You are on page 1of 14

IOP PUBLISHING JOURNAL OF GEOPHYSICS AND ENGINEERING

J. Geophys. Eng. 6 (2009) 412–425 doi:10.1088/1742-2132/6/4/009

Fractal-based stochastic inversion of


poststack seismic data using very fast
simulated annealing

Downloaded from https://academic.oup.com/jge/article/6/4/412/5127411 by National Geophysical Research Institute user on 24 May 2022
R P Srivastava1 and M K Sen2
1
National Geophysical Research Institute, Fractals in Geophysics Group, Uppal Road, Hyderabad,
500 007, India
2
Institute for Geophysics, J J Pickle Research Campus, Bldg 196, 10100 Burnet Road (R2200),
Austin, TX 78758-4445, USA
E-mail: ravi prakash@ngri.res.in and mrinal@ig.utexas.edu

Received 2 May 2009


Accepted for publication 21 September 2009
Published 23 October 2009
Online at stacks.iop.org/JGE/6/009

Abstract
Seismic data do not contain some low- and high-frequency information because of the
band-limited nature of the source wavelet. A deterministic inversion of such band-limited
seismic data produces smooth models which are devoid of high-frequency variations observed
in well logs. Stochastic inversion methods often based on random Gaussian priors can have a
limitation of producing high frequencies in the desired model particularly the frequency band
not constrained by the input seismic data. In this paper, we propose a new stochastic poststack
inversion algorithm where fractal models constructed from statistical properties of well logs
are used to generate a priori models. This provides a high-resolution model without injecting
spurious high-frequency estimates in model space. Stacked seismic data are used in the
inversion in which a suitable objective function is minimized using a nonlinear optimization
method called ‘very fast simulated annealing’. We demonstrate the effectiveness of our
method for the estimation of acoustic impedance with the application to a field dataset.

Keywords: seismic, inversion, stochastic, fractal, acoustic impedance


(Some figures in this article are in colour only in the electronic version)

1. Introduction the number and spatial distribution of wells over a reservoir.


Therefore, seismic being the most continuous information
Inversion of seismic data plays a vital role in reservoir available (although at a lower vertical resolution), inversion
characterization. High-resolution inversion methods add of seismic data adds great value in reservoir characterization
significant value to the inversion results and increase the (Francis 2006a, Pendrel 2001). Seismic inversion is a
confidence level in the interpretation of seismic data (Sen process which converts seismic information into petrophysical
and Stoffa 1991, 1995, Sen 2006, Rowbotham et al 2003). properties such as acoustic and shear impedances, P- and S-
The well logs provide the most accurate information on the wave velocities and density (Dimri 1992, Sen 2006, Vedanti
petrophysical properties of a subsurface reservoir. However, and Sen 2009) which are then used to understand the reservoir
a spatially continuous description of a reservoir at the well characteristics.
log scale is not available due to the availability of well Compared to raw seismic traces, an inversion
data at sparse locations. One approach to obtain a 3D provides more high-resolution results which facilitate better
description of a reservoir is to use some type of geostatistical interpretation, for instance, of net pay thickness and porosity
interpolation, extrapolation or simulation (Hass and Dubrule (Torres-Verdin et al 1999, Francis 2006a, 2006b). Despite
1994). Accuracy of such geostatistical models is limited by obtaining higher resolution images compared to the original

1742-2132/09/040412+14$30.00 © 2009 Nanjing Institute of Geophysical Prospecting Printed in the UK 412


Fractal-based stochastic inversion

seismic, inversion results are devoid of some low and high


frequencies. Due to the band limitation of the seismic data,
inversion estimates fall way below the resolution desired
by the reservoir engineers. Regularization methods can be
employed to address some of these issues but even with
regularization such as sparsity constraints, blocky-ness and
maximum entropy constraints, inversion results do not match
the frequency band of the well logs (Russell 1988, Russell
and Hampson 1991, Rietsch 1977, 1978, 1988). In other

Downloaded from https://academic.oup.com/jge/article/6/4/412/5127411 by National Geophysical Research Institute user on 24 May 2022
words, the low- and high-frequency parts of a reservoir model
reside in the null space. The only approach to increase the
resolution of seismic inversion is to combine well log data as
a priori information with seismic data during inversion—such
an inversion is termed a ‘stochastic inversion’. A stochastic
inversion incorporates prior models which control the high-
and low-frequency bands missing in the seismic data. Even
though the high-frequency variation of elastic properties does
affect seismic amplitudes, estimation of such high-frequency
part of a model is highly non-unique. Thus it becomes very
important to select/generate a priori models which contain
a realistic frequency band and do not inject spurious high-
frequency model parameters in the null space which cannot be
constrained by the input seismic data. Generally, stochastic
inversion methods employ random Gaussian prior with a white
spectrum unlike the observed logs or those created by a fractal
method which shows a scaling spectrum. Thus, random
Gaussian prior has enough scope to leave high- and low-
frequency model estimates unconstrained by the band-limited
seismic data.
In this paper, we report on the development of a new
stochastic inversion method where we describe our a priori
probability density function (pdf) using a fractional Gaussian
pdf which represents the statistics of a well log better than
a normal distribution does (Hewett 1986, Hardy 1992). A Figure 1. Flow chart of the general methodology used for
fractional Gaussian pdf is described by three parameters, deterministic inversion.
namely mean, variance and Hurst coefficient which can be
easily estimated from available well logs. The prior generated normal incidence of the seismic wave. Also, it is assumed that
in this way facilitates high-resolution consistent estimates of data are noise-free. Further, the reflectivity (rk ) is related to
the model parameters. We demonstrate the feasibility and the acoustic impedance by
usefulness of our method with application to 2D field seismic Ik+1 − Ik
data. A comparison of results obtained with our algorithm rk = , (2)
Ik+1 + Ik
and a widely used deterministic algorithm shows that the
where Ik = ρ k vk , ρ k is the density and vk is the velocity of the
stochastic inversion yields high-resolution estimates. The
kth layer.
confidence level of stochastic model estimates is determined
Thus poststack inversion can be thought of as estimation
by a statistical analysis.
of reflectivity or acoustic impedance. Thus, based on
equation (1), estimation of reflectivity is a linear inverse
2. Theory: poststack inversion problem whereas estimation of acoustic impedance is a
nonlinear inverse problem (equation (2)). To solve the
Poststack inversion is a process to analyse stacked seismic nonlinear problem of acoustic impedance estimation, we use
traces for the estimation of the earth parameters. The a global optimization technique called very fast simulated
estimation process relies on the fundamental relationship annealing (VFSA). The innovative part of this paper is the
between the earth’s reflectivity and observed seismic data application of fractal theory to generate realistic prior models
given by a 1D convolution model (Yilmaz 2000): as intelligent initial guesses to the VFSA optimization module.
x(t) = w(t) ∗ r(t) + n(t), (1)
2.1. Deterministic inversion
where x(t) is the observed seismic trace, w(t) is the source
wavelet, r(t) is the earth’s reflectivity and n(t) is the noise. A nice and concise treatment of deterministic seismic inversion
In this model, multiples are not considered and it is valid for can be found in Hampson and Russell software documentation.

413
R P Srivastava and M K Sen

(a)

Downloaded from https://academic.oup.com/jge/article/6/4/412/5127411 by National Geophysical Research Institute user on 24 May 2022
(b)

(c)

Figure 3. Flow chart of the general methodology used for


fractal-based stochastic inversion.

model constraints (Shewchuk 1994, Minkoff 1996, Vedanti


et al 2005). We have used the STRATA module of Hampson
and Russell software for the deterministic seismic inversion.
(d ) There are several options available in the deterministic seismic
inversion which differ in their treatment of input wavelet and
model constraints to address the non-uniqueness problem.
The following procedures are performed prior to a
deterministic poststack inversion.
(1) Well tie and extraction of seismic wavelet using both
seismic as well as well log data.
(2) Pick horizons to decide the zone of inversion analysis.
(3) Interpolate the available well logs guided by the picked
horizons. Further apply a low-pass filter, typically 10–
12 Hz, to generate a smooth low-frequency acoustic
Figure 2. Frequency spectrum of (a) observed log, (b) fractal-based
impedance model.
log, (c) random prior and (d) spectrum of observed seismic data; (4) Invert the seismic traces using a recursive inversion
these show that any model estimate below 25 Hz and above 90 Hz procedure. This inversion result gives a middle-
frequency will not be constrained by the input seismic data. frequency band (approximately 25–70 Hz, depending on
the frequency content of the seismic data) estimate of the
acoustic impedance.
Most of the deterministic inversion algorithms are gradient (5) Finally combine the low-frequency estimate derived in
based with the variation in treatment of the first- and second- step 3 with the band-limited estimate of step 4 to generate
order derivative calculations and smoothing/roughening a final inversion output.

414
Fractal-based stochastic inversion

Downloaded from https://academic.oup.com/jge/article/6/4/412/5127411 by National Geophysical Research Institute user on 24 May 2022
Figure 4. Test of fractional Gaussian noise characteristic of the given well log. (a) Given acoustic impedance log. (b) Histogram of the well
log. (c) Co-variance showing power law behaviour. (d) Semi-variogram showing power law behaviour. (e) Log power spectrum showing
power law. (f ) Hurst coefficient (H = 0.80) showing the fGn characteristic of the well log.

Before applying the inversion algorithm to invert an entire the seismic during inversion. This fact is demonstrated by
seismic volume, an analysis is performed to tune the the frequency spectrum shown in figure 2 for real well log
parameters and scaling factors by examining the inversion and seismic trace, fractal-based and random Gaussian-based
results at a well location. A flow chart of general methodology models of the well log.
followed for deterministic inversion is shown in figure 1. The novel approach of stochastic inversion algorithm used
in this paper is to make use of the fractal nature of the well logs.
2.2. Stochastic inversion We use the concept of fractal theory to exploit the fact that well
logs follow fractal behaviour. Using this concept we generate
The band-limited characteristic of seismic data and its poor
an a priori model which is input to our VFSA inversion module
vertical resolution decaying with time entail more focus on
as a starting solution. The prior model generated by the fractal
developing an inversion algorithm that takes care of these
method contains statistical properties of the available well log
limitations. By definition, a stochastic inversion is not
and thus it serves as an intelligent initial guess to our inversion
an estimation process—rather it is a statistical process in
algorithm. The prior generated by our algorithm contains a
which several possible model realizations are generated which
similar frequency band to that of the available well log. Thus,
satisfy the observed data. The beauty of stochastic inversion
there is no scope that it can leave spurious frequencies in the
algorithms is that they can incorporate various geophysical
information with seismic to enhance the inversion results model space even if it is not constrained by the input seismic
(Contreras et al 2006, Varela et al 2006). Often well log data. Also we impose constraints on some of the statistical
information is used in conjunction with the seismic data parameters such as the mean, variance and Hurst coefficient
to enhance vertical resolution as well as the low-frequency (H) in our inversion algorithm.
content of the model parameters. Standard stochastic inversion The input to our stochastic algorithm is the same as that
algorithms use random Gaussian prior as an initial model discussed in the deterministic case. However, in this case,
which contains the white spectrum. The inversion algorithm we use an interpolated well log corresponding to each seismic
does not constrain the high- and low-frequency components trace for the estimation of statistical properties, namely mean,
of the initial model as the input seismic is devoid of the high- variance and Hurst coefficient. In our method, we do not need
and low-frequency bands. Thus, random Gaussian prior can to combine the low-frequency model estimate derived from
leave undesired frequency in model space unconstrained by the well log with the model estimate obtained from inversion.

415
R P Srivastava and M K Sen

Harte 1987, Peitgen and Saupe 1988, Percival and Walden


1993, Caccia et al 1997). We closely follow a method proposed
by Caccia et al (1997) to generate the exact fractional Gaussian
noise (fGn) for 1D time series using the fractional Gaussian
process (fGp). To make the paper self-contained, we describe
below the fGp and its use to generate fGn.

2.3.1. The fGp algorithm. Fractional Gaussian noise


represents stochastic processes which can be simulated as time

Downloaded from https://academic.oup.com/jge/article/6/4/412/5127411 by National Geophysical Research Institute user on 24 May 2022
series realizations with expected auto-covariance determined
by intermittence coefficient also known as the Hurst coefficient
(H). Fractional Gaussian noise consists of normally distributed
random variables with zero mean and an auto-covariance
function given by

R(τ ) = 0.5σ 2 (|τ + 1|2H − 2|τ |2H + |τ − 1|2H ), (3)

where τ is the time separation of random variables.


Note that Gaussian white noise is fGn with H = 0.5 and
R(τ ) = 0 for τ  1; thus, Gaussian noise is a special case of
the fractional Gaussian noise.
The fGp generates exact fGn so that both the mean and the
auto-covariance function for the generated time series follow
the mean and auto-covariance of the input time series (well
log), respectively. The algorithm transforms independent
identically distributed (iid) standard normal random variables
Figure 5. Poststack seismic data used in inversion. Red curves into correlated or anti-correlated series based on their Hurst
show three available well logs. Horizon top and bottom marked by coefficient, dictated by equation (3). Using the fast Fourier
blue lines show the zone of interest. Seismic data between 980 and transform of the auto-covariance function, fGp operates on
1110 ms have been used for inversion (courtesy D Hampson and N log2 N operations. A mathematical framework of fGn
B Russell).
generation and estimation of the Hurst coefficient is given
in appendix A.
The general framework of the algorithm used in this paper
is shown in figure 3. The details of the theoretical background
of generating fractal-based prior model and its application are 2.4. Very fast simulated annealing (VFSA)
described in the following subsections.
A detailed description of simulated annealing (SA) and
very fast simulated annealing (VFSA) can be found in
2.3. Fractal characteristics of well logs Kirkpatrick et al (1983), Ingber (1989) and Sen and Stoffa
Analysis of well logs shows that most of the well logs (1995). However, for completeness, we discuss in brief the
exhibit power law (scaling) behaviour often termed as fractal VFSA method used in this paper. We closely follow the
behaviour (Hewett 1986, Emanual et al 1987, Hardy 1992, notation and description given by Sen and Stoffa (1995).
Dimri 2000, 2005). Similar is the case with a reflectivity The concept of SA is an extension of statistical mechanics
sequence which unlike the common assumption of whiteness concept cast in the form of an optimization problem. It is
follows fractal behaviour (Todoeschuck and Jensen 1988, essentially a Monte Carlo approach for minimizing a function
Saggaf and Toksoz 1999, Saggaf and Robinson 2000, Toverud of a large number of parameters. Simulated annealing (SA)
et al 2001). A time/space series (in our case well logs) is said is analogous to the annealing of solids. When a solid is
to exhibit a fractal characteristic if its co-variance, variogram heated beyond its melting point and then cooled slowly to
and log power spectrum follow power behaviour and the Hurst form crystals, the process is known as annealing. The critical
coefficient (H) lies between 0.5 and 1.0. Analysis for the parameter in this process is the rate of cooling and initial
fractal characteristic of a well log used in our study is shown temperature. If cooling is slow enough (annealing) then the
in figure 4. solid forms pure crystals referring to the state of minimum
Fractal characteristic of well log data and seismic energy (global minima). On the other hand if the melt is cooled
reflectivity sequence motivated us to generate trace-by-trace rapidly (quenching) then it does not form crystals, rather it ends
1D realization of an acoustic impedance model as an input to up forming a glass, which refers to a local minimum solution.
our stochastic inversion algorithm. There are various methods The description of the VFSA algorithm, model update and
available to generate synthetic fractal time series (Davies and cooling schedule used in our study is given in appendix B.

416
Fractal-based stochastic inversion

Downloaded from https://academic.oup.com/jge/article/6/4/412/5127411 by National Geophysical Research Institute user on 24 May 2022
(a)

(b)

Figure 6. (a) Computed acoustic impedance (right panel) using density log (left panel) and P-wave (middle panel) at well 1 located at
Xline-42, Inline-13. (b) Computed acoustic impedance (right panel) using density log (left panel) and P-wave (middle panel) at well 2
located at Xline-42, Iline-27. The zone of interest is between 980 and 1120 ms. The analysis for inversion parameters is done using this well
log. (c) Computed acoustic impedance (right panel) using the density log (left panel) and P-wave (middle panel) at well 3 located at
Xline-42, Iline-39. Logs shown in (a)–(c) have been used to generate interpolated well logs corresponding to each trace in seismic data
guided by the picked horizons shown in figure 5.

3. Numerical examples are shown in figure 5. The spectra of the seismic and log data
are shown in figure 2 which show the band-limited character
3.1. Data of the seismic data. Three well logs available along selected
cross line with computed acoustic impedance are shown in
We test our algorithm on a known dataset—poststack seismic figures 6(a)–(c). We picked two horizons shown as top and
data and well logs given as a demo dataset in the Strata bottom horizons in the seismic data for our analysis (figure 5).
module of Hampson Russell software. We selected a cross After a horizon pick, we performed well tie at all three
line (number 42) that has three wells which can be used to well locations available across the seismic line. Subsequent
constrain the analysis. The seismic data along cross line 42 to well tie, we extracted a wavelet using seismic and well

417
R P Srivastava and M K Sen

Downloaded from https://academic.oup.com/jge/article/6/4/412/5127411 by National Geophysical Research Institute user on 24 May 2022
(c)

Figure 6. (Continued.)

Figure 7. Interpolated acoustic impedance logs corresponding to each trace in cross-line no 42 of the given seismic data.

log data. As a part of data preparation for a model-based 3.2. Examples


inversion, available acoustic impedance logs were interpolated
The result from deterministic inversion analysis is shown in
and extrapolated along the picked horizons to generate acoustic
figure 8, which shows model estimate and data match at the
impedance corresponding to each seismic trace. This provides
well location. Finally inversion was applied to the entire
interpolated well logs at each trace location (figure 7). Also we seismic volume to estimate the earth’s acoustic impedance
generated a 10–12 Hz smooth model from interpolated well using the same scaling constants and parameters used at
logs. Interpolated well logs serve as input to our inversion the well location. A deterministic inversion estimate of
algorithm for estimating statistical properties, namely mean, acoustic impedance constrained with three wells is shown in
variance and Hurst coefficient in order to generate fractal- figure 12(a). It is obvious from figures 8(a) and 12(a) that
based prior model for use in our stochastic inversion. Smooth the deterministic inversion estimates are smooth and represent
logs are used to choose model bounds in the stochastic an average estimate of the model parameters. Further, the
inversion and starting models in a deterministic inversion. deterministic inversion results from figure 12(a) show that

418
Fractal-based stochastic inversion

Downloaded from https://academic.oup.com/jge/article/6/4/412/5127411 by National Geophysical Research Institute user on 24 May 2022
(a)

(b)
Figure 8. (a) Estimate of acoustic impedance using a deterministic inversion algorithm at well location. (b) Data match and residual at a
well location using our stochastic inversion algorithm. First panel shows observed seismic trace at well location, second panel is seismic
obtained from the best-fit model using stochastic inversion and third panel shows data residual. A single trace is displayed five times.

broad features from seismic data have been picked as can be estimate are shown in figure 10 which shows 25 realizations
seen at times 1040, 1060 and 1070 ms. The deterministic of the inverted acoustic impedance at the well location. The
estimates are average and smooth estimates by way of their mean impedance curve is shown by a thick black line. The
mathematical formulation. Thus they often fail to meet the inverted and observed seismic data along cross line 42 are
resolution demanded by reservoir engineers. shown in figure 11 using both stochastic and deterministic
Stochastic inversion result at a well location is shown inversion to check the data estimates from inversion. An
in figures 9(a)–(c); this shows the model estimate (acoustic inverted acoustic impedance map, which is a mean of the
impedance), data fit and comparison of stochastic model 25 realizations generated by the stochastic inversion method
estimate with deterministic model estimate at the well location for the entire line is shown with the deterministic inversion
shown in figure 6(b). Multiple realizations of the model results in figure 12 for comparison.

419
R P Srivastava and M K Sen

Downloaded from https://academic.oup.com/jge/article/6/4/412/5127411 by National Geophysical Research Institute user on 24 May 2022
(a)

(c)

Figure 9. (Continued.)

(b)

Figure 9. (a) Estimate of acoustic impedance using our


fractal-based stochastic inversion algorithm at a well location. Note
the high resolution characteristic of the results which picks major
peaks at 1040, 1055 and 1068 ms. At the well location, inversion
analysis was performed between 980 and 1120 ms. The entire
seismic line was inverted between 980 and 1110 ms. (b) Data match
and residual at a well location using our stochastic inversion
algorithm. The left panel shows observed seismic trace at well
location, the middle panel is the seismic obtained from the best-fit
model using stochastic inversion and the right panel shows data
residual. A single trace is displayed five times. (c) Comparison of
stochastic (blue) model estimate with the deterministic model (red).
Observed model computed from the well log is shown as a black
curve. Figure 10. Multiple realizations of acoustic impedance estimates at
a well location using our stochastic inversion algorithm. A standard
deviation analysis based on these realizations shows 15–20%
4. Discussions variation in the model estimate. The thick black line shows the
mean of all the realizations.
In our stochastic inversion analysis, we generated several
realizations to appraise the sensitivity of the solution. It is a stochastic inversion has remarkable agreement with the
clear from figure 9(a) that acoustic impedance derived from measured values of acoustic impedance from the well log.

420
Fractal-based stochastic inversion

(a)

Downloaded from https://academic.oup.com/jge/article/6/4/412/5127411 by National Geophysical Research Institute user on 24 May 2022
(b)

Figure 11. (a) The input seismic data (observed seismic) are shown in the upper panel; the lower panel shows synthetic seismic data
corresponding to the best-fit model derived by our stochastic inversion algorithm. (b) The input seismic data (observed seismic) are shown
in the upper panel; the lower panel shows synthetic seismic data corresponding to the best-fit model derived by the deterministic inversion
algorithm.

421
R P Srivastava and M K Sen

Downloaded from https://academic.oup.com/jge/article/6/4/412/5127411 by National Geophysical Research Institute user on 24 May 2022
Figure 12. Comparison of the results obtained using stochastic (upper panel) and deterministic (lower panel) inversion algorithms. Note the
vertical stripes at trace numbers 10 and 40 in the deterministic result due to the addition of low-frequency estimate by the algorithm. Also
poor continuity of the event at 1030 ms after trace number 70 is obvious. Colour bar shows acoustic impedance values (in m s−1 × g cc−1).

The result shows high-resolution estimates which have picked stochastic inversion results is consistent with the well log
most of the realistic peaks available in the well log. Further, measurements and helps to delineate thin beds. The well logs
the estimated result at the well location has similar variations shown in figures 6(a)–(c) correspond to the trace numbers 13,
as those in the observed well log. This ensures that the inverted 27 and 39, respectively. The inversion analysis was done at the
model estimate contains frequencies similar to those available well corresponding to the trace number 27. Note the quality
in the observed log. The maximum standard deviation of of results away from wells in stochastic (figure 12(top)) and
the acoustic impedance of a layer computed from multiple deterministic (figure 12(bottom)) results. It is obvious that the
realizations is below 20%. This analysis shows that all the deterministic result after the trace numbers 70–119 has poor
results are realistic and increases the level of confidence in the continuity and a low resolution (figure 12), whereas stochastic
estimated model parameters. inversion result has remarkable continuity. It is obvious that
The deterministic inversion result at a well location the vertical resolution is greatly enhanced in the stochastic
(figure 8(a)) shows that the estimated model neither captures inversion results. Thus this is a successful demonstration of
correct amplitudes nor the frequency content. Rather it the fact that a stochastic inversion enforces a tight connection
gives a mean value of the model estimate. Also the between seismic amplitude and well log information whereas
deterministic solution suffers from the bias forcefully included a deterministic inversion does not.
in the form of low-frequency estimate derived from well
logs. The low frequency bias appears in the form of 5. Conclusions
vertical bands in deterministic inversion results. A close
look at figure 12(bottom) shows clear vertical bands between A stochastic inversion using a fractal-based prior facilitates
trace numbers 10 and 40 and another band between trace achieving models with a realistic frequency band similar to
numbers 40 and 70. Careful observation shows that the those observed at a well location. This comes from the fact
estimated acoustic impedance volume using the deterministic that a fractal-based prior contains a similar frequency band as
algorithm (figure 12(bottom)) has low resolution and many that of the real model. It is shown in the case of poststack
subtle events are not resolved, for instance a small event at seismic inversion that it is possible to accurately detect and
1065 ms between trace numbers 100 and 110 (figure 12(top)). delineate spatial reservoir units (thin beds beyond the seismic
Further, the continuity of the events starting at 1020, 1040 resolution limit) by our method, which would otherwise go
and 1060 ms is better defined in the stochastic inversion undetected using a deterministic seismic inversion. Also a
results (figure 12(top)) than those in the deterministic fractal-based prior provides an intelligent initial guess to the
results (figure 12(bottom)). Stochastic inversion results VFSA modelling module which facilitates faster convergence.
(figure 12(top)) have higher resolution compared to the Our approach facilitates the incorporation of constraints like
deterministic results shown in figure 12(bottom). A the Hurst coefficient and covariance computed using the Hurst
comparison of results using stochastic and deterministic coefficient, which enforces convergence, and it is also cheap to
inversions is shown in figure 12. The vertical resolution in our compute. One assumption in our approach is the stationarity of

422
Fractal-based stochastic inversion

the time series meaning that the fractal parameters are the same Estimation of Hurst coefficient: rescaled range analysis
in the entire well log used in our analysis. We also assume that
One of the crucial parameters to generate a fractal-based signal
the fractal parameters computed from interpolated well logs
is the Hurst coefficient. There are several methods available
are representatives of true fractal nature. Thus, the success of
for the estimation of the Hurst coefficient (Chamoli et al 2007,
our method is also dependent on the horizon-guided well log
Turcottee 1997, Caccia et al 1997). It is known that if the
interpolation.
input signal has a sufficient number of samples then re-scaled
range (RS) analysis works well. In this paper, we have used
Acknowledgments the re-scaled range analysis to compute the Hurst coefficient.
Hurst et al (1965) found empirically that many datasets in

Downloaded from https://academic.oup.com/jge/article/6/4/412/5127411 by National Geophysical Research Institute user on 24 May 2022
RPS would like to thank the Department of Science and nature satisfy the power law relation:
 H
Technology (DST), Govt of India for financial support in the R N
form of BOYSCAST fellowship. Thanks to two anonymous = , (A.4)
S 2
reviewers for their constructive comments which improved
where H is known as the Hurst coefficient, N is the number
the manuscript. The help by Eike Rietsch in using Seislab
of data points, and R and S stand for the range and standard
is greatly appreciated. Discussions with Samik Sil, Armando
deviation obtained from the R/S analysis.
Sena and Nimisha Vedanti were very fruitful. The R/S analysis is easily extended to discrete time series.
The analysis is based on the computation of range and standard
Appendix A deviation of the detrended time series. The running sum of the
series relative to its mean is computed for several segments of
The following steps describe the generation of fGn. the time series, which is given by
Compute the variance (σ 2) and the Hurst coefficient (H) 
n

of given acoustic impedance (AI) log; then using equation (1) yn = (yi − ȳn ), n = 2, N − 1. (A.5)
i=1
compute the auto-covariance. Without loss of generality, let
N be the power of 2 samples in log (say M = 2N). For j = The range is defined by
0,1, . . . , M/2, compute the exact spectral power S(k) from the RN = (yn )max − (yn )min, with SN = σN, (A.6)
discrete Fourier transform of the sequence R: R(0), R(1), . . . ,
where SN is the standard deviation of each segment of the time
R(M/2 − 1), R(M/2), R(M/2 − 1), . . . , R(1) given by
series (yn ).

M/2 The plot of log(R/S) versus log(N/2) is a straight line, and
S(k) = R(τ ) e−i2πk(τ/M) the slope gives the value of the Hurst coefficient H.
τ


M−1
Appendix B
+ R(M − τ ) e−i2πk(τ/M) . (A.1)
τ =M/2+1 The annealing concept has been nicely framed in a
Check that all the values of the auto-covariance function mathematical framework using the analogy between model
S(k)  0 for all k. Negative values indicate that the sequence parameters of an optimization problem and particles of an
is not valid. idealized heated solid material. The formation of crystals in
the annealing process occurs at the minimum energy state.
Further, generate W (k), k = 0, 1, . . . , M − 1, iid
The lowest energy state of the physical material is analogous
Gaussian random numbers with zero mean and unit
to the minimum error in an optimization problem. In terms
variance.
of stochastic process each configuration of physical particles
Calculate randomized spectral amplitude, V(k), such that is referred to as a state. As temperature is lowered, at each
√ ⎫
V (0) = W (0) S(0), ⎪ temperature, the solid is allowed to reach thermal equilibrium
√ ⎪

V (k) = [W (2k − 1) + iW (2k)] 0.5 S(k), 1  k < M/2 where the probability of being in a state i with energy Ei
√ , (equivalent to error function in optimization) is given by the
V (M/2) = W (M − 1) S(M/2), ⎪


V (k) = V ∗ (M − k), M/2 < k  M − 1 following Gibb’s or Boltzmann probability density function as
described in Sen and Stoffa (1995),
(A.2) Ei
 
exp − KT 1 Ei
where ∗ denotes that V(k) and V(M − k) are complex P (Ei ) = E
= exp − , (B.1)
exp j Z(T ) KT
conjugates. j ∈S
KT

Finally, use the first N elements of the discrete Fourier


where the set S consists of all possible configurations, K is
transform of V(k) to compute the simulated series Y(k):
Boltzmann’s constant, T is temperature and Z(T) is a partition
1  function given by
M−1
Y (k) = √ V (k) e−i2πk(c/M) , c = 0, 1, . . . , N − 1.   
M k Ej
Z(T ) = exp − . (B.2)
j ∈S
KT
(A.3)

423
R P Srivastava and M K Sen

The temperature is reduced gradually after thermal equilibrium a probability governed by the temperature and error function
has been reached such that in the limit T → 0, the minimum given as
energy state becomes highly probable. Based on the cooling  
E
schedule, there are several variants of simulated annealing P = exp − , (B.6)
T
(Metropolis et al 1953, Geman and Geman 1984, Szu and
where E is the cost function or error.
Harley 1987, Ingber 1993).
In the beginning, when T is high, the probability of
The variants to SA depend on random move to search
accepting solutions with E  0 is more. However, as
the model space and cooling schedule to change the control the temperature decreases, the algorithm moves towards the
parameter ‘T’ referred to as temperature. VFSA is one such greedy mode, where it selects only those solutions which have

Downloaded from https://academic.oup.com/jge/article/6/4/412/5127411 by National Geophysical Research Institute user on 24 May 2022
variant of SA proposed by Ingber (1989, 1993). The model E  0.
acceptance rule in VFSA is the same as that used in SA; Occasional selection of models with higher energy states
however, it differs in the following counts, which makes it (E  0) makes the algorithm robust and increases the chance
more robust and faster. that the algorithm does not get trapped in the local minimum.
• In VFSA, each model parameter in NM dimensional Since this is a probabilistic estimation, it requires several runs
space can have different finite range of variations. Thus to generate a sufficient number of realizations to assess the
each model parameter is allowed to have different degree uncertainty in the final model estimates. A normal practice is
to do a standard deviation analysis of the multiple realizations
of perturbations from their current position. Further it
to assess the consistency of the model estimates.
requires a temperature for each model parameter and these
can be different for each model parameter.
• It also requires a global temperature to be used in References
acceptance criteria, which can be different from the model
Caccia D C, Percival D, Cannon M J, Raymond G and
parameter temperatures. Bassingthwaighte J B 1997 Analyzing exact fractal time series:
• It uses a product of NM one-dimensional Cauchy-like evaluating dispersional analysis and rescaled range methods
distributions as it is difficult to generate NM dimensional Physica A 246 609–32
Chamoli A, Bansal A R and Dimri V P 2007 Wavelet and rescaled
Cauchy distribution for calculating random numbers.
range approach for the Hurst coefficient for short and long time
• To achieve these conditions, VFSA uses model parameter series Comput. Geosci. 33 83–93
perturbations according to the following formula: Contreras A, Torres-Verdin C and Fasnacht T 2006 AVA

simultaneous inversion of partially-stacked seismic amplitude
mk+1
i = kik + yi mmax i − mmin
i , (B.3) data for the spatial delineation of lithology and fluid units of
deepwater hydrocarbon reservoirs in the central Gulf of
Mexico Geophysics 71 E41–8
 
1 |2ui −1| Davies R B and Harte D S 1987 Tests for Hurst effect
yi = sgn(ui − 0.5)Ti 1+ −1 , (B.4) Biometrika 74 95–101
Ti Dimri V P 1992 Deconvolution and Inverse Theory: Application to
Geophysical Problems (Amsterdam: Elsevier) p 230
with the following cooling schedule: Dimri V P 2000 Application of Fractals in Earth Sciences
ed V P Dimri (New Delhi: Balkema) p 238
Ti (k) = T0 exp(−ci k 1/NM ), (B.5) Dimri V P 2005 Fractal Behaviour of the Earth System
ed V P Dimri (New York: Springer) p 207
where mi k is the model parameter at the kth iteration, u is a Emanual A S, Alameda G D, Bchrens R A and Hewett T A 1987
random number between [0, 1], Ti is the temperature of the ith Reservoir performance prediction methods based on fractal
parameter in the kth iteration, T0 is the global temperature and geostatistics SPE 16971, 62nd Ann. Tech. Conf. of SPE
Francis A 2006a Understanding stochastic inversion: part 1 First
ci is the decay parameter.
Break 24 69–77
To implement this algorithm, a valid starting model is used Francis A 2006b Understanding stochastic inversion: part 2 First
and also all the model parameters are assigned some initial Break 24 79–84
temperature. Also a global temperature is set depending on Geman S and Geman D 1984 Stochastic relaxation, Gibbs
the problem; it is desirable to start with a higher temperature. distributions, and the Bayesian restoration of images IEEE
Trans. Pattern Anal. Mach. Intell. 6 721–41
Another important parameter is defining the range of all model Hardy H H 1992 The fractal character of photos of slabbed cores
parameters, known as bounds. Simulation starts with a random Math. Geol. 24 73–97
walk of the function governed by Cauchy-like distribution Hass A and Dubrule O 1994 Geostatistical inversion—a sequential
with constraints that a randomly selected model lies within method for stochastic reservoir modeling constrained by
seismic data First Break 12 561–9
the bounds. Once a new model is generated then the error
Hewett T A 1986 Fractal distribution of reservoir heterogeneity and
is calculated between the previous model and the new model their influence on fluid transport SPE 15386, 61st Ann. Tech.
(often known as the cost function or the error function). If Conf. of SPE
error (or cost) 0, then the new model is selected; however, Hurst H E, Black R P and Simaika Y M 1965 Long-Term Storage
this may correspond to a local minimum. Most of the nonlinear (London: Constable)
Ingber L 1989 Very fast simulated re-annealing Math. Comput.
global optimization problems have several local minima, and Model. 12 967–73
to avoid the optimization algorithm getting trapped into such Ingber L 1993 Simulated annealing: practice versus theory Math.
local minima, even solutions with error 0 are accepted with Comput. Model. 18 29–57

424
Fractal-based stochastic inversion

Kirkpatrick S, Elatt C D and Vecchi M P 1983 Optimization by Sen M K 2006 Seismic Inversion (USA: Society of Petroleum
simulated annealing Science 220 671–80 Engineers)
Metropolis N, Rosenbluth A W, Rosenbluth M N, Teller A H and Sen M K and Stoffa P L 1991 Nonlinear one-dimensional seismic
Teller E 1953 Equation of state calculations by fast computing waveform inversion using simulated annealing
machines J. Chem. Phys. 21 1087–92 Geophysics 56 1624–38
Minkoff S E 1996 A computationally feasible approximate Sen M K and Stoffa P L 1995 Global Optimization Methods in
resolution matrix for seismic inverse problems Geophys. J. Geophysical Inversion (Amsterdam: Elsevier)
Int. 126 345–59 p 279
Peitgen H O and Saupe D 1988 The Science of Fractal Images Shewchuk J R 1994 An introduction to the conjugate gradient
(New York: Springer) method without the agonizing pain http://www.cs.cmu.
Pendrel J 2001 Seismic inversion-still the best tool for reservoir edu/∼quake-papers/painless-conjugate-gradient.pdf

Downloaded from https://academic.oup.com/jge/article/6/4/412/5127411 by National Geophysical Research Institute user on 24 May 2022
characterization CSEG Recorder publication Feature article Szu H and Harley R 1987 Fast simulated annealing Phys. Lett. A
January 2001 122 157–62
Percival D B and Walden A T 1993 Spectral Analysis for Physical Todoeschuck J P and Jensen O G 1988 Joseph geology and seismic
Applications: Multitaper and Conventional Univariate deconvolution Geophysics 53 1410–4
Techniques (Cambridge: Cambridge University Press) Torres-Verdin C, Victoria M, Merletti G and Pendrel J 1999
Rietsch E 1977 The maximum entropy approach to inverse Traced-based and geostatistical inversion of 3-D seismic data
problems J. Geophys. 42 489–506 for thin-sand delineation: an application in San Jorge Basin,
Rietsch E 1978 Extreme models from the maximum entropy Argentina The Leading Edge 18 1070–7
formulation of inverse problems J. Geophys. 44 273–5 Toverud T, Dimri V P and Ursin B 2001 Comparison of
Rietsch E 1988 The maximum entropy approach to the inversion of deconvolution methods for scaling reflectivity J. Geophys.
one-dimensional seismograms Geophys. Prospect. 36 365–82 XXII 117–23
Rowbotham P S, Marion D, Lamy P, Insalaco E, Swaby P A and Turcotte D L 1997 Fractals and Chaos in Geology and Geophysics
Boisseau Y 2003 Multidisciplinary stochastic impedance (Cambridge: Cambridge University Press) p 398
inversion: integrating geological understanding and capturing Varela O J, Torres-Verdı́n C and Sen M K 2006 Enforcing
reservoir uncertainty Pet. Geosci. 9 287–94 smoothness and assessing uncertainty in nonlinear
Russell B 1988 Introduction to Seismic Inversion Methods (SEG one-dimensional pre-stack seismic inversion Geophys.
Course Notes Series) (Tulsa, OK: SEG) p 90 Prospect. 54 239–59
Russell B and Hampson D 1991 A comparison of post-stack seismic Vedanti N and Sen M K 2009 Seismic inversion tracks in-situ
inversion methods 61st Ann. Int. Meeting SEG Expanded combustion: a case study from Balol oil field India
Abstracts pp 876–8 Geophysics 74 B103–12
Saggaf M M and Toksoz M N 1999 An analysis of deconvolution: Vedanti N, Srivastava R P, Sagode J and Dimri V P 2005 An
modeling reflectivity by fractionally integrated noise efficient 1D Occam’s inversion algorithm using analytically
Geophysics 64 1093–107 computed first and second order derivatives for DC resistivity
Saggaf M M and Robinson E A 2000 A unified framework for the soundings Comput. Geosci. 31 319–28
deconvolution of traces of nonwhite reflectivity Yilmaz O 2000 Seismic Data Analysis: Processing, Inversion and
Geophysics 65 1660–76 Interpretation of Seismic Data (Tulsa, OK: SEG)

425

You might also like