Вы находитесь на странице: 1из 13

LS-DYNA Theory Manual

Implicit

34. IMPLICIT
Implicit solvers are properly applied to static, quasi-static, and dynamic problems with a
low frequency content. Such applications include but are not limited to

Static and quasi-static structural design and analysis


Metal forming, especially, the binderwrap and springback

An advantage of the implicit solver on explicit integration is that the number of load or time
steps are typically 100 to 10000 times fewer. The major disadvantage is that the cost per step is
unknown since the speed depends mostly on the convergence behavior of the equilibrium
iterations which can vary widely from problem to problem.
An incremental-iterative numerical algorithm is implemented in LS-DYNA. The method
is stable for wide range of nonlinear problems that involve finite strain and arbitrarily large
rotations. Accuracy consideration usually limits the load increment or time step size. An
inaccurate solution will often not converge. Nine iterative schemes are available including the
full Newton method and eight quasi-Newton methods. These are:

Full Newton
BFGS (default),
Broyden,
Davidon-Fletcher-Powell (DFP) [Schweizerhof 1986],
Davidon symmetric, [Schweizerhof 1986],
modified constant arc length with BFGS,
modified constant arc length with Broydens,
modified constant arc length with DFP,
modified constant arc length with Davidon.

A line search is combined with each of these schemes along with automatic stiffness
reformations, as needed, to avoid non-convergence. LS-DYNA defaults to the BFGS quasi Newton method which is the most robust although the other methods are sometimes superior.
Generally, the quasi-Newton methods require fewer iterations than the modified Newton method
since they exhibit superlinear local convergence due to the rank one or rank two updates of the
stiffness matrix as the iterations proceed. In this chapter, important aspects of the static
algorithm are explained, hopefully, in a way that will be understandable to all users. The arc
length methods are generally used in solving snap through buckling problems.
Discretization leads to the matrix equations

where

Mx&& ( t ) + F ( x, x& ) P ( x, t ) = 0
%%
% % %
% %
%
&&
x, x& , x = acceleration, velocity, coordinate vectors
% % %
M = mass matrix
%

34.1

(34.1)

Implicit

LS-DYNA Theory Manual

P = body force and external load vector


%
F = Lt B t dVe
%
e % Ve % %
A diagonal lumped mass matrix is obtained by row summing according to Equation (34.2).
M ii = i i dv = i dv

(34.2)

The primary nonlinearities, which are due to geometric effects and inelastic material behavior,
are accounted for in F .
%
(34.3)
Fe ( x, x& ) = B t ( , & ) dVe
% % % Ve % % % %
where and & are the strains and strainrates. If linear
%
%
Fe ( x, x& ) = Cx& + K x
% % %
%% % %

(34.4)

where K , C , and x are the stiffness matrix, damping matrix, and displacement vector.
% %
%
Additional nonlinearities arise in P due to geometry dependent applied loads.
%
Regardless of whether an implicit or explicit integration scheme is used, we require that

If linear

Q = Mx&& + F P = 0
%% % % %
%

(34.5)

Q = Mx&& + Cx& + K x P = 0
% % %% % % % %
%

(34.6)

Explicit integration trivially satisfies these equations since the calculation of the
acceleration guarantees equilibrium, i.e.,
&&
x = M 1 P n F n
%
%
% %

(34.7)

The explicit update of the velocities and coordinates is given by:


x& n +1/ 2 = x& n 1/ 2 + t n &&
xn
%
%
%

(34.8)

x n +1 = x n + t n +1/ 2 x& n +1/ 2


%
%
%

(34.9)

Stability places a limit on the time size. This step size may be very small and, consequently, a
large number of steps may be required.
For the implicit solution the residual vector becomes an implicit function of x n +1 only.
%
We seek the vector x n +1 such that
%

34.2

LS-DYNA Theory Manual

Implicit
Q ( x n +1 ) = 0
%
% %

(34.10)

Assume an approximation x nk +1 to x n +1 for k = 1, 2, 3...etc. In the neighborhood of x nk +1 we use


%
%
%
the linear approximation to Q ( x n +1 ) = 0 given by
%
% %
Q ( x n +1 ) = Q ( x kn +11 ) + J ( x kn +11 ) x k
% %
%
% %
% %

(34.11)

x nk +1= x kn +11+ x k
%
%
%

(34.12)

Q
J ( x nk +11 ) = % x nk +11
% %
x %
%

(34.13)

and iterate for the solution:

The Jacobian matrix is expressed as

x F x& F P
&&
x
x&
P
J = M % + % % + % % = M % + C % + Kt %
% x x& x x x
% x ~ x % x
%
%
% %
%
%
%
%
%

(34.14)

F
F
where C = % is the tangent damping matrix, and K t = % is the tangent stiffness matrix.
%
% x&
x
%
%
To obtain the solution at load increment n+1 given the solution at load increment n,
linearized equations of the form
K t ( x n ) uo = P ( x n )
% %
% %
%

n +1

F ( xn )
% %

(34.15)

are assembled where


Kt
%
uo
%

P ( xn )
% %

n +1

F ( xn )
% %

Positive-definite tangent stiffness matrix.

Desired increment in displacements.

External load vector at n+1 based on geometry at time n.

Stress divergence vector at time n.

The displacement vector is updated:

and equilibrium iterations begin:

x1n +1 = x n + s0 u0
%
%
%

34.3

(34.16)

Implicit

LS-DYNA Theory Manual

K t j ui = P ( xin +1 )
% %
% %

n +1

F ( xin +1 ) = Qin +1
% %
%

(34.17)

where the subscripts i denotes the iterate and j < i and s0 is a parameter between 0 and 1 found
from a line search. After each iteration, convergence is checked. Convergence is assumed if the
conditions
u i
% < d
(34.18)
u max
%
and
u itQi
% % < e
u 0t Q0
% %

(34.19)

are satisfied. If convergence is not attained, the displacement vector is updated:


x in++11 = x in +1 + si u i
%
%
%

(34.20)

and another iteration is performed. A lack of convergence within the allowable number of
iterations (default = 10) or divergence
Q 0n +1 < Q in++11
%
%

(34.21)

causes K t to be reformed. Termination occurs if the allowable number of reformations is


%
reached (default =15).
The foregoing integration method with K t held constant is called the modified-Newton
%
method. When convergence problems arise, the stiffness matrix is reformed.
Four methods, called quasi-Newton methods, for updating the stiffness matrix are also
available:
Broydens first method
Davidon
DFP
BFGS
These methods involve
Line searches
Rank 1 or rank 2 stiffness updates
and are more expensive than modified Newton but result in a more robust program.
The secant matrices K i are found via the quasi-Newton equations
K i ui 1 = Qi

where

34.4

(34.22)

LS-DYNA Theory Manual

Implicit

Qi = Qi 1 Qi = ( Pi 1 Fi 1 ) ( Pi Fi )

(34.23)

Figure 34.1 below illustrates the calculation of the secant stiffness.

Figure 34.1. The same sign indicates a stable softening system and that s > 1. No line search
is necessary.

Two classes of matrix updates that satisfy the quasi-Newton equations are of interest:

Rank one update:

Ki = Ki 1 + z x t

(34.24)

Rank two update:

Ki = Ki 1 + z x t + v y t

(34.25)

Substituting K i = K i 1 + z x t into Equation (34.22) gives


K i 1ui 1 + z x t ui 1 = Qi

(34.26)

By choosing
1
x ui 1

1.

2.

z = Qi K i 1ui 1

(34.27)

(34.28)

Equation (34.22) is satisfied. Note that x is an arbitrary vector but is restricted such that
uit1 x 0

34.5

(34.29)

Implicit

LS-DYNA Theory Manual

Broyden set x = ui 1 and obtained the update


K i = K i 1 +

Q K i 1ui 1

( Q Ki 1ui 1 )

ui 1

( Q Ki 1ui 1 )

resulting in nonsysmmetric secant matrices. The inverse forms are found by the ShermanMorrison formula:

(A+ ab )

t 1

A1a b t A1
=A
1 + bt A1a
1

(34.30)

where A is a nonsingular matrix. The inverse form for Broydens update can be found by letting
a=

Qi K i 1ui 1
t
ui 1 ui 1

b = ui 1

(34.31)

in the Sherman-Morrison formula. Therefore,

1
i

=K

1
i 1

( u
+

i 1

K i11Qi )( ui 1 K i11Qi )

uit1 K i11 Qi

(34.32)

Again, recalling the quasi-Newton equation, Equation (34.22), and substituting Equation (34.25)
gives:
K i 1ii 1 + z x t ii 1 + v y t ii 1 = Qi

(34.33)

Set
1.

1
x ui 1

(34.34a)

2.

z = K i 1ui 1

(34.34b)

3.

4.

v = Qi

1
y ui 1

(34.34c)

(34.34d)

Here x and y are arbitrary vectors that are nonorthogonal to ui 1 , i.e.,


uit1 x 0

(34.35)

34.6

LS-DYNA Theory Manual

and

Implicit

uit1 y 0

(34.36)

In the BFGS method:


x = K i 1ui 1

y = Qi

(34.37)

which leads to the following update formula


Qi Qit K i 1ui 1uit1 K i 1
K i = K i 1 +

uit1 Qi
uit1 K i 1ui 1

(34.38)

that preserves the symmetry of the secant matrix. A double application of the Sherman-Morrison
formula then leads to the inverse form.
Special product forms have been derived for the DFP and BFGS updates and exploited by
Matthies and Strang [1979].
K i 1 = ( I + wi v it ) K i11 ( I + vi w it )

(34.39)

The primary advantage of the product form is that the determinant of K i and therefore, the
change in condition number can be easily computed to control updates. The updates vectors are
easely defined. Let,

i = ui 1

(34.40)

i = Qi 1 Qi

(34.41)

u 2
vi = Qi Qi 1 1 + i i
i Qi 1

wi =

i
i i

(34.42)

(34.43)

Noting that the determinant of K i is given by:


det ( K i ) = det ( K i 1 ) (1 + w t v ) 2

it can be shown that the change in condition number, C , is

34.7

(34.44)

Implicit

LS-DYNA Theory Manual

v t v 12 w t w 12 + v t v w t + 4 1 + v t w 2
) ( )
(
)
(
C=
2
4 (1 + v t w )

(34.45)

Following the approach of Matthies and Strang [1979] this condition number is used to decide
whether or not to do a given update.
With quasi-Newton updates line searches are usually necessary to achieve convergence.
The incremental displacement is uses the variable, s , which is determined during the line search,
in the calculation of ui :
ui = si di

(34.46)

d i = K i1 Qi

(34.47)

where

Therefore, the update of the displacement, ui +1 , becomes:


ui +1 = ui + si di

(34.48)

The line search computes si such that the component of P F in the search direction is zero:
di P ( xi +1 ) Fi ( xi +1 ) = 0

(34.49)

In LS-DYNA, s is not allowed to exceed unity. In a system which is softening such as that in
Figure 34.2 s would be greater than 1 so the line search is skipped. A line search is performed if
the stiffness of the structure is increasing, i.e., s is less than unity, also See Figure 34.3:
0 < s <1

(34.50)

Since the line search involves additional right hand side evaluations it is very expensive.
Improvements in the convergence behavior, however, more than justifiy the expense. See Figure
34.4.
The implementation of Broyden's method illustrates how the quasi-Newton method
works. The inverse of K i1 for the ith iterate is:
K i1 = ( I + i 1wi 1ui 1 ) ... ( I + 1w1u1t )( I + 0 w0 u0t ) K

where

34.8

1
0

(34.51)

LS-DYNA Theory Manual

Implicit
wi = ui K

1
i

Qi + K

1
i

Qi +1

(34.52)

i =

1
u ( ui wi )
t
i

are the update factors. Broydens algorithm can be outlined as follows:


1.

Solve d = K 01Qi

2.

Do

i = 1, k 1

= u it1d
Recall
d = d + i 1wi 1

3.

wk 1 = ui 1 uk 1
s+d

4.

k 1 =

t
k 1

1
( uk 1 wk 1 )

= ukt 1d
5.

d k 1 = d + k 1wk 1

34.9

Implicit

LS-DYNA Theory Manual

Figure 34.2. The sign change indicates a hardening system and the need for a line search
(0 < s <1).

Figure 34.3. Line search with linear interpolation.

34.10

LS-DYNA Theory Manual

Implicit

In the implementation note that:


The inverse matrices are not computed in full form
The solution is restricted to one backward substitution plus several vector
operations for each update.
The update parameters and update vectors, two per iteration, are stored in
core.
When the automatic time step option is turned the following steps are taken if convergence fails:
DT is set to DTMIN
LS-DYNA backs up to beginning of step
Solution is retried
If convergence again fails, solution is retried with a larger step
And so on.
An optimal number of iteration, IOPT, per step is chosen. ILS is the number of iterations during
the last increment. If
ILS < IOPT

DT = DT + ( DTMAX DT ) Min .50, ( IOPT / ILS ) 2 1.

(34.53)

ILS > IOPT


DT = DT ( DT DTMIN ) 1. ( IOPT / ILS ) 2

(34.54)

where DT, DTMAX, and DTMIN are the current step size, the maximum step size permitted,
and the minimum step size permitted, respectively.
34.1 Implicit Dynamics
For the implicit structural dynamics problem, the linearized equilibrium equations may
be written in the form

where

Mu&&n +1 + Du& n +1 + K t ( x n ) u0 = P ( x n )
%%
%%
% %
% %
%

M
%

n +1

F ( xn ) ,
% %

= mass matrix

D = damping matrix
%
u n +1 = x n +1 X
%
%
%

= nodal displacement vector

u& n +1 = nodal point velocities at time n+1


%
u&&n +1 = nodal point accelerations at time n+1.
%
34.11

(34.55)

Implicit

LS-DYNA Theory Manual

Equation (34.55) is solved by the unconditionally stable, one-step, Newmark- time


integration scheme
u
u&&n
1 1
n
u&& n +1 =

%2
% u&& ,
%
t t 2
%

(34.56a)

u& n +1 = u& n + t (1 ) u&& n + tu&& n +1


%
%
%
%

(34.56b)

x n +1 = x n + u
%
%
%

(34.56c)

Here, t is the time step size, and and are the free parameters of integration. For = 12 and
=14 the method becomes the trapezoidal rule and is energy conserving. If
1
2

> ,
(34.57)

>

1 1

+
4 2

numerical damping is induced into the solution leading to a loss of energy and momentum.
Substitution of Equations (34.56) into Equation (34.55) leads to an equation that is
similar in form to Equation (34.1):
K * u0 = P ( x n )
%
% %
%

n +1

F * ( xn )
% %

(34.58)

where
1

K * = Kt
M+
D,
2
%
% t % t %

u& n



1 1

F * = F M % + u&&n D 1 u& n + t
1 u&&n
%
%
% t 2
%
%
%
2
%

(34.59a)

(34.59b)

The solution of Equation (34.58) yields u0 , the displacement, velocity, and acceleration vectors
%
are updated
u
u& n 1
u&&1n +1 = % 02 %
%
t
t

1
n
,
u&&
2
%

(34.60a)

u& 1n +1 = u& n + t (1 ) u&& n + t u&&1n +1 ,


%
%
%
%

(34.60b)

34.12

LS-DYNA Theory Manual

Implicit

x 1n +1 = x n + u0 ,
%
%
%

(34.60c)

and equilibrium iterations begin with

where

K * ui = P ( xin +1 )
%
%
%

F * ( xin +1 )
% %

(34.61)

F * = F ( xin +1 ) + Mu&&in +1 + Du&in +1


%
% %
%%
%%

(34.62)

n +1

At this point the method is essentially the same as the static algorithm.

34.13

Вам также может понравиться