See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/287929162

Artificial Viscosity Methods forModelling Shock Wave Propagation

Article · June 2009

DOI: 10.1007/978-1-4419-0727-1_19

CITATIONS READS

5 763

2 authors:

James Campbell Rade Vignjevic

Brunel University London Brunel University London
61 PUBLICATIONS   1,095 CITATIONS    106 PUBLICATIONS   1,497 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

PhD Project View project

NetDeliver: Network on numerical modeling and simulation of anticancer drug delivery in liver View project

All content following this page was uploaded by Rade Vignjevic on 01 August 2019.

The user has requested enhancement of the downloaded file.

Artificial Viscosity Methods for Modelling Shock
Wave Propagation

James Campbell and Rade Vignjevic

Abstract The paper gives an overview of the artificial viscoity method widely used
today to alow the simulation of problems containg shock waves. The development of
the most common basic form of the viscosity term is summarised and its behaviour
is illustrated through simulations of a 1D piston problem. Test problems that are
commonly used to test different viscosity formulations are then discussed to further
illustrate the method. Finally other shock viscosity forms such as edge and tensor
viscosities are briefly discussed.

1 Introduction

The requirement to model shock wave propagation has been around since the earliest
days of hydrocodes, with finite difference simulations used at Los Alamos during the
Manhattan project in order to study the behaviour of shock waves. An understanding
of shock propagation was critical for the design of the atomic bomb.
A shock wave has a thickness of the order of a few molecular mean free paths.
This is a very small dimension, much smaller than the typical length scale consid-
ered in continuum mechanics simulations. It is completely impractical to consider
modelling a macroscopic problem with a mesh size small enough to resolve this.
The result is that a shock represents a discontinuity in the solution, a surface over
which there is a jump in velocity, pressure, density and energy.
One possible approach is to treat the shock as a interface between two regions of
the flow. In this approach the Hugoniot equations along with the material’s Equation

James Campbell
Cranfield University, School of Engineering - Applied Mechanics, Cranfield, Bedford, UK, e-mail:
[email protected]
Rade Vignjevic
Cranfield University, School of Engineering - Applied Mechanics, Cranfield, Bedford, UK, e-mail:
[email protected]

1
2 James Campbell and Rade Vignjevic

of State (EOS) can be used to solve for the jump in solution variables across the
shock. These values are then applied as boundary conditions to the two regions of
flow. This approach can be used in one dimensional simulations, but in two- or
three- dimensional simulations it is impractical as the shock represents a moving
interface of potentially complex shape. Tracking this potentially arbitrary boundary
accurately, and applying the necessary jump conditions over the interface represents
a complex numerical and algorithmic challenge.
Two well established methods exist today for the treatment of shocks within nu-
merical simulations:
• Artificial viscosity
• Godunov’s method
The artificial viscosity concept, developed by von Neumann and Richtmyer [1],
allowed the first practical numerical simulations of problems containing strong
shocks. The concept involves the introduction into the numerical scheme of a viscos-
ity like term that acts to spread the thickness of any shock wave over several mesh
cells. As the shock no longer represents a discontinuity in the solution, standard
numerical methods can be used to simulate the shock propagation. This approach
has proved to be simple and robust, accounting for its continuing use today. The
drawbacks of the method is that it can introduce an unacceptable degree of mesh
sensitivity into the solution and there is a trade-off between shock thickness and
degree of oscillation behind the shock.
Godunov’s method [2] and the class of numerical methods developed from it rep-
resent an entirely different approach. In summary Godunov’s method assumes that
all solution variables within a cell are constant at the start of a step, with discontinu-
ities occurring at the edges of a cell treated through the solution of a Riemann prob-
lem. This approach allows the shocks in the solution to be physically and naturally
treated. The drawback is the high numerical cost of solving the Riemann problem.
This has required the use of approximate Riemann solvers which introduce fur-
ther approximations into the numerical method. Even approximate solutions of the
Riemann problem remain expensive when complex equations of state are required,
which has effectively limited the common application of the methods to fluid me-
chanics simulations. In addition, like the artificial viscosity method, it introduces
mesh sensitivity into the solution. Further information on Godunov’s method and
Riemann solvers can be found in Toro [3].

2 The Von Neumann - Richtmyer viscosity

The governing equations of Lagrangian hydrodynamics are the momentum (1), the
energy (2) and the continuity (3) equations:

dv
ρ = ∇·σ , (1)
dt
Artificial Viscosity Methods 3

de
ρ =σ :D, (2)
dt
dρ
= −ρ ∇ · v (3)
dt
where ρ is density, σ is the stress tensor, v is the velocity vector, e is the specific
internal energy. D is the rate-of-deformation tensor defined as the skew symmetric
part of L, the velocity gradient tensor:
1
L = (∇v)T and D = (L + LT ) (4)
2
The approach taken by Von Neumann and Richtmyer [1] for one dimensional
shock wave calculations was to modify the momentum and energy equations by
adding a dissipative, viscosity like, term q to the stress tensor. Introducing this term
into the governing equation acts to smear out the shock so as to produce a thick-
ness of the order of the resolution length of the computational mesh. A particular
feature of this approach is that the term is added to the equations throughout the
computational domain, not just where a shock is present, removing the need to track
shocks. As the dissipative term is added for purely mathematical reasons, it can be
any function that satisfies the following constraints [1]:
1. The modified conservation equations (1-3) must possess solutions without dis-
continuities.
2. The thickness of a shock must everywhere be of the order of the resolution length
of the mesh, independent of the strength of the shock.
3. The dissipative term must be negligible outside of the shock wave.
4. The Hugoniot equations must hold when all other dimensions are large compared
to the shock thickness.
The expression proposed by Von Neumann and Richmyer for their viscosity is
written for the one-dimensional case as
 
2 ∂v ∂v
 
q = −ρ (c∆ x) , (5)
∂x ∂x

where c is a dimensionless constant. The viscous term q is then included in the

solution by replacing the stress, σ , in the governing equations by (σ −qI). This term
is quadratic in the velocity gradient and is positive in compression and negative in
tension. As the viscosity term is not required in expansion it is common to set q = 0
when ∂∂ xv > 0.
The Von Neumann-Richtmyer q shown in equation (5) does work effectively,
and with a value of c of the order of 2 will spread a shock over three to five mesh
cells [4]. Its disadvantage is that any oscillation that does occur behind the shock
is only slowly damped out and lowering the value of c to reduce the thickness of
the shock results in an overshoot that produces oscillation. To address this problem
Landshoff [5] proposed an expression for q that was linear in the velocity gradient.
This term vanishes less rapidly behind a shock front and so rapidly damps out the
4 James Campbell and Rade Vignjevic

oscillation, but as the term is smaller in the shock front itself much larger overshoots
occur. Landshoff recommended that the two terms be combined to produce a q that
combines the best features of each:
   
∂v ∂v ∂v
q = −ρ cL ∆ x a   − ρ cQ (∆ x)2   , (6)
∂x ∂x ∂x

where a is the local speed of sound and cL and cQ are dimensionless constants that
multiply the linear and quadratic terms respectively. This basic form of viscosity
proved effective and is still widely used today.
Extending the one-dimensional form of the viscosity (6) to two or three dimen-
sions requires appropriate definitions of the velocity gradient and the characteristic
length. It is common to follow the original idea of Von Neumann and Richtmyer
[1] by replacing the velocity gradient term with the trace of the rate-of-deformation
tensor:
q = −ρ cL l a Ḋkk − ρ cQ l 2 Ḋ2kk if Ḋkk < 0
(7)
q=0 if Ḋkk ≥ 0
The definition of the characteristic grid length, l, is not so simple. Ideally the char-
acteristic length used would be the element thickness in the shock propagation di-
rection, but calculating this for all elements every step is difficult and costly.√The
3
usual approximation used in three dimensions is the √ cube root of the volume, V ,
and in two dimensions the square root of the area, A . These values are simple and
quick to calculate and provide a good estimate of the critical length provided the
aspect ratios of the elements are close to one. As the element aspect ratio becomes
poor the use of these estimates leads to increasing unphysical behaviour and even
numerical problems.

2.1 Demonstration

The behaviour of the von Neuman-Richtmyer viscosity will now be illustrated

through simulations of a 1D piston problem. In this problem a box with an initial
length of 1 is filled with a cold perfect gas with γ = 5/3 and initial density ρ0 = 1.0.
The right hand end of the box is fixed, while the left hand end is a piston that moves
into the box with a fixed velocity of 1.0. A shock of infinite strength (P0 = 0) is
generated by the piston motion and moves into the gas ahead of the piston, figure 1.
An exact solution to this problem can be obtained using the Hugoniot relations
that connect the state ahead of the shock with the state behind:
(v1 − v0 )ρ1
vs = , (8)
ρ1 − ρ0

P1 − P0 = ρ0 vs (v1 − v0 ), (9)
Artificial Viscosity Methods 5

Fig. 1 Diagram of the 1D piston problem, showing the state of the gas ahead and behind the shock.

 
P1 + P0 1 1
e1 − e0 = − . (10)
2 ρ0 ρ1
These equations link the density, ρ , pressure, P, specific internal energy, e, and
U1 = 4travelling with
particle velocity, v, across a shock vs velocity U0 state
1.3333 vs . The = 1 ahead of the
shock is denoted with subscripte1 =0 and
0.5 behind the shock with subscript
e0 = 0 1.
P1 = for
To calculate the exact solution 1.3333 P0 = is
this problem a fourth equation 0 required, the
perfect gas equation of state:v1 = 1 v0 = 0
P = (γ − 1)ρ e. (11)
Piston Shock
These equations can now be solved explicitly to derive an expression for the pressure
jump across the shock in terms of the conditions ahead of the shock and the change
in particle velocity across the shock:
s 
(γ + 1) γ +1 2
P1 = P0 + ρ0 (∆ v) + ρ0 |∆ v|
2
(∆ v)2 + a20 , (12)
4 4

where a is the local speed of sound given by

s
γP
a= (13)
ρ

The conditions ahead and behind the shock and the shock speed are given in 1.
Figures 2 to 6 show results from several simulations of this problem with vary-
ing cL and cQ . All the simulations were performed using the DYNA Lagrangian hy-
drocode [6], using the standard form of viscosity given by equation (7). The model
consists of 100 elements along the length of the box. Eight node hexahedral con-
tinuum elements were used, along with symmetry boundary conditions to enforce a
state of√
uni-axial strain. All the elements are perfect cubes at the start of the problem
so l = 3 V is a good measurement of the critical length. All the results are shown
as plots of element pressure vs. element coordinate at time t = 0.7. At this point the
piston coordinate 0.7 and the shock coordinate is 0.93̇ .
The first simulation, figure 2, shows a result using the quadratic term only with
a small coefficient (cL = 0.0 and cQ = 0.5). The overshoot and oscillation behind
the shock can be clearly seen and is only slowly damped. In a practical analysis
the level of noise in this solution would not be acceptable. Increasing the value of
the quadratic coefficient reduces the level of overshoot and hence the oscillation
6 James Campbell and Rade Vignjevic
1.6667

1.3333

1.0000

Pressure
0.6667

0.3333

0.0000
0.7 0.8 0.9 1.0
Coordinate

Fig. 2 Pressure profile for piston problem at time t = 0.7, with cL = 0.0 and cQ = 0.5.

1.6667

1.3333

1.0000
Pressure

0.6667

0.3333

0.0000
0.7 0.8 0.9 1.0
Coordinate

Fig. 3 Pressure profile for piston problem at time t = 0.7, with cL = 0.0 and cQ = 2.0.

1.6667

1.3333

1.0000
Pressure

0.6667

0.3333

0.0000
0.7 0.8 0.9 1.0
Coordinate

Fig. 4 Pressure profile for piston problem at time t = 0.7, with cL = 0.1 and cQ = 0.5.

as can be seen in figure 3 (cL = 0.0 and cQ = 2.0). Here the solution behind the
shock is good, but at the cost of a wider shock front. The effect of introducing the
linear term is to more rapidly damp out the oscillations behind the shock as can
be seen in figure 4 (cL = 0.1 and cQ = 0.5). Here even a fairly small value for the
linear coefficient results in a much more rapid damping behind the shock with only
a small increase in the width. Increasing the linear coefficient eventually results in
Artificial Viscosity Methods 7
1.6667

1.3333

1.0000

Pressure
0.6667

0.3333

0.0000
0.7 0.8 0.9 1.0
Coordinate

Fig. 5 Pressure profile for piston problem at time t = 0.7, with cL = 1.0 and cQ = 0.5.

a monotonic profile with no overshoot behind the shock but again at the cost of a
wider front, figure 5.
Today, in a 1D calculation like this, the width of the shock front is not a problem
as the computational cost of each simulation is negligible and a high spatial resolu-
tion can be used. In 3D models it is still desirable to keep the width of the shock as
small as practical as increasing the spatial resolution can easily result in models that
are too computationally expensive to use.

1.6667

1.3333

1.0000
Pressure

0.6667

0.3333

0.0000
0.7 0.8 0.9 1.0
Coordinate

Fig. 6 Pressure profile for piston problem at time t = 0.7, with cL = 0.06 and cQ = 1.5.

There is a further and more serious problem with artificial viscosity that must
be considered when selecting the values of the two coefficients. This problem was
studied by Noh who called it shockless Q heating [7]. The viscosity term is non-zero
everywhere where compression occurs and this leads to error in the solution where
the assumption that the dissipative term is negligable outside the shock wave does
not hold. Due to this error, as the linear term is larger than the quadratic term when
Ḋkk is small, it is common to keep the value of the linear coefficient small. The
default values used in the DYNA code are cL = 0.06 and cQ = 1.5 [6], experience
has shown that these values provide a reasonable balance between shock thickness
and oscillation behind the shock for many cases. Results for the piston problem
using these coefficients are shown in figure 6. The value of the quadratic coefficient
8 James Campbell and Rade Vignjevic

results in a reasonably sharp shock front with a small overshoot and small linear
term damps out the oscillation behind the shock.

2.2 Wall Heating

This is an error that occurs on shock formation, for example at the start of the piston
problem. It manifests in the solution as region where the density is underestimated
and the internal energy overestimated. While this problem has been known and stud-
ied since the early days of shock computations, the now common name wall heating
comes from the study by Noh [7]. He showed that this error is unavoidable as it is
present in the solution of the governing differential equations contining a q term.
The presence of this error is often ignored as often it only affects the solution in a
small region and does not threaten the overall stability of the calculation. However
in some particular applications, such as when the shock is generated at the centre
of convergent geometry, the error can be significant and so solutions have been pro-
posed, an example is the artificial heat flux term developed by Noh [7] to smear out
this error. This error is still investigated; see for example Rider [8] for a more recent
study.

3 Test problems for shock viscosity formulations

Since its original development many different forms for the artificial viscosity term
have been proposed. The motivation has included improving the solution near the
shock and in particular reducing the mesh sensitivity introduced when the original 1-
D formulation has been extended to 2- and 3- dimensions. In principle any problem
involving shock propagation can be used to investigate the properties of a shock
viscosity, but one result of the continued development has been the emergence of
certain test problems that are more commonly used to test or illustrate different
formulations. Three problems will now be considered in more detail: the Sod shock
tube [9], the Noh problem [7] and the Saltzman piston problem [11]. Other test
problems that are used include the Sedov blast wave [10], uniform compression [7]
and the Coggeshall adiabatic ompression probelm [10].

3.1 Sod shock tube

In addition to the piston problem used previously another commonly used 1D prob-
lem is the Sod shock tube problem, named after Gary Sod who used this problem to
investigate the performance of several numerical methods [9]. This problem consists
of two regions of perfect gas with different initial densities and pressures, figure 7.
Artificial Viscosity Methods 9

In both regions the gas is initially at rest. The solution consists of a rarefaction wave
that travels to the left, a contact discontinuity and a shock that travels to the right.
In both regions the ratio of specific heats is γ = 1.4. It is common to take the initial
position of the contact discontinuity as x = 0.5. The exact solution for this problem
at t = 0.25 is shown in figure 8. As with the piston problem this can be used to
investigate the effect of the viscosity formulation on the shock front, but in addi-
tion the viscosity should not affect the solution at the contact discontinuity or in the
rarefaction wave.

Left Right
UL = 1.0 UR = 0.125
PL = 1.0 PR = 0.1
vL = 0.0 vR = 0.0

Fig. 7 Initial conditions of the Sod shock tube problem.

1.25 1.25

1.00 1.00
Velocity
Density

0.75 0.75

0.50 0.50

0.25 0.25

0.00 0.00
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
X X

1.25 3.00

1.00 2.75
Pressure

Energy

0.75 2.50

0.50 2.25

0.25 2.00

0.00 1.75
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
X X

Fig. 8 Exact solution for Sod shock tube problem showing density, velocity, pressure and energy
at time t = 0.25.

The Sod shock tube problem is an example of a 1D Riemann problem for the
Euler equations which are commonly used for testing fluid dynamics codes [3].

3.2 Noh generic constant velocity shock

This problem, usually just called the Noh problem, has become a widely used test
for hydrocodes since it was first described by Noh [7]. While it is a 1D problem there
10 James Campbell and Rade Vignjevic

are three variants, one in planar geometry, one in axi-symmetric geometry and the
final in spherically symmetric geometry. In all three cases the problem consists of a
region of cold ideal gas, γ = 5/3, with ρ0 = 1.0 and P0 = 0.0. The initial velocity
v0 = −1.0 everywhere. In all three cases the shock speed is vs = 1/3, and all have
constant post-shock conditions.

20
64

16

48
12
Density

Density
32
8

16
4

0 0
0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4
R R

Fig. 9 The exact solution for density at t = 0.6 for Noh’s axi-symmetric problem (left) and spher-
ical problem (right).

The planar version is identical to the piston problem considered previously; the
change is that is in a frame of reference where the piston is at rest. Exact results for
the density at t = 0.2 is shown in figure 9 for the axi-symmetric and spherical cases.
In both of these cases there is a region of uniform compression ahead of the shock,
and especially with the spherical case the shockless Q heating error can be seen in
this region.
Noh originally developed these problems to investigate the wall heating error. An
example of the wall heating error can be seen in figure 10, showing results for the
axi-symmetric problem calculated on a polar mesh. The effect of wall heating can be
seen in the significant under-estimate of the density near the point of convergence
in the numerical solution. The results shown use the edge viscosity formulation
developed by Caramana [10].

20

16

12
Density

8
Numerical
Exact
4

0
0.0 0.1 0.2 0.3
R

Fig. 10 Example axi-symmetric Noh problem results on a polar mesh (left) showing the density
error due to wall heating (right). Solution time t = 0.2.
Artificial Viscosity Methods 11

Another application of the Noh problem is to investigate the effect of mesh sensi-
tivity when the shock propagates through a non-uniform mesh. The example shown
in figure 10 used a polar mesh where the element edges are aligned with the flow
and the mesh reflects the symmetry of the flow. For practical reasons it is rare to
achieve this and using a rectangular mesh where all elements are initially square for
the Noh problem can be used to investigate the consequences.
Figure 11 shows the results from a DYNA simulation of the axi-symmetric Noh
problem. The elements are initially uniform cubes with symmetry boundary con-
ditions used to enforce plane strain. The simulation used the standard viscosity (7)
with the default values for the viscosity coefficients. The mesh sensitivity of the so-
lution can be seen in both the mesh plot, especially along the 45 degree line, and in
the scatter of the density results. The Noh problem does represent a tough problem
for shock codes as it involves an infinitely strong shock and hence the level of mesh
sensitivity is severe, but it must be understood that mesh sensitivity is present in all
simulations that use the artificial viscosity method to capture the shock behaviour.

20

16

12
Density

4 Numerical
Exact
0
0.0 0.1 0.2 0.3
R
a b

Fig. 11 Simulation results for the axi-symmetric Noh problem on a Cartesian mesh using the Von
Neumann - Richtmyer viscosity at solution time t = 0.2. (a) Final mesh. (b) Plot of element density
vs. radius for all elements.

3.3 Saltzman piston

The Saltzman piston problem [11] is a problem that has been widely used to test
shock viscosities [10, 11, 12, 13, 14]. It tests the ability of a code to propagate a
one dimensional shock through a two dimensional mesh. The initial conditions and
analytical solution are identical to the piston problem considered earlier, however
the initial mesh is different, figure 12. The mesh fills a rectangular domain that is
1.0 long by 0.1 high with 100 elements along the long edge and 10 elements along
the short edge. The initial x coordinate of each node is defined in terms of their
logical coordinates i and j as
12 James Campbell and Rade Vignjevic
 
π (i − 1)
x(i, j) = (i − 1)dx + (11 − j) sin dy, (14)
100

where dx = dy = 0.01. The result of this is a distorted mesh, figure 12, although all
elements have an aspect ratio close to one.

Fig. 12 Initial mesh for Saltzman piston problem

The standard viscosity is unable to preserve the one-dimensional solution, with

the distortion of the mesh behind the shock clear in figure 13.

Fig. 13 Mesh at t = 0.7 for the Saltzman piston problem using the default DYNA viscosity.

The standard viscosity is unable to preserve the one-dimensional solution, with

the distortion of the mesh behind the shock clear in figure. However the numerical
solution remains a reasonable approximation to the analytical solution. Two variants
of the Saltzman piston have been proposed, both alter the height of the rectangular
domain leaving all other parameters unchanged. By altering the initial element as-
pect ratio both represent a tougher challenge for a viscosity than the standard test.
In the first variant the height of the domain is increased by a factor of 100. This vari-
ant was originally proposed by Margolin [11] to investigate how very high aspect
ratio elements affect the solution. The standard viscosity does not perform well on
this problem [12] as the characteristic length calculated from the element volume
Artificial Viscosity Methods 13

is much larger than the thickness of the element in the shock direction. In the sec-
ond variant the height of the domain is reduced by a factor of 4. This variant was
originally proposed by Campbell and Shashkov [14] to investigate how a moderate
element aspect ratio affects the solution and to provide a greater degree of discrimi-
nation between viscosity forms. A further option with the Saltzman piston problem
is to run the calculation past time t = 0.75. At this time the shock reaches the fixed
end of the piston and is reflected. The reflected shock now propagates through mesh
that has been distorted by the initial shock. Again this provides a greater challenge
to the analysis code.

4 Alternative forms of artificial viscosity

The problems and errors resulting from the use of artificial viscoity have lead to
the development of many different forms with varying properties. This section will
briefly discuss two basic forms: the edge centred and the tensor viscosities, illustrat-
ing each with an example.
All forms of viscosity should satisfy the four conditions set out by Von Neu-
mann and Richtmyer, see section 2 of this paper. More recently Caramana et al. [10]
specified five additional properties that an artificial viscosity should possess, these
are
1. Dissipativity: The artificial viscosity must only act to decrease kinetic energy.
2. Galilean invariance: The viscosity should vanish smoothly as the velocity field
becomes constant.
3. Self-similar motion invariance: The viscosity should vanish for uniform contrac-
tion and rigid rotation.
4. Wave-front invariance: The viscosity should have no effect along a wave front of
constant phase, on a grid aligned with the shock wave.
5. Viscous force continuity: The viscous force should go to zero continuously as
compression vanishes and remain zero for expansion.
A viscosity that satisfies these conditions will not suffer from the shockless Q heat-
ing error and should show reduced mesh sensitivity over the standard form.

4.1 Edge centred viscosity

In an edge centred viscosity the viscosity force is calculated at an element edge

rather than centred within an element. The strength of this approach is that the un-
certainty over the choice of characteristic length in higher dimensions is removed,
the length of the edge is now the natural choice. Each element edge connects two
nodes and the forces resulting from the viscosity are directly applied to them. These
forces should be applied in the direction of the relative velocity of the two nodes,
14 James Campbell and Rade Vignjevic

not along the line joining then. This change significantly improves the results and
according to Margolin [11] is known as the Barton fix.
The viscosity developed by Caramana et al. [10] is an example of a modern edge
viscosity that includes limiter terms. It is based on an alternative to equation (6) that
was investigated by Wilkins [4] who attributed it to Kurapatenko [15].
 s  
 (γ + 1) 2 
γ +1
qKur = ρ c2 |∆ v| + c22 (∆ v)2 + c21 c2s |∆ v| (15)
 4 4 

where c1 and c2 are non-dimensional constants, γ is the ratio of specific heats and
∆ v = ∂∂ xv ∆ x. This expression was derived from the pressure jump across a shock in
an ideal gas. Wilkins shows that when simulating an ideal gas, using this viscosity
removes the overshoot behind a shock.
In two dimensions the viscosity force for edge k of element e, that connects two
nodes b and c, is

d
 (1 − ψk )qKur (∆ vk · sec )∆cvk if (∆ vk · sec ) < 0
fk = (16)

0 if (∆ vk · sec ) ≥ 0 .

fk then contributes to the total force at points b and c. Vector sec is a unit vector in
the direction normal to the line connecting the mid-point of edge k to the centre of
element e. The velocity difference for the edge is ∆ vk = vb − vc , and ∆cvk is the unit
vector in the direction of this velocity difference. For the edge the density and sound
speed are
2ρb ρc
ρk = , cs,k = min(cs,b , cs,c ) . (17)
ρb + ρc
The density and sound speed at a node is the volume weighted average of the sur-
rounding elements. The function ψk is defined as

ψk = max[0, min(0.5(rl,k + rr,k ) , 2rl,k , 2rr,k , 1)] , (18)

 
∆ vk+1 · ∆cvk |∆ vk | ∆ vk−1 · ∆cvk |∆ vk |
rr,k = , rl,k = . (19)
∆ xk+1 · ∆cxk |∆ xk | ∆ xk−1 · ∆cxk |∆ xk |
Subscripts l and r refer to the left and right edges respectively. The left edge is
found by considering all other edges that connect to node b and selecting the edge
that forms the largest angle with edge k. The right edge is found in the same manner
by considering edges that connect to node c.
The term ψ can be considered as a multi-dimensional form of a one dimensional
TVD advection limiter. It acts to switch off the viscosity when the second derivative
of the velocity field is zero and ensures that the viscosity satisfies the conditions of
self-similar motion invariance and wave front invariance.
Artificial Viscosity Methods 15
0.1
0.09
0.08
0.07
0.06
0.05

Y
0.04
0.03
0.02
0.01
0
0.7 0.8 0.9 1
X

Fig. 14 Mesh at t = 0.7 for the Saltzman piston problem using the Caramana edge viscosity.

Results for the Saltzman piston problem using the Caramana edge viscosity are
shown in figure 14, and shows a smoother solution than that obtained using the
standard form, figure 13.

4.2 Tensor viscosity

In a tensor viscosity the scalar q term is replaced by a tensor Q. The potential benefit
of this change is that the effect of the viscosity can be directionally dependent,
like an edge viscosity, while still being element centred. In addition there is no
requirement that Q be symmetric, and a nonsymmetric viscosity can remove mode
conversion [11]. This means that for a shear flow in which all velocities are parallel,
the viscous force will only act in the velocity direction. With a symmetric tensor
viscosity the force would have a component perpendicular to the velocity direction.
The tensor viscosity developed by Campbell and Shashkov [14] is an example of
a modern tensor viscosity that includes limiter terms. It assumes a form similar to
physical viscosity, but based on L rather than D and so is not symmetric:

Q = µ LT . (20)

µ is a scalar coefficient defined as

 s  
 (γ + 1) 2 
γ +1
µ = (1 − ψ )ρ c2 |∆ v| + c22 (∆ v)2 + c21 c2s l (21)
 4 4 

where ψ is a limiter function similar to the function used in the Caramana edge
viscosity although using the value of ∆ v in four directions rather than two.
Results for the Saltzman piston problem using this tensor viscosity are shown
in figure 14, again showing a smoother solution than that obtained using the stan-
dard form. It should be noted that both the Caramana edge viscosity and this tensor
viscosity are formulated within the framework of mimetic finite difference meth-
ods although there is no reason why they could not be extended to other numerical
approaches.
16 James Campbell and Rade Vignjevic
0.1
0.09
0.08
0.07
0.06
0.05

Y
0.04
0.03
0.02
0.01
0
0.7 0.8 0.9 1
X

Fig. 15 Mesh at t = 0.7 for the Saltzman piston problem using the Campbell and Shashkov tensor
viscosity.

5 Summary

This paper discusses the artificial viscosity method for the simulation of shock
waves that is widely used today, especially for solid mechanics applications. The be-
haviour of the most widely available form of artificial viscosity is illustrated through
a set of examples covering test problems that are commonly used to test different
viscosity formulations. Finally other shock viscosity forms such as edge and tensor
viscosities are briefly discussed.

References

1. Von Neumann J., Richtmyer R.D. (1950) A method for the calculation of hydrodynamic
shocks. J. Appl. Phys. 21, 232-237.
2. Godunov S. K. (1959) A difference scheme for numerical computation of discontinuous so-
lutions of equations in fluid dynamics. Mat. Sb. 47, 271-306.
3. Toro E.F. (1999) Riemann Solvers and Numerical Methods for Fluid Dynamics. Springer-
Verlag Berlin Heidelberg
4. Wilkins M.L. (1980) Use of artificial viscosity in multidimensional fluid dynamic calcula-
tions. J. Comput. Phys. 36, 281-303.
5. Landshoff R. (1955) A numerical method for treating fluid flow in the presence of shocks.
LA-1930, Los Alamos National Laboratory.
6. Lin J.I. (2004) DYNA3D: A nonlinear, explicit, three-dimensional finite element code for solid
and structural mechanics. UCRL-MA-107254, Lawrence Livermore National Laboratory.
7. Noh W.F. (1987) Errors for calculations of strong shocks using an artificial viscosity and an
artificial heat flux. J. Comput. Phys. 72, 78-120.
8. Rider W.J. (2000) Revisiting wall heating. J. Comput. Phys. 162, 395-410.
9. Sod G.A. (1978) A survey of several finite difference methods for systems of nonlinear hy-
perbolic conservation laws. J. Comput. Phys. 27, 1-31.
10. Caramana E.J., Shashkov M.J., Whalen P.P. (1998) Formulations of artificial viscosity for
multi-dimensional shock wave computations. J. Comput. Phys. 144, 70-97.
11. Margolin L.G. (1988) A centered artificial viscosity for cells with large aspect ratios. UCRL-
53882, Lawrence Livermore National Laboratory.
12. Benson D.J. (1991) A new two-dimensional flux-limited shock viscosity for impact calcula-
tions. Comput. Methods Appl. Mech. Engrg. 93, 39-95.
13. Benson D.J., Schoenfeld S. (1993) A total variation diminishing shock viscosity. Comput.
Mech. 11, 107-121.
14. Campbell J.C., Shashkov M.J. (2001) A tensor viscosity using a mimetic finite difference
algorithm. J. Comput. Phys. 172, 739-765.
 Artificial Viscosity Methods 17

15. Kurapatenko V.F. (1967) In: Difference methods for solutions of problems of mathematical
physics, I (Editor: N.N. Janenko). Amer. Math. Soc., Providence, R.I., p. 116.

View publication stats