy ⇒ F = 4Aσy ⇒ u = . (1.2.7)
2E
After that it is not possible to increase the load further. A nonlinear load-displacement
relation is derived when the equations (1.2.4),(1.2.6) and (1.2.7) are combined, see figure
1.10. Now the nonlinearity stems from the nonlinear elastoplastic constitutive relation.
F
Aσy
4

1 2 3 uE
lσy

Figure 1.10: example 1.4: load-displacement relation

1.3 Nonlinear Boundary Conditions

Nonlinear boundary conditions occur when the boundary conditions change with the
deformation state of the system. This happens when structures come into contact or
when the external forces are deformation dependent (e.g. water / air pressure).

9
1 Introduction - Nonlinear Phenomena

Example 1.5: Bars in contact

A simple example is considered that consists of two bars with the stiffness EA, as illus-
trated above. One bar is loaded and for larger loads the two bars come in contact, they
can not penetrate each other. The load-displacement relation is derived in two steps, for
small and large loads. When the force is small there is no contact between the bars. Bar
F
EA
4

3 with contact
u1 u2
2
F
E, A E, A 1
without contact
u1
l l δ l δ 2δ l
l l
Figure 1.11: example 1.5: system and loading and the resulting force displacement relation

2 is not deformed (since it is not loaded) and the displacement at the end of bar 1 can be
calculated by the elastic constitutive response
Fl
if u1 − u2 < δ : u2 = 0, u1 = (1.3.1)
EA
An increase of the load leads to contact of the two bars, which changes the boundary
conditions for both bars. Since penetration of the bars is not possible it holds that

u1 − u2 = δ. (1.3.2)

The individual values of the displacements u1 and u2 depend on the stiffnesses of the bars
(statically undetermined system) and are derived in terms of the still unknown ’contact
force’ K as
Fl K 2l Kl
u1 = − and u2 = (1.3.3)
EA EA EA
Combination of equation (1.3.3) and (1.3.2) yields the contact force as
 
1 δEA
K= F− (1.3.4)
3 l

which is then inserted into equation (1.3.3) and results in the following relation of force
F and displacement u1 , provided that u1 > δ.
1Fl 2
u1 = + δ. (1.3.5)
3 EA 3

10
 1.3 Nonlinear Boundary Conditions

The overall load-displacement answer is shown in the diagram in figure 1.11.

All discussed examples show that the sources of nonlinear behaviour can be quite dif-
ferent. Geometrical effects, material properties or changing boundary conditions yield
nonlinear behaviour. Often in technical applications different nonlinearities are combined
(e.g. contact + inelasticity + geometrical nonlinearities).

11
1 Introduction - Nonlinear Phenomena

1.4 Linear vs. Nonlinear FEM

up to now: linear FEM

• equilibrium in undeformed configuration

• strains as a linear function of displacements

• stress as linear function of strain (linear elasticity)

⇒ linear system of equations

f int − f ext = 0 with f int = K · u

⇒ direct solution for unknown displacements

K · u = f ext → u = K −1 · f ext

• compare the linearised equation from example 1.1

c l
ϕ − F =0⇒ϕ= F
l
|{z}
|{z} c
external force
internal force

now: nonlinear FEM

• equilibrium in deformed configuration (geometrically nonlinear)

• strain as a nonlinear function of displacements (geometrically nonlinear)

• stress as nonlinear function of strain (physically nonlinear)

⇒ nonlinear system of equations

f int − f ext = 0 with f int = f int (ϕ)

⇒ iterative solution for unknown nodal deformations ϕ

• compare the nonlinear equation from example 1.1

c ϕ
F l cos(ϕ) = c ϕ ⇒ − F = 0
l cos(ϕ) |{z}
| {z } external force
internal force

in this course:

• geometrical nonlinearities; finite deformations

• physical nonlinearities limited to (nonlinear) hyperelasticity

12
2 Basic Equations in Continuum Mechanics

This chapter summarizes the basic equations of nonlinear continuum mechanics which are
the foundation for the finite element implementation. A detailed introduction to nonlinear
continuum mechanics can be found in the books of Holzapfel [6] or Bonet & Wood [4].

2.1 Kinematics

The kinematic relations describe the motion and deformation of a body without taking
into account the forces. Relevant quantities are the deformation map and its derivatives in
space and time. Furthermore, strain measures will be introduced. The kinematic relations
are derived for the geometrically nonlinear setting, that is for arbitrary large strains.

deformation map

In accordance with the continuum theory we consider a body as a composition of a set

of particles. The placements of these particles at an initial time t0 are specified by the
position vectors X and define the reference or material configuration B0 of the body. At
a certain time t the current positions of the particles are described by the position vectors
x and define the spatial or current configuration Bt .
We define a nonlinear deformation map ϕ, which describes the motion of the body. The
deformation map has to be unique, continuous and differentiable and maps the points X
in the material configuration to the places x in the spatial configuration

x = ϕ(X, t) with ϕ : B0 → Bt . (2.1.1)

This description is common in solid mechanics and is denoted as the Lagrangian descrip-
tion of motion: the motion is characterised with respect to the material coordinates and
one follows the movement of a particle in time.
In the same manner the inverse motion can be introduced, which reverses the deformation
map ϕ and thus maps the spatial to the reference configuration

X = Φ(x, t) with Φ : Bt → B0 , (2.1.2)

see figure 2.1. This Eulerian description of motion is common in fluid mechanics: the
motion is characterised with respect to the spatial coordinates and movement is described
at a spatial point in time. We will use the Lagrangian description in the following.

13
2 Basic Equations in Continuum Mechanics

ϕ(X, t)

F
X B0 x Bt
F −1

Φ(x, t)

Figure 2.1: configurations and motion of a body

The displacement field u is determined as the difference of spatial and material coordinates
as

u(X, t) = y(X, t) − X = x − X. (2.1.3)

deformation gradient

The deformation gradient F , which is a fundamental kinematical quantity, is defined as

the gradient of the deformation map with respect to the material coordinates

F = ∇X ϕ(X, t) = 1 + ∇X u(X, t) with F : T B0 → T Bt . (2.1.4)

Therefore, F describes the linear tangent map from the material tangent space T B0 to
the spatial tangent space T Bt . The determinant of the deformation gradient is denoted
Jacobian

J = det(F ) > 0. (2.1.5)

The condition J > 0 implies that the body can not penetrate itself during the defor-
mation. To ensure the existence of the inverse deformation gradient, F must not be
singular, i.e. the Jacobian is not equal zero. The inverse deformation gradient is then de-
fined by means of the gradient of the inverse deformation map with respect to the spatial
coordinates x as

F −1 = ∇x Φ(x, t) with F −1 : T Bt → T B0 . (2.1.6)

The deformation gradient presents a linear tangent map of a line element in the material
configuration dX ∈ T B0 to a line element in the spatial configuration dx ∈ T Bt as
shown in figure 2.2.

dx = F · dX. (2.1.7)

The deformation gradient can further be used to transform area and volume elements

14
 2.1 Kinematics

ϕ(X, t)

dA F da
B0 X Bt x dx
dX
F −1 dv
dV

Φ(x, t)

Figure 2.2: transformation of line, area and volume elements

from the undeformed to the deformed configuration. The transformation of area elements
is described by Nanson’s formula

n da = J F −t · N dA, (2.1.8)

whereby N and n denote the outward unit normal vectors in the material or spatial
configuration. An infinitesimal volume element is transformed by means of the Jacobian
as

dv = J dV. (2.1.9)

strain measures

In the geometrically nonlinear theory different strain measures can be defined, related to
the material or spatial configuration. The strain measures have to vanish in all components
for rigid body motions. In the material configuration the Green-Lagrange strain is defined
as
1 T  1
E := F · F − 1 = [C − 1] , (2.1.10)
2 2
whereby 1 denotes the second order unity tensor and C = F T · F is the right Cauchy-
Green tensor. The Green-Lagrange strain describes the change ∆s in the (squared) length
of line elements from B0 to Bt .
1
∆s = [dx · dx − dX · dX] = 12 [F · dX · [F · dX] − dX · dX]
2
  (2.1.11)
= dX · 21 F T · F − 1 · dX = dX · E · dX

To compare the Green-Lagrange strain with the linear theory, equation (2.1.10) is formu-
lated in terms of the displacement gradient using equation (2.1.4)

1
   
E = 2
F T · F − 1 = 12 [1 + ∇X u]T · [1 + ∇X u] − 1
  (2.1.12)
= 21 ∇X uT + ∇X u + ∇X uT · ∇X u

15
2 Basic Equations in Continuum Mechanics

Comparison with the definition of the linear strain tensor

1 
ε = ∇X usym = ∇X u + ∇X uT (2.1.13)
2
shows that higher order terms are neglected in the linear theory which is sufficient if
||∇X u|| ≪ 1.
A strain tensor defined in the spatial configuration is the Euler-Almansi strain
1  1 
e := 1 − F −T · F −1 = 1 − b−1 , (2.1.14)
2 2
with the left Cauchy-Green tensor b = F · F T . The Euler-Almansi strain describes the
same change in the squared line elements ∆s as the Green-Lagrange strain, but with
respect to the spatial line elements dx

1
∆s = [dx · dx − dX · dX] = 21 [dx · dx − F −1 · dx · [F −1 · dx]]
2
  (2.1.15)
= dx · 12 1 − F −T · F −1 · dx = dx · e · dx

The strain measures (and also other vector and tensor fields) in the material and spatial
configuration are related by push forward and pull back operations. The Euler-Almansi
strain e is obtained by the push forward of the Green-Lagrange strain

e = F −T · E · F −1 (2.1.16)

and vice versa by means of the pull back operation

E = FT · e · F. (2.1.17)

This operations do not change the physical meaning of the strain measure but only their
configurations.

time derivatives

Time derivatives of the kinematical quantities have to be considered if either dynamical

effects have to be taken into account or if the material behavior is rate-dependent. The
velocity of a material point with respect to the material configuration is defined by the
material time derivative (by holding X fix)

V (X, t) = ϕ̇(X, t). (2.1.18)

The material time derivative is denoted by means of an over-dot. In the spatial configu-
ration the velocity of a point is given by

v(x, t) = v(ϕ(X, t), t) = V (X, t) (2.1.19)

16
 2.2 Stress Measures

The acceleration is given analogously by means of the second derivative in time

A(X, t) = V̇ (X, t) = ϕ̈(X, t). (2.1.20)

The acceleration parametrized with spatial coordinates follows from equation (2.1.19) and
application of the chain rule
∂v
a(x, t) = v̇(x, t) = + ∇x v · v, (2.1.21)
∂t
whereby the first term is the local part and the second part is called the convective part of
the acceleration. The local part ∂v/∂t is calculated by holding x fix and is therefore de-
noted as spatial time derivative. The material time derivative of the deformation gradient
is calculated as

Ḟ = ∇X ϕ̇(X, t) = ∇X V = ∇x v · F (2.1.22)

The spatial velocity gradient l = ∇x v can therefore be calculated as l = Ḟ · F −1 .

2.2 Stress Measures

Prescribed external forces or deformations result in an internal stress state in a body B.

To analyse the stress state at a certain point the body is cut by a plane surface, compare
figure 2.3. The resultant force on an area element can be measured in the material or in

N n t
T
X x
dA da
B0 Bt

Figure 2.3: tractions in material and spatial configuration

the spatial configuration which gives rise to the definition of two different stress measures.

df = t(x, t, n)da = T (X, t, N )dA. (2.2.1)

t denotes the Cauchy (or true) traction vector, which measures the force per unit area
in the spatial configuration. The traction vector T is denoted as Piola traction vector

17
2 Basic Equations in Continuum Mechanics

and measures the force per unit area in the undeformed configuration. The vectors t
and T point in the same direction, i.e. the direction of the force element df , but have in
general different absolute values. The Cauchy traction t describes the true intensity of
the stresses.
It follows from Cauchy’s stress theorem that there exist unique second-order tensor fields σ
and P such that the traction vectors are linear functions of σ or P and the corresponding
normal vectors

t(x, t, n) = σ(x, t) · n
(2.2.2)
T (X, t, N ) = P (X, t) · N ,

whereby the Cauchy stress tensor σ is a symmetric spatial tensor field and the Piola stress
P is a generally not symmetric two-point tensor field. It follows from equation (2.2.1)
and the Nanson’s formula (2.1.8) that the two stress measure are related by

σ · n da = P · N dA ⇒ σ · JF −T · N dA = P · N dA
(2.2.3)
⇒ P = Jσ · F −T or σ = J −1 P · F T .

A purely material stress measure is obtained by the pull back of the Cauchy stress tensor
to the material configuration

S = JF −1 · σ · F −T . (2.2.4)

This 2nd Piola-Kirchhoff stress tensor S does not permit a physical interpretation but
is often used in constitutive theory since it is symmetric and energy-conjugated to the
Green-Lagrange strain.

2.3 Constitutive Equations: Hyperelasticity

The constitutive theory describes the behavior of materials which can be observed in
experiments. In a phenomenological description a functional relation between the stresses
and the deformation history is required that represents the experimental observations.
Since material behavior can be very complex some approximations and simplifications are
usually necessary. Furthermore, the definition of constitutive relations has to follow basic
principles from continuum mechanics, which are summarised in the following:

• principle of determinism
stresses are determined by means of deformation history, no randomness
• principle of equipresence
all constitutive equations depend on the same set of variables
• principle of local action
the material response at a point depends only on information in an infinite small
neighbourhood of that point

18
 2.3 Constitutive Equations: Hyperelasticity

• principle of frame invariance

constitutive equations have to be invariant with respect to a change of the observer
• principle of irreversibility
the second law of thermodynamics has to be satisfied.

We focus here on isotropic hyperelastic material behavior which is a suitable model for
many materials that undergo finite strains, e.g. rubber, biomaterials or foam. Hyperelas-
ticity reduces to the classical Hooke’s law of linear elasticity for small strains. The main
characteristics of hyperelasticity are

• the current stress state depends only on the current strain state

• the stress state is path independent (same path for loading and unloading)

• the stress state can be derived as a function of a stored strain energy density

For isotropic materials (same material response in different directions) the strain energy
density depends on the right Cauchy Green tensor

ψ = ψ(C) (2.3.1)

and the stresses are calculated as derivatives of ψ with respect to certain deformation
measures
∂ψ(C) ∂ψ(E) ∂E ∂ψ(E)
S = 2 =2 : =
∂C ∂E ∂C ∂E
ψ(F ) (2.3.2)
P =
∂F
∂ψ(C) ∂ψ(F )
σ = J2 F · · F T = J1 · FT
∂C ∂F
Usually, the relation between the stresses and strains is nonlinear. The increase in the
2nd Piola-Kirchhoff stresses ∆S due to an increase in the Green-Lagrange strains ∆E can
be obtained by linearising the stress strain relation (2.3.2).1. The concept of linearization
is described in detail in section 2.6. The material (or Lagrangian) fourth order elasticity
tensor L, which depends on the deformation, is obtained as
∂S ∂ 2ψ
∆S = L : ∆E with L = = (2.3.3)
∂E ∂E 2
Using index notation yields
∂Sij ∂ 2ψ
∆Sij = Lijkl ∆Ekl with Lijkl = = . (2.3.4)
∂Ekl ∂Eij ∂Ekl
The spatial (or Eulerian) elasticity tensor E is derived by the push forward of L, which
is defined in index notation for simplicity
1
Eijkl = FiI FjJ LIJKL FKk FLl . (2.3.5)
J

19
2 Basic Equations in Continuum Mechanics

In the following two hyperelastic constitutive equations are specified.

St.-Venant-Kirchhoff material (isotropic & compressible)

The St.-Venant-Kirchhoff hyperelastic constitutive law constitutes the direct transfer of
Hooke’s law of linear elasticity to the finite strain setting: the linear strain tensor in the
classical Hooke’s law is replaced by the Green-Lagrange strain. The St.-Venant-Kirchhoff
material gives only meaningful results for moderate strain values (up to 5%), but can be
used e.g. in problems with large rotations.
The strain energy function is defined in terms of the Lamé parameters µ and λ as

ψ V K (E) = 21 λ[E : 1]2 + µE : E

(2.3.6)
ψ V K (E) = 12 λ[Epp ]2 + µElm Elm

The 2nd Piola-Kirchhoff stress tensor is calculated as

∂ψ V K
S = = λ[E : 1]1 + 2µE
∂E (2.3.7)
∂ψ V K
Sij = = λEkk δij + 2µEij
∂Eij

and the Lagrangian elasticity tensor is given by

∂ 2ψV K
L = = λ1 ⊗ 1 + 2µ Isym
∂E 2 (2.3.8)
∂ 2ψV K
Lijkl = = λδij δkl + µ[δik δjl + δil δjk ]
∂Eij ∂Ekl
The St.-Venant-Kirchhoff material model is only nonlinear due to the nonlinear strain
measure E. The 2nd Piola-Kirchhoff stress is linear in E and the Lagrangian elasticity
tensor L is constant and similar as the elasticity tensor in linear elasticity.
Neo-Hooke material (isotropic & compressible)
Isotropy is defined by requiring the constitutive behaviour to be identical in any material
direction. Therefore, the relationship between ψ and C (or b) must be independent of
the material axis chosen and consequently ψ must only be a function of the invariants of
C (or b). The Neo-Hooke material model is formulated in terms of the invariants of C
(or b).
The invariants of the deformation tensors C and b are given as

IC = C:1 = b:1 = Ib
1 1
IIC = 2
[tr2 (C) − tr(C 2 )] = 2
[tr2 (b) − tr(b2 )] = IIb (2.3.9)
IIIC = det(C) = J2 = det(b) = IIIb

The partial derivatives of the invariants with respect to the deformation measures are

20
 2.3 Constitutive Equations: Hyperelasticity

calculated as

∂IC = 1 ∂Ib = 1
∂C ∂b
∂IIC = IC 1 − C T ∂IIb = I 1 − bT (2.3.10)
b
∂C ∂b
∂IIIC = IIIC C −1 ∂IIIb = III b−1
b
∂C ∂b
The material description of strain energy function is defined in terms of the Lamé param-
eter λ and µ

µ √ √
ψ N H (C) = 2 [IC − 3] − µ ln( IIIC ) + λ2 ln 2
( IIIC )
(2.3.11)
µ
= 2 [IC − 3] − µ ln(J) + λ 2
2 ln (J).

The calculation of the 2nd Piola-Kirchhoff stresses requires the derivative of the strain
energy density with respect to C

∂ψ N H (C)
SNH = 2
∂C
∂ψ N H ∂IC ∂ψ N H ∂IIC ∂ψ N H ∂IIIC (2.3.12)
= 2 +2 +2
∂IC ∂C ∂IIC ∂C ∂IIIC ∂C

⇒ S N H = µ [1 − C −1 ] + λ ln(J)C −1

and the second derivatives are required for determining the Lagrangian elasticity tensor

∂S ∂C −1
LN H = 2 = [−2µ + 2λ ln(J)] + λ C −1 ⊗ C −1 (2.3.13)
∂C ∂C
with
∂Cij−1 1 −1 −1
= − [Cik Cjl + Cil−1 Cjk
−1
] (2.3.14)
∂Ckl 2
The spatial description of the stress and the elasticity tensor is obtained by means of the
push forward of the material quantities

1 µ λ ln(J)
σN H = F · S · F T = [b − 1] + 1 (2.3.15)
J J J
λ µ − λ ln(J) sym
EN H = 1⊗1+2 I = λ̃1 ⊗ 1 + 2µ̃ Isym (2.3.16)
J J
with λ̃ = Jλ and µ̃ = µ − λ ln(J) .
J
In case of small strains with J ≈ 1 it follows that λ̃ → λ and µ̃ → µ. Then the standard
4th order elasticity tensor from linear elasticity is recovered, compare also equation (2.3.8).

21
2 Basic Equations in Continuum Mechanics

1D example: tension test

The diagram 2.4 shows the relation of the Cauchy stress σ and the stretch α = l/L which
is equal to the deformed length l devided by the undeformed length L, calculated in
a one-dimensional setting. The St.-Venant-Kirchhoff model gives unphysical results for
α → 0, i.e. the stress required for a compression towards zero volume converges to zero.
In contrast, for the Neo-Hooke model the stress tends towards −∞ if α → 0.

10

Neo-Hooke
8
St.-Venant-Kirchhoff
6

2
σ

-2

-4

-6

-8
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
α

Figure 2.4: Comparison of Neo-Hooke and St.-Venant-Kirchhoff hyperelasticity in 1D

2.4 Balance Equations

The balance equations are the fundamental relations in continuum mechanics, which have
to be fulfilled. The general format of the differential equations states that the time rate
of change of a balance quantity (•)0 or (•)t is equal to the sum of source terms ()0 or
()t and flux terms (⋆)0 · N or (⋆)t · n, compare figure 2.5. The general or master balance

(⋆)0 · N B0
N

V0
()0 ∂V0
(•)0

Figure 2.5: generic balance equation

equation can be formulated in the material or the spatial configuration, i.e. in a cut-out

22
 2.4 Balance Equations

out region V0 or Vt . In the material configuration the following relations hold

Z ˙ Z Z
(•)0 dV = ()0 dV + (⋆)0 · N dA
VZ0 VZ
0 ∂V
Z0
⇒ ˙ dV =
(•) ()0 dV + Div(⋆)0 dV
0 (2.4.1)
VZ0 V0 V0

⇒ ˙ − ()0 − Div(⋆)0 ] dV = 0,
[(•) 0
V0

whereby the divergence theorem was applied to obtain the result in the second line.
Equation (2.4.1) has to hold for an arbitrary region V0 which leads to the local statement
˙ − ()0 − Div(⋆)0 = 0 ∀ X.
(•) (2.4.2)
0

The derivation of the master balance equation in the spatial configuration requires the
time derivative of an integral over the time dependent integration region Vt . Therefore,
the integral is transferred to the material configuration by a pull back, followed by taking
the time derivative and a subsequent push forward. Thus, the time derivative of a spatial
quantity χ(x, t) integrated over Vt is calculated as
Z ˙ Z
D
Z
D
χ(x, t) dv = Dt χ(x, t) dv = χ(ϕ(X), t) J(X, t) dV
Dt
V
Zt Vt V0

= ˙
[χ̇(ϕ(X), t) J(X, t) + χ(ϕ(X), t) J(X, t)] dV (2.4.3)
V0
Z " ˙
# Z
J(X, t)
= χ̇(x, t) + χ(x, t) dv = [χ̇(x, t) + χ(x, t) divv] dv
J(X, t)
Vt Vt

With this relation the generic balance equation in the spatial configuration follows as
Z ˙ Z Z
(•)t dv = ()t dv + (⋆)t · nda
V
Zt Vt ∂V
Zt Z
⇒ ˙ + (•)t divv] dv =
[(•) ()t dv + div(⋆)t dv
t (2.4.4)
V
Zt Vt Vt

⇒ ˙ t + (•)t divv − ()t − div(⋆)t ] dv

[(•)
Vt

which leads to the local format of the balance equation in the spatial configuration
˙ t + (•)t divv − ()t − div(⋆)t = 0 ∀x.
(•) (2.4.5)

The relevant balance quantities are the mass, the linear and angular momentum and the
energy. In the following the local formats of the balance equations are derived in the
material and spatial setting by utilising their general formats (2.4.2) and (2.4.5).

23
2 Basic Equations in Continuum Mechanics

balance of mass

We consider only closed systems, that means systems where the mass m is conserved.
The source and the flux terms vanish, the balance quantity (•) is the density ρ and the
balance equations take the following simple local formats

ρ˙0 = 0 ∀X
(2.4.6)
ρ̇t + ρt divv = 0 ∀x

The balance of mass in the spatial description is known as the continuity equation in fluid
mechanics.

balance of linear momentum

The time rate of change in linear momentum is equal to the sum of all external forces
acting on a body. The linear momentum density ρv is the balance quantity. The external
forces are composed of a source term which are the volume forces b, for instance gravity,
and the flux terms which are the surface tractions T = P · N or t = σ · n. The local
format of the balance of momentum in the material configuration follows as

ρ0˙V − b0 − Div(P ) = 0 ⇒ ρ0 A − b0 − Div(P ) = 0 ∀X, (2.4.7)

whereby the balance of mass, i.e. ρ˙0 = 0, was introduced to obtain the final format. In
the spatial configuration the following formulation is obtained

ρt˙v + ρt v div(v) − bt − div(σ) = 0

⇒ ρ̇t v + ρt v̇ + ρt v div(v) − bt − div(σ) = 0
(2.4.8)
⇒ −ρt div(v)v + ρt v̇ + ρt v div(v) − bt − div(σ) = 0
⇒ ρt a − bt − div(σ) = 0 ∀x.

Again the balance of mass was introduced to simplify the equation. The resulting equation
is also denoted as Cauchy’s equation of motion.

balance of angular momentum

The time rate of change in angular momentum is equal to the sum of all external moments
acting on a body. The angular momentum density r ×ρv is the balance quantity, whereby
r is the position vector r = x − x0 . The sum of external moments consists of the source
term, i.e. the moments due to the body forces r × b, and the flux term, the moments due
to the surface forces r×T or r×t. The balance of angular momentum results after several
manipulations and introduction of the balance equations of mass and linear momentum

24
 2.4 Balance Equations

in the following equations

F · PT = P · FT ∀X
(2.4.9)
σ = σT ∀x

The consequence of the balance of angular momentum is that the Cauchy stress tensor
is symmetric and the Piola stress tensor is generally not symmetric. If the Cauchy stress
tensor is symmetric the balance of angular momentum is automatically satisfied.

balance of mechanical energy

We consider only mechanical energy, other forms of energy such as thermal, chemical or
electrical energies are neglected. In this case the balance of energy is not an additional
statement to be satisfied but is a consequence of the balance of linear momentum. The
change in time of the total energy E is equal to the mechanical power of the external forces
P ext . The total energy density e is composed of the internal energy density i and the ki-
netic energy density k, whereby the internal energy denotes the sum of all microscopic
energy forms (random velocities of molecules, energy of chemical bonds, interaction po-
tentials, ...). The external mechanical power is composed of the power of the body forces
(source term) and the power of the surface forces (flux term) as

()0 = V · b0 and (⋆)0 · N = V · [P · N ]

(2.4.10)
()t = v · bt and (⋆)t · n = v · [σ · n].

The balance of energy in the material configuration in its local format results as

ė0 = i0 +˙ k0 = V · b0 + Div(V · P )
˙
⇒ i̇0 + 12 ρ0 V 2 = V · [b0 + Div(P )] + Ḟ : P
⇒ i̇0 = V · [−ρ0 A + b0 + Div(P )] +Ḟ : P (2.4.11)
| {z }
=0

⇒ i̇0 = Ḟ : P

When only mechanical energy is considered, the time rate of change of the internal energy
is equal to the internal mechanical power, i.e. the time rate of change of the internal
mechanical work done by the stresses
Z Z
i̇0 dV = Ḟ : P dV = P int (2.4.12)
B0 B0

The balance of energy in the spatial formulation is derived by means of the general balance
law (2.4.5). The first two terms of this equation can be simplified by means of the balance

25
2 Basic Equations in Continuum Mechanics

of mass

ėt + et div(v) = [i̇t + it div(v)] + [k̇t + kt div(v)]

= [i̇t + it div(v)] + [ 21 ρ̇t v 2 + ρt v · v̇ + 21 ρt v 2 div(v)] (2.4.13)

= [i̇t + it div(v)] + ρt v · v̇

Introducing this result and the flux and source terms (2.4.10) into the general balance
equation leads to the local format of the mechanical energy balance in the spatial config-
uration

[i̇t + it div(v)] + ρt v · v̇ = v · bt + div(v · σ)

⇒ [i̇t + it div(v)] = v · [−ρt v̇ + bt + div(σ)] +l : σ (2.4.14)
| {z }
=0

⇒ [i̇t + it div(v)] = l : σ

The time rate of change of the internal energy I or the internal power P int can be written
in the material or spatial format as
Z Z Z
˙I = P int = i̇0 dV = Ḟ : P dV = Ḟ : [Jσ · F −T ] dV
B
Z0 B0 Z B0 Z
−1
= Ḟ · F : σ dv = l : σ dv = [i̇t + it div(v)] dv (2.4.15)
Bt Z Bt Bt

= ... = Ė : S dV
B0

The double contractions of the respective rate of deformation tensors and stress tensors
describe the physical power, that is the rate of the internal mechanical work. In this sense
the pairs (Ḟ , P ), (Ė, S) and (l, σ) are said to be energy conjugated.
In the following we focus on the balance of linear momentum. The balance of mass is au-
tomatically satisfied if the density ρ0 is kept constant. The balance of angular momentum
is fulfilled by claiming that the Cauchy stress tensor is symmetric. And the balance of
mechanical energy follows from Cauchy’s equation of motion. When thermo-mechanical
problems are considered, the external heat power from heat fluxes and heat sources has
to be taken into account in the balance of energy. Then, the balance of energy specifies
an additional differential equation which has to be solved.

2.5 Weak Form of Equilibrium

In this section the weak forms of the balance of linear momentum are derived in the
material and the spatial formulation. The weak form of equilibrium is necessary for the

26
 2.5 Weak Form of Equilibrium

subsequent discretization and solution with the finite element method.

material description

• bal. of lin. momentum ρ0 ϕ̈ = Div P + b0 in B0

• Dirichlet bc ϕ = ϕ̄ on ∂B0ϕ
• Neumann bc T = P · N = T̄ on ∂B0σ
0. starting point: local form of material balance of momentum

−ρ0 ϕ̈ + Div P + b0 = 0
−ρ0 ϕ̈i + PiJ,J + b0 i = 0i
1. scalar product with test function η and integration over B0
Z Z Z
− η · ρ0 ϕ̈ dV + η · Div P dV + η · b0 dV = 0
ZB0 Z B0 B0
R
− ηi ρ0 ϕ̈i dV + ηi PiJ,J dV + B0 ηi b0 i dV = 0
B0 B0

2. partial integration of divergence term

Z Z Z
η · Div P dV = Div [ η · P ] dV − ∇X η : P dV
Z B0 Z B0 Z B0
ηi PiJ,J dV = [ ηi PiJ ],J dV − ηi,J : P iJ dV
B0 B0 B0

3. Gaussian integral theorem

Z Z
Div [ η · P ] dV = η · P · N dA
Z B0 Z∂B0
[ ηi PiJ ],J dV = ηi PiJ NJ dA
B0 ∂B0

4. boundary conditions

ϕ = ϕ̄ on ∂B0ϕ η = 0 on ∂B0ϕ T = P · N = T̄ on ∂B0σ

ϕi = ϕ̄i on ∂B0ϕ ηi = 0i on ∂B0ϕ Ti = PiJ · NJ = T̄i on ∂B0σ
Z Z Z
η · P · N dA = η · P · N dA + η · T̄ dA
Z∂B0 Z∂B0ϕ Z∂B0σ
ηi PiJ NJ dA = ηi PiJ NJ dA + ηi T̄i dA
∂B0 ∂B0ϕ ∂B0σ

• resulting weak form

Z Z Z Z
η · ρ0 ϕ̈ dV + ∇X η : P dV − η · T̄ dA − η · b0 dV = 0
Z B0 Z B0 Z∂B0 Z B0
ηi ρ0 ϕ̈i dV + ηi,J : PiJ dV − ηi T̄i dA − ηi b0 i dV = 0
B0 B0 ∂B0 B0

27
2 Basic Equations in Continuum Mechanics

with ∇X η : P = ∇X η : F · S = [F t · ∇X η]sym : S

Z Z Z Z
η ·ρ0 ϕ̈ dV + [F t ·∇X η ]sym :S dV − η ·T̄ dA− η ·b0 dV = 0
Z B0 Z B0 Z∂B0 Z B0 (2.5.1)
ηi ρ0 ϕ̈i dV + [FIit ηi,J ]sym :SIJ dV − ηi T̄i dA− ηi b0 i dV = 0
B0 B0 ∂B0 B0

interpretation as principle of virtual work

with η ← δϕ ... virtual displacement
dyn
δW0 = δW0 + δW0int − δW0ext = 0

with
Z
dyn
δW0 = δϕ · ρ0 ϕ̈ dV
Z B0 Z
δW int 0 = ∇X δϕ : P dV = δE : S dV
Z B0 Z B0
δW0ext = δϕ · T̄ dA + δϕ · ρ0 b dV
∂B0 B0

whereas δE = [ F t · δF ]sym ... virtual strain tensor

28
 2.5 Weak Form of Equilibrium

spatial description

• bal. of lin. momentum ρt ϕ̈ = div σ + bt in Bt

• Dirichlet bc ϕ = ϕ̄ on ∂B⊔ϕ
• Neumann bc t = σ · n = t̄ on ∂B⊔σ

0. starting point: local form of spatial balance of momentum

−ρt ϕ̈ + div σ + bt = 0
−ρt ϕ̈i + σij,j + bt i = 0i

1. scalar product with test function η and integration over Bt

Z Z Z
− η · ρt ϕ̈ dv + η · div σ dv + η · bt dv = 0
Z Bt Z Bt Z Bt
− ηi ρt ϕ̈i dv + ηi σij,j dv + ηi bt i dv = 0
Bt Bt Bt

2. partial integration of divergence term

Z Z Z
sym
η · div σ dv = div [ η · σ ] dv − ∇x η : σ dv
Z Bt Z Bt Z Bt
sym
ηi σij,j dv = [ ηi σij ],j dv − ηi,j : σij dv
Bt Bt Bt

3. Gaussian integral theorem

Z Z
div [ η · σ ] dv = η · σ · n da
Z Bt Z∂Bt
[ ηi σij ],j dv = ηi σij nj da
Bt ∂Bt

4. boundary conditions

ϕ = ϕ̄ on ∂Btϕ η = 0 on ∂Btϕ t = σ · n = t̄ on ∂Btσ

ϕi = ϕ̄i on ∂Btϕ ηi = 0i on ∂Btϕ ti = σij · nj = t̄i on ∂Btσ

Z Z Z
η · σ · n da = η · σ · n da + η · t̄ da
Z∂Bt Z∂Btϕ Z∂Btσ
ηi σij nj da = ηi σij nj da + ηi t̄i da
∂Bt ∂Btϕ ∂Btσ

• resulting weak form

Z Z Z Z
sym
η · ρt ϕ̈ dv + ∇x η : σ dv − η · t̄ da − η · bt dv = 0
Z Bt Z Bt Z∂Bt Z Bt (2.5.2)
sym
ηi ρt ϕ̈i dv + ηi,j : σij dv − ηi t̄i da − ηi bt i dv = 0
Bt Bt ∂Bt Bt

29
2 Basic Equations in Continuum Mechanics

interpretation as principle of virtual power

with η ← δv ... virtual velocity
dyn
δWt = δWt + δWtint − δWtext = 0

with
Z
dyn
δWt = δv · ρt ϕ̈ dv
Z Bt Z
sym
δW int
t = ∇x δv : σ dv = δlsym : σ dv
Z Bt Z Bt
δWtext = δv · t̄ da + δv · ρt b dv
∂Bt Bt

with δl = ∇x δv ... virtual velocity gradient

2.6 Linearisation

To solve the nonlinear equations of continuum mechanics the linearisation of the equations
is required (e.g. in iterative solution schemes). The general concept of linearisation is first
explained by means of a scalar function f which is n + 1-times continuously differentiable
at the position x̄. The Taylor series of f (x) at point x̄ takes the following format

1 ∂f 1 ∂ 2f 2 1 ∂ nf
Tf (x) = f (x̄) + |x=x̄ [x − x̄] + | x=x̄ [x − x̄] + ... + |x=x̄ [x − x̄]n (2.6.1)
1! ∂x 2! ∂x2 n! ∂xn
We consider only the first two terms of the Taylor series
∂f
Lf (x) = f (x̄) + |x=x̄ [x − x̄]. (2.6.2)
∂x
When x̄ is fixed and x is the independent variable in (2.6.2), the function Lf (x) is the
tangent on the function f in x = x̄. The linearised change in the function value ∆f , or
its linearization at x = x̄, is calculated as
∂f ∂f
∆f (x̄) = Lf (x) − f (x̄) = |x=x̄ [x − x̄] = |x=x̄ ∆x (2.6.3)
∂x ∂x
These one-dimensional results can be extended to scalar valued functions of points in the
three-dimensional space. Consider the function f = f (x). Its linearisation is given as its
directional derivative, defined as

∂f (x) d
∆f (x̄) = |x=x̄ · ∆x = (f (x̄ + ǫ∆x)) |ǫ→0 (2.6.4)
∂x dǫ
The directional derivative measures the incremental change of the function value f at
point x̄ in the direction of a straight line x̄ + ǫ∆x, whereby ǫ is a scalar parameter. The
directional derivative of vector- and tensor-valued functions is defined analogously.

30
 2.6 Linearisation

linearisation of kinematical quantities

displacements u = ϕ(X) − X

∆u = ∂u · ∆ϕ = d [u(ϕ + ǫ∆ϕ)]|ǫ→0
∂ϕ dǫ (2.6.5)
= d [ϕ + ǫ∆ϕ − X]|ǫ→0 = ∆ϕ
dǫ
deformation gradient F = ∇X ϕ

∆F = ∂F · ∆ϕ = d [F(ϕ + ǫ∆ϕ)]|ǫ→0
∂ϕ dǫ (2.6.6)
= d [∇X ϕ + ǫ∇X ∆ϕ]|ǫ→0 = ∇X ∆ϕ
dǫ
Green-Lagrange strain E = 12 [FT · F − 1]

∆E = ∂E · ∆ϕ = d [E(ϕ + ǫ∆ϕ)]|ǫ→0
∂ϕ dǫ
= d 21 [FT (ϕ + ǫ∆ϕ) · F(ϕ + ǫ∆ϕ) − 1] |ǫ→0
 
dǫ
= d 21 [(∇XT ϕ + ǫ∇XT ∆ϕ) · (∇X ϕ + ǫ∇X ∆ϕ) − 1] |ǫ→0
 
dǫ (2.6.7)
= 21 [∇XT ∆ϕ · (∇X ϕ + ǫ∇X ∆ϕ) + (∇XT ϕ + ǫ∇XT ∆ϕ) · ∇X ∆ϕ]|ǫ→0
1
= [∇T ∆ϕ
2 X
· ∇X ϕ + ∇XT ϕ · ∇X ∆ϕ]
1
= 2
[∆FT · F + FT · ∆F] = [∆FT · F]sym
alternative derivation of E
∆E = ∂E : ∆F = d [E(F + ǫ∆F)]|ǫ→0
∂F dǫ
= d 1 T
[(F + ǫ∆F T
) · (F + ǫ∆F) − 1]

|ǫ→0
dǫ 2 (2.6.8)
= 21 [∆FT · (F + ǫ∆F) + (FT + ǫ∆FT ) · ∆F]|ǫ→0
1
= 2
[∆FT · F + FT · ∆F] = [∆FT · F]sym

The evaluation of ∆E at F̄ = 1 leads to the linear strain tensor (with ∆u = u)

1 1 1
∆E = [∆FT + ∆F] = [∇XT ∆u + ∇X ∆u] = [∇XT u + ∇X u] = ε. (2.6.9)
2 2 2
linearisation of constitutive equations
St.-Vernant hyperelasticity (compare section 2.3)
2nd Piola-Kirchhoff stress tensor S

S = λ[E : 1]1 + 2µE (2.6.10)

linearisation of 2nd Piola-Kirchhoff stress ∆S

∆S = ∂S : ∆E = d [S(E + ǫ∆E)]|ǫ→0
∂E dǫ
= d [λ[(E + ǫ∆E) : 1]1 + 2µ(E + ǫ∆E)] |ǫ→0 (2.6.11)
dǫ
= [λ 1 ⊗ 1 + 2µ Isym ] : ∆E = L : ∆E

31
2 Basic Equations in Continuum Mechanics

linearisation of the weak form

The weak formulation in the material description is given in equation (2.5.1). This equa-
tion is nonlinear in the deformation map ϕ. The solution of nonlinear equations requires
iterative methods, commonly the Newton-Raphson method is applied, which will be in-
troduced in chapter 4. Starting at an approximate solution ϕ0 (e.g. the last equilibrium
point), an improvement to this approximate solution is iteratively calculated by means of
the linearisation of the nonlinear equation.
∂δW0
LδW0 (ϕ) = δW0 (ϕ0 ) + |ϕ=ϕ0 · ∆ϕ = 0 (2.6.12)
∂ϕ
The improved approximation on the solution is then updated as

ϕ1 = ϕ0 + ∆ϕ (2.6.13)

and the next iteration steps follow by evaluating equation (2.6.12) at point ϕk and up-
dating the solution according to equation (2.6.13) until a convergence criterion is fulfilled.
Thus, application of the Newton method requires the linearisation of the weak form at
the points ϕk . Neglecting the dynamic part, the weak formulation is given in the format
of the virtual work as
Z Z Z
T
δW0 = [∇X δϕ · F ] : S dV − δϕ · ρ0 b dV − δϕ · T̄ dA = 0. (2.6.14)
B0 B0 ∂B0

We assume that the external loads are conservative (i.e. do not depend on the deforma-
tion). Then, only the linearisation of the internal virtual work is required. It is calculated
according to equation (2.6.4) as
Z Z
∆δW0 = T
[∇X δϕ · ∆F ] : S dV + [∇XT δϕ · F ] : ∆S dV
B0Z Z B0

= δ∆E : S dV + δE : L : ∆E dV (2.6.15)
B0 B0
| {z } | {z }
geometrical part material part

Here, the variation of the Green-Lagrange strain δE and its linearisation ∆δE = δ∆E
were introduced. The variation of a function δ(•) is formally calculated in the same way
as its linearisation:

δf d (f (x̄ + ǫδx))|
= ǫ→0
dǫ (2.6.16)
δF = ∇X δϕ; δE = [δF T · F ]sym ; δ∆E = [∆F T · δF ]sym

The two parts of the linearisation ∆δW0 are denoted as the geometrical and the material
part. The geometrical part vanishes if the strains and rotations are small. The second

32
 2.7 Nonlinear Analysis of a Three-Hinged Arch

part which contains the elasticity tensor is the material part. In a geometrically linear
analysis only the material part exists.
The linearisation of the weak form in the spatial configuration is derived by means of the
push forward of the linearised material weak form. By application of the transformation
rules for the volume elements and the stress and strain tensors, the following result is
obtained
Z Z
T sym
∆δWt = [∇x η · ∇x ∆ϕ] : σ dv + ∇xsym η : E : ∇xsym ∆ϕ dv (2.6.17)
Bt Bt

Again, we can identify a geometrical and a material part of the linearisation.

remarks:
The discretisation of the linearised equations as given above does not coincide in all cases
with the linearisation of the discrete equations, which result from a finite element discreti-
sation. For a finite element discretisation with standard volume elements the equivalence
of both approaches can be shown. In general, the weak form should be discretized first
(→ section 3.2) and then linearized.
The material and spatial descriptions yield similar results. The individual integral terms
can be either evaluated in the material or the spatial form. Constitutive equations de-
scribed in the material form (e.g. St.-Venant-Kirchhoff) motivate a material description
of the weak form, and spatial constitutive laws (e.g. Neo-Hooke) motivate a spatial de-
scription.

2.7 Nonlinear Analysis of a Three-Hinged Arch

In the present section the equations of nonlinear continuum mechanics are formulated and
evaluated for a simple, quasi one-dimensional example. The geometry of the deformed
and undeformed configuration are given in figure 2.6. It is assumed that the deformation
2f ext
L 2f ext
w l
X
E mod
A x

B B B

Figure 2.6: Three-hinged arc - geometry

in each bar is homogeneous, i.e. the deformation gradient F (which is a scalar in 1D) is
constant. Due to symmetry only one half of the system is considered. First the kinematic
relations have to be defined
√ √
L = B2 + X 2 l = B 2 + x2 (2.7.1)

33
2 Basic Equations in Continuum Mechanics

Due to the quasi one-dimensional setting, it has to be distinguished carefully between l,

which denotes the elongation of the bar in the direction of the bar, and X or x, respectively,
which denote the vertical positions of the center point of the arch in the undeformed or
deformed configuration.
The deformation gradient and the strain measures are defined for the one-dimensional
setting as
l l 1 1 l 2 − L2 1 2
F = ; J = F = ; E = [F T F − 1] = 2 = 2
2 [x − X ] (2.7.2)
L L 2 2 L 2L
For later use the linearisations and variations of the kinematical quantities are derived.
p
∆l = d (l(x + ǫ∆x))|ǫ→0 = d ( [B 2 + [x + ǫ∆x]2 ])|ǫ→0
dǫ dǫ
1 2 2 − 21
= [ 2 [B + [x + ǫ∆x] ] 2[x + ǫ∆x]∆x]|ǫ→0 = [B 2 + x2 ]− 2 x ∆x
1

= 1 x ∆x
l
δl = d (l(x + ǫδx))|ǫ→0 = 1 x δx
dǫ l
d l + ǫ∆l)
∆F = (F (l + ǫ∆l))|ǫ→0 = d ( L )|ǫ→0 = ∆l L
dǫ dǫ
δF = d (F (l + ǫδl))|ǫ→0 = L δl (2.7.3)
dǫ
∆E = d (E(F + ǫ∆F ))|ǫ→0 = d ( 21 [(F + ǫ∆F )(F + ǫ∆F ) − 1])|ǫ→0
dǫ dǫ
1
= 2 [∆F (F + ǫ∆F ) + (F + ǫ∆F )∆F ]|ǫ→0 = 12 l∆l
L
δE = d (E(F + ǫδF ))|ǫ→0 = 2 lδl 1
dǫ L
δ∆E = d (∆E(l + ǫδl))|ǫ→0 = d ( 12 (l + ǫδl)∆l)|ǫ→0
dǫ dǫ L
= 1 δl∆l
L2
The material behaves hyperelastic and the St.-Venant-Kirchhoff model is applied. In the
one-dimensional setting the elasticity tensor and the stress tensor become scalar variables,
so that the constitutive equation can be written as

S = E mod E, (2.7.4)

whereby E mod denotes the Young’s modulus. The principle of virtual work is used to derive
the weak formulation of the problem (which is not necessary here, one could also directly
set up the strong form and solve that, it is just for training purposes...). Dynamical effects
are neglected, that is ϕ̈ → 0. The principle of virtual work in the material configuration
reads

δW0 (x) = δW0int (x) − δW0ext = 0 ∀ δx

ZL
int
δW0 (x) = δE E mod E A dS (2.7.5)
0

δW0ext (x) = −δx f ext .

34
 2.7 Nonlinear Analysis of a Three-Hinged Arch

Thereby, the volume element is expressed by means of the cross section area A as
dV = A dS, and the external force f ext replaces the integral of tractions over the cross
section area. Now the variations of the kinematic quantities (2.7.3) are introduced
ZL
1 mod 1
δW0 (x) = 2 x δx E [x2 − X 2 ] AdS + δx f ext = 0
L 2L2
 0 
ZL
1 3
⇒ [x − x X 2 ]E mod AdS + f ext  δx = 0
2L4 (2.7.6)
0 
 1 3 2 mod

ext 
⇒
 2L3 [x − x X ]E A +f  δx = 0 ∀δx
| {z }
f int

The resulting equation is nonlinear in x and an iterative solution scheme has to be ap-
plied to solve for the unknown x for a given external force f ext . However, due to the
one-dimensional setting it is possible to plot the nonlinear load-displacement relation by
precribing different deflections w and computing the related external forces f ext . The
load-displacement relation is plotted in figure 2.7, formulated in terms of the load factor
λ and the deflection w = X − x. The height is chosen as H = X = 1. Equation (2.7.6) is
rearranged to obtain

2L3
f ext = x X 2 − x3 ⇒ λ = [X − w]X 2 − [X − w]3 .
mod (2.7.7)
|E {zA }
=λ

0.8

linear solution
0.6

0.4
load factor λ

0.2

-0.2

-0.4
0 0.5 1 1.5 2
deflection w

Figure 2.7: Three-hinged arc - load-displacement relation

35
3 Finite Element Discretisation

In the following chapter the numerical solution of the weak form by means of the finite
element method is derived. The chapter is focussed on the essential details needed to
perform a standard nonlinear analysis using finite elements. All relevant equations are
presented, but some basic operations, which are known from linear finite element meth-
ods, such as the assembly process, are not discussed.
The geometry of the problem is discretised with finite elements and the unknown field
values (deformation, strains, stresses) are approximated by means of elementwise polyno-
mial shape functions. The real geometry of the body in the reference configuration B0 is
subdivided into and approximated by nel finite elements
n[
el −1

B0 ≈ B0h = B0e . (3.0.1)

e=0

Overlaps or gaps between elements are not allowed. The boundary of the body is subdi-
vided as
n[
be −1

∂B0 ≈ ∂B0h = ∂B0e , (3.0.2)

e=0

whereby nbe denotes the number of element edges which form the boundary. This is
usually an approximation of the real boundary of the geometry, compare figure 3.1.

∂B0

∂B0h
B0e

Figure 3.1: subdivision into finite elements

3.1 Isoparametric Concept

Within the finite element method ansatz functions have to be chosen to approximate the
unknown primary variable, here the deformation. These ansatz (or shape or interpola-

37
3 Finite Element Discretisation

tion) functions are defined on each element and are usually polynomials. The idea of
the isoparametric concept is that the geometry and the unknown deformation are ap-
proximated by the same ansatz functions. The concept is well suited for geometrically
nonlinear problems since the undeformed and the deformed configuration can be approx-
imated by the same ansatz functions. Furthermore, isoparametric elements allow for a
sufficiently accurate approximation of the geometry.
In order to introduce the concept of the later used vector-valued ansatz functions, first
the elementwise approximation of a scalar variable T is considered. On each element e
the variable T is approximated as

nX
en −1
e
T = Ni (ξ)Ti (3.1.1)
i=0

with the scalar polynomial ansatz functions Ni , which are usually defined in the local
coordinates ξ on a reference element B , and the nodal values Ti . The number of element
nodes is denoted by nen and the sum starts from 0 due to C++ convention. For a scalar
variable the number of degrees of freedom per element ndof is equal to nen . For the here
considered mechanical problems the approximations of vector-valued variables, e.g. the
geometry or the deformation, are required and done by means of vector-valued ansatz
functions N i . In this case, the index i counts the degrees of freedom, not the nodes.
These vector-valued ansatz functions consist of the previously introduced scalar-valued
ansatz functions Nen(i) multiplied with a unit vector edir(i) , such that all entries in the
vectors N i are equal to zero except the one coefficient in the direction of the corresponding
degree of freedom. The indices en(i) and dir(i) denote the element node and the direction
that belong to degree of freedom i.
As an example the vector-valued ansatz functions for a two-dimensional element with
three element nodes, i.e. six degrees of freedom, are specified: The range of the index i is
i = 0, ..., 5. There are three scalar-valued ansatz functions, related to the element nodes
N0 , N1 , N2 . The six vector-valued ansatz functions take the following format
     
N0 0 N1
N 0 = N0 e 1 =   , N 1 = N 0 e2 =   , N 2 = N 1 e1 =  ,
0 N0 0
      (3.1.2)
0 N2 0
N 3 = N1 e 2 =   , N 4 = N 2 e1 =   , N 5 = N 2 e1 =  
N1 0 N2

These vector-valued ansatz functions simplify the discretization of the weak form and are
in agreement with the notation used in the finite element library deal.ii, which is applied
in the computer exercises.
The elementwise approximations of the material and spatial coordinates take the following

38
 3.1 Isoparametric Concept

format
ndof −1 ndof −1
X X
e e
X = N i (ξ)Xi and x = N i (ξ)xi , (3.1.3)
i=0 i=0

whereby Xi , xi denote the coefficients of the nodal coordinates in the material or spatial
configuration, respectively.
For each element a transformation has to be performed which relates the global coordi-
nates X e , xe with the local coordinates ξ on the reference element. The transformation
has to satisfy the following conditions:

• for each point in the reference element B , exists one and only one point in B0e

• the nodal points in B are related to the nodal points in B0e

• each part of the boundary of B defined by some nodal points, is related to a part
of the boundary of B0e , defined by the same nodal points

These assumptions ensure that the shape of the reference element is preserved during
the isoparametric transformation, e.g. triangle remains triangle. The introduction of a
reference element simplifies the formulation of the ansatz functions and the integration
of the weak form. The isoparametric mapping is illustrated for one element in figure 3.2.
The transformation from the reference element to the undeformed or deformed element

ϕe (X), F e
X2 B0e Bte x2
X1 x1

X e (ξ), J e xe (ξ), j e

ξ2
ξ1

B

Figure 3.2: isoparametric concept, mapping from reference element

requires the gradients of the material and spatial coordinates with respect to the reference

39
3 Finite Element Discretisation

coordinates ξ which are defined by use of equation (3.1.3) as

ndof −1
e X
J e
= ∂X = ∇ξ N i Xi
∂ξ
i=0
ndof −1 (3.1.4)
e X
je = ∂x = ∇ξ N i xi .
∂ξ
i=0

The 2nd order tensor ∇ξ N i denotes the gradient of the vector-valued shape function N i
with respect to the local coordinates ξ. With these definitions at hand, the gradients of
the ansatz functions with respect to the spatial or material coordinates can be derived

∂ξ
∇X N i = ∂N i · = ∇ξ N i · [J e ]−1
∂ξ ∂X e (3.1.5)
∂N ∂ξ
∇x N i = i · = ∇ξ N i · [j e ]−1 .
∂ξ ∂xe

In the following the approximation of kinematic quantities in terms of the polynomial

ansatz functions N i is specified. The elementwise deformation map ϕe is introduced to
relate the element in the undeformed configuration with the element in the deformed
configuration. Furthermore, the deformation gradient for this element F e = ∇X ϕe is
defined
ndof −1
X
e e
ϕ =x = N i (ξ) ϕi
i=0
ndof −1 ndof −1 (3.1.6)
X X
e e e −1
F = ∇X ϕ = ∇X N i ϕi = ∇ξ N i · (J ) ϕi
i=0 0=1

whereby ϕi denote the (unknown) nodal coefficients of the deformation. Following the
Bubnov-Galerkin method, the same ansatz functions as for the approximation of the
deformation are used to approximate the test function (or virtual deformation):
ndof −1
X
e
η = N i (ξ) ηi
i=0
ndof −1 ndof −1 (3.1.7)
X X
∇X η e = ∇X N i ηi = ∇ξ N i · (J e )−1 ηi
0=1 i=0

with ηi being the nodal coefficients of the test function.

The ansatz functions have to satisfy some requirements to yield convergence of the finite
element solution to the exact solution. For the here considered volume elements the
ansatz function have to be C 0 continuous. Classical beam and shell elements require
higher continuity since the underlying partial differential equations are of higher order,
but these elements are not discussed within the course. The ansatz functions have to be
complete polynomials and should be interpolating, i.e. 1 at one node, and 0 at the other

40
 3.1 Isoparametric Concept

nodes: Nen(i) (ξ en(j) ) = δij . There exist different possibilities to satisfy these conditions,
one is the use of Lagrange polynomials. In one dimension they are defined as
nY
ne −1
ξj − ξ
Ni (ξ) = , (3.1.8)
j=0, j6=i
ξj − ξi

whereby ξi denotes the the local coordinate at node i. In the following, the isoparametric
scalar-valued ansatz functions for one, two and three-dimensional elements are formulated
and the numerical integration within the reference element is introduced.

One-dimensional approximation

The one-dimensional ansatz functions are derived by inserting the local coordinate
ξ ∈ [−1, 1] into equation (3.1.8). The approximation of the geometry and the deformation
is then obtained as
ndof −1 ndof −1
X X
e e
X = Ni (ξ)Xi and ϕ = Ni (ξ)ϕi . (3.1.9)
i=0 i=0

The ansatz functions can be specified for different polynomial degrees, whereby the num-
ber of nodes increases with the polynomial degree.

• linear ansatz functions (ξ0 = −1, ξ1 = 1)

1 1
N0 = [1 − ξ] and N1 = [1 + ξ] (3.1.10)
2 2

• quadratic ansatz functions (ξ0 = −1, ξ1 = 0, ξ2 = 1)

1 1
N0 = − ξ[1 − ξ], N1 (ξ) = 1 − ξ 2 and N2 = ξ[1 + ξ] (3.1.11)
2 2

N0 N1 N0 N2
N1 node
Gauss point
0 1 0 1 2
ξ ξ

Figure 3.3: one-dimensional linear and quadratic ansatz functions

It can be easily verified that the ansatz functions are interpolating. The linear and
quadratic one-dimensional reference elements together with the related ansatz functions
are illustrated in figure 3.3.

41
3 Finite Element Discretisation

example: one-dimensional linear element:

With the specification of the ansatz functions the approximation of the geometry and of
the gradients can be derived. This is carried out in the following examplarily for a linear
element with the nodal coordinates [X0 , X1 ] = [0, Le ].
The derivatives of the linear ansatz functions with respect to the local coordinates are
derived as
1 1
∇ξ N0 = − , ∇ξ N1 = .
2 2
The approximation of the geometry is specified by means of equation (3.1.3) as
1
X 1
Xe = Ni (ξ)Xi = [1 + ξ] Le
i=0
2

and the gradient J e is either derived by directly taking the derivative of X e with respect
to the local coordinate ξ or by application of equation (3.1.4) as
1
X Le
Je = Xi ∇ξ N i = .
i=0
2

The derivatives with respect to the material coordinates are then obtained by means of
equations (3.1.5)

∇X N0 = ∇ξ N0 (J e )−1 = − 12 2
Le = − L1
e
∇X N1 = ∇ξ N1 (J e )−1 = 1 2 = 1
2 Le L e

Assuming that the nodal deformations [ϕ0 , ϕ1 ] are known the deformation gradient is
computed by means of equation (3.1.6)
1
e
X ϕ1 − ϕ0
F = ϕi ∇X Ni =
i=0
Le

and is constant within the element.

Numerical integration on the reference element
The evaluation of the weak form requires the integration within each element. Using
the isoparametric concept these integrations are carried out within the reference element.
The integral terms have to be transformed from the undeformed or deformed configuration
to the reference element. For an arbitrary function g(X) this transformation takes the
following format
XZ
1 +Le Z1 Z1
dX
g(X)dX = g(ξ) dξ = g(ξ)J e (ξ)dξ. (3.1.12)
dξ
X1 −1 −1

42
 3.1 Isoparametric Concept

In general the occuring integral terms can not be solved analytically and therefore a
numerical integration is carried out. For standard finite elements with polynomial ansatz
functions Gauss integration is applied due to its accuracy and efficiency. The integral
is approximated by the sum of function values, evaluated at the ngp Gauss points ξg ,
multiplied with weight factors wg

Z1 ngp −1
X
e
g(ξ)J (ξ)dξ ≈ g(ξg )J e (ξg ) wg (3.1.13)
−1 g=0

A polynomial of degree p = 2ngp − 1 is exactly integrated with ngp Gauss points. The
Gauss points and related weighting factors in one dimension are specified in table 3.1 for
one, two and three Gauss points.

ngp ξg wg
1 ξg0 = 0 wg0 = 2
2 ξg0 = − √1 , ξg1 = √1 wg0 = 1, wg1 = 1
3 3
3 ξg0 = − √ , ξg1 = 0, ξg2 = √3
3 wg0 = 59 , wg1 = 98 , wg2 = 59
5 5

Table 3.1: Gauss points and weighting factors in one dimension

Two-dimensional approximation

In two dimensions different shapes of elements can be defined, the most common ones are
triangles and quadrilaterals. In the following the bilinear and biquadratic ansatz functions
for triangular and quadrilateral elements are specified.
Triangular elements
The simplest two dimensional element is the linear triangle (or constant strain triangle)
with three nodes (and 6 degrees of freedom). The scalar-valued ansatz functions are
defined by means of the local coordinates of the reference element as depicted in figure
3.4, left, and take the following format

N0 = 1 − ξ 1 − ξ2 , N1 = ξ1 , N2 = ξ 2 (3.1.14)

with ξ1 , ξ2 ∈ [0, 1]. It can be easily verified that the ansatz functions are interpolating and
that their derivatives with respect to the local coordinates are constant. This element
is simple and robust but shows poor approximation properties. It behaves too stiff,
especially in bending and compression, meaning that the numerical results obtained with
this element show smaller deformations than the analytical solution.

43
3 Finite Element Discretisation

2 2

B 4
5
ξ2 ξ2 B

0 ξ1 1 0 ξ1 3 1

Figure 3.4: linear and quadratic triangular reference elements

A better approximation is obtained with the quadratic triangle, compare figure 3.4, right,
with the ansatz functions

N0 = ξ3 [2ξ3 − 1], N1 = ξ1 [2ξ1 − 1], N2 = ξ2 [2ξ2 − 1]

(3.1.15)
N3 = 4ξ1 ξ3 , N4 = 4ξ1 ξ2 , N5 = 4ξ2 ξ3

whereby the abbreviation ξ3 = 1 − ξ1 − ξ2 was introduced.

Quadrilateral elements
The simplest quadrilateral element has four nodes (that means 8 degrees of freedom)
and bilinear ansatz functions. The scalar-valued ansatz functions are constructed by the
product of the one-dimensional Lagrange polynomials (3.1.8), defined either in ξ1 or in ξ2 .
The reference element is depicted in figure 3.5, left, and the interpolating ansatz functions
are defined as

N0 = 41 [1 − ξ1 ][1 − ξ2 ], N1 = 41 [1 + ξ1 ][1 − ξ2 ]
(3.1.16)
N2 = 14 [1 + ξ1 ][1 + ξ2 ], N3 = 14 [1 − ξ1 ][1 + ξ2 ]

with the local coordinates ξ1 , ξ2 ∈ [−1, 1]. A biquadratic quadrilateral element has 9

3 2 3 6 2
ξ2 ξ2

7 5
ξ1 ξ1
B B

0 1 0 4 1

Figure 3.5: bilinear and biquadratic quadrilateral reference elements

nodes as illustrated in figure 3.5, right, and the ansatz functions can be derived by the use
of the product formula, introducing the quadratic one-dimensional langrange polynomials.

44
 3.1 Isoparametric Concept

Numerical integration on the reference element

For the quadrilateral elements the integration of the weak form is carried out within
the reference element, similar to the one-dimensional case, and the integrals have to be
transformed from the undeformed or deformed configuration to the reference element.
The integral of an arbitrary function g(X) is transformed as
Z Z Z1 Z1
e
g(X)dA = g(ξ) detJ dA = g(ξ) detJ e (ξ) dξ1 dξ2 . (3.1.17)
B0e B ξ1 =−1 ξ2 =−1

Please note that in this two-dimensional case the area element dA is not an element of
the surface area, but rather a volume element with an assumed thickness of 1. Therefore,
it holds that 1 dA = dV = detJ e dV = detJ e 1dA .
The integral is evaluated by numerical Gauss integration as
Z1 Z1 ngp −1
X
e
g(ξ) detJ (ξ) dξ1 dξ2 ≈ g(ξ g ) detJ e (ξ g ) wg (3.1.18)
g=0
ξ1 =−1 ξ2 −1

The Gauss points and weighting factors for one and 4 Gauss points on quadrilateral
elements are given in the following table 3.2.

ngp ξg wg
1 ξ g0 = (0, 0) wg0 = 4
4 ξ g0 = (− √1 , − √1 ), ξ g1 = ( √1 , − √1 ) wg0 = wg1 = wg2 = wg3 = 1
3 3 3 3
1 1 1
ξ g2 = ( √ , √ ), ξ g3 = (− √ , √ )1
3 3 3 3

Table 3.2: Gauss points and weighting factors for quadrilateral elements

The transformation of the integrals to the reference element is different for triangular
elements and defined as
Z Z1 1−ξ
Z 1
g(X)dA = g(ξ) detJ e (ξ) dξ1 dξ2 (3.1.19)
B0e ξ1 =0 ξ2 =0

The numerical integration of (3.1.19) results in

Z1 1−ξ
Z 1 ngp −1
X
e
g(ξ) detJ (ξ) dξ1 dξ2 ≈ g(ξ g ) detJ e (ξ g ) wg (3.1.20)
g=0
ξ1 =0 ξ2 =0

and requires different Gauss points and weighting factors as specified in table 3.3.

45
3 Finite Element Discretisation

ngp ξg wg
1 ξ g0 = ( 31 , 31 ) wg0 = 12
3 ξ g0 = ( 61 , 61 ), ξ g1 = ( 16 , 32 ), ξ g2 = ( 23 , 61 ) wg0 = wg1 = wg2 = 61

Table 3.3: Gauss points and weighting factors for triangular elements

Three-dimensional approximation

The formulation of isoparametric elements in three dimensions follows the same strategy
as in one and two dimensions. Tetrahedral or hexahedral reference elements are defined
as illustrated in figure 3.6 and interpolatory polynomial ansatz functions are specified.
The numerical integration is again carried out on the reference element. For the explicit

ξ2
2 3 2
6
7
ξ2 ξ1
0 1 ξ1 ξ3
0 1

3 4 5
ξ3

Figure 3.6: linear tetrahedral and trihexahedral reference elements

form of the ansatz functions and the required Gauss points and weighting factors it is
referred to the textbook of Wriggers [9].

3.2 Discretisation and Linearisation of the Weak Form

The weak formulation of the governing equations which was derived in section 2.5 is
now discretised with finite elements. Therefore, the approximations of the test function
and other kinematic quantities, which were derived in the last section, are introduced
into the weak form. The dicrete weak forms are derived in the material and the spatial
configuration.

46
 3.2 Discretisation and Linearisation of the Weak Form

Discretisation of the weak form in the material description

Starting point is the continuous weak form in the material configuration, given in equation
(2.5.1), which is recapitulated here for the sake of clarity
Z Z Z Z
T
δW0 = η · ρ0 ϕ̈dV + [F · ∇X η] sym : SdV − η · T̄ dA − η · b0 dV = 0. (3.2.1)
B0 B′ ∂B0σ B0

Since the approximations are defined elementwise the integral over the body B0 has to be
subdivided into integrals over the particular finite elements B0e
Z Z nel −1 Z nel −1 Z
(...)dV ≈ (...)dV = A (...)dV = A (...)detJ e dV , (3.2.2)
e=0 e=0
B0 B0h B0e B

nel −1
whereby the operator
e=0
A denotes the assembly of all element contributions (compare
linear finite elements). To derive the discrete weak formulation the approximations of the
test function and its gradient are introduced
ndof −1 ndof −1
X X
e e
η = N i ηi and ∇X η = ∇X N i ηi
i=0 i=0

to obtain

nel −1 Z Z
A [F T · ∇X N i ηi ]sym : SdV

δW0h =  N i ηi · ρ0 ϕ̈dV +
e=0
B0e B0e
 (3.2.3)
Z Z

− N i ηi · T̄ dA − N i ηi · b0 dV  = 0,
∂B0e B0e

whereby the sum signs were omitted for clearer notation, but summing over the index i is
assumed. This equation has to be valid for arbitrary nodal coefficients of the testfunction
ηi and is rearranged as

nel −1 Z Z
δW0 = A ηi  N i · ρ0 ϕ̈dV + [F T · ∇X N i ]sym : SdV
h 
e=0
B0e B0e
 (3.2.4)
Z Z

− N i · T̄ dA − N i · b0 dV  = 0. ∀ηi
∂B0e B0e

The coefficients of the test function have to be arbitrary. Therefore, the sum of the
integral terms in (3.2.4) has to vanish. The integral terms can be summarized to obtain
the following scalar-valued equations of equilibrium

rI = fIdyn + fIint − fIext = 0 ∀I = 0, ..., ngf − 1 (3.2.5)

47
3 Finite Element Discretisation

with
nel −1 Z
fIdyn = A
e=0
N i · ρ0 ϕ̈dV
B0e
nel −1 Z
fIint = A [F T · ∇X N i ]sym : SdV (3.2.6)
e=0
B0e
nel −1 Z Z
fIext = A
e=0
N i · T̄ dA + N i · b0 dV.
∂B0e B0e

Thereby, ngf is the total number of degrees of freedom in the finite element mesh, and
nel −1
the assembly operator A
e=0
ensures that the contributions of the element degrees of free-
dom (indicated with small index i ) are assigned to the correct global degrees of freedom
(indicated with upper-case index I ). Equation (3.2.5) is in general a nonlinear system of
equations in the unknown nodal coefficients of the deformation ϕJ .

Linearisation of the weak form in the material description

The efficient solution of the nonlinear equation by means of the Newton-Raphson method
requires its linearisation. Different solution methods for time independent and time de-
pendent problems are discussed in the following chapters, here only the linearisation of
the discrete weak form is derived. The linearisation of the residuum in equation (3.2.5)
is obtained as
∂rI (ϕJ )
∆ rI = ∆ϕJ = KIJ ∆ϕJ (3.2.7)
∂ϕJ
in which we have introduced the coefficients of the global tangential stiffness matrix KIJ
and the incremental update of the solution coefficients ∆ϕJ . The stiffness matrix can be
obtained in two different ways, either as the linearisation of the discrete weak form (3.2.5)
or by discretisation of the linearised weak form (2.6.15). For the introduced isoparametric
finite elements, both possibilities result in the same tangential stiffness matrix. In general,
the linearisation of the discrete weak form gives the correct tangential stiffness matrix.
The dynamic part of the stiffness matrix is obtained as
nel −1 Z
fIdyn = A
e=0
N i · ρ0 ϕ̈ dV
B0e
nel −1 Z nel −1 Z
∂ ϕ̈ ∂ ϕ̈
∆fIdyn = A
e=0
N i · ρ0
∂ϕ
· ∆ϕ dV = A
e=0
N i · ρ0
∂ϕ
· N j ∆ϕj dV
(3.2.8)
B0e B0e
 
nel −1 Z
∂ ϕ̈
= A  N i · ρ0 dyn
 
· N j dV  ∆ϕJ = KIJ ∆ϕJ .
e=0 ∂ϕ
B0e

48
 3.2 Discretisation and Linearisation of the Weak Form

The explicit format of the dynamic stiffness matrix depends on the discretisation in time,
i.e. a certain ansatz for the accelerations has to be introduced to compute the partial
derivative ∂ ϕ̈/∂ϕ. Different approaches are discussed in chapter 5.
The internal part of the stiffness matrix is derived as

nel −1 Z
fIint = A [F T · ∇X N i ]sym : S dV
e=0
B0e
 
nel −1 Z Z
A [∆F T · ∇X N i ]sym : S dV + [F T · ∇X N i ]sym : ∆S dV
 
∆fIint = 
e=0
e e
B0 B0

nel −1 Z Z
= A  [∆ϕj ∇XT N j · ∇X N i ]sym : S dV + [F T · ∇X N i ]sym : L : ∆E dV
 
e=0
e B0e
B0
nel −1 Z
= A  [∇XT N j · ∇X N i ]sym : S ∆ϕj dV

e=0
B0e

Z
[F T · ∇X N i ] : L : [F T · ∇X N j ] ∆ϕj dV

+
B0e

nel −1 Z
A [∇XT N j · ∇X N i ]sym : S dV

= 
e=0
B0e

Z
[F T · ∇X N i ] : L : [F T · ∇X N j ] dV  ∆ϕJ

+
B0e

= KIJ ∆ϕJ
(3.2.9)

The internal stiffness matrix is composed of a geometrical and a physical part, as was
already noted in the previous chapter. For geometrically linear problems the geometrical
part vanishes.
Conservative loading is assumed and the linearization of the external forces vanishes

nel −1 Z Z
fIext = A
e=0
N i · T̄ dA + N i · b0 dV
∂B0e B0e (3.2.10)
∆fIext = KIJ
ext
∆ϕJ ext
with KIJ = 0.

If deformation dependent loading is considered the linearisation o the external forces gives
an additional contribution to the tangential stiffness matrix.

49
3 Finite Element Discretisation

Discretisation of the weak form in the spatial description

Starting point is the continuous weak form in the spatial configuration, given in equation
(2.5.2), which is recapitulated here for the sake of clarity
Z Z Z Z
sym
δWt = η · ρt ϕ̈ dv + ∇x η : σdv − η · t̄da − η · bt dv = 0 (3.2.11)
Bt Bt ∂Bt σ Bt

Again, the integral is partitioned into integrals over the elements and the approximation
of the test function and its gradient with respect to the spatial coordinates
ndof −1
X
e
∇x η = ∇x N i ηi (3.2.12)
i=0

are introduced to obtain


nel −1 Z Z
δWt = A

 N i ηi · ρt ϕ̈dv + [∇x N i ηi ] : σdv
e=0
Bte Bte
 (3.2.13)
Z Z

− N i ηi · t̄da − N i ηi · bt dv  = 0
∂Bte Bte

which has to be valid for arbitrary nodal coefficients of the test function ηi and is rear-
ranged similarly as the material weak form

nel −1 Z Z
δWt = A ηi  N i · ρt ϕ̈dv + ∇x N i : σdv

e=0
Bte Bte
 (3.2.14)
Z Z

− N i · t̄da − N i · bt dv  = 0 ∀ ηi .
∂Bte Bte

The integral terms are summarised to obtain the same discrete equilibrium equation as
in the material case

rI = fIdyn + fIint − fIext = 0 ∀I = 0, ..., ngf (3.2.15)

whereby the particular forces are calculated by means of spatial quantities

nel −1 Z
fIdyn
= A N i · ρt ϕ̈dv
e=0
Bte
nel −1 Z
fIint = A
e=0
∇x N i : σdv (3.2.16)
Bte
nel −1 Z Z
fIext = A
e=0
N i · t̄da − N i · bt dv.
∂Bte Bte

50
 3.3 Implementation of the FE method

Linearisation of the weak form in the spatial description

The linearisation of the discrete weak form in the spatial description is obtained by a
push forward of the linearization of the discrete weak form in the material configuration.
The following expressions are obtained for the contributions of the stiffness matrix

∆fIdyn = KIJdyn
∆ϕJ
nel −1 Z
∂ ϕ̈ (3.2.17)
dyn
KIJ = A N i · ρt · N j dv
e=0 ∂ϕ
Bt

∆fIint = KIJint
∆ϕJ
nel −1 Z Z
(3.2.18)
int
KIJ = A T
[∇x N j · ∇x N i ] : σdv + ∇x N i : E : ∇x N j dv
e=0
Bt Bt

∆fIext = KIJ
ext
∆ϕJ
(3.2.19)
ext
KIJ = 0
The particular parts of the residuum or the stiffness matrix can be either computed in
the material or the spatial configuration, the results are the same.

3.3 Implementation of the FE method

The general structure of a nonlinear Finite Element code is subdivided into three parts:
preprocessing, processing and postprocessing.

• PREPROCESSING
type of analysis; mesh generation; element types; material models/properties;
boundary conditions

• PROCESSING
the actual computation of the residuum, the stiffness matrix and the solution incre-
ment

• POSTPROCESSING
output and interpretation of results: deformation, strains, stresses

The implementation of a nonlinear FE-code is done in the computer exercises by means

of the open source FE-library deal.ii. Details and examples are discussed in the exercises.

51
4 Solution Methods for Time-Independent Problems

The mathematical modelling in solid mechanics leads to systems of nonlinear partial

differential equations. The discretisation in space by means of the finite element method
results in an ordinary differential equation in time. If time-dependency can be neglected
(static, quasi-static problems), a nonlinear algebraic system results

r(ϕ) = 0 (4.0.1)

which has to be solved for the unknown nodal deformations ϕ. In this chapter, the indices
indicating the FE nodes will be ommitted and if not stated otherwise the global vectors
are considered. The general question of solvability of this equation requires methods
of nonlinear functional analysis and is not covered within this lecture. We assume in
the following that solutions of equation (4.0.1) exist in the considered regions. A direct
solution of the nonlinear equation is in general not possible, iterative solution procedures
have to be applied. The fundamental questions which have to be clarified and constitute
the sucess of an iterative method are

• Does the method converge to the solution?

• How fast is convergence? Does the convergence rate depend on the problem size?

• How efficient is the method (numerical operations within one iteration, storage
requirements)?

The range of problems in nonlinear solid mechanics is wide and the sources and charac-
teristics of nonlinearities are different. Therefore, up to now no iterative method exists
which can be applied to all problems in an efficient and robust way. Due to that, several
methods will be introduced in the following. An essential ingredient of nonlinear solution
methods are linear equation solvers, since the nonlinear solution is derived via the solu-
tion of linear subproblems, i.e. often in each iteration a linear problem has to be solved.
Direct and iterative solvers for linear equations are not discussed with in the course, and
it is referred to section 5.2 of the textbook [9].

4.1 Newton-Raphson Method (Load Control)

The most frequently applied solution method is the Newton-Raphson scheme, which relies
on a Taylor series expansion of the nonlinear equation (4.0.1) at a known state, i.e. at

53
4 Solution Methods for Time-Independent Problems

the last equilibrium state. The Taylor series is truncated at the linear term
∂r k
r k+1 = r k + · ∆ϕ = r k + K k · ∆ϕ = 0,
∂ϕ
whereby the index k indicates the iteration. The unknown increment of the solution ∆ϕ
is obtained by rearranging and solving the above equation

r k + K k · ∆ϕ = 0 ⇒ K k · ∆ϕ = −r k .

The resulting linear system of equations can be solved by direct or iterative solvers. The
updated nodal deformation is then derived as

ϕk+1 = ϕk + ∆ϕ.

With the updated deformation the residuum is computed again and the procedure is re-
peated until the norm of the residuum is approximately 0.
The term load control means that the external forces are prescribed and increased step-
wisely. For each load step, the unknown deformations are computed. The algorithm of
the Newton-Raphson method is given as

• loop over all load steps n

– prescribe f ext
n+1

– compute residuum r 1n+1

• loop over iteration steps k

– compute tangent stiffness matrix K kn+1
– solve linear system of equations K kn+1 · ∆ϕ = −r kn+1
– update solution ϕk+1 k
n+1 = ϕn+1 + ∆ϕ

– compute residuum r k+1

n+1

– check convergence

k+1
|r n+1 | ≤ tol → next load step
|r k+1
n+1 | > tol → next iteration

The Newton-Raphson method shows quadratic convergence near the solution point. This
convergence behavior is characterized by the following inequality

||ϕk+1 − ϕ|| < c ||ϕk − ϕ||2

whereby ϕ denotes the converged solution of equation (4.0.1) and c is a positive constant.
This convergence behaviour is of advantage since it often leads to converged solutions

54
 4.2 Modified Newton-Raphson Method (Load Control)

f ext

ext
fn+1
3
r2 r
K2

r1

K1
ϕ

∆ϕ1 ∆ϕ2 ∆ϕ3 ∆ϕ4

Figure 4.1: convergence of the Newton-Raphson method

after a few iterations. The drawback of the Newton-Raphson method is that the tangent
stiffness matrix has to be composed and a linear system of equations has to be solved
in every iteration step. A qualitativ representation of the convergence behavior (for a
one-dimensional problem) is given in figure 4.1. The red curve represents the unknown
solution, i.e. the load-displacement diagram.

4.2 Modified Newton-Raphson Method (Load Control)

The modified Newton-Raphson method renders a simplification of the Newton-Raphson

method in the way that the stiffness matrix is not changed in every iteration step. The
stiffness matrix is computed and inverted only once at the beginning of each load step and
then applied in every iteration. The algorithm of the modified Newton-Raphson method
is given as

• loop over all load steps n

– prescribe f ext
n+1

– compute residuum r 1n+1

– computation of tangent stiffness K n+1

• loop over iteration steps k

– solve linear system of equations K n+1 · ∆ϕ = −r kn+1
– update solution ϕk+1 k
n+1 = ϕn+1 + ∆ϕ

– compute residuum r k+1

n+1

55
4 Solution Methods for Time-Independent Problems

– check convergence

|r k+1
n+1 | ≤ tol → next load step
|r k+1
n+1 | > tol → next iteration

The modified Newton-Raphson method saves computing time in each iteration since the
stiffness matrix is only computed once and in the linear system of equations which has
to be solved in every iteration changes only the right hand side. However, the number of
iterations is much higher as compared to the standard Newton-Raphson scheme, especially
for strongly nonlinear problems. The method shows linear convergence to the solution and
f ext

ext
fn+1
3
r2 r

r1

K1
ϕ
∆ϕ3
∆ϕ1 ∆ϕ2 ∆ϕ4

Figure 4.2: convergence of the modified Newton-Raphson method

is mostly applied for problems with weak nonlinearities. The qualitative representation
of the convergence behavior for a one-dimensional problem is given in figure 4.2.

4.3 Quasi-Newton Methods

The large group of Quasi-Newton methods resembles the Newton-Raphson method but the
explicit computation of the tangential stiffness matrix is avoided. Instead, the inverse of
the tangential stiffness is substituted by an approximation, which is continuously updated
during the solution process. The various Quasi-Newton methods differ by means of various
update formulas. The solution update is calculated by
∆ϕ = −H kQN · r k (4.3.1)
whereby H kQN denotes the Quasi-Newton approximation of the inverse tangent stiffness,
which has to satisfy the following equation
ϕk − ϕk−1 = −H kQN · [r k − r k−1 ] (4.3.2)

56
 4.3 Quasi-Newton Methods

and has to be symmetric and positive definite. This relation clearly shows that instead
of the tangent stiffness matrix a secant stiffness matrix is applied, which can be com-
puted from the known states at previous iterations. Different updates for H QN were
developed such that equation (4.3.2) is satisfied. One possibility is the rank 2 BFGS
(Broyden-Fletcher-Goldfarb-Shanno) update, which is described e.g. in [2] or [9], and is
not discussed in detail here.
The algorithm for the Quasi-Newton methods is the same as for the Newton-Raphson
scheme, but K −1 is replaced by H QN . The Quasi-Newton methods show a super linear
convergence behaviour. The qualitative convergence behavior of the method is shown in
figure 4.3.
f ext

ext
fn+1
3
r2 r
K2

r1

K1
ϕ

∆ϕ1 ∆ϕ2 ∆ϕ3 ∆ϕ4

Figure 4.3: convergence of the Quasi-Newton-Raphson method

Comparison of the Newton-Raphson method, modified Newton-Raphson

method and Quasi-Newton method

The example of the three-hinged arch with one degree of freedom (compare section 2.6)
is solved with the three introduced methods. The following parameters are applied:
E mod A = 10000kN, B = 10m, H = X = 0.5m. The residuum rn+1 k
, given in section
k k
2.6, is recapitulated here and the tangent stiffness Kn+1 = ∂rn+1 /∂xk is computed as
mod
k
rn+1 = E 3A [(xk )3 − (xk ) X 2 ] − fn+1
ext
2L (4.3.3)
mod
k
Kn+1 = E 3A [3(xk )2 − X 2 ]
2L
One load step of ∆f ext = 100N is applied and the convergence behaviour of the three
methods is studied by means of the residuum, the solution increment and the solution.
The results are presented in the tables 4.1-4.3.

57
4 Solution Methods for Time-Independent Problems

Newton-Raphson method
k |rk | ∆ϕ ϕk
1 0.1 -0.04015 0.45985
2 0.011723 -0.0061223 0.45373
3 0.00025643 -0.00014004 0.45359
4 1.3296e-007 -7.2684e-008 0.45359
5 3.5832e-014 -1.9588e-014 0.45359

Table 4.1: convergence of the Newton-Raphson method for one load step

modified Newton-Raphson method

k |rk | ∆ϕ ϕk
1 0.1 -0.04015 0.45985
2 0.011723 -0.0047066 0.45514
3 0.0028623 -0.0011492 0.45399
4 0.00074478 -0.00029903 0.4537
5 0.00019674 -7.8989e-005 0.45362
6 5.217e-005 -2.0946e-005 0.4536
7 1.3849e-005 -5.5603e-006 0.45359
8 3.6772e-006 -1.4764e-006 0.45359
9 9.7645e-007 -3.9204e-007 0.45359
10 2.5929e-007 -1.0411e-007 0.45359
11 6.8855e-008 -2.7646e-008 0.45359
12 1.8285e-008 -7.3413e-009 0.45359
13 4.8555e-009 -1.9495e-009 0.45359

Table 4.2: convergence of the modified Newton-Raphson method for one load step

The Newton method and the Quasi-Newton method converge in a few iterations whereby
as expected the modified Newton method requires a high number of iterations.
In the next example the number of iterations for different load increments is compared.

58
 4.3 Quasi-Newton Methods

Quasi-Newton method
k |rk | ∆ϕ ϕk
1 0.1 -0.04015 0.45985
2 0.011723 -0.04015 0.45985
3 0.0017085 -0.0053317 0.45452
4 3.8685e-005 -0.0009096 0.45361
5 1.3304e-007 -2.1074e-005 0.45359
6 1.0424e-011 -7.2721e-008 0.45359

Table 4.3: convergence of the Quasi-Newton method for one load step

Note that the nonlinearity of the equilibrium path becomes more pronounced for higher
loads, i.e. close to the maximum possible force.

f ext ϕ NR mod NR QN
40 0.48309 4 8 5
80 0.46411 5 11 6
120 0.44218 5 15 6
160 0.41559 5 22 7
200 0.37988 6 35 7
240 -0.57744 517 229 -

Table 4.4: number of iteration for different load step sizes

Again, the Newton-Raphson and the Quasi-Newton method show a good convergence
behaviour, whereby the modified Newton method requires many iterations, especially
for larger load steps. When the load is increased to 240N the maximum point of the
equilibrium path is exceeded and it is not possible to follow the equilibrium path any
further with a load controlled scheme. Methods to compute the load-displacement path
in the post-critical state (behind such critical points) will be introduced in the following.
By coincidence two iterative schemes find, after some hundred iterations, the next possible
equilibrium point with the solution ϕ = −0.57744, i.e. the snap-through of the structure
has already occured.

59
4 Solution Methods for Time-Independent Problems

4.4 Newton-Raphson Method (Displacement Control)

Load-controlled methods can not pass so-called limit points of the equilibrium path as
was highlighted in the last example and is indicated in figure 4.4. One possibility to
f =0
f ext
L1

L2 r=0

Figure 4.4: limit points for load control (L1 ) and displacement control (L2 ), constraint path for arc-length
methods

overcome this problem and to compute equilibrium points in the post-critical state is to
prescribe and control the displacement at certain nodes instead of the load. The solution
of the discrete problem requires then the partitioning of the solution vector in unknown
and prescribed deformation increments

∆ϕ = [∆ϕu ∆ϕp ]T . (4.4.1)

The linear system of equations which has to be solved in every iteration step takes the
following format
     
K K up ∆ϕu r
 uu ·  = − u , (4.4.2)
K pu K pp ∆ϕp rp

whereby the prescribed deformation increment is applied in the first increment as ∆ϕk=0
p
k6=0
and set to 0 in the following iterations, i.e. ∆ϕp = 0:

1. iteration
K uu · ∆ϕk=1
u = −[r u + K up · ∆ϕk=0
p ]

k. iteration
K uu · ∆ϕk+1
u = −r u

The displacement control can also be applied in combination with the modified Newton-
Raphson or the Quasi-Newton methods. With a displacement-controlled algorithm the
equilibrium path in figure 4.4 can be computed up to point L2 . At this point a snap-back

60
 4.5 Arc-Length Methods

takes place and both the displacement and the loading at a certain point decrease. To
reach the equilibrium path behind this point path-following or arc-length methods have
to be introduced.

4.5 Arc-Length Methods

The idea of the arc-length methods is to apply an incremental force, depending on an a

priori unknown load factor λ, which is determined by means of an additional constraint.
As indicated in figure 4.4 the equilibrium path (r=0) and the constraint path (f=0) have
an intersection point which constitutes the solution point for the actual load step. As
a motivation for this type of solution methods the following example of a star-shaped
structure, as illustrated in figure 4.5, is considered. The structure is built of pin-jointed

top view

front view

Figure 4.5: star-shaped structure: geometry and load-displacement relation, from [9]

straight nonlinear trusses. The constitutive behavior of the truss elements is modelled by
the St.-Venant-Kirchhoff model and the structure is simply-supported at the outer nodes.
A load is applied at the node in the middle of the structure. The load can be increased
a bit, then a snap-through of the inner part of the structure is oberserved, as illustrated
in the load-displacement diagram in figure 4.5, right. Then a complex snap-through be-
haviour of the outer parts together with a snap-back of the inner part is observed which
leads to the loop-like part of the equilibrium path. Finally, the inner part snaps through
again and then a stable behaviour of the strucure is reached. To be able to follow such
equilibrium paths (e.g. to compute how a strucutre behaves after failure) arc-length meth-
ods are required.

61
4 Solution Methods for Time-Independent Problems

Various arc-length methods exist which differ in terms of the particular constraint equa-
tion. But the general structure of the methods and solution procedure are the same. The
external load at load step n + 1 is defined as
ext
f ext
n+1 = λn+1 f̄ (4.5.1)
ext
whereby f̄ is a prescribed value and λn+1 the still unknown load factor. An additional
equality constraint is formulated which usually depends on the deformation and on the
load factor f (ϕ, λ) = 0. The enlarged system of equations which has to be solved is
composed of the discrete equilibrium equation and the constraint equation
   
r(ϕ, λ) 0
 =  (4.5.2)
f (ϕ, λ) 0

Linearisation of the enlarged system leads to

         
∂r ∂r
 ∂ϕ ∂λ   ∆ϕ K K ∆ϕ r
 ∂f ∂f  ·  =  ϕϕ ϕλ
·  = −  (4.5.3)
∆λ K λϕ Kλλ ∆λ f
∂ϕ ∂λ
Particular parts of the linearisation are already known or can be calculated without
specifying the constraint equation: K ϕϕ is the usual tangent stiffness matrix, and
ext
K ϕλ = −f̄ . The two other parts have to be evaluated with respect to a particular
constraint equations. The matrix of the enlarged system of equations is usually not sym-
metric and therefore a partitioning technique (block elimination) is applied to use the
symmetry of K ϕϕ . The first equation of the matrix equation (4.5.3) is rewritten as
ext
K ϕϕ · ∆ϕ − f̄ ∆λ = −r
ext
⇒ ∆ϕ = −K −1 −1
ϕϕ · r + ∆λK ϕϕ · f̄
(4.5.4)

= ∆ϕϕ + ∆λ∆ϕλ

with
ext
∆ϕϕ = −K −1
ϕϕ · r and ∆ϕλ = K −1
ϕϕ · f̄ (4.5.5)

This result is introduced into the second equation of (4.5.3) which is then rearranged to
obtain the increment ∆λ

K λϕ · ∆ϕ + Kλλ ∆λ = −f
⇒ K λϕ · ∆ϕϕ + [K λϕ · ∆ϕλ + Kλλ ]∆λ = −f (4.5.6)
f + K λϕ · ∆ϕϕ
⇒ ∆λ = − K · ∆ϕ
λϕ λ + Kλλ

The procedure yields the load factor increment and the displacement increment whereby
only symmetric matrices are involved. In contrast to the standard Newton-Raphson

62
 4.5 Arc-Length Methods

method, the arc-length schemes need a predictor step. One possibility is to prescribe an
arc-length ∆s, calculate the solution of the linear system of equation
ext
K ϕϕ · ∆ϕλ = f̄ (4.5.7)

and compute the first load increment ∆λ0 by scaling the tangent vector

±∆λ0 |∆ϕλ | = ∆s. (4.5.8)

Thereby, the sign of ∆λ0 depends on the current tangent of the solution path and can be
computed by means of the current stiffness parameter κn /κ0
" ext #
f̄ · ∆ϕn
κn /κ0 = /κ0 , (4.5.9)
∆ϕn · ∆ϕn

with
ext
f̄ · ∆ϕ0n+1
κ0 = . (4.5.10)
∆ϕ0n+1 · ∆ϕ0n+1

When κn /κ0 is positive the plus sign and otherwise the minus sign is used for ∆λ0 .
The increment of the arc-length ∆s should be adjusted during the simulation to achieve
convergence at critical points. A simple indicator to adjust the increment is the number
of Newton iterations. When it increases e.g 9, the increment size is divided in half and
when the number decreases below 5 the increment size is doubled. Before specifying
different constraint equations the general algorithm of the arc-length method is stated in
the following.

• loop over all load steps n

– predictor step: prescribe arc-length ∆s, compute initial value λ0n+1

ext
K 0ϕϕ · ∆ϕλ = f̄
λ0n+1 = λn + ∆λ0 = λn ± ∆s
|∆ϕλ |

– compute residuen r 1n+1 , fn+1

1

• loop over iteration steps k

ext
– compute tangents K kϕϕ , K kλϕ , Kλλ
k
, (K kϕλ = f̄ is given)
– compute deformation increments
ext
K kϕϕ · ∆ϕϕ = r kn+1 and K kϕϕ · ∆ϕλ = f̄

63
4 Solution Methods for Time-Independent Problems

– compute load increment

k
fn+1 + K kλϕ · ∆ϕϕ
∆λ = −
K kλϕ · ∆ϕλ + Kλλ
k

– update deformation and load factor

∆ϕ = ∆ϕϕ + ∆λ∆ϕλ
ϕk+1 k
n+1 = ϕn+1 + ∆ϕ

λk+1
n+1 = λkn+1 + ∆λ

– compute residuen r k+1 k+1

n+1 , fn+1

– check convergence

|[r k+1 k+1

n+1 fn+1 ]| ≤ tol → next load step
|[r k+1 k+1
n+1 fn+1 ]| > tol → next iteration step

Constraint equations

Different constraint conditions can be specified for the arc-length method. The most
simple ones are load control or displacement control which were already introduced before.
The constraint equation for load control depends only on the load factor
f (λ) = λ − λ̄ = 0 (4.5.11)
whereby λ̄ is a given value. For displacement control the constraint equation depends
only the the deformation
f (ϕ) = ϕI − ϕ̄ = 0 (4.5.12)
whereby ϕI denotes the component of ϕ which is prescribed and ϕ̄ is a given value. In
both cases the equilibrium and the constraint equation are uncoupled and can be solved
in the way we have introduced before. For both approaches limit points exist which can
not be passed. Therefore, two more general constraint equations will be introduced.

Iteration on a fixed normal plane

The constraint equation describes a plane which is perpendicular to the tangent of the
equilibrium path at the last equilibrium state
k+1
fn+1 (ϕ, λ) = [ϕ1n+1 − ϕn ] · [ϕk+1 1 1 k+1 1
n+1 − ϕn+1 ] + [λn+1 − λn ][λn+1 − λn+1 ] = 0 (4.5.13)
The constrain equation is linear in the deformation and the load factor. The solution is
given as the intersection between the normal plane and the equilibrium path as indicated
in figure 4.6, left.

64
 4.5 Arc-Length Methods

Iteration on a sphere

The constraint equation describes a sphere with the radius ∆s around the last equilibrium
point. That ensures that at least one intersection point with the equilibrium path exists.
The drawback is that in most cases two solutions exist and the correct one has to be
chosen. The constraint equation is quadratic in the deformation and the load factor

k+1
fn+1 (ϕ, λ) = [[ϕk+1 k+1 k+1 k+1
n+1 − ϕn ] · [ϕn+1 − ϕn ] + [λn+1 − λn ][λn+1 − λn ]]
0.5
− ∆s = 0 (4.5.14)

In figure 4.6 the qualitative convergence behaviour of this arc-length method is specified.

f ext f ext
.
∆s2
.

∆s2
.

∆s1 ∆s1

ϕ ϕ

Figure 4.6: , right, arc-length methods: iteration on normal plane and sphere

65
5 Solution Methods for Time-Dependent Problems

In many engineering applications the change of state variables and deformation in time
has to be considered. Then, the inertia terms in the balance of momentum can not be
neglected. Examples for such applications are vibration analysis, impact problems or
car crash simulations. These problems require the solution of an initial boundary value
problem and the integration in time of the governing equations. Starting point is the
discrete form of the balance of momentum (3.2.5)

r = f dyn + f int − f ext = 0. (5.0.1)

This equation has to be discretised and integrated in time.

5.1 Time Discretisation with Finite Differences

The time interval of interest T is decomposed into nstep time steps ∆tn
nsteps −1
[
T = [tn , tn+1 ] with ∆tn = tn+1 − tn > 0. (5.1.1)
n=0

An approximation of the time derivatives, the velocity ϕ̇ and the acceleration ϕ̈, within
one time step is required. These approximations of velocity and acceleration at the time
tn+1−α usually depend on known and/or unknown deformations, velocities and accelera-
tions

ϕ̇n+1−α = ϕ̇n+1−α (ϕn−1 , ϕn , ϕ̇n , ϕ̈n , ϕn+1 , ...)

(5.1.2)
ϕ̈n+1−α = ϕ̈n+1−α (ϕn−1 , ϕn , ϕ̇n , ϕ̈n , ϕn+1 , ...).

The discrete balance of momentum (5.0.1) is solved at the time tn+1−α

r n+1−α = f dyn int ext

n+1−α + f n+1−α − f n+1−α = 0. (5.1.3)

The choice of the the parameter α, 0 ≤ α ≤ 1 defines if an explicit or implicit integration

scheme is used. An explicit method follows if α = 1. Explicit methods are easy to
implement since the solution at time tn+1 depends only on quantities at time tn . Explicit
schemes are very efficient especially in combination with a lumped mass matrix. The
disadvantage is the limitation of time step size due to stability reasons. Implicit schemes
are obtained by setting α = 0. Then the approximations of the time derivatives depend
on unknown quantities and the solution of the discrete equilibrium equation at time
tn+1 requires the solution of a nonlinear system of equations. The advantage of implicit
methods is that they can be constructed such that they are unconditionally stable.

67
5 Solution Methods for Time-Dependent Problems

5.2 Explicit Time Integration Methods

A common choice for the approximation of the time derivatives in explicit methods is the
central difference scheme. The time derivatives at time tn are approximated by means of
the deformations at the previous, current and next time step

1 [ϕ
ϕ̇n = 2∆t n+1 − ϕn−1 ]
(5.2.1)
ϕ̈n = 1 2 [ϕn+1 − 2ϕn + ϕn−1 ].
∆t
The solution of the discrete balance of momentum is computed at time tn

r n = f dyn
n + f int ext
n − fn = 0 (5.2.2)

with
nel −1 Z
dyn
fI,n = A
e=0
N i · ρ0 ϕ̈n dV
B0e
nel −1 Z
int
fI,n = A [F Tn · ∇X N i ]sym : S n dV (5.2.3)
e=0
B0e
nel −1 Z Z
ext
fI,n = A
e=0
N i · T̄ n dA + N i · b0 dV.
∂B0e B0e

The acceleration ϕ̈n which occurs in the dynamic force vector is approximated in space
by means of the same shape functions as the deformation
ndof −1
X
ϕ̈n = N j ϕ̈j n . (5.2.4)
j=0

Then, the internal force vector can be written as the product of the mass matrix and the
nodal acceleration vector as
 
nel −1 Z nel −1 Z
dyn
= A N i · ρ0 N j ϕ̈j n dV = A  N i · ρ0 N j dV  ϕ̈J n = MIJ ϕ̈J n .
 
fI,n (5.2.5)
e=0 e=0
B0e B0e

in which MIJ denote the coefficients of the global mass matrix M . The global system of
equations takes the following format

r n = M · ϕ̈n + f int ext

n − f n = 0. (5.2.6)

The approximation of the acceleration (5.2.1) is introduced and yields a linear system of
equations for the unknown deformations ϕn+1
1 int ext
rn = 2 M · [ϕn+1 − 2ϕn + ϕn−1 ] + f n − f n = 0. (5.2.7)
∆t

68
 5.2 Explicit Time Integration Methods

Note, that the internal force vector that usually contains the nonlinearities depends only
on known quantities at time tn .
Explicit methods are especially efficient if a lumped mass matrix is used, that is a matrix
with a diagonal structure. The system of equations (5.2.7) reduces then to a number of
uncoupled equations whose solution is trivial. The lumped mass matrix must satisfy the
conservation of mass. The mass matrix for one element is derived as
Z Z ngp −1
X
e
Mij = N i · ρ0 N j dV = N i ρ0 · N j detJ dV ≈ N i ρ0 · N j detJ e wg . (5.2.8)
g=0
B0e B

One possibility to derive a diagonal structure is to use quadrature points at the element
nodes instead of the usual Gauss points. Then, the ansatz functions are either 0 or 1 and
a diagonal structure for M is obtained

 ρ0 detJ e wg if i = j
diag
Mij = (5.2.9)
 0 if i 6= j.
Explicit integration schemes are very efficient since the solution of a system of equations is
not necessary. The disadvantage is that they are not unconditionally stable, but stability
depends on the time step size. An estimate for the critical time step size can be derived
from the analysis of linear problems, where the critical time step depends on the element
size and the speed of a wave travelling through a solid
h
∆tcrit = , (5.2.10)
cL
whereby h denotes a characteristic size of the smallest element and cL , the velocity of a
compression wave, is computed for a linear elastic material as c2L = 3[K[1 − ν]]/[ρ0 [1 + ν]]
with the bulk modulus K and Poisson’s ratio ν. For nonlinear problems an estimation of
the critical time step size is given by means of a reduction factor δ, that has to be selected
considering the nonlinearity of the problem
h
∆tcrit = δ with 0, 2 ≤ δ ≤ 0, 9, (5.2.11)
cL
which is obviously a difficult task. Explicit method are ideally suited for engineering
problems where the time step size has to be small to capture the physics of the problem.
This is the case when high frequencies have to be resolved, e.g. in impact problems or car
crash simulations.
It is known from experimental evidence that damping effects occur in dynamic problems.
They are for example related to viscous material behavior or internal friction. Damping
is often taken into account by means of a phenomenological approach, assuming that the
effects are proportional to the velocity. An additional damping term is then added to the
discrete equation of equilibrium, compare equation (5.2.6),

M · ϕ̈n + C · ϕ̇n + f int ext

n − f n = 0, (5.2.12)

69
5 Solution Methods for Time-Dependent Problems

whereby the damping matrix C is chosen to be a constant combination of the mass and
the initial stiffness matrix

C = δM M + δK K, (5.2.13)

whereby the parameter δM and δK have to be identified by experiments. The introduction

of the damping term increases stability of the solution method.
The algorithm for an explicit solution scheme is specified in the following.

• prescribe initial values ϕ0 , ϕ̇0 , ϕ̈0

• extrapolate starting value ϕ−1 (by second order accurate Taylor expansion)

1
ϕ−1 = ϕ0 − ∆tϕ̇0 + ∆t2 ϕ̈0
2

• compute mass matrix M

• loop over time steps n

– prescribe external force vector f ext
n

– compute internal force vector f int

n

– compute new deformation ϕn+1 from

M · ϕn+1 = ∆t2 [−f int ext

n + f n ] + M · [2ϕn − ϕn−1 ]

– update velocity and acceleration

1 [ϕ
ϕ̇n+1 = 2∆t n+1 − ϕn−1 ]

ϕ̈n+1 = 1 2 [ϕn+1 − 2ϕn + ϕn−1 ]

∆t

5.3 Implicit Time Integration Methods

A common implicit time integration method is the Newmark method, which suggests the
following approximations of velocity and acceleration values at the time tn+1

γ γ−β γ − 2β
ϕ̇n+1 = [ϕ − ϕn ] − ϕ̇n − ∆tϕ̈n
β∆t n+1 β 2β
(5.3.1)
= 1 [ϕ 1 ϕ̇ − 1 − 2β ϕ̈
ϕ̈n+1 n+1 − ϕn ] −
β∆t2 β∆t n 2β n

70
 5.3 Implicit Time Integration Methods

whereby the parameter γ and β are the Newmark parameter, which are specified to ensure
stability and accuracy of the method. Within this implicit integration scheme the discrete
equation of equilibrium is solved at time tn+1

r n+1 = f dyn int ext

n+1 + f n+1 − f n+1 = 0 (5.3.2)

with
nel −1 Z
dyn
fI,n+1 = A
e=0
N i · ρ0 ϕ̈n+1 dV
B0e
nel −1 Z
int
fI,n+1 = A [F Tn+1 · ∇X N i ]sym : S n+1 dV (5.3.3)
e=0
B0e
nel −1 Z Z
ext
fI,n+1 = A
e=0
N i · T̄ n+1 dA + N i · b0 dV.
∂B0e B0e

The internal force vector depends nonlinearly on the unknown deformation ϕn+1 and an
iterative solution of the nonlinear system of equations is required in every time step. The
dynamic force vector can again be expressed by means of the global mass matrix M as
nel −1 Z
f dyn
I,n+1 = A
e=0
N i ρ0 · N j ϕ̈j n+1 dV = MIJ ϕ̈J,n+1 . (5.3.4)
B0e

Then the approximation of the acceleration (5.3.1) is introduced. Application of the

Newton-Raphson method to solve the nonlinear system of equations requires the lineari-
sation of (5.3.2). This was derived in section 3.2, compare equations (3.2.8)-(3.2.10). The
dynamic part of the stiffness matrix can now be specified as
 
nel −1 Z nel −1 Z
∂ ϕ̈ 1
KIJ = A  N i · ρ0
dyn
· N j dV  = A
 
N i ρ0 · N j dV, (5.3.5)
e=0 ∂ϕ e=0 β∆t2
B0e B0e

whereby the partial derivative ∂ ϕ̈n+1 /∂ϕn+1 follows from equation (5.3.1).
The accuracy and the stability of the Newmark scheme depend on the choice of the
parameter γ and β. The following results are valid for linear problems. The method is
unconditionally stable for β ≥ 1/4[γ + 1/2]2 and accurate of 2nd order for γ = 1/2. Other
parameters reduce either the stability or the accuracy properties. Numerical damping of
higher frequencies can be obtained by choosing γ > 1/2. This is often advantageous since
the spatial finite element discretisation resolves lower eigenmodes much better than higher
ones, and implicit methods are often applied to engineering problems whose response is
governed by lower frequencies. But this choice of the Newmark parameter leads to a
reduced accuracy of the method which is the reason for the developement of modified
Newmark methods which will be introduced in the following section.

71
5 Solution Methods for Time-Dependent Problems

The Newmark approximation is often rewritten, such that deformation and velocity are
defined by means of acceleration terms
2
ϕn+1 = ϕn+1 + ∆tϕ̇n+1 + ∆t 2 [[1 − 2β]ϕ̈n + 2β ϕ̈n+1 ] (5.3.6)
ϕ̇n+1 = ϕ̇n + ∆t[[1 − γ]ϕ̈n + γ ϕ̈n+1 ]

The discrete equilibrium equation is then solved for the unknown accelerations ϕ̈n+1 .
The algorithm of the Newmark scheme is summarised in the following.

• prescribe initial values ϕ0 , ϕ̇0 , ϕ̈0

• loop over the time steps n

– prescribe external force f ext
n+1

– compute residuum r 1n+1

• loop over the iteration steps k

– compute tangent K kn+1
– solve linear system of equations

K kn+1 · ∆ϕ = −r kn+1

– update deformation

k+1
ϕn+1 = ϕkn+1 + ∆ϕ

– compute residuum r k+1

n+1

– test convergence

k+1
|r n+1 | > tol → next iteration step
|r k+1
n+1 | ≤ tol → update velocity and acceleration

• update velocity and acceleration

γ γ−β γ − 2β
ϕ̇n+1 = [ϕ − ϕn ] − ϕ̇n − ∆tϕ̈n
β∆t n+1 β 2β
ϕ̈n+1 = 1 [ϕ 1 ϕ̇ − 1 − 2β ϕ̈
n+1 − ϕn ] −
β∆t2 β∆t n 2β n

→ next time step

72
 5.4 Modifications of the Newmark Method

5.4 Modifications of the Newmark Method

The Newmark method loses accuracy if numerical damping is introduced. To overcome

this problem different modifications of the method have been proposed. All of the methods
use the Newmark approximation of the velocity and the acceleration at time tn+1 , but
solve the equilibrium equation at a generalised mid-point tn+1−α . The variables at this
generalised mid-point are computed as

ϕn+1−αf = [1 − αf ]ϕn+1 + αf ϕn
ϕ̇n+1−αf = [1 − αf ]ϕ̇n+1 + αf ϕ̇n (5.4.1)

ϕ̈n+1−αm = [1 − αm ]ϕ̈n+1 + αm ϕ̈n

whereby the values at time tn+1 are approximated as given in (5.3.1). The equation of
equilibrium at the generalised mid-point is defined as

r n+1−α = f dyn int ext

n+1−αm + f n+1−αf − f n+1−αf = 0 with 0 ≤ αf,m ≤ 1 (5.4.2)

Three modifications of the Newmark method are distinguished, depending on the choice
of αm and αf :
Newmark: αm = αf = 0
Bossak-α: αm 6= 0, αf = 0
Hilber-α: αm = 0, αf 6= 0
generalised-α: αm 6= 0, αf 6= 0
All methods result in a nonlinear system of equations for the unknown deformations ϕn+1 .
The evaluation of the internal force vector at time tn+1−αf requires the computation of
the field values related to the deformation ϕn+1−αf

F n+1−αf = [1 − αf ]F n+1 + αf F n
E n+1−αf = E(F n+1−αf )
S n+1−αf = S(E n+1−αf ) (5.4.3)
nel −1 Z
f int
n+1−αf = A F n+1−αf · S n+1−αf · ∇X Ni dV.
e=0
B0e

The mid-point values also have to be taken into account within the linearisation of the
discrete equilibrium equation.
The generalised-α methods are unconditionally stable for linear problems if
β ≥ 1/4 + 1/2[αf − αm ] and 2nd -order accurate if γ = 1/2 − αm + αf . Numerical damping
of higher frequencies can be added in an optimal sense by a special choice of αm and αf ,
depending on the damping coefficient 0 < ρ∞ ≤ 1. No numerical damping is applied if

73
5 Solution Methods for Time-Dependent Problems

ρ∞ = 1, for ρ∞ ≤ 1 numerical damping is considered. The other parameters are then

defined as
2ρ∞ − 1 ρ∞ 1 1
αm = , αf = , γ = − αm + αf , β = [1 + αf − αm ]2 (5.4.4)
1 + ρ∞ 1 + ρ∞ 2 4

The related algorithm resembles the one of the Newmark method but the computation of
the mid-point values has to be considered.

5.5 Conserving Algorithms

The introduced implicit methods were originally developed for linear problems. When
they are applied to nonlinear problems it can be observed that certain physical quantities
might not be preserved during the computation: the linear and angular momentum or
the energy of the system. Conservation of the first two quantities is mainly related to
the accuracy of the methods, whereby the violation of energy conservation can lead to
stability problems. It has been shown that especially for problems with finite rotations
and for long time integration the application of conserving algorithms is important. The
idea of the conserving algorithms is to ensure the conservation of linear and angular
momentum and energy by selecting suitable parameters and by substituting the stresses
in the internal force vector with an algorithmic stress tensor. The resulting method is
called the energy-momentum-method, which is introduced in the following.
The starting point for the development of conserving time integration schemes is the
generalised-α method. It can be shown that conservation of linear momentum is satisfied
if αf = αm = α and 0 ≤ α ≤ 1. Conservation of angular momentum is achieved for
α = 1/2. The proof of these results is based on the continuous weak form of equilibrium,
compare equation (2.5.1), considering no external forces, no prescribed deformations, no
damping and hyperelastic material behaviour
Z Z
δϕ · ρ0 ϕ̈dV + [∇tX δϕ · F ] : SdV = 0. (5.5.1)
B0 B0

Exemplarily, the conservation of linear momentum can be shown by choosing

δϕ = ω = const. ⇒ ∇tx δϕ = 0. For the Newmark parameter γ = 1/2 and β = 1/4 and
α = 1/2 the approximation of the acceleration at tn+1/2 is given as

1
ϕ̈n+1/2 = [ϕ̇ − ϕ̇n ]. (5.5.2)
∆t n+1
Introducing this relation and ∇tx δϕ = 0 into equation (5.5.1) at time tn+1/2 results in
Z Z
1
ω · ρ0 ϕ̈n+1/2 dV = ω· [ρ0 ϕ̇n+1 − ρ0 ϕ̇n ]dV = 0 (5.5.3)
∆t
B0 B0

74
 5.5 Conserving Algorithms

The term in the last integral can be identified as the difference of the linear momentum
density at time tn+1 and tn , which verifies that the linear momentum is conserved. The
proof of the conservation of angular momentum has the same structure, but a different
test function is specified. Both, conservation of linear and angular momentum are inde-
pendent of the computation of the stress tensor. Conservation of energy means that the
total energy is constant: En+1 = En . It can be shown that the kinetic energy is exactly
approximated by the generalised-α method, but the internal energy term has to be con-
sidered in more detail. The time rate of change of the internal energy or the internal
power are defined in (2.4.15) as
Z Z
I˙ = P =
int
Ė : S dV = F · S : ∇X v dV (5.5.4)
B0 B0

Time integration with α = 1/2 yields

Z
In+1 − In = ∆t F n+1/2 · S alg : ∇x v n+1/2 dV, (5.5.5)
B0

whereby an algorithmic stress tensor S alg was introduced (but not further specified). The
deformation gradient and the velocity gradient at the mid-point are evaluated as

F n+1/2 = 21 [F n+1 + F n ]
(5.5.6)
∇x v n+1/2 = ∆t1 [F
n+1 − F n ],

which leads after some manipulations to the following equation

Z
In+1 − In = S alg : [E n+1 − E n ] dV, (5.5.7)
B0

which describes the change of the internal energy depending on an algorithmic stress
tensor. This algorithmic symmetric stress tensor S alg has to be computed from the
constitutive law.
Let us consider the St.-Venant-Kirchhof hyperelastic law, which is specified by an energy
density that is quadratic in the Green-Lagrange strains and the stresses are a therefore a
linear function of the strains
1
ψ V K (E) = E : L : E and S = L : E. (5.5.8)
2
The internal energies at times tn+1 and tn can be computed by integrating the energy
density ψ over the volume
Z  
VK VK 1 1
In+1 − In = E n+1 : L : E n+1 − E n : L : E n dV
2 2
B
Z0
1 (5.5.9)
= [E n+1 + E n ] : L : [E n+1 − E n ]dV
2
B0

75
5 Solution Methods for Time-Dependent Problems

Comparison of equations (5.5.7) and (5.5.9) shows that the algorithmic stress tensor for
the St.-Venant-Kichhoff law has to be computed as
1 1
S alg = [E n+1 + E n ] : L = L : [E n+1 + E n ] (5.5.10)
2 2
such that energy conservation is guaranteed. This result is unexpected since
1 [E
2 n+1 + E n ] 6= E(ϕn+1/2 ) and since the internal force vector is evaluated at the mid-
point it would be natural to compute the state variables related to the deformation ϕn+1/2 .
However, energy conservation is only guaranteed if S alg is used instead of S n+1/2 .
For more general hyperelastic constitutive laws, it can be shown that the algorithmic
stress tensor has to be computed with respect to a certain strain or deformation measure
that is a linear combination of the values at tn+1 and tn , e.g.

C n+1−β = [1 − β]C n+1 + βC n and S alg = S alg (C n+1−β ) (5.5.11)

whereby β has to be computed, depending on the constitutive law, such that energy con-
servation is satisfied.
The introduced energy-momentum method is unconditionally stable in the nonlinear
regime since conservation of linear and angular momentum and energy are exactly satis-
fied.

76