CWP-444

Numerical modeling of waves incident on slip

discontinuities

Matthew Haney and Roel Snieder

Center for Wave Phenomena, Colorado School of Mines, Golden, CO 80401

ABSTRACT
Interfaces in elastic media need not be in welded contact. For instance, fractures
allow a small amount of slip to occur along their surfaces during the passage
of a seismic wave. By slip, we mean that the displacement across the interface
can be discontinuous. Reflection and transmission of plane waves in this case
are frequency dependent.
Previous numerical studies have chosen the finite-difference method to simu-
late slip discontinuities. The approach suffers from difficulties in incorporating
boundary conditions into the strong form of the equations of motion. We show
in detail that only a finite-element formulation of the σ-v (stress and veloc-
ity) equations can overcome these problems. Numerical examples illustrate the
method for an SH-wave normally incident on a slip discontinuity and a P -wave
incident at an angle.
1 INTRODUCTION Boore, one of the pioneers of the finite-difference method
in seismology, summarized the problem by stating that
Full-waveform forward modeling in seismology has tra- in finite-difference schemes “the interface displacements
ditionally been dominated by finite-difference methods. must satisfy the continuity of displacement and stress,
Perhaps the intuitive appeal of such methods has led but are not explicitly required to satisfy the equation of
to their popularity. Kelly and Marfurt (1990), in re- motion” (Boore, 1970).
viewing the numerical literature from the exploration The motivation of this paper is to explore the ad-
geophysics community, cited at least four possible rea- vantages of the finite-element method for the modeling
sons for the relative neglect of other methods, specifi- of slip discontinuities. The slip discontinuity has been
cally finite-elements. We complement their list by stat- proposed by Schoenberg (1980) as an interface condi-
ing that finite-elements require additional numerical de- tion applicable to cracks. The boundary condition can
tail to implement properly. For instance, explicit finite- be thought of as a generalized interface condition since
difference algorithms do not require the inversion of a it supplies a parameter, the compliance η, that, over its
matrix. In contrast, all finite-element methods eventu- range of physical values, takes an elastic interface from
ally lead to a matrix equation that must be solved. a welded contact (η = 0) to a free surface (η = ∞).
Finite-elements offer more flexibility than finite- Mathematically, a slip discontinuity for an SH-wave is
differences with respect to the shape of the numerical expressed as
mesh; however, it is less widely accepted that finite-
elements are superior to finite-differences in the way u+ −
y − uy = ησyz , (1)
they account for boundary conditions. Such an advan- + −
σyz − σyz = 0, (2)
tage can easily be seen in the formula for integration
by parts: an integral is replaced by another integral and where (-) refers to the side of the interface on which
a term representing the values of the function at the the wave is incident, (+) the other side of the inter-
boundaries. This basic fact shows up in the formulation face, uy refers to the displacement normal to the plane
of finite-elements since they approximate the integrated, of propagation, and σyz is the shear stress. In this pa-
or weak, equations of motion. per, a slip discontinuity extends infinitely; we do not
Finite-differences, as approximations to the strong consider crack-tips. Physically, boundary conditions (1)
form of the equations of motion, require ad hoc tech- and (2) model the transmission and reflection of a thin,
niques to take into account, for instance, a welded low shear zone (Schoenberg, 1980). Faults generally fall
boundary between two different elastic media. David within this category.
2 M. Haney & R. Snieder
In this paper, we will argue three main points:
1. Finite-element modeling is needed to incorporate in-
terface boundary conditions into the equations of mo- 1
tion.
2. The equations of motion should be the first-order
σ-v (stress and velocity) system of partial differential
equations instead of the second-order wave equation in
u (displacement) to avoid any finite-difference-like ap-
proximations to spatial derivatives.
3. Only an implicit, unconditionally stable time in- z k−1 zk z k+1
tegration scheme can handle the zero-length elements Figure 1. The basis functions we use for finite-elements.
needed to accurately model a slip discontinuity.
We conclude this discussion of the finite-element Fig. 1 shows these piecewise linear functions graphically.
method with numerical examples of the reflection and The Galerkin formulation of the finite-element
transmission of plane SH- and P -waves at a slip discon- method seeks to minimize the weighted average of the
tinuity. error induced by the incompleteness of the finite set
of basis functions in equation (5) The essence of the
Galerkin method is that it weights the errors with the
2 A FINITE-ELEMENT SCHEME FOR THE same basis functions used in the approximation of the
1D SCALAR WAVE EQUATION displacement, u (Marfurt, 1984)
The beginning of any numerical study must be an anal- L 
∂2

N
ysis of the two competing limitations imposed on finite c2 2 αK (t)φK (z)−
systems: numerical dispersion and stability. An under- ∂z K=1
0
standing of these concepts should steer the subsequent 
∂2

N
implementation of a particular numerical scheme. Ide- αK (t)φK (z) − s φJ (z)dz = 0
ally, any numerical simulation should satisfy a stability ∂t2 K=1
criterion with the least amount of dispersion possible. for J=1,2,...,N . (6)
In this section, we present the most simple case: a
finite-element implementation of the scalar wave equa- Rearranging the order of the sums and integrals in this
tion in 1D. We choose this problem since, in 1D, the equation yields
 L 
essence of the finite-element scheme is not complicated

N  2
∂ φ K
by mesh generation or subsequent “assembly” of the 0 = −c 2
αK  φJ dz  +
mass or stiffness matrix. Assembly of these matrices is K=1
∂z 2
0
directly related to the notion of wrapping up, or order-  L 

N  L
ing, a higher-dimensional set of points into a 1D vector. ∂ 2 αK  
In 1D, this is not necessary. φ φ
K J dz + sφJ dz. (7)
K=1
∂t2
We consider the 1D scalar wave equation with 0 0

source term in a homogeneous medium Since the basis functions are piecewise linear over
their range, the first integral on the r.h.s. of equa-
∂2u ∂2u
c2 2
− 2 − s = 0, (3) tion (7), containing a second derivative of the basis func-
∂z ∂t
tion, must be altered (Haltiner and Williams, 1980). It
and assume Dirichlet (essential) boundary conditions at turns out that in altering this term by integrating by
z = 0 and z = L. The displacement, u, is approximated parts, boundary terms appear explicitly in the equation
by a finite series of spatial basis functions with coeffi- of motion
cients that depend on time L L
L 2
∂ φK ∂φK  ∂φK ∂φJ


N
φJ dz = φJ  − dz. (8)
u(z, t) = αK (t)φK (z). (4) ∂z 2 ∂z 0
∂z ∂z
0 0
K=1
To implement Dirichlet boundary conditions at z = 0
In this paper, the nodal basis will always be used. For
and z = L, the weighting function, φJ , is set to zero
a spatial discretization step h and a uniform partition
at the boundaries (Wang, 2000). Hence, the boundary
of the interval [0,L] into N + 1 subintervals, these basis
term in equation (8) disappears and the elements φ are
functions are mathematically defined as
 fixed to be zero on the boundaries. More complicated
 (z − zK−1 )/h if zK−1 ≤ z ≤ zK , boundary conditions, such as absorbing boundary con-
φK (z) = (zK+1 − z)/h if zK ≤ z ≤ zK+1 , (5) ditions, are not discussed here.
 Without the boundary term, inserting equation (8)
0 otherwise.
 Numerical modeling of waves incident on slip discontinuities 3
into equation (7) yields the weak form of the scalar wave example, we use an explicit time integration known as
equation central differences (Zienkiewicz and Taylor, 2000) to ap-
 L  proximate the time behavior of α 

N  
∂φ K ∂φJ
0=c 2
αK  dz  +  n+1 − 2
α αn + α n−1
∂z ∂z 0 = c2 S
αn + M (16)
K=1 0 ∆t2
 L 

N  L where the subscripts refer to the time step n + 1, to be
∂ 2 αK  
φK φJ dz + sφJ dz. (9) calculated, and the previous time steps n and n−1. The
∂t2
K=1 0 0 time discretization interval is shown as ∆t. Notice that
equation (16) is symmetric in the n + 1 and n − 1 terms.
The bracketed terms in equation (9) can be represented
This ensures that the resulting dispersion relation on
as matrices multiplying the vectors α  and α¨ . The ma-
the numerical grid is real-valued. Such a property is de-
¨ , is usually
trix multiplying the second time derivative, α
sirable since a complex dispersion relation would yield
called the mass matrix, M , and the matrix acting on α  is
solutions to the wave equation that grow or decay ex-
referred to as the stiffness matrix, S. From equation (5),
ponentially with distance. In contrast, real-valued solu-
the entries of these matrices can be calculated exactly.
tions do not grow exponentially, so long as they satisfy
For the mass matrix, MK,J , the non-zero entries are
a stability criterion.
L To calculate the stability criterion of this explicit
φK φJ dz = 2h/3 for K = J (10) time integration scheme, first use equations (10), (11),
0
(12), and (13) to carry out the matrix multiplications
in equation (16). What results is the so-called finite-
L element stencil
φK φJ dz = h/6 for K = J + 1, J − 1. (11) h 2h
0
0 = αm−1,n+1 + αm,n+1 +
6∆t2 3∆t2

This means that M is symmetric and tridiagonal. Simi- h h c2
α m+1,n+1 − + αm−1,n −
larly, the stiffness matrix, SK,J , is also symmetric tridi- 6∆t2 3∆t2 h
agonal and the non-zero entries are  
4h 2c2 h c2
− αm,n − + αm+1,n +
L 3∆t2 h 3∆t2 h
∂φK ∂φJ
dz = 2/h for K = J (12) h 2h
∂z ∂z αm−1,n−1 + αm−1,n−1 +
0 6∆t2 3∆t2
h
L αm−1,n−1 , (17)
∂φK ∂φJ 6∆t2
dz = −1/h for K = J + 1, J − 1. (13)
∂z ∂z where the first subscripts refer to gridpoints in space and
0
the second subscripts are still the time steps. To investi-
The finite-element scheme can now be written in matrix gate the stability of the central difference scheme, insert
form a harmonic function for α (Alterman and Loewenthal,
1970)
0 = c2 S ¨ .
α + Mα (14)
αm,n = eikmh ζ n ζ = eb∆t , z = mh, and t = n∆t, (18)
Note that the source term has been omitted in equa-
tion (14). In our numerical simulations, we use an ini- where the integers m and n multiply the spatial and
tial condition as a source instead of a forcing term in temporal interval lengths h and ∆t, and k is the
the equations of motion. To include a forcing term, s wavenumber.
has to be approximated in the nodal basis Inserting equation (18) into equation (17), and or-
ganizing terms yields


N

s(z, t) = sK (t)φK (z). (15) h
K=1 0 = (e−ikh + 4 + eikh )ζ 2
6∆t2
Substituting this into equation (9), it can be seen that c2 −ikh
the mass matrix multiplies the source vector. − (e − 2 + eikh )ζ
h
h
− (e−ikh + 4 + eikh )ζ
3∆t2 
3 STABILITY ANALYSIS OF AN h −ikh ikh
+ (e + 4 + e ) eikmh ζ n−1 . (19)
EXPLICIT FINITE-ELEMENT SCHEME 6∆t2
The finite-element approach has served to integrate the In order to have non-trivial solutions, the terms in-
above matrix equation in space, but a suitable time in- side the brackets in equation (19) must equal zero. The
tegration scheme has yet to be determined. As a first resulting equation contains all the information about
4 M. Haney & R. Snieder
the dispersion and stability properties of the numerical Above, we stated that equation (19) contains all the
scheme. Note that in higher dimensions, equation (19) information about the dispersion and stability proper-
is a matrix equation and the condition for non-trivial ties of the numerical scheme. We have the stability con-
solutions is that the determinant of the matrix equals dition - what about the dispersion relation? To obtain
zero. this, set b in equation (18) to iω so that ζ = eiω∆t .
Dividing equation (19) by the coefficient of the ζ 2 Substituting this into equation (19), setting the terms
term and setting the terms inside the brackets to zero in the brackets to zero, and simplifying a bit yields the
gives dispersion relation
 
2 6c2 ∆t2 e−ikh − 2 + eikh 3∆t2 c2 cos(kh) − 1
0=ζ − ζ − 2ζ + 1. (20) cos(ω∆t) = 1 +
h2
. (28)
h2 e−ikh + 4 + eikh cos(kh) + 2

Using the Euler formula and the relation cos(kh) − 1 = For wavelengths much larger than the grid-spacing, ω∆t
−2 sin2 (kh/2), we can simplify equation (20) and kh are small parameters. Keeping the lowest order
  2   terms in a series expansion of equation (28) gives the
2h
∆t2
− 4ch + 3∆t 4h
2 sin2 (kh/2) well-known relation
0=ζ −
2  ζ + 1. (21)
h 2h
− 3∆t 2
2 sin (kh/2)
ω 2 = c2 k2 . (29)
∆t2
In other words, the finite-element scheme should do an
For the numerical scheme to be stable, the roots of
excellent job propagating waves well-sampled by the nu-
equation (21) should have magnitudes less than or equal
merical grid.
to 1. This means that the solutions are not exponentially
growing in time. For a quadratic of the form
0 = x2 − zx + 1, (22) 4 ZERO-LENGTH ELEMENTS FOR
FRACTURES
the magnitude of the two roots, x1 and x2 , are both less
than or equal to 1 if The finite-element example explored above shows the
z essence of the numerical implementation of the weak
−1 ≤ ≤ 1. (23) form of the equations of motion. It is relatively easy
2
to code since all that is needed is the inversion of a
Inserting equation (21) into this expression yields tridiagonal matrix. Modeling interfaces in elastic media,
 2  though, introduces several complications.
2h
∆t2
− 4ch + 3∆t 4h
2 sin2 (kh/2) To model true boundaries, some of the elements
−1 ≤ h 2h 2
≤ 1. (24)
2( ∆t 2 − 3∆t2 sin (kh/2)) should have zero length. In this way, two points on
opposite sides of a crack, but infinitesimally close to
Scanning over all possible values of sin2 (kh/2), the one another, could be modeled precisely. Explicit finite-
strongest condition on the spatial and temporal dis- elements of the second-order wave equation fail this re-
cretizations h and ∆t occurs when sin2 (kh/2) = 1 quirement for two reasons. First, entries in the stiffness
(Alterman and Loewenthal, 1970). In this case, equa- matrix, equations (12) and (13), go to infinity as h goes
tion (24) becomes: to zero. Second, even if the stiffness matrix existed, the
2h 4c2 CFL-condition, equation (27), would demand that the
3∆t2
− h
−1 ≤ 2h
≤ 1. (25) time discretization ∆t go to zero if the length of any
3∆t2 element went to zero. This statement assumes that the
CFL-condition acts at each point in the computational
From equation (25), the right inequality:
domain for an irregular grid (which has been verified
2h 4c2 2h numerically).
− ≤ , (26)
3∆t2 h 3∆t2 In addition, the boundary conditions in elastic me-
dia concern the continuity or relationship of displace-
is trivially true. The left inequality provides a more
ment and stress at an interface. Discrete approximations
meaningful relation:
of the stress would have to do for an implementation of
2h 2h 4c2 the wave equation in displacement. By formulating a
− ≤ − ,
3∆t2 3∆t 2 h finite-element analysis for the σ-v (stress and velocity)
4c2 4h system of equations, the stress would be available in
≤ , the numerical scheme without approximation. Further-
h 3∆t2
√ more, it turns out that the mass and stiffness matrices
c∆t 3 ≤ h. (27)
in the σ-v formulation can accommodate elements of
This is the Courant-Friedrichs-Levy (CFL) stability zero length. The next section will explore this in more
condition for this implementation of explicit finite el- detail.
ements (Marfurt, 1984). The solution to the problem of the CFL condition
 Numerical modeling of waves incident on slip discontinuities 5
for an element of zero length is to use an implicit time the resulting weak form
integration scheme that is unconditionally stable. To il-  L 

N 
lustrate unconditional stability, consider this alternative ∂αK 
time scheme for equation (16) φK φJ dz  =
K=1
∂t
0
  L 
0 = c2 S
α
 n+1 + α
 n−1 
1

N
+ ∂φ J
2 − βK  φK dz  . (36)
 ρ K=1 ∂z
 n+1 − 2
α αn + α n−1 0
M , (30)
∆t2
In this equation, the mass matrix (multiplying α ˙ ) is
where now the value of α at the time n is replaced by a identical to the one for the scalar wave equation; how-
simple average of its values at times n+1 and n−1. Note ever, the “stiffness” matrix (multiplying β) has a dif-
again that this expression is symmetric in the terms at ferent form. Using equation (5), the entries of this new
n + 1 and n − 1 so that the dispersion relation is real- stiffness matrix, SK,J , can be calculated
valued. By going through the same steps outlined pre-
L
viously to obtain the stability of this time integration ∂φJ
scheme (which we call forward-backward time integra- φK dz = −1/2 for K = J + 1 (37)
∂z
tion), the CFL-like condition that results is 0

L
−2h2 ≤ 3c2 ∆t2 , (31) ∂φJ
φK dz = 1/2 for K = J − 1. (38)
∂z
which is always true. Hence, the forward-backward time 0

scheme is unconditionally stable. This sort of time

The matrix form of equation (36) is thus
scheme makes the modeling of zero-length elements pos-
sible. 1 
˙ = − S β.
Mα (39)
ρ
Note that the entries of this stiffness matrix are con-
stants, independent of the spatial discretization h. This
is in contrast to the expression for the second-order wave
equation. The independence of this stiffness matrix from
5 A FINITE-ELEMENT SCHEME FOR h allows zero length elements to be modeled in the finite-
STRESS-VELOCITY IN 1D element implementation of the σ-v system.
For the second equation in the σ-v system, equa-
In 1D, all three elastic wave types decouple since they
tion (33), we apply the Galerkin method but stop short
are normally incident on interfaces. Here, we focus the
of integrating by parts
analysis to the SH-wave. The system of partial differ-
 L 
ential equations for velocity and stress in this case is

N 
∂βK 
φK φJ dz 
∂vy 1 ∂σyz ∂t
− = 0, (32) K=1 0
∂t ρ ∂z  L 

N 
∂φ K
∂σyz
−µ
∂vy
= 0. (33) =µ αK  φJ dz  . (40)
∂t ∂z K=1
∂z
0

These equations are recognized as Newton’s and Hooke’s The term on the l.h.s. of equation (40) is similar to
Laws, respectively. The same nodal basis as above is what has been encountered up to this point. The new
used to approximate the velocity and stress challenge is to properly treat the first term on the r.h.s.,
since it contains information about slip discontinuities.


N
That is the subject of the next section.
vy (z, t) = αK (t)φK (z), (34)
K=1

N
σyz (z, t) = βK (t)φK (z), (35) 6 THE INCLUSION OF A SLIP
K=1
DISCONTINUITY
and Dirichlet boundary conditions are assumed to apply The first step in evaluating the term on the r.h.s. of
on the ends of some domain z = 0 to z = L. Applying equation (40) involves breaking the integral into two -
the Galerkin method to equation (32), as was done to one running up to the fault (or fracture), and the other
the scalar wave equation, and integrating by parts gives continuing on the other side of the fault. In the limit of
6 M. Haney & R. Snieder
f take the limit as f − → f + , creating a zero-length ele-
ment at the fault. The leftmost term of equation (43)
becomes


N
 
−µ αK φK (f + ) − φK (f − ) . (44)
K=1

The series in this equation is the slip in the nodal basis.

Taking the time-derivative of equation (1), we replace
− +
f f the terms inside the series with the nodal basis repre-
sentation of the time-derivative of the stress at the fault
multiplied by the compliance, η

∂

N
−µη βK φK (f ), (45)
∂t K=1

where f is located halfway between f − and f + (see

Fig. 2). Since we have put two nodes centered at f −
Figure 2. The limiting procedure we use in creating a slip and f + , the φK in equation (45) are all zero except for
discontinuity. The fault, f , is placed between two nodes the two that straddle the fault. When evaluated at f ,
whose positions eventually coincide with each other. both of these nodal basis functions are equal to 1/2,
making the previous equation a simple average
the fault having no thickness (a slip discontinuity) µη ∂  f − +

 L  − β + βf , (46)
 2 ∂t

N
∂φ K
µ αK  φJ dz  = where the superscripts f − and f + denote that these
K=1
∂z
0 terms are non-zero for only the two nodes nearest the
 −  fault. In taking the limit as the fault thickness goes to

N f L
 ∂φK ∂φK  zero, these two nodes represent displacement and stress
µ αK  φJ dz + φJ dz  . (41)
∂z ∂z infinitely close to, but on opposite sides of, a slip dis-
K=1 0 f+ continuity.
After integrating by parts and combining the resulting Substituting equation (46) into equations (43)
two integrals back into a single integral from z = 0 to and (40), we obtain the weak form of the second equa-
z = L, we obtain for the r.h.s. of equation (41) tion in the σ-v system
 L 
 

µη ∂  f − 
N
−
N
∂βK  +
µ αK φK φJ |f0 + φK φJ |L
f+ − φK φJ dz  = − β + βf
K=1 K=1
∂t 2 ∂t
 0
L L
∂φJ 

N
∂φJ
φK dz . (42) −µ αK φK dz. (47)
∂z ∂z
0 K=1 0
For simplicity at this point, assume Dirichlet bound-
Putting this equation into matrix notation using the
aries at z = 0 and z = L, and distribute the sums
mass and stiffness matrices, the finite-element system


N
  in the σ-v formulation is
µ αK φK (f − )φJ (f − ) − φK (f + )φJ (f + ) −
1 
K=1 ˙ = − S β,
Mα (48)
ρ


N L
µ αK φK
∂φJ
dz. (43) ˙ = −µS µη ∂  f − +

∂z Mβ α− β + βf . (49)
K=1 0
2 ∂t
The term with the integral is ready to go into the equa-
tions of motion. Notice that it has the same stiffness
matrix as in equation (36). The other term represents
7 A NUMERICAL EXAMPLE IN 1D
the slip discontinuity explicitly. To evaluate this term,
we first place the nodes of our basis functions at f − In implementing equations (48) and (49), we chose an
and f + , (Fig. 2). The weighting functions φJ in equa- implicit time integration scheme so that zero-length el-
tion (43) are continuous at the fault, and we set them ements at the slip discontinuity could be modeled with-
to their values at the nodes f − and f + (unity). We then out stability issues. Denoting the values of displacement
 Numerical modeling of waves incident on slip discontinuities 7

velocity along z−axis the boundary conditions are integrated in equations (50)
and (51).
Figure 3 shows the velocity and stress, calculated
t=5s from equations (50) and (51), at several moments in
time. The simulation had an initial condition represent-
ing a plane wave incident from the left. At the center,
t = 10 s
z = 50 m, is a slip discontinuity with compliance η equal
to .5 s2 m2 /kg. The media are identical on either side of
t = 15 s the slip discontinuity. For a compliance of zero, the two
media would be in welded contact and there would be
no reflection. The wave reflected from the slip disconti-
t = 20 s
nuity is the derivative of the incident wave, as pointed
out by Widess (1973), and hence has a higher frequency
t = 25 s content. The slip discontinuity is actually an infinitesi-
mally thin bed; in the parlance of scattering theory, it
is a 1D Rayleigh point scatterer.
Assuming a high velocity thin bed, Widess con-
0 20 40 60 80 100 cluded that the reflection of the thin bed would become
position (m)
negligible for a bed thickness less than one-eighth of a
shear stress along z−axis wavelength (Widess, 1973). So why is it that the slip
discontinuity simulated here gave such a strongly re-
flected wave? From Schoenberg (1980), we know that a
t=5s slip discontinuity models the reflection from a thin, low
shear layer. For a sufficiently low shear velocity, there
t = 10 s is no lower limit to the thickness of a bed that would
produce a significant S-wave reflection. Extremely low
shear wave velocities have been observed in laboratory
t = 15 s measurements of fluid pressurized sands (Zimmer et al.,
2002) and can be expected to occur in “blown” faults
t = 20 s that have high pore pressures in the fault itself. Hence,
the slip discontinuity should be an excellent model for
a “blown” fault.
t = 25 s Note that the stress is continuous and the veloc-
ity discontinuous across the imperfect interface. This
has been accomplished with the zero-length elements
at the slip discontinuity. Figure 4 shows detail of the
0 20 40 60 80 100 wavefield near the fracture. Plotted along the horizon-
position (m)
tal axis is the position. The zero length element shows
Figure 3. Snapshots of the velocity and stress fields at five up as a kink in the stress and a velocity jump in this
different times. There is a slip discontinuity at z = 50 m. plot. We have benchmarked the results of the numeri-
cal simulation with the analytic expressions for reflec-
tion and transmission coefficients at a slip discontinuity
and stress at a future timestep by n + 1 and at previous (Schoenberg, 1980; Pyrak-Nolte et. al, 1990). We obtain
timesteps as n and n − 1, the system of equations is excellent agreement over the frequency band used in the
simulation.
 n+1 − α
α  n−1 1 βn+1 + β
n−1
M =− S , (50)
2∆t ρ 2
8 EXTENSION TO THE P-SV CASE IN 2D
n+1 − β
β n−1 α
 n+1 + α
 n−1 Of more interest to seismic surveys is the reflection and
M = −µS +
2∆t 2 transmission of P − SV -waves at a slip discontinuity.
− − + +
f f f f Two compliances, the normal and tangential compli-
µη βn+1 − βn−1 µη βn+1 − βn−1
− − . (51) ances (ηN and ηT ), characterize a crack in this case.
2 2∆t 2 2∆t
Analogous to equations (1) and (2), the interface bound-
This system of equations is solved in Matlab using LU ary conditions are
decomposition to obtain the values of stress and dis-
u+ −
x − ux = ηT σxz , (52)
placement at time n + 1. Note that no additional con-
+ −
straints to account for boundary conditions are needed - σxz − σxz = 0, (53)
8 M. Haney & R. Snieder

velocity near crack at t = 15 s The horizontal slowness, px , does not change when the
background velocity changes. With this information
0.02
∂σxx 1 ∂σxx
=− , (62)
0 ∂t px ∂x
∂vx ∂vx
−0.02 = −px , (63)
∂x ∂t
40 45 50 55 60
shear stress near crack at t = 15 s and substituting this into equation (60) yields
0.05 ∂σxx ∂vx ∂vz
− p2x (λ + 2µ) − px λ = 0. (64)
∂x ∂t ∂z
0 Equation (64) can be used to eliminate σxx in equa-
tion (56). Notice that we have also replaced a derivative
−0.05 w.r.t. x with a t-derivative in equation (63). Using this
40 45 50 55 60 same trick on equations (57)-(59), P − SV propagation
position (m) in a c(z) medium can be written as a system of 4 PDEs
with derivatives in t and z only
Figure 4. Detail of the velocity and stress fields near the
crack at 15 s. The discontinuity is easily seen in the velocity ∂vx 1 ∂σxz λ ∂vz
at the crack, whereas the stress is continuous.
(1 − p2x Vp2 ) − + px = 0, (65)
∂t ρ ∂z ρ ∂z
∂σxz ∂vx ∂vz
−µ + px µ = 0, (66)
∂t ∂z ∂t
u+ −
z − uz = ηN σzz , (54)
∂vz 1 ∂σzz 1 ∂σxz
+
σzz −
− σzz = 0. (55) − + px = 0, (67)
∂t ρ ∂z ρ ∂t
In what follows, we assume that ηN = 0, i.e. the crack ∂σzz ∂vz ∂vx
is a thin low shear zone, but not a zone of low bulk − (λ + 2µ) + px λ = 0, (68)
∂t ∂z ∂t
modulus.

The equations of motion for P − SV -waves couple where Vp = (λ + 2µ)/ρ is the velocity of the P -wave.
two components of displacement with three stresses Note that for normal incidence (px = 0), the system of
  4 PDEs decouples into 2 systems of 2 PDEs in the form
∂vx 1 ∂σxx ∂σxz
− + = 0, (56) of equations (32) and (33).
∂t ρ ∂x ∂z
Applying the same Galerkin method to equa-
 
∂σxz ∂vx ∂vz tions (65)-(68), we use the expansions
−µ + = 0, (57)
∂t ∂z ∂x


N
  vx (z, t) = αK (t)φK (z), (69)
∂vz 1 ∂σxz ∂σzz
− + = 0, (58) K=1
∂t ρ ∂x ∂z


N
∂σzz ∂vx ∂vz σxz (z, t) = βK (t)φK (z), (70)
−λ − (λ + 2µ) = 0, (59)
∂t ∂x ∂z K=1

∂σxx ∂vx ∂vz

− (λ + 2µ) −λ = 0. (60)

N
∂t ∂x ∂z vz (z, t) = γK (t)φK (z), (71)
Since the quantity σxx does not enter into the bound- K=1

ary conditions, it makes sense to eliminate it from this

system of equations. To do so, we use the fact that, in a

N
σzz (z, t) = δK (t)φK (z). (72)
medium with 1D variation with depth (a c(z) medium), K=1
the horizontal component of the slowness
sinθ Including a slip discontinuity from equations (52) and
px = , (61) (53), the matrix equations are
c
for the incident angle θ from the vertical and medium ve- α n+1 − α
 n−1
(1 − p2x Vp2 )M =
locity c, is constant. Symbolically, this means that plane 2∆t
wave solutions for any of the displacements or stresses 1 βn+1 + β
n−1 λ γn+1 + γn−1
in a c(z) medium would have the form S(t − pxx − pz z). − S + px S , (73)
ρ 2 ρ 2
 Numerical modeling of waves incident on slip discontinuities 9

velocity along z−axis

Pinc Sref Pref
Ptra
Sref
0.02
2
S
1 1 Pref tra

0
x
4
z

3 −0.02
Ptra

Stra 20 30 40 50 60 70 80
position (m)
Figure 5. A P -wave is incident at an angle of 45◦ with the Figure 6. Snapshot of the vertical and horizontal velocity
vertical on a horizontal slip discontinuity. Four waves radiate fields after the incident P -wave has interacted with a slip dis-
from the imperfect interface. The compliance of the crack continuity. The vertical velocity is shown as a solid line, with
does not affect the directions of the reflected, transmitted, the horizontal velocity a dashed line. The slip discontinuity
and converted waves, only their magnitudes and frequency is at z = 50 m, marked by a vertical dashed-dot line. The
content. incident wave is a P -wave incident at an angle of 45◦ with
the vertical.
n+1 − β
β n−1 α
 n+1 + α
 n−1
M = −µS
2∆t 2 cident P -wave should not change direction. To check
γn+1 − γn−1
−px µM that the numerical simulation preserved the incident
2∆t direction, note how the vertical and horizontal veloci-
− − + +
β f − βn−1
f
β f − βn−1
f
ties overlap for the transmitted P -wave in Fig. 6. This
−µηT n+1 − µηT n+1 , (74) overlap occurs for the propagation angle 45◦ , the same
2∆t 2∆t
as the incident P -wave. Also shown in Fig. 6 are the
γn+1 − γn−1 1 δn+1 + δn−1 converted waves Sref and Stra . They have a higher fre-
M =− S quency content than the incident wave and the signs of
2∆t ρ 2
their components determine their respective polariza-
1 n+1 − β
β n−1
− px M , (75) tions. These polarizations agree with the S-waves ex-
ρ 2∆t pected from Fig. 5.

δn+1 − δn−1 γn+1 + γn−1

M = −(λ + 2µ)S
2∆t 2
 n+1 − α
α  n−1
−px λM . (76)
2∆t
These equations are implicit expressions for the time
step n + 1 and can be solved by LU decomposition for
the components [ n+1 , γn+1 , δn+1 ].
αn+1 , β
We show an example where the simulation for the
P − SV case is initialized with an incident P -wave at
an angle of 45◦ . Depicted in Fig. 5 are the directions of
the reflected and transmitted waves. In specifying initial
conditions, the wavelet for the vertical and horizontal
velocities, w, had to be projected into the vertical di-
rection, and the waveforms for the 2 stress components
were computed from Hooke’s Law.
A result of the numerical P − SV simulation is
shown in Fig. 6. The tangential compliance, ηT , equals
.5 s2 m2 /kg, as in the previous example. Since the slip
discontinuity is between two identical media, the in-
9 CONCLUSIONS
The exercise of realizing slip discontinuities in a sim-
ple 1D finite-element algorithm should assist in the im-
plementation of slip interfaces in 2D numerical codes,
which are further complicated by meshing and assem-
bly issues. We find that proper modeling of fractures
demands an implicit, unconditionally stable time inte-
gration scheme. Future work should explore if an explicit
time scheme is at all possible, since this would increase
the computational efficiency needed for an extension to
a 2D or 3D code. In 2D or 3D, the repeated use we made
of integration by parts in this paper would be replaced
by Green’s Theorem. Simulations in a higher dimension
would allow the modeling of crack tips.
In the numerical examples, we kept the background
medium homogeneous except for the slip discontinuities.
The finite-element scheme can also be worked out in the
10 M. Haney & R. Snieder
case of smoothly-varying changes in the material prop-
erties. An example of this would be a zone of linearly
increasing velocity bounded by two slip discontinuities.
A zero-length element is also necessary in simulating a
welded interface.
We limited our study to the reflection and transmis-
sion of a pure slip (displacement) discontinuity. Schoen-
berg (1980) has also suggested the possibility of a vis-
cous slip condition. Other, more exotic interface con-
ditions exist, such as the Maxwell and Kelvin versions
of the combined displacement and velocity discontinu-
ity (Pyrak-Nolte, 1996). This interface condition models
the reflection and transmission from a thin, low shear
zone with attenuation and has been validated in lab-
oratory experiments (Pyrak-Nolte et. al, 1990). The
same finite-element methods we have described here are
needed to numerically model these other types of inter-
faces.
Shampine, L. F., Allen, Jr., R. C., and Pruess, S. 1997. Fun-
damentals of Numerical Computing, Wiley, New York.
Smith, W. 1975. The Application of Finite Element Analysis
to Body Wave Propagation Problems, Geophys. J. R.
astr. Soc., 42, 747.
Wang, J. 2000. Finite Element Methods: Theory and Imple-
mentation, Course Notes, Colorado School of Mines.
Widess, M. B. 1973. How Thin is a Thin Bed? Geophysics,
38, 1176.
Zhu, Y., and Snieder, R. 2002. Reflected and transmitted
waves from fault zones, SEG Int’l Exposition and 72nd
Annual Meeting, Salt lake City, Utah, 2273.
Zienkiewicz, O. C., and Taylor, R. L. 2000. The Finite Ele-
ment Method, v.1, Butterworth-Heinemann, Oxford.
Zimmer, M., Prasad, M., and Mavko, G. 2002. Pressure and
porosity influences on VP − VS ratio in unconsolidated
sands. The Leading Edge, 21, 178.

ACKNOWLEDGMENTS
We would like to thank Yaping Zhu and Jon Sheiman for
discussions on fault zones. Shell has generously provided
funding for this work.

REFERENCES
Alterman, Z. S., and Loewenthal, D. 1970. Seismic Waves
in a Quarter and Three-Quarter Plane, Geophys. J. R.
astr. Soc., 20, 101.
Boore, D. M. 1970. Finite difference solutions to the equa-
tions of elastic wave propagation, with applications to
Love waves over dipping interfaces, Ph.D. thesis, Mas-
sachusetts Institute of Technology.
Coates, R. T., and Schoenberg, M. 1995. Finite-difference
modeling of faults and fractures, Geophysics, 60, 1514.
Haltiner, G. J., and Williams, R. T. 1980. Numerical Pre-
diction and Dynamic Meteorology, Wiley, New York.
Kelly, K. R., Ward, R. W., Treitel, S., and Alford, R. M.
1976. Synthetic seismograms: A finite-difference ap-
proach, Geophysics, 41, 2.
Kelly, K. R., and Marfurt, K. J., eds. 1990. Numerical Mod-
eling of Seismic Wave Propagation, SEG Geophysics
reprint series No. 13.
Lysmer, J., and Drake, L. A. 1972. A finite element method
for seismology, in: Methods in computational physics
v.11, ed. B. A. Bolt, Academic Press, New York.
Marfurt, K. J. 1984. Accuracy of finite-difference and finite-
element modeling of the scalar and elastic wave equa-
tions. Geophysics, 49, 533.
Pyrak-Nolte, L., Myer, L., and Cook, N. 1990. Transmission
of Seismic Waves Across Single Natural Fractures, JGR,
95, 8617.
Pyrak-Nolte, L. 1996. The seismic response of fractures and
the interrelations among fracture properties, Int. Journ.
of Rock Mechanics and Mining Sci. and Geomech. Ab-
str., 8, 785.
Schoenberg, M. 1980. Elastic wave behavior across linear slip
interfaces, J. Acoust. Soc. Am., 68, 1516.