You are on page 1of 4

Inversion of Amplitude-versus-OffsetData Using a Genetic Algorithm

SL2.5

SubhashisMaHick, WesternGeophysical

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/

SUMMARY

values for every model generated in the courseof the run are
stored in the model-space. At the end of the GA run, these
I cast the inversion of amplitude-versus-offset(AVO) data into fitness values are normalized by their respectivecumulative fit
the framework of Bayesian statistics, where prior informatiox nessvalues to get the estimates of the a-Posteriori Probability
on model parameters and the physics of the forward problen Density (PPD) functions.
are used to form synthetic data that are matched with the
observed data to obtain an r+Posteriori Probability Den.+
EXAMPLES
(PPD) function in the model space. Genetic Algorithm (GA:
usesa directed random search technique to estimate the shape Figure la showsa synthetic seismogramcomputed for a model
of the PPD. Unlike the classicalinversion methods, GA is no1 consisting of a single interface separating two media with Pdependent on the choiceof the initial model and is well suitet wave velocities 2350 and 2730 m/s, densities2.2 and 2.15 g/cm
for the AVO inversion. The inversion results using GA on syn. and Poissons ratios 0.4 and 0.95. Thii synthetic seismogram
thetic and real data show that the method works well, ever is computed using ANIVEC software developedby Mallick and
when the signal-tonoise ratio is low. Comparisons on syn Frazer (1987, 1988, 1999, 1991). The two arrivals appearing
thetic data also indicate that GA inversion obtains more accu. in the seismogramare the P-P and P-W reflections, respecrate results than do the inversion techniquesusually practiced tively. Figure lb shows the instantaneous amplitudes of the
in the industry.
seismogramof Figure la, and Figure lc showsthe normaliied
INTRODUCTION
Genetic Algorithm (GA) is documented in Goldbergs 1989
book. Applications of GA to geophysicsare given by Frazer
et al. (199Q), Stoffa and Sen (1991), Sen and Stoffa (1992a,
b), and Smith et al. (1992). In principle, GA consists oi
the four basic steps: (1) parameter coding, (2) reproduction,
(3) crossover,and (4) mutation.

P-P instantaneousamplitudes. The AVO responsefor the P-P


reflection in Figure la, b, and c is representative of a class1 gas-sand reflection (Rutherford and Williams, 1989). The
normalized instantaneousamplitudes of Figure lc are used as
observed data in the Genetic Algorithm. Since instantaneous
amplitudes do not have polarity information, GA is given the
additional information that the contrast in Poissonsratio, Au,
between the lower and the upper medium is negative ( i.e. the
amplitudes are actually decreasingwith the angle of incidence).
The results of the inversion are shown in Figure Id - f. While
the inversion is not able to find the absolute values of model
parameters (e.g., the P-wave velocity of layer 1, shown in Figure Id), it uniquely finds I&, the reflection coefficientat normal
incidence (Figure le), and Au, the contrast in Poissons ratio
(Figure If). This result is consistent with the observation by
others (e.g., see Hampson, 1991). In the presenceof noise as
high as 75% of peak signal amplitude (Figures 2a, 2b), GA is
still able to find the PPD of & and Au with reasonableaccuracy (Figures 2c, 2d). Table 1 comparesthe GA inversion
with linearized inversionsusing the approximate P-P reflection
coefficientsgiven by Shuey (1985), and Hilterman (1990), and
demonstrates that the GA can give more accurate parameter
estimates than the AVO inversiontechniquesusually practiced
in the industry. In addition, note that to perform AVO inversion using Shuey (1985) or Hilterman (199Q), the seismic
data must be calibrated such that the amplitudes correspond
to the true reflection coefficients. Also, Shuey (1985) requires
the knowledge of the P-wave velocities, and the average Poissons ratio, and Hilterman (1990) requires the knowledge of
P-wave velocities for the AVO inversion (see Mallick, 1992, for
details). The GA on the other hand, does not require any such
a priori information, nor does it require that the seismic data
be absolutelycalibrated. GA works as long as the relative variations of amplitude are preserved.

In parameter coding,each model parameter is coded as an unsigned binary string. In the case of a simple AVO problem
consisting of a reflection from a single interface, the model
parameters are the P-wave velocity, Poissons ratio, density
for the medium above the interface, and the contrasts of these
quantities across the interface. The coded binary strings for
eachparameter of a particular model are concatenatedto form
a chromosome.Each chromosometherefore representsa model.
Once a random population of chromosomesis generated, each
member of the population is decodedfor the actual parameter
values. Next, the physicsof the forward problem,i.e. the exact
ormula for the P-P reflection coefficient (Aki and Richards,
1980,vl, p 144150) is used to obtain the syntheticAVO data
br each model (chromosome). These synthetic data are then
:ross-correlatedwith the observedAVO data to obtain a fitncs.9for each chromosome. In reproduction,chromosomesare
limply reproduced in number proportional to their fitness. In
Erossower,two chromosomesfrom the reproduced population
ue randomly chosen as parcnt.9. With a given probability of
:rossover,the gene (bit) contents between the two parents are
martiallyswapped to produce two children. Finally, in mutu!ion, a single gene in each child is randomly changed, with
L specified probability of mutation. Reproduction, crossover,
md mutation are the building blocksof GA which take one
rcncration of chromosomesinto the next, and these three pro
:essesare repeated until a generation with many chromosomes Figure 3a shows an angle gather computed for marine data
itting data with reasonableaccuracyare obtained. The fitness (Todd and Backus, 1985). Proper care was taken to preserve

844

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/

InversionAVO

relative amplitude information while processingthe data set.


To enhance signal-to-noise ratio, three adjacent angle gathers
were stacked together to produce the seismogram. The data
show substantial AVO anomalies in the time interval of 2000
ms and 2200 ms. The amplitudes as functions of incidence
angle for events marked 1, 2, and 3 on Figure 3a are shown
in Figures 3b, 3c, and 3d, respectively. Although the angle
gather was computed to an angle of 32O,the amplitudes could
be picked reliably only to 20 for event 1, and to 25Ofor events 2
and 3. Figure 4 gives the inversion results for these three r+
Section events. Figure 4 showsthat GA ia able to estimate the
PPD for & and Au as long as the amplitudes can be reliably
picked from the observeddata. For example, a visual comparison of amplitudes picked for the events 1,2, and 3 in Figure 3a
(see Figures 3b, 3c, and 3d, respectively) indicates that the
amplitudes for event 2 are more irregular than those for the
events 1 and 3. This irregularity is reflected in the inversion
results for Au, shown in Figure 4. Notice that the PPD for
Au for event 2 (Figure 4d) has a wider peak than those for the
events1 and 3 (Figures 4b and 4f).
ACKNOWLEDGEMENTS
I thank Western Geophysical for permission to present this
paper, and Gee-Pacific Corporation for permission to use the
ANIVEC software. I also thank Bill Dragoset, Craig Beasley,
Frank Levin, Wendell Wiggins, Ron Chambers, Alfonso Gonsales, Mike Reed, Uwe Albertin, and Dan Wisecup for many
interesting discussionsand useful suggestionsthroughout the
preparation of this work.
REFERENCES
Aki, K., and Richards, P.G., 1980, Quantitative Seismology,1,
W.H. Freeman and Co.
Fraser, L.N., Basu, A., and Lou, J., 1990, Geophysical inver;ion using simulated annealing and genetic algorithms, EOS
Trans. Am. Geophys. U., 71, 1477.
Goldberg, D.E., 1989, Genetic algorithms in search optimize
tion and machine learning: Addison-Wesley Pub. Co., Inc.
Hampson, D., 1991, AVO inversion, theory and practice: The
Leading Edge, 10, No. 6, 3442.
Hilterman, F., 1996, Is AVO the seismicsignature of lithology?
A casehistory of ship-shoal-southaddition: The Leading Edge,
8, No. 6, 15-22.
Mallick, S., and Fraser, L.N., 1987, Practical aspectsof reflectivity modeling: Geophysics,52, 13551364.
Mallick, S., and Frazer, L.N., 1988, Rapid computation of multioffset vertical seismic profile synthetic seismogramsfor layered media: Geophysics,S&479-491.
Mallick, S., and Frazer, L.N., 1990, Computation of synthetic
seismogramsfor stratified asimuthally anisotropic media: J.
Geophys. Res., 95,8513-8526.

Mallick, S., and Frazer, L.N., 1991, Re&ction/Transmizsion


coefficients and azimuthal anisotropy in marine seismic studies, Geophys. J. Int., 105,241-252.
Mallick, S., 1992, A simple approximation to the P-wave relIection coefficient and its implication in the inversionof amplitude
versusoffset data: Submitted to Geophysics.
Rutherford, S.R., and Williams, R.H., 1989, Amplitude-versusoffset variations in gas sands: Geophysics,54,68&688.
Sen, M.K., and Stoffa, P.L., 1992a, Rapid sampling of model
space using genetic algorithms: examples from seismic waveform inversion: Geophys. J. Int. (In Press).

Sen,M.K., and Stoffa, P.L., 1992b, Genetic inversion of AVO:


The Leading Edge, 11, No. 1,27-29.
Shuey,R.T., 1985, A simpli6cation of the Zoepprite equations:
Geophysics,50,6C%614.
Smith, M.L., Scales, J.A., and Fischer, T.L., 1992, Global
search and genetic algorithms: The Leading Edge, 11, NO.
1, 22-26.
Stoffa, P.L., and Sen, M.K., 1991, Nonlinear multiparameter
optimization using genetic algorithms: Inversion of plane-wave
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, 586588.

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/

Inversion AVO

Fig. 2. (a) Same as Figure la, but with random noisehaving


a peak amplitude equal to 75% of the peak signal amplitude.
[b) Normalized P - P instantaneousamplitude (event, marked
by an arrow in Figure 2a). (c) PPD for I&J, (d) PPD for Au.

it

Fig. 1. (a) Synthetic seismogramsfor the model consisting


3f two layers with P wave velocities 2350 and 2730 m/s, densities 2.2 and 2.15 g/cmS, and Poissons ratios 0.4 and 0.05. (b)
Instantaneousamplitudes of the seismogramshown in Figure
la. (c) Normalized instantaneousamplitude of the P - P re
Bection (event marked by arrows in Figures la and lb). (d)
Normalized Posteriori Probability Density (PPD) for the Pwavespeed of layer 1. (e) PPD for the reflection coefficient at
normal incidence (Ii&). (f) PPD for the contrast in Poissons
ratio (Au).

846

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/

InversionAVO

Fig. 4. (a) PPD for the reflection coefficient at normal incidence (RJ), obtained from inverting the amplitude response
shown in Figure 3b. (b) PPD for the contrast in Poissonsratio
(Au), obtained from inverting the amplitude reaponseshown
in Figure 3b. (c) Same as (a), but for the amplitude response
shown in Figure 3c. (d) Same as (b), but for the amplitude
responseshown in Figure 3c. (e) Same aa (a), but for the amplitude responseshown in Figure 3d. (f) Same as (b), but for
the amplitude responseshown in Figure 3d.

Fig. 3. (a) Angle stack computed from a marine data set.


(b) Normalized amplitude 88 a function of offset for the event
1 marked on (a). (c) Normalized amplitude aa a function of
offsetfor the event 2 marked on (a). (d) Normalized amplitude
as a function of offset for the event 3 marked on (a).

Exact Values
Ra
I
A0
0.063 [ -0.35

Noise Free Synthetics


AU
Ra
E&t 1Err (%) Est 1Err (%)

Synthetics with 75% Noise


AU
&l
Est 1Err (%)
Est 1Err (%)

Shuey (1985)

0.063

-0.20

43

0.072

12

-0.16

54

Hilterman (1990)

0.063

-0.18

49

0.072

12

-0.15

59

GA

0.063

-0.34

2.9

0.060

-0.31

11

Table 1. Comparisons of linearized inversionson the synthetic data using Shuey (1985)
and Hilterman (1990) with GA inversion. For any quantity A with the exact value A,.
and the estimated value Aelt, the error A,,, in the estimate of A is computed using
A,,, = 100 1A,, - Atrt I / I -4,sI.

847

You might also like