SESSION TITLE:

SOLUTION METHODS AND STABILITY ANALYSIS

Topics:
•Introduction and Overview
•Explicit time integration method
•Implicit time integration methods and solution of equilibrium
equations
•Stability of solutions
•Numerical stability
•Material stability
INTRODUCTION AND OVERVIEW:
• Solution procedures for nonlinear finite element analysis of
continua and structures are presented.
• Explicit and implicit methods for transient problems and methods
for the solution of equilibrium problems are described
• Stability is an important topic in NLFEA.
• Stability of solutions, stability of numerical procedures and material
stability are presented.
• In Explicit time integration, we focus on the central difference
method.
• We next describe implicit time integration methods with the
Newmark -method as our focus
• Methods for solution of static problems, i.e., the solution of
equilibrium problems, are developed simultaneously.
• Newton method description also includes techniques used for
convergence checks and line search method.
• A critical step in the implicit scheme and static problems is the
LINEARIZATION of the governing equations. Linearization
procedures for the equations of motion and the Equilibrium
Equations are described.
• In nonlinear problems, stability of solution is of considerable
interest. Methods for examining stability are described.
• We then examine numerical stability of time integration methods
and analysis of central difference method and the Newmark, -
method are performed.
• The topic of material stability concludes the session.
Explicit Time Integration Methods
• The central difference method is among the most popular of the
explicit methods in computational mechanics. The central
difference method is developed from central difference formulae
for the velocity and acceleration. We consider here its application
to Lagrangian meshes. Methods for Eulerian and ALE meshes
are discussed separately. Geometric and material nonlinearities
are included in the equations of motion, and in fact have little
effect on the time integration algorithm.
• Notation: Let the time of the simulation 0  t  tE be subdivided
into time steps tn, n = 1 to nTS where nTS is the number of time
steps and tE is the end of simulation. tn and dn  d(tn) are the time
and nodal displacements, respectively at time step n.
• The algorithm with a variable time step is considered. This is
necessary in practical calculations since the stable time step
changes as the FE mesh deforms and wave speed changes due to
stress.
• The time increments as
1 1
n n
Δt n  t  t 2 2

• The central difference formula for the velocity

d 
1 1
n n 1
d 2
V 2
 1
n 1
 dn
n
Δt 2

Which is converted to an integration formula:

dn+1 = dn + tn+1/2.Vn+1/2
• The acceleration and the corresponding integration formula are
 n  12 n
1

d n  a n   V  V 
2

 n 1 n 1 
 t 2 t 2 
 
1 1
n n
V 2
V 2
 Δt n a n
d  a
n n

d n 1
 2d n  d n 1 
Δt  n 2
• The equations of motion at time step n are :

   
M a n  f n  f ext d n , t n  f int d n , t n    Eqn(A)
Subject to constraint s :
 
g I d  0, I  1 to n c
n

We have to find dn so as to satisfy Eqn(A) and the constraints.

These are Ordinary Differential Equations (ODE’s) of second order
in time, often called semidiscrete since they have been discretized in
space but not in time. Generalized representation of the nc
displacement boundary conditions and other constraints on the
model are also appended. Note that these constraints are linear or
nonlinear algebraic functions of the nodal displacements. M the
mass matrix is a constant for Lagrangian mesh.
• The internal and external nodal forces are functions of the nodal
displacements and the time.
• The external loads are prescribed by the analyst as function of
time. They may also depend on nodal displacement because they
may also depend on the deformation configuration of the
body/structure, as when pressure forces are applied to surfaces,
which undergo large deformations.
• The dependence of internal nodal forces on displacements is
obvious. The nodal internal forces depend on the stresses. Stresses
depend on the strains and strain rates by the constitutive equations;
the strains and strain rates in turn depend on the displacements and
their derivatives. The internal nodal forces can also depend directly
on time, e.g., when the temperature is prescribed as a function of
time, then the stresses and hence internal nodal forces are directly
functions of time.
• The equations for UPDATING the nodal velocities and nodal
displacements are:
1 1
n n
V 2
V 2
 Δt n M 1 f n    Eqn(B)
1 1
n n
n 1
d  d  Δt
n 2
V 2

At any time step n, the displacements dn are known. The nodal

forces fn can be determined by sequentially evaluating the strain-
displacement equations, the constitutive equations expressed in
terms of Dn-1/2 OR En and the nodal external forces. Thus the entire
RHS of Eqn(B) can be evaluated, and Eqn(B) used to compute
Vn+1/2. The displacement dn+1 can then be determined.
• The update of the nodal velocities and nodal displacements can
be accomplished without SOLVING any equations provided that
the mass matrix M is diagonal. This is the salient characteristic
of an EXPLICIT METHOD. In an explicit method, the time
integration of the discrete momentum equations does not require the
solution of any equations. However, this hinges critically on the use
of a diagonal mass matrix. The required inversion of diagonal mass
matrix is trivial.
• There are exceptions. For example, the consistent mass matrix is
sometimes used with the central difference method in WAVE
PROPAGATION problems, and the update then does involve the
solution of equations!!
Software Implementation:
Box 6.1 gives the flowchart for explicit time integration. This
flowchart includes nonzero initial conditions, a variable time step,
elements with more than one quadrature point, and DAMPING.
• Damping is modeled by a linear viscous force.
fdamp = Cdamp.V
so that the total force in eqn (B) is f - Cdamp.V
• The implementation of velocity update is broken into two sub
steps by
Vn = Vn-1/2 + (tn – tn-1/2) an
Vn+1/2 = Vn + (tn+1/2 – tn). an
This enables energy balance to be checked at integer time steps.
• The cardinal dependent variables in this flowchart are the
velocities and the Cauchy stress. Initial conditions must be
given for the velocities, the Cauchy stress, and all material state
variables. The initial displacements are assumed to vanish
(initial displacements are meaningless in NLFEA except for a
hyperelastic material since the stress depends on the history of
deformation).
• The main part of the explicit time integration procedure is the
calculation of the nodal forces, which is done in get force. The
major steps in the subroutine get force are
1. Extract the nodal displacements and velocities of the element
from the global arrays by the GATHER operation.
2. The strain measures are computed at each quadrature point of the
element.
3. The stresses are computed at each quadrature point using the
constitutive equations
4. Evaluate the nodal internal forces by integrating the product of
the B matrix and the Cauchy stress over the domain of the
element.
5. The nodal forces of the element are SCATTERED into the global
array.
• In the first step, (n=1), the strain measures and the Cauchy stresses
are not computed, instead, as shown in the flowchart, the nodal
internal forces are calculated directly from the initial stresses.
• In the flow chart, the matrix form of the nodal internal force
calculation, in which the stress tensor is stored as a square matrix. To
change to Voigt form, replace B by B and the square matrix of
stresses by the column matrix {}.
• Most essential boundary conditions are easily handled in explicit
codes. For example, if the velocity OR displacement is prescribed
as functions of time along any boundary, then the prescribed
velocity/displacement boundary conditions can be enforced by
setting the nodal velocities according to the data:
ViIn+1/2 = Vi(XI, tn+1/2)
•Any damping in the system lags by a halftime step. This also holds
for any rate-dependent terms in the constitutive equation evaluation
phase of the step getforce. The time lag is unavoidable if the
implementation is to be fully explicit, i.e., not require the solution of
any equations. However, this decreases the stable time step for the
method.
•The explicit time integration method is easily implemented and is
very robust, by which we mean that the procedure seldom aborts due
to failure of the numerical algorithm. The salient disadvantage is
the conditional stability. If the chosen time step t exceeds a
critical value tcrit, the solution will grow unboundedly!!
•The critical time step also called stable time for a mesh as constant
strain elements and with rate-independent material models is given
by t = .tcrit
tcrit = min (le/Ce) 2/max
• when max is the maximum frequency of the linearized system, le is a
characteristic length of the element e, ce is the current wave speed in
element e, and  is a reduction factor that accounts for the destabilizing
effects of nonlinearities. A good choice for  is 0.80.98.
•The critical time step decreases with mesh refinement and increasing
stiffness of the material. The cost of an explicit simulation is independent of
the frequency range that is of interest and depends only on the size of the
FE model and the number of time steps.
•An interesting question for elastic-plastic materials is whether the slower
wave speed in the plastic response enables one to increase the time step.
Based on experience, the answer appears to be negative. An elasto-plastic
material can unload at any moment, and in numerical solutions unloading
often occurs due to numerical noise. During elastic unloading, the critical
time step depends on the elastic wave speed, and a time step that exceeds
the critical time step results in instability.
•The mesh time step is obtained from element time steps.
Energy Balance
• Certain instabilities can lead to a large over prediction of
displacements. They are easily detected by an energy balance check.
Any instability results in the spurious generation of energy, which
leads to a violation of the conservation of energy. Therefore, whether
stability was maintained during a nonlinear computation can be
established by checking energy balance.
• The internal and external energies are integrated as follows:
n 1 1 T n
W  W  Δd f int  f int
int
2
n
int 
n 1

n 1 n 1 T n
2

n 1
Wext  Wext  Δd f ext  f ext 
n 1
where Δd  d  d n
The kinetic energy is given by

W n
kin  V
2
 
1 n T
M Vn

Note that integer time steps are used for the velocities.
The internal energies can also be computed on the element OR
quadrature point level.
• Energy conservation requires that

Wkin  Wint  Wext  ε max Wext , Wint , Wkin 

where  is small tolerance of the order of 10-2

Mass scaling, subcycling, and dynamic relaxation:
• When a FE model contains a few small OR stiff elements, the
efficiency of explicit time integration method is compromised
severely, since the time step of the entire mesh is set by these very
stiff elements. Several techniques are available for circumventing
this difficulty.
a) Mass scaling: the masses of stiffer elements are increased so that
the time step is not decreased.
b) Subcycling: a smaller time step is used for the stiffer elements.
• Mass scaling should be used for problems where high frequency
effects are not important. For example, in sheet metal forming
simulation, which is essentially a static process, it causes no
difficulties. On the other hand, if high frequency response is
important, mass scaling is not recommended.
• Subcycling was introduced by Belytschko in 1979. In this
technique, the FE model is split into subdomains and each is
integrated with its own stable time step. The crucial issue in
subcycling is the treatment of the interface between subdomains.
• Dynamic relaxation is often used in explicit codes to obtain static
solutions. The basic idea is to apply the load very slowly and solve
the dynamic system equations with enough damping so that
oscillations are minimized. In path-dependent materials, dynamic
relaxation often yields poor solutions. Furthermore, it is very slow.
Newton methods combined with effective iterative solvers, such as
preconditioned conjugate gradient or multi-grid methods are much
faster and more accurate.
Implicit Time Integration Methods and Solution of Equilibrium
Equations:
• We will combine the description of the solution of equilibrium
equations with time integration by implicit methods since they
share many common features.
• To begin, we write the discrete momentum equation at time step
n+1 in a form applicable to both equilibrium and dynamic
problems:

0  r d n 1
,t n 1
 S
D Ma n 1
f int
d , t 
n 1 n 1

f ext
d n 1
,t n 1
     1
where SD is a switch which is set by:
SD = 0 for a static (equilibrium) problem.
SD = 1 for a dynamic (transient) problem.
The column matrix r (dn+1, tn+1) is called a RESIDUAL FORCE.
The discrete equilibrium equations are nonlinear algebraic
equations in the nodal displacements, dn+1.
The discrete equations for the implicit update of the equations of
motion are also nonlinear algebraic equations in the nodal
displacements, dn+1.
Equilibrium solutions and equilibrium points
•When the accelerations vanish or are negligible, a system is in static
equilibrium and the solution of (1) is called an EQUILIBRIUM
SOLUTION. The equations of equilibrium are given by (1) with SD  0.

     
0  r d n 1 , t n 1  f int d n 1 , t n 1  f ext d n 1 , t n 1    2
• In equilibrium problems, the residuals correspond to the out-of-
balance forces. Problems in which the accelerations can be neglected
are called static problems. A solution of the above is called an
equilibrium point and a continuous locus of solutions is called an
equilibrium path.
• In equilibrium problems with rate-independent materials, t need not
be the real time. In stead, it can be any monotonically increasing
parameter.
• If the constitutive equation is a differential or integral equation, it
must also be discretized in time to obtain a set of algebraic equations
Newmark  Equations:
• We use a popular class of time integration methods called
newmark  method. The updated displacements and velocities are

n 1 ~ n 1 2 n 1
d  d  β Δt a
Δt 2
~ n 1
where d  d n  Δt V n  1  2β a n    3
2
~ n 1
V  V  γ Δt a n 1
n1

where V  V n  1  γ  Δt a n        4
~ n 1

Where t = tn+1 – tn
 and  are parameters whose useful values and stability
properties are summarized in Box 6.2
• The parameter  controls artificial viscosity, a damping introduced
by the numerical method. It is used to suppress noise in the solution,
when  = ½, the Newmark method adds no damping for  > ½
artificial damping proportional to -1/2 is added by the integration
method.
• Eqn (3) can be solved for the updated acceleration, giving
a  n 1 1
β Δt 2

n 1

d  d when β  0    5
~ n 1

When substituted into eqn (1) gives

0r
SD
β Δt 2
M d n 1

~ n 1
d f int

d n 1 n 1
, t  f ext
d ,t 
n 1 n 1
   6

which is a set of nonlinear algebraic equations in the nodal

displacement dn+1.
• Equation (6) applies to both static and dynamic problems. In both
the cases, the problem is to find dn+1 so that r(dn+1, tn+1)  0
subject to g(dn+1,tn+1)  0
• Equations (6) are the Discrete Equations of motion/Equilibrium
Newton’s Method: The Newton Raphson Algorithm
• The most widely used and most robust algorithm for the solution
of the set of nonlinear algebraic equations (6) is Newton’s
method. It is often called Newton-Raphson Iterative method in
computational mechanics.
• The Newton-Raphson Iterative procedure is illustrated in Fig 6.1
• The solution of equations (6) by the NR method is an
ITERATIVE PROCEDURE. The iteration number is indicated by
Greek subscript: dn+1  d is the displacement in ITERATION 
at time step n+1; the time step number n+1 will be omitted for
brevity.
•To begin the iterative procedure, a starting value for the unknown
must be chosen; usually the solution dn from the last time step is
selected, so d0 = dn. In dynamic problems, a better starting value is
~ n 1 . A Taylor expansion of the residual r about the current value
d
of nodal displacement d and setting the resulting residual equal to
zero gives
r d 
r d  
d
 
Δd  0 Δd  0      7 
2

The matrix r/d is called the system Jacobean matrix denoted by

A.
We then get
r + A. d = 0 ---------------------------------(8)
which is the linear model of the nonlinear equations. The linear model
is difficult to picture for problems with more than one unknown.
Figure 6.2 shows an example of the first component of the residual for
a function of two unknowns. The linear model is a plane tangent to the
nonlinear function r1(d1,d2). The other residual component is another
nonlinear function r2(d1,d2), which is not drawn.
• The increment in the nodal displacements in the NR-Iterative
procedure is obtained by solving eqn (8), which yields a system of
linear algebraic equations.
A. d = - r (d,tn+1) -------------- (9)
Once the increments in nodal displacements have been calculated, they
are added to the previous iterate
d+1 = d +  d
The new nodal displacements is checked for convergence. If the
convergence criterion is not met, a new linear model is constructed
and the process is repeated. The iterative procedure continues until
the chosen convergence criterion is met.
•In computational mechanics, the Jacobian matrix A is called the
Effective Tangent Stiffness matrix and the contributions of the
INERTIAL, INTERNAL and EXTERNAL forces are linearized
separately.
•The Jacobian is written as
r SD f int f ext
A  M  for β  0
d β Δt 2
d d
• For equilibrium problems; SD  0;
f int
f int
is called the Tangent stiffness matrix denoted by K 
int
d d
f ext
f ext
is called the Load stiffness matrix denoted by K ext 
d d
• For dynamic problems; SD  1; the jacobian matrix is
SD
A M  K int
 K ext

β Δt 2

• Note that this also applies to static problems with SD  0.

• The Jacobians can be used to relate differentials of the nodal forces
to differentials of nodal displacements by
df int
 K dd
int

df  K d d
ext ext

dr  A d d
Software Implementation
• Flow charts for Implicit integration and Equilibrium solutions are
given in Boxes 6.3 and 6.4, respectively. Both the dynamic problem
and the equilibrium problems are solved by time stepping. The
external loads and other conditions are described by the user as
functions of time, which is incremented over the range of interest. In
equilibrium problems, tine is often replaced by a monotonically
increasing parameter. Solutions of equilibrium processes obtained in
this manner are called Incremental Solutions.
• The flow chart shows the full Newton algorithm, where the
Jacobian matrix is evaluated and inverted in every iteration of the
solution procedure.
•Many programs offer a modified Newton algorithm, in which the
Jacobian matrix is assembled and triangulated only at the beginning
of a step or intermittently during a step. For example, the Jacobin
matrix may be triangulated only when the ITERATIVE procedure
does not seem to be converging well. The modified Newton algorithm
is faster but less robust.
•The Implicit scheme begins with the imposition of the initial
conditions. The initial displacements are considered to be zero. The
initial accelerations are computed as shown in steps 2 and 3 (Box
6.3).
•The nodal displacements dn+1 are obtained by an iterative procedure.
To begin the iteration, a starting value of d is needed; usually the
solution from the preceding time step is used. The residual r is then
calculated for this starting value. In an equilibrium solution, the
residual depends only on the internal and external nodal forces; which
is obtained in the module getforce. In transient implicit solution, the
residuals also depend on the acceleration.
Constraints
• There are four options for treating the constraints.
gI(dn) = 0, I = 1 to nc (no. of constrants)
1. Penalty.
2. Lagrange multiplier.
3. Augmented Lagrangian
4. Perturbed Lagrangian.

These methods originate in optimization theory. They are adopted to

the solution of discrete momentum and equilibrium equations in
nonlinear FEA programs.
 Convergence Criteria
• The termination of iterations in Newton algorithms is
determined by convergence criteria.
• Three types of convergence criteria are:
1. Based on the magnitude of the residual forces r
2. Based on the magnitude of displacement increment d.
3. Energy error.
Implicit Time Integration Method: Accuracy and Stability
•For linear transient dynamic problems, suitable implicit integration
schemes are unconditionally stable.
•The major restriction on the time step in implicit methods arise from
accuracy requirements and decreasing robustness.
•Large time steps also impair the convergence of Newton method,
particularly in problems with very rough response, such as Contact
Impact. Small time steps improve the robustness of the Newton
algorithm.
•In return for there enhanced stability implicit methods exact a
significant price; they require the solution of a set of nonlinear
algebraic equations in each time step.
• Implicit solutions are now feasible in many problems where they
failed before. We are certain that further improvements are
imminent. Nevertheless, high cost and insufficient robustness still
plague Newton method.
Newton Iteration: Convergence and Robustness:
• The rate of convergence of the ITERATIONS in the Newton
method (or N-R algorithm) depends on the Jacobian matrix A
satisfying the following conditions.
 1. The Jacobian matrix A should be sufficiently smooth function
of the nodal displacements d.
2. The Jacobian matrix A should be regular (invertible) and well
conditioned in the entire domain in the displacement space that
the iterative procedure traverses.
• These conditions are not satisfied by engineering problems. For
example, in elastic-plastic material, the residual r is not a
continuously differentiable function of the nodal displacement d:
when an element quadrature point changes from elastic to plastic
or vice versa, the derivative is discontinuous.
• In contact-impact problems with Lagrange multiplier methods,
the residual also lacks smoothness.
• Yet, N-R algorithm is still remarkably effective. It converges!!
• In some nonlinear problems, the conditions for quadratic
convergence of the N-R method are fully satisfied. For example,
in a smoothly loaded FE model with a Mooney-Rivlin material
where load increment is small enough so that equilibrium
conditions are stable.
• In equilibrium problems, the NR method fails to converge in the
vicinity of unstable states.
• In dynamic problems, the convergence difficulties of the NR
iteration are ameliorated by the mass matrix. For many problems,
use of the Newton method is totally unsuccessful, and further
enhancements such as the arc length method are needed.
IMPLICIT VS EXPLICIT
• The selection of time integration scheme depends on:
1. The type of Partial Differential Equation (PDE).
2. The smoothness of the data.
3. The response of interest.
• For parabolic PDE’s, implicit methods are preferred. The
solutions of parabolic systems are smooth even when the data is
rough. The stable time step in a parabolic system decreases by a
factor of four each time the element size is halved, so that
refinement in an EXPLICIT method becomes prohibitively
expensive. Parabolic systems should never be integrated by
explicit schemes.
•There are exceptions where explicit methods are preferred even
when the system is partially or fully parabolic. For example in car
crash simulation (using shell elements) the equations for the shells
which model the sheet metal response are parabolic, yet because the
noise introduced by Contact-Impact algorithm, explicit methods are
preferred.
•Similarly, in complex heat conduction problems, it is often
impossible to take advantage of the large time step permitted by
implicit methods explicit methods are preferred.
•In hyperbolic systems, the choice depends on the response of
interest. Hyperbolic PDE’s govern wave propagation problems and
inertial problems (structural dynamics).
• In structural dynamics problems, the frequency spectrum of the
input is far below the resolution limits of the FE mesh; the FE mesh
refinement is dictated by the need for very high accuracy in the
frequency band of interest. Here, IMPLICIT integration methods
are definite choice. Examples include seismic response also.
• Wave propagation problems are those in which relatively high
frequency parts of the spectrum are of interest. For example, waves
in cellular phones in a drop test, the simulation needs a small time
step to tract the high frequency portion of the ‘response’, and
explicit methods are preferred.
LINEARIZATION
Linearization of the Internal nodal forces
• We now derive expressions for the tangent stiffness matrix Kint
for Lagrangian elements. As will be seen part of the expression
can be derived independently of the material response.
Linearization of the constitutive equation is carried out in two
ways:
1. With the continuum tangent moduli, which does not account for
the actual constitutive update algorithm; the resulting material
tangent stiffness is called the TANGENT STIFFNESS MATRIX.
2. with the algorithmic tangent moduli, which gives rise to the so
called consistent tangent stiffness.
• The choice rests on considerations related to ease of
implementation and on the smoothness of the problem.
• The tangent stiffness matrix will be expressed in terms of CSE, the
tangent moduli relating the rate of the PK2 stress to the rate of the
Green strain in the Total Lagrangian Formulation and in terms of C,
the tangent moduli relating the Truesdell rate of the Cauchy stress to
the rte-of-deformation in the Updated Lagrangian Formulation. These
tangent moduli are related: if the current configuration is chosen to
be the reference configuration, C = CSE.
• We will develop the tangent stiffness matrix by relating the rates of
 .The procedure
the internal nodal forces f int to the nodal velocities d
is identical to relating an infinitesimal increment of internal nodal
forces dfint to an infinitesimal increment of nodal displacements d d
•The internal nodal forces in the TLF are given by
 f int   B T P dΩ         A 
Ω

OR
N I
f int
 Pji dΩ 
X j
iI
Ω

Where P is the nominal stress tensor with components Pji, NI are

the nodal shape functions, and
N I
B 


X j
jI

• We have opted the TLF because this leads to the simplest

derivation. In the TLF, the only variable, which depends on the
deformation is the nominal stress, i.e., it is the only variable which
varies with time.
• In the ULF, the domain of the element (or body), the spatial
derivatives NI/xj and the Cauchy stress depend on the
deformation and hence on time, which complicates the derivation of
the tangent stiffness matrix.
•Taking the material time derivative of eqn (A),

 0
f int  B T P dΩ
Ω0
0

Since B0 and d0 are independent of deformation or time.

• To get the stiffness matrix Kint, it is now necessary to write the
nominal stress rate P  in terms of the nodal velocities via the
constitutive equations and the strain measure.
• The material time derivative of the PK2 stress is related to the
material time derivative of the nominal stress P by taking the time
derivative of the transformation P = S.FT.
P  S FT  S F T
• Finally eqn A becomes
f int
iI 
N I 
X j
 
S jr Fir  S jr Fir dΩ 0
Ω0

N I
f int
 dS jr Fir  S jr dFir dΩ0
X j
iI
Ω0

• The above shows that the rate ( or increment) of the internal nodal
forces consists of two parts.
  , and thus depends on
1. The first part involves the rate of stress S
the material response and leads to what is called the material
tangent stiffness matrix which we denote by Kmat
2. The second part involves the current state of stress S, and
accounts for geometric effects of the deformation (including
rotation and stretching). This term is called the geometric
stiffness. It is also called the initial stress stiffness matrix to
indicate the role of the existing state of stress. It is denoted by
Kgeo.
• The changes in the internal nodal forces due to these two effects are
given analogous names, so
f int  f mat  f geo
OR f int  f mat  f geo
iI iI iI
where
N I 
fiImat  Ω X j Fir S jr dΩ0
0

N I
f geo 
iI Ω X j ir 0
S jr
 dΩ
F
0
Material Tangent Stiffness Matrix
• We rewrite the rate of the internal nodal forces, in Voigt notation


f mat   BT0 S dΩ 0
Ω0

Where
format.
S is the rate of the PK2 stress in Voigt column matrix
• The constitutive equation in rate form is

S  CSE E
ij ijkl kl

or   
S  CSE E

Where E  is the rate of the Green strain.
• We now relate the rate of the Green strain to the nodal velocities
in Voigt notation,
E  B 0

d
• Finally we get the material rate of the internal nodal forces

Ω0
  
f int  B T CSE B dΩ d
mat 0 0 0

OR
fmat
int
 
  B T0 CSE B 0 dΩ 0 dd
Ω0

• So, the material tangential stiffness matrix is

K mat
  B C B dΩ
T
0
SE
0 0
OR Ω0

K mat
IJ   B C B dΩ
T
0I
SE
0J 0
Ω0
• The material tangent stiffness relates increments (OR rates) of the
internal nodal forces to the increment (or rate) of the nodal
displacements due to material response, which is reflected in the
tangent modulii CSE. Its form is identical to the stiffness matrix in
linear finite element analysis.
Geometric Stiffness: Initial Stress Stiffness Matrix
• The geometric stiffness matrix is obtained as follows.
From the definition

N I
B  0

X i
iI

We can write
 ii  Ij jr ir 0
f geo  B0 S F dΩ
Ω0

  B S jr B dΩ 0 u i,J
0
Ij
0
rj
Ω0

  B S jr B dΩ 0 δ ik u kJ
0
Ij
0
rj
Ω0

• Writing the above in matrix form, it gives

f geo  K geo u
I IJ J

K geo  
T
IJ I B 0I S B 0J dΩ 0
Ω0
•Note that the PK2 stress S is in tensor form, i.e., a square matrix.
•The geometric stiffness matrix Kgeo is invariant with rotation.
•The above in Total Lagrangian Formulation are easily converted
to ULF (Updated Lagrangian Formulation) by letting the current
configuration be the reference configuration:

K mat
IJ  B CT
I  B dΩ
στ
j
Ω

 
K mat   BT C στ B j dΩ
Ω

K geo
IJ  I  B σ BJ dΩ
σ
I
Ω
• The tangent modulii [CSE] and [C] for a wide variety of
engineering materials are presented in the session: Constitutive
Models.
• Tangent material stiffness matrices for specific finite elements
are presented later in this session.
• The numerical values of the above matrices Kmat and Kgeo in the
TLF and ULF are identical, and the choice of which to use is a
matter of convenience.
• The Issue of symmetry of the tangent stiffness matrix deserves
attention. Symmetry speeds the solution of the set of nonlinear
algebraic equations reduces storage requirements and simplifies
stability analysis.
• The material tangent stiffness matrix is symmetric when the
Voigt form of the tangent modulii matrix [CSE] is symmetric.
•The geometric stiffness matrix is always symmetric. Therefore, the
tangent stiffness matrix Kint is symmetric whenever the tangent
modulii possess major symmetry.
•See box 6.5 for Jacobian of Internal Nodal Forces (tangent stiffness
matrix).
External load stiffness
•Follower loads are loads that change with the configuration of the
body (elements). They appear in geometrically nonlinear problems
only. Pressure is a common example of Follower load. A pressure
force is always normal to the surface, so as the surface MOVES the
external nodal forces change even if the magnitude of pressure is
constant. These effects are accounted for in the Jacobian matrix of
external nodal forces, denoted by Kext, which is also called the
LOAD STIFFNESS .
• The load stiffness Kext relates the rate of external nodal forces to
the nodal velocities.
• Consider a pressure load, p(x,t), the external nodal forces on a
surface of element e are given by letting t  p n

f I
ext
  N I p n dΓ
Γ
• Let the surface  be parametized by two variables  and . For
quadrilateral surface element, these variables can be the parent
element natural coordinates on the biunit square, since

n dΓ  x ,ξ x ,η dξ dη we can write

1 1
f I
ext
    pξ, η N I ξ, η x ,ξ x ,η dξ dη
1 1
Taking the time derivative of the above gives
1 1
fIext     N I p x ,ξ x ,η  p V,ξ x ,η  p x ,ξ V,η dξ dη
1 1

•The first term is the rate of change of the external nodal forces
due to the rate of change of the pressure. In many problems, the
rate of change of pressure is prescribed as part of the problem
definition.
•In other problems, such as in Fluid-Structure Interaction
problems, pressure may arise from changes of the geometry: these
effects must then be linearized and added to the Load Stiffness. We
omit this aspect in the following.
•The last two terms represent the changes in the external nodal
forces due to the changes in the direction of the surface and the
area of the surface. These are related by the external load stiffness,
so the RHS of the above equation becomes
 1 1

IK VK     p N I V,ξ x , η  x ,ξ V, η dξ dη
K ext
1 1

We can finally write f Iext  K ext

IK VK

  0 z ,η  y ,η   0  z ,ξ  y ,ξ  
1 1     
and K ext
IJ     p N I  N J, ξ  z ,η 0 x ,η   N J, η  z ,ξ 0 x ,ξ   dξ dη
 
1 1

 y ,η
  x ,η 0   y ,ξ
  x ,ξ 0  

as the load stiffness matrix.

• It is immediately apparent that the submatrices of the load
stiffness matrix are not symmetric. So the Jacobian matrix in
general is not symmetric in the presence of follower forces.
However, it can be shown that for a closed structure in a constant
pressure field, the assembled external load stiffness is symmetric.
TWO-NODE ROD ELEMENT
• Consider the two-node rod element shown in Figure 6.9. The ROD
is in a state of uniaxial stress. The x̂ -axis lies along the axis of the
rod and rotates with the rod, i.e., it is a corotational coordinate. The
only nonzero Cauchy stress component is σ̂11  σ̂ xx . The tangent
stiffness and load stiffness matrices are derived in the ULF, i.e., in
the current configuration.
• Material tangent stiffness matrix (ULF)
 
K̂ mat   B̂T Ĉστ B̂ dΩ
Ω

 1 0  1 0
στ 
AE  0 0 0 0
K̂ mat 
l  1 0  1 0
 
0 0 0 0
• This is identical to the linear stiffness matrix for a rod if E is
replaced by Young’s modulus E.
• The material tangent stiffness matrix relating global components
of Internal Nodal forces and Nodal Velocities is

f Imat  K mat VI
K mat  T T K̂ mat T
Where T is given by
 cosθ sin θ 0 0 
 sin θ cosθ 0 0 
T 
 0 0 cosθ sin θ 
 
 0 0  sin θ cosθ 
• Geometric Stiffness Matrix (ULF)
The geometric stiffness matrix is given by
K̂ IJ  Ĥ IJ I
Ĥ   B̂T σ B̂ dΩ
Ω

1   1
    σ̂ xx  1  1 dΩ
1
l  1
Ω 
l
σ̂ xx A  1  1
Ĥ 
l   1  1
 1 0  1 0 
 0  1 0  1
Aσ̂ xx  
K̂ geo 
l  1 0  1 0 
 
 0  1 0  1
K geo  T T K̂ geo T
• The total tangent stiffness matrix is the sum of the material and
geometric stiffnesses.
Kint = Kmat + Kgeo
Load stiffness
The load stiffness for the 2-node ROD element is developed here.
First it is evaluated in the corotational coordinate system x̂, ŷ

 0  1 0  1
 1 0  1 0 
 pa  
K̂ ext 
2  0  1 0  1
 
 1 0  1 0 
K ext  T T K̂ ext T  K̂ ext
Material tangent stiffness matrix in TLF

 
K mat   BT0 CSE B0 dΩ 0
Ω0

cos 2θ cosθ sin θ  cos 2θ  cosθ sin θ 

2 
A 0 E SE l  sin θ
2
 cosθ sin θ  sin θ 
2
  
l0  0
l  cos 2
θ cosθ sin θ 
 
 sin θ 
2

Note the material constant ESE relates the rate of the PK2 stress to
the rate of the Green strain in a uniaxial state of stress.
Geometric Stiffness Matrix in TLF
 1 0 1 0 
 
A 0 Su  0 1 0  1
K geo 
l 0  1 0 1 0 
 
 0  1 0 1 
THREE-NODE TRIANGULAR ELEMENT
• We consider the 3-node triangular element in two dimensions in a
state of plane strain.
Material tangent Stiffness Matrix

K mat   BT Cστ B dA 

A

Where A denotes current area of the element. We have set the

thickness a  1.
The tangent modulus matrix in Voigt matrix form is
 C1111
στ στ
C1122 στ
C1112 
 στ 
C 
στ στ
C C C
στ
  στ 2211 2222 2212 
 C1211 C1222
στ στ
C1212 
 
 
Using the known B matrix, since the integrand is frequently
constant, the tangent stiffness matrix is the product of the integrand
with the area.
Geometric stiffness matrix

K geo  2x2  I σ B J dA I 2x2 H IJ

T
IJ I B
A

1  y 23 y 31 y12 
when B 
2A  x 32 x13 x 21 
 y 23 x 32 
1   σ xx σ xy   y 23 y 31 y12 
H x13  
σ yy   x 32 x 21 
 y 31
4A  σ xy x13
 y12 x 21 

The geometric stiffness matrix is independent of material response.

It depends only on the current stress state and the current geometry
of the element
Physical Stability and Continuation Methods:
•In nonlinear problems, stability of solutions is of considerable
interest. However, we will focus on the FEM and application to
nonlinear analysis.
•Consider a beam loaded axially (fig 6.7 (a)). If we perturb the
location of the load P by a distance  or 2, then the equilibrium
paths are as shown in Figure 6.7(c). it can be seen that when the load
is below the buckling load, the path for different initial conditions
remain close to AC. However when the load exceeds the buckling
load, the solution for different initial conditions diverge. Therefore,
any process in which the applied load exceeds the buckling load is
unstable.
•As an aside, we remark that when the beam is straight, the numerical
solution usually remains on the path AC shown in Figure 6.7(b). The
lateral displacement  is zero even when the applied load P exceeds
the buckling load! If you do not believe this, try it. In a simulation, a
straight beam will usually not buckle in an Incremental Static
Solution or a Dynamic Solution, regardless or whether explicit or
implicit integration is used. Only when round off error introduces a
“numerical imperfection” or when an imperfection is introduced in
the INPUT DATA will the straight beam buckle in a simulation i.e., an
imperfection is needed to break the symmetry.
• It is widely believed among structural mechanicians that the
difficulties associated with unstable behavior can be circumvented by
obtaining a dynamic solution. When a structure is loaded above its
limit point or a bifurcation point in a dynamic simulation, the
structure passes dynamically to the nearest stable branch. However,
the instability is not readily apparent, and the possibility of
imperfection sensitivity is not clear.
• Therefore, to understand the behavior of a structure thoroughly, its
equilibrium behavior should be ecarefully
iI
examined. Many vaganes
of structural behavior may be hidden by dynamic simulations.
•For example, for a cylindrical shell under axial load the intersecting
branch is asymmetric as shown in Figure 6.7(e). A structure with
asymmetric branch is highly sensitive to imperfections, as can be
seen by the large change in the maximum load with the change in the
imperfection in Figure 6.7(e).
•The theoretical bifurcation point of a perfect structure is then not a
realistic measure of the strength: the actual structure may buckle at a
much lower load than the theoretical value because imperfections are
unavoidable in a real structure. A single numerical simulation could
miss this sensitivity completely.
•This sensitivity to imperfections for cylindrical shells was
analyzed by Koiter and is a classical example of imperfection
sensitivity. To ascertain this type of behavior, the equilibrium
branches must be known.
•The behavior of the shallow truss exhibits a turning or Limit Point,
as can be seen from Figure 6.9; Points A and B are turning points.
The branch after the first turning point, point A, is unstable. While
the branch after the second turning point, Point B, is stable!
•The behavior of the beam shown in Figure 6.7(a) is a classic
example of a bifurcation. The point B where the two branches
intersect is the point of bifurcation. Subsequent to the bifurcation
point, the continuation BC of the fundamental branch AB becomes
unstable. Point B, the bifurcation point, corresponds to the buckling
load of the Euler beam.
• Methods for tracing equilibrium branches are called continuation
methods. The tracing of equilibrium branches is often quite
difficult: Robust, automatic procedures for continuation are not yet
developed. A continuation method called arc length method is
implemented in commercial FEA programs. The arc length method
is often called the Rik’s method in structural mechanics literature.
NUMERICAL STABILITY
•A numerical procedure is stable if small perturbations of initial data
result in small changes in the numerical solution.
•General results for numerical stability analysis of the time
integration of the equations of motion are of interest.
•However, we stress that at the present time there is no stability
theory, which encompass the nonlinear problems, which are routinely
solved by Nonlinear FEA programs.
•Difference between physical stability and numerical stability:
Physical stability pertains to the stability of the solution of a model
whereas numerical stability pertains to the stability of the numerical
method. Numerical stability is usually only examined for processes,
which are physically stable.
•Stability analysis of widely used Time Integration Methods are well
documented: example: Explicit Central Difference method; Implicit
Newmark  method.
MATERIAL STABILITY
•An important issue in nonlinear continuum mechanics is the
stability analysis of the material model. We now describe criteria for
material stability. Some remedies for the numerical difficulties
incurred by material instabilities are also discussed.
•Material instabilities are usually associated with a localized growth
of the deformation. This is called localization. It corresponds to
phenomena which are observed in experiments: For certain stress
states, metals, rocks and soils will exhibit narrow bands of intense
deformation; these are often called SHEAR BANDS, since the
deformation mode in these bands is usually shear.
•Example: Rate-independent material model.
•Shortly after NLFA Programs became available in the 1970’s,
computational analysts began to include unstable material models
both intentionally and inadvertently, and they discovered many
difficulties. Numerical solutions became often unstable and it was
discovered that the results depend very much on the mesh!
•At that time, it was argued by some mechanicians that material
models that violate the stability postulates should never be used. This
argument had merit for unless a constitutive law is carefully designed
to become unstable only in the rare situations when the material is
unstable, many difficulties are encountered. However, there is no
way to replicate experimentally observed phenomena such as shear
banding without a material model that exhibits strain softening.
•Several techniques have evolved to improve the coarse-mesh
accuracy of the finite element models for unstable material models.
These involve the embedment of discontinuities or enriched fields in
the element.
 Ortiz Leroy and Needleman (1987) were the first to modify an
element at the point of material instability: they embedded
discontinuities in the strain field of the 4-node quadrilateral
elements, when the acoustic tensor indicated a material instability in
the element.
Belytschko, Fish and Englemann (1988) embedded a
displacement discontinuity by enriching the strain field with a
narrow band in which the material is unstable.
•Just as shear bands can be viewed as the outcome of a material
instability in the shear stress component, fracture can be considered
as the outcome of a material instability in the stress components
normal (and tangential in the case mode II fracture) to the
discontinuity (i.e., a crack). The relation between damage and
fracture has long been noted where fracture is assumed to occur when
the damage variable reaches a critical value.
•Damage mechanics the modeling of fracture by constitutive models
with damage poses same difficulties encountered in shear band
modeling, since the material model becomes unstable, when the
damage exceeds threshold value.
•These difficulties were resolved in a novel manner. The idea is to
match the energy dissipation in cracking to the energy dissipated in
the element, which exceeds the stability threshold. This is
accomplished by treating the fracture energy as a material parameter
and the energy dissipation in the strain-softening element is equaled
to the fracture energy (a material parameter).
 Lemaitre J and Chaboche JL (1990) Mechanics of Solid
Materials, Cambridge University Press, Cambridge