Vous êtes sur la page 1sur 4

Downloaded 01/08/15 to 5.22.98.42. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.

org/

Waveform Inversion of Amplitude-Versus-Offset Data Using


a Genetic Algorithm

SL2.7

Subhashis Mallick, Western Geophysical


SUMMARY
I cast inversion of amplitude-versus-offset (AVO) data into a
framework of Bayesian statistics. Under such a framework, the
model parameters and the physics of the forward problem are
used to generate synthetic data. The synthetic data are then
matched with the observed data to obtain an a-Posteriori Probability Density (PPD) function in the model space. Genetic
algorithm (GA) is a directed random search technique that utilizes the natural analogy of the survival of the fittest in directing
its search process. GA is not dependent upon the choice of an
initial model and is therefore well suited for the AVO inversion.
The AVO waveform inversion where all or a part of the
prestack data are inverted, is nonunique for the direct estimation
of the absolute values of velocities, Poissons ratios, and densities. However, from simplified approximations to the P-wave
reflection coefficient, it can be shown that the normal incidence
the contrast in the acoustic impedance
reflection coefficient
and the gradient in the reflection coefficient(G) can be
estimated from such inversion. From the GA estimated valand G, and from reliable estimates of velocity
ues of
and Poissons ratio at the start time of the input data, an inverted model can be obtained. I apply this procedure to marine data and demonstrate that the synthetics computed from
such inverted model match the input data to reasonable accuracy. Comparison of the log data from a nearby well shows
that the GA inversion obtains both the low and the high frequency trends (within the bandwidth of seismic resolution) of
the P-wave acoustic impedance. In addition, GA also obtains an
estimate of the Poissons ratio, an extremely important parameter for the direct detection of hydrocarbons from the seismic
data.
INTRODUCTION
Amplitude-versus-offset (AVO) analysis for the direct detection of hydrocarbons from seismic data has been of growing interest to exploration seismologists over the past few years. Reflection records of prestack seismic data contain valuable amplitude information that can be related to the subsurface lithology. With the increasing popularity of AVO, much work has
also been done on AVO inversion; and the fundamental problem
of nonuniqueness associated with such inversion is now wellrecognized (Hampson, 1991; Sen and Stoffa, 1992a).

as the P-wave velocities, Poissons ratios, densities, and thicknesses. The physics of the forward problem is expressed by an
exact or an approximate elastic wave theory. The synthetic data
are the synthetic seismograms computed for a given model using the physics of the forward problem, and the observed data
are the prestack seismic data. To estimate the PPD, I compute an objective fitness function, such as a normalized cross
correlation giving the degree of similarity between the observed
and the synthetic data. This fitness function is a measure of
the absolute frequency for a given model parameter in its model
space. The absolute frequencies of many models, when normalized by their respective cumulative absolute frequencies, give
the relative frequencies for each model parameter. Provided
each model parameter is accurately sampled in its domain of
definition, these relative frequencies are proportional to their respective a-Posteriori Probability Density (PPD) functions. The
objective of an inverse method is to find the global maxima of
the PPD function for each model parameter to get a maximum
likelihood estimate of the true model.
Genetic Algorithm (GA) is documented in Goldberg (1989).
Applications of GA to geophysics are given by Stoffa and Sen
(1991), Sen and Stoffa (1992a, b), and Mallick (1992). In principle, GA consists of the four basic steps: (1) parameter coding,
(2) reproduction, (3) crossover, and (4) mutation. A comprehensive account of each of the above steps are given in the above
references and are not repeated here. I must point out that in
addition to the above steps, an additional operation, namely update, introduced by Stoffa and Sen (1991), and Sen and Stoffa
(1992a, b) is required for a rapid convergence of GA. Also, to
avoid genetic drift, caused by the stochastic errors due to finite
population sizes, I implemented a parallel GA in which many
GA are run in parallel and output from each such run are combined while estimating the PPD.
EXAMPLE
For the waveform inversion of AVO, I use an angle gather
(Todd and Backus, 1985) as the observed data. Each trace
in the angle gather represents a plane-wave propagating at a
constant angle-of-incidence. Angle gathers are therefore easily
interpretable as the variations of the plane-wave reflection amplitudes as functions of time and angle-of-incidence. Proper data
processing, e.g., a prestack migration can take care of any two or
three dimensional effect on the angle-gather data reducing each
model in the GA inversion to a horizontally layered earth model;
each layer being described by its P-wave velocity, Poissons ratio, density, and thickness. Finally, I use the reflectivity method
(Fuchs and Mller, 1971), modified to compute synthetic angle
gathers as the physics of the forward problem.

In AVO waveform inversion all or a part of the prestack data


are inverted for travel-times as well as wave-shapes. Such an
inversion, like most other geophysical inverse methods, can be
cast into the framework of Bayesian statistics, where the model
parameters and the physics of the forward problem are used to
obtain synthetic data. These synthetic data are then matched
with the observed data to estimate an a-Posteriori Probability
Figure la shows an angle gather (between a time interval of
Density (PPD) in the model-space. For the AVO waveform in- 1800 and 2200 ms) for a marine data set. Frequency analysis
version, the model parameters describe the earth models given showed that the data contain significant energy between 5 and
750

Downloaded 01/08/15 to 5.22.98.42. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Waveform inversion of AVO data


85 Hz frequency. In the following example, I estimated the
source wavelet by trial and error, although procedures of automatic estimation of source wavelets can easily be incorporated
into the algorithm.
Figure lb shows the synthetic angle gather computed using the estimated model obtained from ten parallel GA runs.
Figure 2a compares the acoustic impedance estimated from ten
GA runs with the log data from a nearby well. Notice the low
frequency trend of the acoustic impedance is not supplied, but
obtained from the GA inversion. Also notice that the GA inversion fails to pick some high frequency oscillations of the acoustic
impedance as they fall beyond the resolution of the seismic data.
Considering that the well is not on the seismic line, the overall match between the well data and the inverted model is quite
satisfactory. In addition to P-wave acoustic impedance, GA also
gives an estimate of the Poissons ratio (Figure 2b), and the Pand S-wave attenuation Q p and Q s (Figure 2c). If I ignore any
attenuation effect, I do not obtain as good a fit between the
well data and the inverted model as the one shown in Figure 2a.
However, since the data were deconvolved and compensated for
Q at the processing stage, the high Q p and Q S shown in Figure 2c may be far from the true ones. Figure 3 compares the
GA estimated and the well-log acoustic impedances at different
frequency bandwidths.

the arrival time for each reflection event is known, the vertical
travel time (t) for each reflecting interface can also be estimated.
Figure 5 shows the PPD for R o AA, G, and t. These PPD distributions of Figure 5 were computed using the same models as
the ones used to generate the PPD of Figure 4. Tight PPD distributions of Figure 5 demonstrates that in the AVO inversion,
R o , AA, G, and t can be unambiguously estimated. Using reliable P-velocity, density, and the Poissons ratio at the start time
of the input data and the GA-estimated Ro AA, and G, I obtain
the estimated models which were used to compute the synthetic
seismograms shown in Figure 1. Note that after the model parameters at the start time are fixed, their relative changes with
time are well-determined.

ACKNOWLEDGEMENTS
I thank Western Geophysical for permission to present this
paper. Many stimulating discussions with Mrinal Sen at the
University of Texas at Austin and with Neil Frazer at the University of Hawaii at Manoa regarding various aspects of Simulated
Annealing and Genetic Algorithm were extremely beneficial. I
thank Bill King for assisting me with the data processing for the
AVO analyses, Bill Dragoset, Craig Beasley, Frank Levin, Wendell Wiggins, Ron Chambers, and Alfonso Gonzalez for many
interesting discussions.

REFERENCES

DISCUSSION
The synthetic angle-gather data shown in Figure lb were
computed using the models estimated from the GA inversion.
Figure 4 shows the estimated PPD for the P-wave velocity, Poissons ratio, density, and layer thickness obtained from such run.
The PPD distributions in Figure 4 are in a standard wiggle trace
format. Wide PPD distributions in Figure 4 demonstrate that
the AVO waveform inversion is nonunique for the estimation of
such absolute values of model parameters as the P-wave velocities, densities etc. Considering that the model parameters are
estimated so poorly in Figure 4, how did the GA-estimated models match so closely to the seismic and well-log data (Figures 1,
2a, and 3) ? To answer this question, note that the angle-gather
data are controlled by the P-wave reflection coefficient at different reflecting interfaces. Although the exact formula for this
coefficient is quite complicated (Aki and Richards, 1980), there
are approximate versions offering physical insights into the parameters playing key roles in the variation of the P-wave reflection coefficient. These approximate formulas relate the P-wave
reflection coefficient R to the angle-of-incidence
as

Ro + G

(1)

in which Ro is the reflection coefficient at normal incidence and


G is a dimensionless gradient term. Equation (1) suggest that
Ro and G should be estimated well in the AVO inversion. Since
R o is related to the contrast in the acoustic impedance
this quantity should also be well estimated. In addition, since

Aki, K., and Richards, P.G., 1980, Quantitative seismology 1:


W.H. Freeman and Co.
Fuchs, K., and Mller, G., 1971, Computation of synthetic seismograms with reflectivity method and comparison with
observations: Geophys. J. Roy. Astr. Soc., 76, 461-481.
Goldberg, D.E., 1989, Genetic algorithms in search optimization
and machine learning: Addison-Wesley Pub. Co., Inc.
Hampson, D., 1991, AVO inversion, theory and practice: The
Leading Edge, 10, No 6, 39-42.
Mallick, S., 1992, Inversion of amplitude-versus-offset data using
genetic algorithm: 62nd Ann. Internat. Mtg., Soc. Expl.
Geophys., Expanded Abstracts, 844-847.
Sen, M.K., and Stoffa, P.L., 1992a, Genetic inversion of AVO:
The Leading Edge, 11, No 1, 27-29
Sen, M.K., and Stoffa, P.L., 1992b, Rapid sampling of model
space using genetic algorithms: examples from seismic
waveform inversion: Geophys. J. Int., 108, 281-292.
Stoffa, P.L., and Sen, M.K., 1991, Nonlinear multiparameter
optimization using genetic algorithms: Inversion of planewave seismograms: Geophysics, 56, 1794-1810.
Todd, C.P., and Backus, M.M., 1985, Offset-dependent reflectivity in a structural context: 55th Ann. Internat. Mtg.,
Soc. Expl. Geophys., Expanded Abstracts, 586-588.

751

Downloaded 01/08/15 to 5.22.98.42. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Waveform inversion of AVO data

Figure 1: (a) Angle gather computed from a marine data set. (b) Synthetic angle gather
computed using a model estimated from ten parallel GA runs.

Figure 2: (a) P-wave acoustic impedance from well data com- Figure 3: Comparison of GA estimated acoustic impedance
pared with the one estimated from ten GA runs. (b) Eswith well acoustic impedance at different frequency bandtimated Poissons ratio from ten GA runs. (c) Estimated
widths: (a) unfiltered, (b) 120Hz cutoff, (c) 80Hz cutoff,
and (d) 40Hz cutoff.
QP and Qs
752

Downloaded 01/08/15 to 5.22.98.42. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Waveform inversion of AVO data

PPD Estimates from the Waveform Inversion of Marine Data

PPD Estimates from the Waveform Inversion of Marine Data

Figure 4: PPD estimates from 10 parallel GA runs in the waveform inversion: (a) P-wave velocity, (b) Poissons ratio, (c)
density, and (d) layer thickness.

Figure 5: PPD estimates from 10 parallel GA runs in the waveform inversion: (a) normal incidence reflection coefficient
(R0), (b) acoustic impedance contrast (AA), (c) gradient
in the reflection coefficient (G), and (d) vertical time in
each layer.

753

Vous aimerez peut-être aussi