Exceptional service in the national interest

National
Laboratories

Estimation of the Critical Time Step

for Peridynamic Models
12th U.S. National Congress on Computational Mechanics
22 July 2013

Jesse Thomas
David Littlewood
Timothy Shelton
SAND2013-XXXXX
U.3. DEPARTMENT OF

ENERGY

Vi'

OSS
Sandia National Laboratories is a multi-program laboratory managed and operated by Sandia Corporation, a wholly owned subsidiary of Lockheed Martin
Corporation, for the U.S. Department of Energys National Nuclear Security Administration under contract DE-AC04-94AL85000. SAND NO. 2011-XXXXP

Advancing Computational Simulation with Peridynamics

Sandia
National
Laboratories

PERIDYNAMICS IS UNIQUE IN ITS ABILITY TO CAPTURE PERVASIVE MATERIAL

FAILURE

Potential to enable rigorous simulation of failure and fracture

Directly applicable to Sandia's national security missions

SIERRA CODE SUITE

Engineering mechanics simulation code suite supporting the

nation's nuclear weapons mission as well as other customers

Advanced Simulation and Computing (ASC) code

Peridynamic modeling for explicit transient dynamics (ETD)

IMPORTANCE OF THE ETD CRITICAL TIME STEP ESTIMATE

Necessary condition for stable simulation

Directly impacts computational expense

Sierra/SolidMechanics

Simulation of brittle fracture

Successful application of peridynamics for engineering analyses

requires an accurate estimate of the critical time step
2

Peridynamic Theory of Solid Mechanics

Sandia
National
Laboratories

Peridynamics is a mathematical theory that unifies the mechanics of

continuous media, cracks, and discrete particles

Peridynamics is a nonlocal extension of continuum mechanics

Remains valid in presence of discontinuities, including cracks

Balance of linear momentum is based on an integral equation

/9(x)ii(x, t)
Divergence of stress replaced with
integral of nonlocal forces.

Peridynamic bonds connect any two material points that interact directly
Peridynamic forces are determined by force states acting on bonds
A peridynamic body may be discretized by a finite number of elements
N

/?(x)iU(x,t) = ^2 {T[x> t] (x' -x) - T'[x',i] (x - x')} AFX/ +b(x,t)

i=0

S.A. Silling. Reformulation of elasticity theory for discontinuities and long-range forces. Journal of the Mechanics and Physics of Solids, 48:175-209, 2000.
Silling,S.A. and Lehoucq, R. B. PeridynamicTheory of Solid Mechanics. Advances in Applied Mechanics 44:73-168, 2010.

Sandia
National
Laboratories

Constitutive Models for Peridynamics

MATERIAL MODEL FORMULATION STRONGLY AFFECTS CRITICAL TIME STEP

Presence of multiple length scales differs from the classical (local) approach
Complex deformation modes possible within a nonlocal neighborhood
Material failure through the breaking of bonds may alter the stable time step
Definitions

Linear Peridynamic Solid 2

Microelastic Material1
Bond-based constitutive model
Pairwise forces are a function
of bond stretch
y-x

s = ---------

Magnitude of pairwise force

density given by
18fc
t=^s

State-based constitutive model

Deformation decomposed into deviatoric and
dilatational components

0 =
(cv x) e dV
mJn

ed = e _ =

bond vector

initial bond length

deformed bond length

bond stretch

bond extension

ed

deviatoric bond
extension

UJ_

influence function

volume

Magnitude of pairwise force density given by

6

3 kO
15/z
d
t =------oo x H---------- oo e
m
m

neighborhod
weighted volume
dilatation

horizon

bulk modulus

shear modulus

pairwise force
density

1.

S.A. Silling. Reformulation of elasticity theory for discontinuities and long-range forces. Journalof the Mechanics and Physics of Solids, 48:175-209, 2000.

2.

S.A. Silling, M. Epton, 0. Weckner, J. Xu, and E. Askari, Peridynamic states and constitutive modeling, Journal of Elasticity, 88, 2007.

Sandia

, ,

SSL

WRAPPER APPROACH RESULTS IN A NON-ORDINARY STATE-BASED MATERIAL MODEL 1

Approximate deformation gradient based on initial and current locations of

material points in family
Approximate Deformation Gradient

F = (Y*X)K_1

Definitions

Shape Tensor

K = X*X

reference position
vector state

deformation
vector state

shape tensor

Kinematic data passed to classical material model

Classical material model computes stress

approximate
deformation gradient

bond

cv

Stress converted to pairwise force density

influence function
Piola stress

T (>=(> ffK-1*

Suppression of zero-energy modes (optional)2

1.

S.A. Silling, M. Epton, 0. Weckner, J. Xu, and E. Askari, Peridynamic states and constitutive modeling, Journal of Elasticity, 88, 2007.

2.

Littlewood, D. A Nonlocal Approach to Modeling Crack Nucleation in AA 7075-T651. Proceedings of the ASME 2011 International Mechanical Engineering
Congress and Exposition, Denver, Colorado, 2011.

Candidate Approaches for Estimating the Critical Time Step ITJKS-k

INVESTIGATE MULTIPLE APPROACHES

Courant-Friedrichs-Lewy (CFL) condition 1

Approach of Silling and Askari for microelastic materials (von Neumann analysis)2 3

Generalized Silling and Askari approach incorporating bond angles

Global estimate using the Lanczos method 1>s

Largest eigenvalue of 3x3 nodal stiffness matrix

MEASURES OF SUCCESS

Accuracy of estimate

Computational expense

STRATEGY FOR ASSESSING CRITICAL TIME STEP ESTIMATES

Evaluate via test simulations

Compare against empirical result

Stable time step determined by numerical experiment

1.

Hughes, T.J.R. The Finite Element Method: Linear Static and Dynamic Finite Element Analysis. Prentice-Hall, Inc., Englewood Cliffs, NJ, 1987.

2.

Silling, S.A. and Askari, E. A meshfree method based on the peridynamic model of solid mechanics. Computers and Structures 83:1526-1535, 2005.

3.

Koteras, J.R. and Lehoucq, R.B. Estimating the critical time-step in explicit dynamics using the Lanczos method. International Journal for Numerical
Methods in Engineering 69:2780-2788, 2007.

Courant-Friedrichs-Lewy (CFL) condition

Sandia
National
Laboratories

Common stability condition for explicit time integration schemes

Heuristic interpretation: size of time step must be less than the time for a wave to
pass to an adjacent grid point

c At
Ax

< 1

where c is the wave speed, At is the time step, and Ax is a characteristic length of
the discretization
Wave speed computed from bulk modulus and density
c

CFL limit is given by

At <

Ax

For peridynamic models, what is the characteristic length scale?

Investigate both the mesh spacing and the horizon

7

Lanczos Global Critical Time Step Estimate

Sandia
National
Laboratories

For numerical solutions (e.g., classical FEM, peridynamics) after spatial

discretization, linearization, and global assembly

Mii + Ku = f
Explicit transient dynamics: central difference time integration; lumped diagonal
mass matrix M; real, symmetric, and positive-definite tangent stiffness matrix K

Global critical time step

Atc

where X is the maximum eigenvalue, and x the associated eigenvector, of the

generalized eigenvalue problem

(K - AM) x = 0
This requires an efficient algorithm to find the maximum global eigenvalue.

Lanczos Global Critical Time Step Estimate (ct'd)

Sandia
National
Laboratories

Lanczos Iteration 1: Define A = M^K, |30= 0, q0= 0, b = arbitrary, and q1 = b/\ |b| |.
For n 1, 2, 3,...
v Aqn
Qn V
v V

fin

f3n-iqn-l

|| v||
V

The largest eigenvalue of the tri-diagonal symmetric matrix with an on the

diagonal and (3n on the off-diagonals converges to the largest eigenvalue of A as
n^oc.
Implementation Efficiency: Each iteration requires Aqn = (M^KJq^ M is diagonal,
and Kqn is simply the (linearized) internal force. Thus for component /:
[Aqn]i
1.

Trefethen, L.N. and Bau, D., III. Numerical Linear Algebra. SIAM 1997.

Sandia

SSL
One-dimensional analysis of Silling and Askari for microelastic material1

A tc

2P
EP vpCiip

where p is the density, p iterates over all bonds at node /, Vp is the volume
associated with each neighbor, and Cip is the micromodulus between nodes / and p

Multiple dimensions and nonlinear materials 1

where the f is the force function of each bond and rj is the relative displacement
of the nodes in the bond
The suggested multi-dimensional measure represents the worst case scenario of
all bonds being aligned.

1.

Silling, S.A. and Askari, E. A meshfree method based on the peridynamic model of solid mechanics. Computers and Structures 83:1526-1535, 2005.

10

Multidimensional Estimate of Critical Time Step for

Microelastic Material

Sandia
National
Laboratories

We investigate using the assembled bond structure in the critical time step
estimate
Atc

2P
Vp^ip

where the for the jk component of Cjp we have

[C'ip]

CipCLjCLk

where Cjp is the scalar linearized bond stiffness and Oj are the direction cosines
associated with the bond p orientation
The norm we use is the maximum principal stiffness, \.e., in three dimensions the
maximum eigenvalue of the assembled 3x3 nodal stiffness

11

Critical Time Step Estimate Based on Probed Nodal

Stiffness Matrix

Sandia
National
Laboratories

Linearized bond-based and state-based peridynamic equation of motion 1>2

pil = [ C (x, q) (u(q, t) u(x, t)) dVq + b(x, t)

where the two theories involve different mathematical objects but similar
formulations

Rewriting the equation of motion

pu = f C (x, q) u(q, t)dVq - P(x)u(x, t) + b(x, t)

where

P(x) = [ c (x, q) dyg

Jgg1 2

1.

S.A. Silling. Reformulation of elasticity theory for discontinuities and long-range forces. Journal of the Mechanics and Physics of Solids, 48:175-209, 2000.

2.

S.A. Silling. Linearized Theory of Peridynamic States, Journal of Elasticity, 99:85-111, 2010.

12

Critical Time Step Estimate Based on Probed Nodal

Stiffness Matrix (ct'd)

Sandia
National
Laboratories

P(x) is seen to be, after discretization, the denominator of the multi-dimensional

bond-based time step estimate

P(x) = [ C(x,q) dVq^J2VpCip

p

Given a displacement field such that at x0 the displacement is e and everywhere

else it is 0, we have the following interpretation 1
The value of the vector P(x0)e is therefore the force density (per unit volume)
at x0 required to displace x0 by the vector e, holding all other points fixed.

i.e. the internal force from a nodal probe at x0 will be P(x0)e.

Probing in all dimensions will allow recovery of P(x0), giving a bond- or state-based
version of the denominator in our multi-dimensional critical time step estimate.
We investigate the applicability of probing to obtain the nodal 3x3 stiffness and
subsequent time step estimate for both bond-based and state-based analyses.

1.

S.A. Silling. Linearized Theory of Peridynamic States, Journal of Elasticity, 99:85-111, 2010.

13

Sandia
National
Laboratories

Test Case: Elastic Wave Propagation

Investigate material models

Microelastic bond-based

Linear peridynamic solid state-based

Wrapped classical elastic model

Investigate critical time step estimates

Material Parameters
Density

7.8 g/cm3

Youngs Modulus

300.0 GRa

Empirical (numerical experiment)

Poissons Ratio

0.25

ID approach of Silling and Askari

Horizon

0.5075 cm

Generalized Silling and Askari

Element time step (3x3 stiffness probe)

Lanczos global estimate

Simulation
Bar Length

10.0 cm

Bar Width

1.0 cm

Initial Velocity

10.0 m/s

Time Step

0.48 ps

Fixed displacement in
longitudinal direction
Initial velocity in
longitudinal direction

---------------- >

Velocity (cm/s)

1500
i!20Q
iSQO

H400
I
0
14

Sandia
National
Laboratories

Microelastic Bond-Based Material Model

Silling and Askari
max. time step = 0.241 ps
max. kinetic energy = 3.51 J

(
CFL Limit (element size)
Time Step

Kinetic Energy

3.51 J

max. time step = 0.329 ps

max. kinetic energy = 3.51 J

f
Generalized Silling and Askari

3.51 J

max. time step = 0.414 ps

max. kinetic energy = 3.51 J

14.1 J

3.51 J

3.51 J

Nodal Stiffness Matrix

max. time step = 0.414 ps

max. kinetic energy = 3.51 J

(
Empirical Observation

max. time step = 0.499 ps

max. kinetic energy = 3.51 J

1.75e+299 J

Global Lanczos

CFL Limit (horizon)

max. time step = 1.00 ps
max. kinetic energy = unstable

Note: Lanczos estimate of critical time step

updated continuously throughout simulation

max. time step = 0.500 ps

max. kinetic energy = 3.51 J

J
15

Linear Peridynamic Solid State-Based Material Model

Sandia
National
Laboratories

Nodal Stiffness Matrix

Time Step

Kinetic Energy

0.1 /is

3.51 J

0.2

/US

3.51 J

0.3 /us

3.51 J

0.4 /us

NaN

0.5 /us

NaN

0.6 /us

NaN

0.7 /us

NaN

0.8 /us

NaN

0.9 /us

NaN

1.0 /us

NaN

max. time step = 0.314 ps

max. kinetic energy = 3.51 J

CFL Limit (element size)

max. time step = 0.329 ps

max. kinetic energy = 3.51 J

Empirical Observation
max. time step = 0.381 ps
max. kinetic energy = 3.51 J

Global Lanczos
max. time step = 0.381 ps
max. kinetic energy = 3.51 J

CFL Limit (horizon)

max. time step = 1.00 ps
max. kinetic energy = unstable
V

16

Sandia
National
Laboratories

Wrapped Classical Material Model

CFL Limit (element size)

max. time step = 0.329 ps
max. kinetic energy = 3.51 J

Time Step

Kinetic Energy

0.1 /is

3.51 J

0.2 /is

3.51 J

0.3

yltS

3.51 J

0.4 ns

3.51 J

0.5 /is

1.33e+135 J

0.6 iis

3.08e+151 J

max. time step = 0.490 ps

max. kinetic energy = 3.51 J
_______________ __________ /

0.7

yltS

1.73e+109 J

------------------------------ \

0.8

yUS

2.17e+124 J

0.9

yUS

2.66e+116 J

1.0 /is

4.86e+208 J

Empirical Observation
max. time step = 0.490 ps
max. kinetic energy = 3.51 J
Global Lanczos

Nodal Stiffness Matrix

max. time step = 0.498 ps

max. kinetic energy = unstable

CFL Limit (horizon)

max. time step = 1.00 ps
max. kinetic energy = unstable

17

The Lanczos Estimate is a Function of Material Deformation

Estimate of the critical time step varies over the course of the simulation

Choice of material model affects behavior of critical time step estimate

Sandia
National
Laboratories

Microelastic material model exhibits greatest degree of variation

Microelastic Material

Time (s)

Linear Peridynamic Solid

Time (s)

Wrapped Classical Model

Time (s)

18

Sandia
National
Laboratories

The Influence Function Affects the Critical Time Step

Choice of influence function affects

critical time step

Parabolic decay
influence function

Lanczos algorithm successfully detects

changes in critical time step
Observation: Influence function that
decays with increasing bond length
results in reduced critical time step

Distance from Node / Horizon

Constant
influence function

J
0

0.2

0.4

0.6

0.1

Distance from Node / Horizon

Peridynamic Linear Solid

Wrapped Classical Material Model

Parabolic decay
influence function

Constant
influence function

Parabolic decay
influence function

Constant
influence function

Max. Lanczos
time step

0.381 ps

0.434 ps

Max. Lanczos
time step

0.490 ps

0.549 ps

Empirical result

0.381 ps

0.434 ps

Empirical result

0.490 ps

0.549 ps

19

Sandia
National
Laboratories

Test Case: Fragmenting Ring

Investigate material models

Microelastic bond-based

Linear peridynamic solid state-based

Wrapped classical elastic model (nosb)

Investigate critical time step estimates

Empirical

ID approach of Silling and Askari

Generalized Silling and Askari

Element time step (3x3 stiffness probe)

Lanczos global estimate

Material Parameters
Density

7.8 g/cm3

Youngs Modulus

300.0 GPa

Poissons Ratio

0.25

Critical Stretch

0.01 cm/cm

Horizon

0.603 cm

Simulation
Ring Diameter

4.5 cm

Ring Width

1 cm

Initial Radial Velocity

200.0 m/s

20

Unstable Time Step Manifests as Increased Bond Failure

laJJESL

Simulation results for

microelastic material
Percentage of
Broken Bonds

Maximum Kinetic
Energy (t > 10 /rs)

0.01 //,s

44.3 %

3.83 kJ

0.1 n s

44.5 %

3.82 kJ

0.2

yUS

44.7 %

3.82 kJ

0.3

yltS

45.3 %

3.82 kJ

0.4

yUS

45.3 %

3.82 kJ

0.5

yUS

45.4 %

3.82 kJ

0.6

yUS

46.7 %

3.81 kJ

0.7

yUS

49.1 %

3.83 kJ

0.8

yUS

73.5 %

3.82 kJ

0.9

yUS

95.3 %

4.39 kJ

1.0

yUS

99.1 %

6.40 kJ

Time Step

Time step = 5.0 ps

46.7% of bonds broken

Time step = 7.5 ps

62.7 % of bonds broken

21

Microelastic Bond-Based
Material Model

Silling and Askari

Sandia
National
Laboratories

max. time step = 0.290 ps

percentage of broken bonds = 44.7 %
max. kinetic energy = 3.82 kJ
CFL Limit (element size)
max. time step = 0.395 ps
percentage of broken bonds = 45.3 %
max. kinetic energy = 3.51 J

Broken Bonds

Maximum Kinetic
Energy (t > 10 yUs)

44.3 %

3.83 kJ

0.1 ,us

44.5 %

3.82 kJ

0.2

yUS

44.7 %

3.82 kJ

0.3

yUS

45.3 %

3.82 kJ

0.4

yUS

45.3 %

3.82 kJ

0.5 /is

45.4 %

3.82 kJ

0.6

yUS

46.7 %

3.81 kJ

Global Lanczos

0.7

yUS

49.1 %

3.83 kJ

0.8

yUS

73.5 %

3.82 kJ

0.9

yUS

95.3 %

4.39 kJ

max. time step = 0.682 ps

percentage of broken bonds = 46.0 %
max. kinetic energy = 3.83 kJ

1.0

yUS

99.1 %

6.40 kJ

Percentage of
Time Step
0.01

yUS

Nodal Stiffness Matrix

max. time step = 0.485 ps
percentage of broken bonds = 45.4 %
max. kinetic energy = 3.82 J
Generalized Silling and Askari
max. time step = 0.486 ps
percentage of broken bonds = 45.4 %
max. kinetic energy = 3.82 kJ

Empirical Observation
max. time step = 0.707 ps
percentage of broken bonds = 50.0 %
max. kinetic energy = 3.83 kJ

Note: Empirical observation corresponds to

the largest constant time step that results in
no more than 50% bond failure

CFL Limit (horizon)

max. time step = 1.19 ps
percentage of broken bonds = 99.1 %
max. kinetic energy = unstable

22

Sandia

SSL
CFL Limit (element size)
max. time step = 0.395 ps
percentage of broken bonds = 41.3%
max. kinetic energy = 3.51 J
Percentage of
Time Step
0.01 /i s

Broken Bonds

Maximum Kinetic
Energy (t > 10 //s)
3.43 kJ

Global Lanczos
max. time step = 0.494 ps
percentage of broken bonds = 42.8 %
max. kinetic energy = 3.43 kJ
Nodal Stiffness Matrix
max. time step = 0.505 ps
percentage of broken bonds = 44.8 %
max. kinetic energy = 3.82 kJ
Empirical Observation
max. time step = 0.509 ps
percentage of broken bonds = 50.0 %
max. kinetic energy = 3.46 kJ
CFL Limit (horizon)
max. time step = 1.19 ps
percentage of broken bonds = 99.1 %
max. kinetic energy = unstable

23

Conclusions

Sandia
National
Laboratories

There are a number of potentially useful ways to estimate the Explicit Transient
Dynamics critical time step with peridynamics
Any estimate should be material-type dependent (or else excessively conservative)

The most accurate and reliable is the Lanczos method

Efficient implementation

Extremely accurate for all material types

Operates on the global linear system and is independent of underlying theory

May be applicable to mixed FEM-peridynamic analyses

Bond-based nodal estimates (ID or multi-dimensional) are accurate for bondbased materials

Probed nodal estimate is accurate for bond-based and ordinary state-based

materials but extremely expensive
Element size CEL limit is conservative but fails to account for material type
Horizon size CEL limit is non-conservative and fails to account for material type

1.

S. A. Silling. Linearized Theory of Peridynamic States, Journal of Elasticity, 99:85-111, 2010.

24

Sandia
National
Laboratories

Questions?

Jesse Thomas
[email protected]

Advanced Simulation and Computing (ASC)

http://www.sandia.gov/asc/

25