APPLICATION OF NEW MONTE CARLO METHOD FOR INVERSION OF

PRESTACK SEISMIC DATA

Yang Xue1, Mrinal K. Sen1 and Zhiwen Deng2

1
Institute for Geophysics, The University of Texas at Austin
2
BGP, CNPC

ABSTRACT

Global optimization methods such as very fast simulated annealing (VFSA)

and a multiple VFSA (MVFSA) inversion have been applied to seismic waveform
inversion and uncertainty characterization respectively. Here we address some of
the limitations of MVFSA by developing a new stochastic method, named greedy
annealed importance sampling (GAIS). GAIS combines very fast simulated
annealing (VFSA) and greedy importance sampling (GIS), which uses a greedy
search in the important regions located by VFSA in order to attain fast
convergence and provide unbiased estimation. We demonstrate the performance
of GAIS with application to pre-stack seismic waveform inversion of angle stack
traces. Initial models containing the fractal behavior as the well logs are used to
gain high resolution estimates of reservoir elastic properties. The results indicate
that GAIS can provide better estimation of both models and their uncertainties
than using VFSA alone. A gas layer can be easily identified using GAIS. The new
MC method demonstrated here could also be applied to reservoir characterization
and reservoir monitoring for accurate uncertainty estimation.

INTRODUCTION

The goal of exploration geophysics is to simulate a log of rock properties as a function of

two-way vertical travel time or depth. The most accurate and high resolution logs can be
generated from well measurements, however, only in vertical direction. Highly heterogeneous
spatial distribution of reservoir elastic properties is not resolvable by seismic data: this causes
multiple models to fit the measurements. A deterministic inversion based on minimization of the
error between the observation and forward modeling only offers one of the best fit models and is
usually band-limited. A complete solution should include both models and their uncertainties,
which requires drawing samples from the posterior distribution. To assess this problem,
stochastic inversion of seismic data with well logs is fairly common.

Conventional stochastic inversion methods based on Monte Carlo Markov Chain (MCMC)
are either computationally intensive or provide biased estimation. The most common MCMC
method is Metropolis-Hastings sampler (Metropolis and Ulam 1949, Hastings 1970), which
generates random samples from a proposal distribution and rejects proposed moves based on
Metropolis criterion. Although this method is proven to asymptotically converge to a stationary
distribution, it is generally very slow. Very fast simulated annealing (VFSA) provides a fast
 Inversion of prestack seismic data

approximation of the expectation value. However, the uncertainty estimation by this method is
biased because of continuous change of proposal distribution with iteration.

Schuurmans and Southey (2000) presented Greedy Importance sampling (GIS), which is a
simple variation of importance sampling and shows an improved inference quality than other
stochastic inference methods. They applied GIS for conducting Monte Carlo inference in a
graphical model and proved that the technique yields unbiased estimates.

Based upon GIS, we further speed up the convergence of Monte Carlo inference by
combining it with VFSA, which is named greedy annealed importance sampling (GAIS).
MVFSA with models drawn from a prior distribution and with varying starting temperature are
employed in order to locate the important regions of the model space which are further explored
by GIS. GAIS is expected to provide an optimal balance between computational efficiency and
accuracy of estimation. Furthermore, fractal initial model (Srivastava and Sen 2009, 2010) is
applied to expand the frequency band of traditional seismic inversion results. The feasibility of
GAIS is tested using prestack seismic from Hampson Russell Strata demo data.

METHODOLOGY

Here we first review the Bayesian formulation of a geophysical inverse problem and then
discuss how to draw samples from a multi-dimensional posterior probability density (PPD) in
model space (Tarantola 1987, Sen and Stoffa 1995) using different stochastic inversion methods.

Bayesian formulation

In Bayesian framework, the degree of belief can be updated by evidence. Represented in

mathematical form, the PPD is proportional to the product of the prior and the likelihood:
p(m)  p(d obs | m) (1)
p(m | d obs )  ,
p(d obs )

where m and d obs represent model parameter and data vectors; p(m) is a priori probability
density function (pdf), p(d obs | m) is the likelihood function and p(m | d obs ) is the conditional
pdf of m given the data d . Denominator p(d obs ) is pdf of observation, or marginal evidence.
It is a constant and independent on m . Normally the likelihood function dominates the much
larger subspaces of the model space than prior pdf. The choice of the likelihood functions
depends on the noise distribution. Gaussian noise is the most common assumption for noise
statistics. Therefore we normally have the likelihood function as shown in equation:
(2)
l (d obs | m)  exp(  E (m)),
where E (m) is the error function given by
E (m)  (d obs  g (m)) / 2T C D1 (d obs  g (m)), (3)

where g (m) is the forward modeling operator and C D is called the data covariance matrix,
which consists observation error and theory error. The marginal PPD of a particular model
parameter, the posterior mean model and the covariance matrix are:

2
 Inversion of prestack seismic data

(mi | d obs )   dm1... dmi 1  dmi 1... dmM (m | d obs ), (4)
m   dm  m  p(m | d obs ), (5)
C'M   dm  (m  m )  (m  m )T  p(m | d obs ). (6)

All of the three equations can be expressed in the form of an integral

(7)
I   dmf (m) p(m | dobs ).
For multi-dimensional problems, evaluation of those integrals is computationally very
expensive; some of the methods for evaluating these integrals are described below.

Simulated annealing

Simulated Annealing (SA) is a global optimization method drawing analogy between the
model parameters and particles in an idealized physical system (Sen and Stoffa 1995). All the
particles are distributed randomly in a liquid phase after being heated to a certain temperature.
The crystallization, or the minimum energy state, occurs if annealing process follows a slow
cooling schedule. Thermal equilibrium is required at each temperature with the probability：

exp(  Ei /( KT )) (8)
P( E i )  ,
 exp(  E j /( KT ))
jS
where E is the energy function. In our application, the set S consists of all possible states
(models) and K is Boltzmann’s constant, which is set equal to 1 in our problem. The energy or
the error function is given by

E (m)  1 / 2(d obs  g (m))T C D1 (d obs  g (m)), (9)

where g (m) is the forward modeling operator and C D is called the data covariance matrix,
which consists observation error and theory error. We can rewrite the Gibbs distribution as PPD
of model parameters: exp(  E (m ) / T )i
P(mi )  .
.  exp(  E (m j ) / T ) (10)
jS
Several modifications are employed by Ingber (1989) to improve the speed of SA, namely
very fast simulated annealing (VFSA). Instead of building NM-dimensional Cauchy distribution,
NM-product of one-dimensional Cauchy distribution is applied (Sen and Stoffa 1995).
The problem is: both SA and VFSA attempt to reach the global minimum using a
temperature-ladder, which has the conflict with the requirement of MCMC that all the states stay
at the same temperature. To be more specific, the proposal distribution becomes sharper and
sharper along the cooling schedule defined by VFSA, which bring biasness to the estimation due
to the short tail (Sen and Stoffa 1996). The mean value of samples is not the true expectation
value, although they are very close to each other. Furthermore, the variance of estimation is
usually underestimated. Greedy importance sampling, described in the following section, avoids
biased sampling by taking fixed steplength and using independent sampling method.

3
 Inversion of prestack seismic data

Greedy Importance Sampling

Importance sampling draws independent samples from a simpler “proposal” distribution Q,

and then weighs these points by the ratio of posterior pdf P and prior pdf Q to obtain a “fair”
representation of PPD. However, if Q misses high probability regions of P, the resulting
estimator will have high variance because the sample will almost always contain
unrepresentative points but is sometimes dominated by a few high weight points (Schuurmans
and Southey 2000).

To overcome this problem, Schuurmans and Southey (2000) presented Greedy Importance
sampling (GIS) based on a simple variation of importance sampling. Starting with independent
sampling from a given prior distribution Q, GIS expands each individual sample to a block of
points by explicitly searching for important regions of the target distribution P. Due to the trend
based search algorithm, each sample block will contain at least one or two important points from
the posterior distribution. The advantage is that even if Q misses high probability regions of P,
the weighted samples from Q still be able to demonstrate a “fair” representation of P. The
procedure of GIS is illustrated below (fig. 1):

 Step 1: draw samples m independently from Q .

 Step 2: For each mi , let mi,1  mi .
 Compute block Bi  {mi,1, mi,2, ..., mi, n } by taking local search steps in the direction of
maximum | f (m) P(m) | until a local maximum or n  1 steps.
 Weigh each point mk  Bi with a weight wi (mk )  P(mk ) i, k / Q(mi ) , where
 ik  1 /(bi 1bi  2 bi  k (k  1)(k  2)) with bi an inward branching factor and
k [0, n  1].
 Step 3: Create the final sample from the blocks of points
m1,1 ,..., m1, n , m2,1 ,..., m2, n ,..., mq,1 ,..., mq, n .
 Step 4: For a random variable f , estimate E P(m) f (m) using the formula
fˆ  iq1 nk1 f (mi,k )wi (mk ).

Figure 1. Workflow of greedy importance sampling (from Shuurmans and Southey 2000).

Therefore we expect smaller uncertainty of the true estimation and faster convergence than a
general importance sampling. Furthermore, compared with VFSA, the prior distribution of GIS
is non-temperature dependent and can be as large as possible, which means a better
quantification of uncertainty than VFSA.

The crucial issue is where to start the local greedy search initialed by mi,1 . Sampling from a
uniform distribution is perhaps logical but it is time consuming to generate blocks with large
samples considering forward modeling is required at each step. An efficient way is to start from
a region near the global minimum error, which can be located by VFSA with a small number of
iterations. We describe this approach in the next section.

4
 Inversion of prestack seismic data

Greedy Annealed Importance Sampling (GAIS)

GAIS avoids being trapped into a local minimum and enables faster convergence by
employing VFSA before GIS. Uniformly drawn temperatures are used as initial temperatures, for
multiple parallel VFSA. In a Bayesian framework, we can consider the temperature to be a
hyper-parameter. Parallel VFSA starting from these temperatures with a small number of
iterations (about 200) is employed to locate the regions near the global minimum error. Then
blocks will be expanded according to the second step of GIS to explore important regions

Seismic Modeling

One of the goals of seismic inversion is to estimate subsurface impedance models (Zp, Zs) . In
order to improve the resolution of our estimation, we use fractal based initial models (Srivastava
and Sen 2009, 2010), which have the same frequency band as that of the well log. Fatti’s
approximation (Fatti. et. al.1994) is employed to calculate the angle dependent reflectivity:
2
Z 
Rpp(θ )  (1  tan ())[ln Z p (i  1)  ln Z p (i )] / 2  8 s  sin 2   [ln Z s (i  1)  ln Z s (i )]
2
 Zp 
 
 2 2 
 tan ()  Z s  2  ρ (11)
 2 sin  ,
 2 Zp   ρ
   
where Z p, s are compressive and shear wave impedance, ρ is the density, and θ is the reflection
angle.

Synthetic seismogram using the convolution of the wavelet and reflectivity are constructed.
The error is evaluated using L2 Norm of the misfit between forward modeling and seismic
observation together with L2 Norm of the misfit between velocity model and well log statistics.

EXAMPLE

Test with Hampson Russell Strata demo data

Here we demonstrate the feasibility of GAIS in seismic inversion based on a 2D line of PP

prestack data set, given as a demonstration data in Hampson and Russell software，which is
available in STRATA module. For the preliminary test and quality control, we employed GAIS
inversion on the seismic traces located at the well location within an angle range of 3 to 24  .
Before inversion, well tie, wavelet extraction and angle gather generation were carried out using
Hampson Russell software package. The well log measurements do not contain shear wave
velocity.

At first, 20 temperatures are randomly drawn from 1 to 100 and treated as initial
temperatures for VFSA. 20% deviation from the well logs is employed as prior constraints for
velocity models. Starting with each initial temperature, parallel VFSA with 200 iterations each is

5
 Inversion of prestack seismic data

applied to update Zp, Zs, ρ from fractal based initial models. Due to limitation of large angles, it
is impossible to estimate reliable density, thus we do not update density models after this stage.
Next we use the best fit modelss of (Zp, Zs) from each VFSA as starting model and expand
those to sample blocks using the criterion defined in step 2 in the workflow of GIS (fig. 1).
Specifically, at each grid point (Zp, Zs) , we define four candidates with perturbations of
(Zp, Zs), (Zp,Zs), (Zp, Zs) and (Zp,Zs) , as the square’s end demonstrated in
fig. 2 (right) in blue circle. Only the one with minimize misfit (red circle in fig. 2 right) will be
chosen for the next grid point. This is repeated 30 times with a fixed step length of
Zp  Zs  50m / s * g / cc .

To summarize, as shown in fig. 2 (left), the middle point locates the best fit model after 200
iterations of VFSA and the green point locates the end model after 30 steps of greedy search. The
magenta points indicate the direction of searching. In the end, all the samples are weighed and
summed to estimate the expectation values of ( Zp, Zs) .

5050 Well log

Samples
5100 Accepted samples

5150

Zs:[m/s*g/cc]
5200

5250

5300

5350

5400

5450

9600 9700 9800 9900 10000

Zp:[m/s*g/cc]

Figure 2. Three marginal probability maps of one layer at the well location (left) with each map
corresponding to one greedy search after independent VFSA. Zoomed one marginal probability map
(right).

RESULTS

Qualify control plot at the well location is shown in fig. 3 and fig. 4. The mean value of
inverted P and S impedances as well as the variance of all accepted impedance samples derived
from GAIS are compared with those derived from the best fit model after 1000 iterations of
VFSA (fig. 3). The same fractal based initial models are used for both methods for comparison.

Based on the estimated impedance model, Fatti’s approximation (Fatti. et. al.1994) is applied
for forward modeling of reflectivity. Synthetic angle gathers using convolution of the reflectivity
with the wavelet are compared with the real measurement (fig. 4).

6
 Inversion of prestack seismic data

Figure 3. Left two columns: Impedance from well logs (blue), initial model (magenta-), the best fit model
from VFSA (green-) with 1000 iterations and the mean model derived from 10 realizations of GAIS
(red). Right two columns: relative variance of all samples generated from VFSA (blue) and GAIS
(red).
Angle gather from observation Synthetic angle gather from best fit VFSA Residuals from best fit VFSA Synthetic angle gather from GAIS Residuals from GAIS
3 6 9 12 15 18 21 24 3 6 9 12 15 18 21 24 3 6 9 12 15 18 21 24 3 6 9 12 15 18 21 24 3 6 9 12 15 18 21 24
560 560 560 560 560

580 580 580 580 580

600 600 600 600 600

TWT (ms)
TWT (ms)

TWT (ms)

TWT (ms)

TWT (ms)

620 620 620 620 620

640 640 640 640 640

660 660 660 660 660

680 680 680 680 680

(left 2) (left 1) (middle) (right 1) (right 2)

 
Figure 4. Seismic angle gathers of 3 to 24 from the measurements (left 2), synthetic angle gather
generated from best fit model of VFSA with 1000 iterations (left 1) and their residuals (middle),
synthetic angle gather derived from 10 realizations of GAIS (right 1) and their residuals (right 2).

After quality control at the well location, we further employ GAIS along the 2D line to run
inversion trace by trace. Fig. 5, fig. 6 and fig. 7 show the inverted Zp, Zs and ZpZs ratio near the
well location (CDP number 71) derived from GAIS and from the mean model of VFSA with 800
iterations.

7
 Inversion of prestack seismic data

GAIS
GAIS VFSA
VFSA
8500
550 550 550
550
8000

7500

600 600 600

600 7000
TWT:[ms]

TWT:[ms]

6500

6000

650 650 650

650 5500

5000

4500
700 700 700
700
4000
20 40 20
604080
60 100 120
80 100120 202040
40 60
60 80
80 100
120
100120
CDP Number
CDP Number CDP Number
CDP Number
Figure 5. inverted Zp using GAIS (left) and compared with mean model of VFSA with 800 iterations
(right).
GAIS GAIS VFSA
VFSA
550 550 5000
550
550 5000

4500 4500

600 600 600

4000600 4000
TWT:[ms]

TWT:[ms]

3500 3500

3000 3000
650 650 650
650
2500 2500

2000 2000

700 700 700

700
1500 1500
204060
2080100
40 120
60 80 100120 20 20
40 40
606080
80 100
100120
120
CDP Number
CDP Number CDP
CDP Number
Number
Figure 6. inverted Zs using GAIS (left) and compared with mean model of VFSA with 800 iterations
(right).
GAIS GAIS VFSA
VFSA
550 550 550550
2.4 2.4

2.2 2.2
600
600 600
600
TWT:[ms]
TWT:[ms]

2 2

1.8 1.8
650
650 650
650
1.6 1.6

700 1.4 1.4

700 700
700
20 40
20 604080100
60120
80 100120 20 20
404060 60 80
80 1 00
100120
120
CDPCDP NumberNumber CDP Number
CDP Number
Fig. 7: ZpZs ratio derived from GAIS (left) and compared with the mean ZpZs value from VFSA with
800 iterations (right).

8
 Inversion of prestack seismic data

DISCUSSION

Based on our test using Hampson Russell Strata demo data, we note that the inverted P and S
impedances from GAIS follow the well log trend better than the best fit model from 1000
iterations of VFSA (fig. 3). Furthermore, as we expected, the variance from VFSA with 1000
iterations is smaller than GAIS because VFSA is temperature dependent and the sampling is
biased towards the best fit model, which results an underestimated variance (Sen and Stoffa
1995). GAIS addresses this problem by further expanding the sample space in the direction of
important region with fixed small step length and assigning the weights based on the ratio
between posterior distribution and prior distribution.

Inverted ZpZs ratio along a 2D seismic line (fig. 7) using GAIS shows more continuity than
the mean model from VFSA with 800 iterations, which helps to identify the gas layer (marked
using black curves) away from the well location more easily and accurately.

CONCLUSIONS

In this paper we investigated the applicability and accuracy of a newly developed GAIS
algorithm to seismic inversion. GAIS starts to seek important regions starting with models that
are close to the important regions already located by VFSA and estimate the expectation value
very accurately. Furthermore, the blocks of samples generated using GIS around the global
minimum error region provides reliable uncertainty of the estimation, which assess the problem
of under estimated variance resulted from typical VFSA. Our test using Hampson Russell Strata
demo data demonstrates a superior performance of GAIS than using VFSA alone.

ACKNOWLEDGMENTS

We thank EDGER forum of the University of Texas at Austin for supporting this research,
Son Phan and Thomas Hess from Institute of Geophysics – University of Texas at Austin for
helping with Hampson Russell software.

REFERENCES

Fatti, J. L., G. C. Smith, P. J. Vail, P. J. Strauss and P. R. Levitt, 1994, Detection of gas in sandstone reservoirs
using AVO analysis: Geophysics, 59,1362–1376.
Gassmann, F., 1951, Elastic waves through a packing of spheres, Geophysics 16, 673–85.
Hastings, W. K., 1970, Monte Carlo methods using Markov chains and their applications, Biometrika, 57, 97-109.
Ingber, L., 1989, Very fast simulated annealing, Mathematical Computer Modeling, 12, 967-993.
Metropolis, N. and S. Ulam, 1949, The Monte Carlo method, J. Acous. Soc. Am., 44, 335-341.
Schuurmans, D. and F. Southey, 2000, Monte Carlo inference via greedy importance sampling, in Proceedings UAI.
Sen, M. K. and P. L. Stoffa, 1995, Global optimization methods in geophysical inversion, Elsevier.
Sen, M. K. and P. L. Stoffa, 1996, Bayesian inference, Gibbs’ sampler and uncertainty estimation in geophysical
inversion, Geophysical Prospecting, 44, 313-350.
Srivastava, R. P. and M. K. Sen, 2009, Fractal-based stochastic inversion of poststack seismic data using very fast
simulated annealing, Journal of Geophysics, 6, 412-425.
Srivastava, R. P. and M. K. Sen, 2010, Stochastic inversion of prestack seismic data using fractal-based initial
models, Geophysics, 75, No. 3, R47-R59.
Tarantola, A. ,1987, Inverse Problem Theory. Elsevier Science.