a r t i c l e i n f o a b s t r a c t
Article history: We propose a Bayesian approach for seismic inversion to estimate acoustic impedance,
Received 7 April 2016 porosity and lithofacies within the reservoir conditioned to post-stack seismic and well
Received in revised form 13 January 2017 data. The link between elastic and petrophysical properties is given by a joint prior
Accepted 4 February 2017
distribution for the logarithm of impedance and porosity, based on a rock-physics model.
Available online xxxx
The well conditioning is performed through a background model obtained by well log
Keywords: interpolation. Two different approaches are presented: in the rst approach, the prior is
Acoustic inversion dened by a single Gaussian distribution, whereas in the second approach it is dened by a
Seismic Gaussian mixture to represent the well data multimodal distribution and link the Gaussian
Monte Carlo components to different geological lithofacies. The forward model is based on a linearized
Rock-physics convolutional model. For the single Gaussian case, we obtain an analytical expression for
Porosity the posterior distribution, resulting in a fast algorithm to compute the solution of the
Lithofacies
inverse problem, i.e. the posterior distribution of acoustic impedance and porosity as well
as the facies probability given the observed data. For the Gaussian mixture prior, it is not
possible to obtain the distributions analytically, hence we propose a Gibbs algorithm to
perform the posterior sampling and obtain several reservoir model realizations, allowing
an uncertainty analysis of the estimated properties and lithofacies. Both methodologies
are applied to a real seismic dataset with three wells to obtain 3D models of acoustic
impedance, porosity and lithofacies. The methodologies are validated through a blind well
test and compared to a standard Bayesian inversion approach. Using the probability of the
reservoir lithofacies, we also compute a 3D isosurface probability model of the main oil
reservoir in the studied eld.
2017 Elsevier Inc. All rights reserved.
1. Introduction
From a mathematical point of view, the estimation of reservoir properties from seismic data is an inverse problem.
The goal of inverse modeling is to predict model variables from a parameterized physical system from observable data,
* Corresponding author.
E-mail address: leandrop.fgr@gmail.com (L.P. de Figueiredo).
http://dx.doi.org/10.1016/j.jcp.2017.02.013
0021-9991/ 2017 Elsevier Inc. All rights reserved.
L.P. de Figueiredo et al. / Journal of Computational Physics 336 (2017) 128142 129
assuming theoretical relations between observable and unobservable parameters, as well as prior information [40]. Seismic
inversion is an important step in the process of modeling and characterization of oil and gas reservoirs, because it allows
estimating the elastic properties of rocks from seismic data [12,37]. Elastic properties are related to petrophysical properties
such as porosity and uid saturation [4,25] that are the main variables in formation evaluation for the quantication of the
hydrocarbon volume within the reservoir. Typically, seismic data are measured according to a regular geometry, for example
seismic traces are measured at several locations every 25 m along orthogonal directions (namely inline and crossline) and
the dataset includes the measurements of seismic traveltime and seismic amplitudes. Well measurements are generally used
in reservoir characterization as a validation of the inversion results. Furthermore, well data provide direct measurements of
rock and uid properties in the borehole, as well as prior information for the wavelet estimation [11] and rock-physics
model calibration [17]. Generally well measurements include measurements of elastic and petrophysical properties such
as P- and S-wave velocity, density, porosity, and uid saturation. The rock-physics model provides a relation between the
petrophysical properties and the elastic properties. It is usually given by a set of equations that can be obtained by a simple
tting of well data or by a more complex theoretical relation [13,17,25]. The distribution of rock and uid properties in
the subsurface depends on the rock type, namely the lithofacies. In our work, a lithofacies (or simply facies) is dened as
a geobody with specic physical and geological characteristics related to the sedimentological and depositional conditions
[34].
The main challenge in seismic inversion is the limited bandwidth of seismic data and consequently the non-uniqueness
of the solution, particularly outside of this band. Thus, it is possible to nd different values of the model parameters that
are compatible with the measured data. Besides, the presence of noise in the data and errors associated to the modeling
process affect the accuracy of the inversion results [38]. These factors motivate the development of stochastic inversion
methods, that can generate several congurations, or stochastic realizations, of subsurface properties with higher resolution
than the input seismic data, providing a quantication of the uncertainty and a set of solutions with different scenarios of
the reservoir model [8]. Stochastic optimization methods, such as simulated annealing, genetic algorithms or particle swarm
optimization, have been proposed to solve geophysical inverse problems [9,22,27]. These methods are based on optimization
procedures and are strongly dependent on computational resources. More ecient techniques based on stochastic Bayesian
formulations using Monte Carlo methods have been proposed in recent publications for the estimation of the posterior
distribution of reservoir properties conditioned to geophysical data [4,7,17,35,36]. For linearized forward models and under
the assumption of Gaussian distribution of the model parameters and observation errors, the posterior distribution can be
analytically computed, leading to very fast algorithms [6,11]. Monte Carlo methods estimate a probability distribution by
numerical random sampling following some criteria. It was rstly introduced by [26] to simulate a liquid in equilibrium
with its gas phase, originating the class of simulations called Metropolis algorithms. The methodology was generalized by
[19], and posteriorly [15] presented a special case of the MetropolisHasting algorithm called Gibbs Sampler. These methods
have been largely applied in physics, chemistry, biology and mathematics [28] in the last decades. The use of Monte Carlo
methods for inverse problems in geophysics was introduced by [20] and [31], as a methodology to estimate the probability
distribution over the model parameter space [7,10,17,33,35,40]. Differently from other stochastic optimization techniques
and Approximate Bayesian Computation algorithms, in Markov chain Monte Carlo methods, the likelihood function is ex-
plicitly computed, allowing the exact estimation of the posterior distribution and a correct quantication of the model
uncertainty.
In this work, we propose a Bayesian seismic inversion based on a rock-physics relation that accounts for the statistical
correlation between porosity and the natural logarithm of acoustic impedance. The rock-physics model is embedded in the
joint prior multivariate normal distribution of both properties. This assumption, together with the linearized convolutional
model and the Gaussian assumption for the errors, enables us to obtain the analytical expression for the conditional dis-
tributions of the model parameters given the observed seismic data. We propose two approaches to obtain the acoustic
impedance, porosity and facies within the reservoir conditioned to the seismic data and the well data. The rst approach
is based on a single Gaussian prior distribution, where the posterior distribution for impedance and porosity as well as the
probability of facies given these properties is analytically derived, and hence calculated with a small computational cost
(Method 1 Analytical). In the second method, we propose a Gaussian mixture model for the prior distribution, and we
introduce a Gibbs algorithm to sample the joint posterior distribution for impedance, porosity and facies (Method 2 Monte
Carlo).
2. Methodology
In this section, we present the seismic forward modeling, the rock-physics model for the prior, and the background prior
model, which are common elements for both proposed inversion methods. We then introduce the analytical formulation for
Method 1 and the statistical sampling algorithm for Method 2.
A seismic dataset includes the measurements of seismic amplitudes for a given time interval during wave propagation at
a set of spatial locations regularly distributed over a surface area corresponding to the subsurface reservoir region. Assuming
130 L.P. de Figueiredo et al. / Journal of Computational Physics 336 (2017) 128142
that the seismic data were processed to remove the multiple reections and other undesirable effects [37], an accurate
approximation of the forward seismic model, for weak elastic contrasts, can be written as a convolutional model:
d(t ) = s( )r (t )d + ed (t ), (1)
where d(t ) is the seismic amplitude at time t, s(t ) is the seismic pulse (namely the seismic wavelet), r (t ) is the weak
contrast reectivity and ed (t ) is a random noise in the observations. The reectivity r (t ) for an incident angle of 0 degree
depends on the acoustic impedance z(t ) (i.e. the product of P-wave velocity and density) above and below the reection
interface through the following relation [37,39]:
1
r (t ) ln( z(t )), (2)
2 t
where t is two way the traveltime of the wave propagation.
In the discretized domain, Equation (1) can be rewritten as:
d = S r + ed , (3)
where the vectors are the discrete representations of the variables, S is the convolutional matrix with the discretized
wavelet s and ed is the seismic noise with zero mean and covariance d [6,11]. If D is a rst order differential operator,
we can dene a linear operator C = 12 S D and a linear relation between the seismic data d and ln( z ) can be obtained:
d = Cln( z ) + ed . (4)
Therefore, if the acoustic impedance z of a sequence of reservoir rocks is known, the seismic response can be computed
by applying Equation (4).
The physical relation between petrophysical and elastic properties is usually given by a set of equations known as
rock-physics model [4,13,17,25]. In this work, we propose a statistical linear correlation between porosity and the natural
logarithm of acoustic impedance. The relation is dened by a joint prior multivariate normal distribution for both properties
(model vector m). Hence, the prior distribution at a given subsurface point is
ln( z) g ( ) + g ( )( ). (7)
Equation (7) provides a linear function with slope A = g ( ) and intercept B = g ( ) g ( ) . Given the slope A, we
can calculate the eigenvectors of the matrix V of the prior covariance matrix using the angle = atan( A ):
sin() cos()
V = . (8)
cos() sin()
The eigenvectors are orthogonal vectors. The eigenvector corresponding to the maximum eigenvalue denes the direction
along which the distribution has the largest variance. By using the eigenvalue decomposition theorem we obtain the prior
covariance matrix:
L.P. de Figueiredo et al. / Journal of Computational Physics 336 (2017) 128142 131
= V V 1 , (9)
where is the diagonal matrix containing the variance for each eigenvector direction.
Finally, we extend this relation to the vectors of the logarithm of impedance ln ( z ) and porosity along the entire
seismic trace, by including a spatial correlation model. Hence, we dene the prior distribution for m as:
ln ( z )ln 11 12
p (m) = p (ln ( z ), ) = N 2n m; m , m = N 2n ; , , (10)
21 22
where
In order to incorporate the prior knowledge of the subsurface in the inversion result, we propose to relate the model
vector m to a known background subsurface model f using the relation:
f =m+ef, (13)
where the variable e f is a noise tolerance with covariance f dening how similar the inversion will be to f . The model f
can be any background model, for example, a low frequency model or some subsurface model from other inversion methods.
In order to incorporate well data information available at the well locations in the inverted model, we build a back-
ground subsurface model f using a geostatistical interpolation of well log data within a stratigraphic grid [3,12], namely
Kriging. The interpolated model is full band, therefore, it has the low frequency information as well as the high resolution
information extracted from the well data. Because, the higher frequencies are reliable only in the regions close to the well
locations, the variance of the error e f is set to be high in regions far from the wells and low in regions close to the wells. A
spatial variance model with these features is the Kriging variance model [12], in which the variance at well locations is zero
or close to zero (due to a nugget effect), whereas the variance far from the wells, at distances larger than the correlation
range, is equal to the variance of e f .
Based on Equation (4), we can write the relation between the model vector m and the seismic data as:
d = Gm + ed , (14)
where
G= C 0 , m = (ln ( z ), ).
Equations (13) and (14) can then be written in the matrix form as:
m 1 0 0 m
f =1 1 0e f . (15)
d G 0 1 ed
These equations dene the linear transformation between the variables (m, f , d) T , and the uncorrelated variables
(m, e f , ed )T . We can then apply the theorems in Appendix B [1,2] to obtain the joint distribution of (m, f , d) T (Theo-
rem 1, Appendix B) and the posterior distribution of the model parameters m, conditioned to the seismic data d and the
background model f (Theorem 2, Appendix B). By introducing the vector o as the vector of observed data, dened by the
seismic data d and the background model f , we can then write the posterior distribution of the inverse problem for a given
seismic trace as:
p (m|o ) = N 2n m; m|o , m|o , (16)
132 L.P. de Figueiredo et al. / Journal of Computational Physics 336 (2017) 128142
where
m|o = m + m,o o 1 o o , (17)
m|o = m m,o o1 o,m , (18)
f m m
o= , o = , o,m = ,
d G m G m
m + f m G T
o = .
G m G m G T + d
If we assume that the random variables ed and e f of Equations (13) and (14) are uncorrelated Gaussian noises, the covari-
ance matrices d and f are diagonal.
As the posterior distribution is a multivariate normal distribution, the maximum a posteriori (MAP) solution for each
seismic trace is the posterior expectation m|o and its covariance is the posterior covariance m|o . The rst component of
m|o is the logarithm of the acoustic impedance, hence the MAP of the impedance is obtained by calculating the mode of a
log-normal distribution exp [m|o diag ( m|o )] [6], whereas the MAP of the porosity is given by the second component of
m|o .
The reference facies model at the well location is generally derived from well log data, by using the observed petrophys-
ical properties to classify different rock types. There are several clustering algorithms in literature for facies classication,
such as Neural Networks, K-Means clustering and Expectation Maximization [18]. In particular, the Expectation Maximiza-
tion method estimates the parameters of a Gaussian mixture model from the experimental data. In our implementation, we
assume that each Gaussian component can be interpreted as a facies [17,23]. Once the facies are classied based on the
training data at the well location, we can dene the probability distribution of the properties mt at any traveltime t in each
facies as a Gaussian distribution:
Both distributions under the integral sign are Gaussian, and the product of two Gaussian distributions is a Gaussian distri-
bution multiplied by a scale factor [30] (see Appendix B for more details). Therefore, Equation (21) becomes
2
p ( t |ot ) p ( t ) exp i s N 2 (mt ; s , s )dmt , (22)
i =1 m
in which
1
i = d ln(2 ) + ln| i | + iT
i
1
i i = { , m|o} ,
2
1
s = d ln(2 ) + ln| s | + sT 1
s s ,
2
and
2 1
2
1
s = i , s = s
i
1
i .
i =1 i =1
L.P. de Figueiredo et al. / Journal of Computational Physics 336 (2017) 128142 133
As the integral of N (m; s , s ) is 1 by denition, and assuming that all the facies are equiprobable a priori, the proba-
bility of a subsurface point t belonging to a facies given the seismic data d is
p ( t |ot ) exp i s . (23)
i
In order to get the exact probability, Equation (23) should be normalized over all the possible facies.
2.6. Method 2 Monte Carlo: posterior distribution of impedance, porosity and facies
The inversion method proposed in Section 2.4 denes a single Gaussian prior distribution based on a linearized rock-
physics model. However, generally, in subsurface studies, different facies show different property values which can not be
described by a single Gaussian distribution. In order to overcome this limitation, we propose a methodology which inte-
grates linear combinations of Gaussian distributions into a Gibbs sampling algorithm to compute the posterior distribution
p (m, |o) for impedance, porosity and facies given the seismic data and the background model.
The Gibbs algorithm consists in computing the desired distribution performing multiple samplings of each variable given
all the other variables [5,14,15]. Hence, in each algorithm iteration, new congurations of the reservoir properties and
facies are generated, leading to multiple realizations of the subsurface model. These realizations can be used for uncertainty
analysis in the reservoir properties [8]. In the Gibbs algorithm, we rst compute the conditional distribution for each variable
given the others. Considering a given conguration for the model vector m = mi at the iteration i in Equation (21), the
distribution p (m|o z ) is a function. Therefore, the probability of a facies given the observed data o and the model mi at
a traveltime t is:
p ( t |ot , mi ) p ( t ) p (m| t )(mi m)dm, (24)
m
m is the distribution in Equation (16) (Section 2.4). Since the prior distribution is dened for a given facies i using the
methodology described in Section 2.5, then:
The steps of Algorithm 1 are described in details below. Steps 1 to 3 are performed only once to dene the initial
congurations to enable the following sampling by the algorithm.
1. Dene the distribution p (m| ) in Equation (19) for all the facies, using a parameterized rock-physics model according
to Subsection 2.2.
2. Compute Equation (16) to obtain the conditional distribution of the properties p (m|o, ) = N (m|o, , m|o, ), for all
the prior distributions dened in step 1, for all the facies.
134 L.P. de Figueiredo et al. / Journal of Computational Physics 336 (2017) 128142
Table 1
Inclusion rock-physics models parameters: compressional modulus of the solid rock, bulk modulus of the uid, density of the solid rock and density of the
uid. Model 2 represents the model within the reservoir (reservoir facies); Model 1 represents the model outside the reservoir (facies 1 and 2).
3. Dene the initial property conguration m0 by arbitrarily choosing one of the means calculated in step 2.
4. For each subsurface point, compute the probability of each facies given the observed data p ( t |ot , mti 1 ) in Equa-
tion (26).
5. Draw the facies conguration i from the calculated probabilities p ( |o, mi 1 ) in step 4. The sampling is performed
using Truncated Gaussian Simulation [12,24].
6. Draw mi conditioned to the facies conguration i from p (m|o, i ) = N (m|o, i , m|o, i ). The multivariate sampling is
performed using FFT-Moving Average [12,32]. First, we draw the properties m| from N (m|o, , m|o, ), by generating
one conguration for the entire dataset for each facies . Then, a conguration is built by setting the properties values
mti at each subsurface point according to the drawn value m| ti for the given facies sampled in step 5. In this step, we
can consider different correlation ranges for each facies to mimic the geological stratigraphy and deposition.
7. Compute the statistics of the sampling to obtain the posterior distribution p (m , |o) and the corresponding mean, most
likely model and associated uncertainty.
We applied the proposed methods to a real case study including post-stack seismic data and well logs from three well
locations. The study area consists in a carbonate reservoir lled by water and hydrocarbon. The model grid includes 312 ms
in two-way traveltime thickness, 450 inlines and 315 crosslines, with vertical sampling rate of 4 ms and lateral sampling
rate of 25 m. A portion of the seismic dataset is shown in Fig. 1, along an arbitrary line passing through two wells. The two
interpreted horizons in Fig. 1 (black lines) delineate the main oil reservoir location.
Well log measurements of porosity and impedance are displayed in Fig. 2A, showing a strong correlation between the
properties of interest. Three facies have been identied at the well locations: the reservoir facies characterized by high
porosity (reservoir facies, or facies 3) and two non-reservoir facies with low porosity (facies 1 and 2), shown in Fig. 2B.
According to the rock-physics literature, an adequate relation to describe the link between porosity and impedance in
carbonate rocks is provided by the Inclusion rock-physics model [21] (see Appendix A). The model requires the knowledge
of mineral and uid parameters such that the compressional modulus M m and density m of the mineral and the bulk
modulus K f and density f of the uid. The compressional modulus is dened as M m = K m + 43 G m , where K m and G m are
the bulk and shear modulus of the mineral. In our application, two uids are present in the reservoir: water and oil. The
uid properties are taken from the rock-physics literature. The rock parameters within the reservoir are calibrated to match
the well log data in the reservoir facies (Model 1; the calibrated curve is shown in red in Fig. 2A and B). Similarly, the rock
parameters outside the reservoir are calibrated to match the well log data in facies 1 and 2 (Model 2; the calibrated curve
is shown in black in Fig. 2B). Table 1 shows the rock-physics parameters obtained for the reservoir and the non-reservoir
facies.
Following the methodology presented in Subsection 2.2, the prior distribution of Equation (5) is calculated using the rock-
physics models. Fig. 2A shows the single Gaussian prior distribution in blue used in the analytical approach of Method 1,
Fig. 1. Arbitrary line of post-stack seismic data used in the inversion methods. The vertical axis is the two way traveltime and the horizontal axis X is the
distance from the rst trace along the arbitrary line. The two horizons in black delineate the main oil reservoir in the eld.
L.P. de Figueiredo et al. / Journal of Computational Physics 336 (2017) 128142 135
Fig. 2. Plot A: Cross-plot of porosity and impedance well logs, together with the theoretical curve obtained using the inclusion rock-physics models and
the prior distribution for the model vector mt = (ln( zt ), t ). Plot B: Cross-plot of porosity and impedance well logs color-coded by facies, together with the
theoretical curves obtained using the inclusion rock-physics models and the prior distribution for the model vector mt = (ln( zt ), t ), computed for each
facies. (For interpretation of the colors in this gure, the reader is referred to the web version of this article.)
and Fig. 2B shows the Gaussian components for each facies of the Gaussian mixture model used in the Gibbs algorithm of
Method 2. The distribution in Fig. 2A is calculated using the rock-physics in red, with mean and variance obtained from the
well log data. For the distributions in Fig. 2B, we use facies-dependent means and variances, and the red curve is used for
the prior distribution of the reservoir facies, whereas the black curve is used for the prior distribution of the non-reservoir
facies. In this application, the range parameter L of the correlation function (Equation (12)) is estimated to be 7 ms.
The background subsurface model f and its variance f are obtained by geostatistical interpolation of well log data in
the stratigraphic grid. The interpolated acoustic impedance and its variance are shown in Figs. 3A and C, and the porosity
and its variance are shown in Figs. 3B and D. According to the variance model, we expect that close to the well, the inversion
result will be similar to the interpolated model f , and consequently to well log data, due to low values of variances e f in
those regions. The wavelet s used in the inversion process is shown in Fig. 4. It is obtained using the smooth part of the
amplitude spectrum of the seismic data, and using well data to optimize the wavelet phase spectrum to match the synthetic
seismic trace to the experimental one [29].
Using the presented data (seismic, prior distribution, wavelet, background model and its variance) the proposed method-
ologies are then applied. In the analytical method (Method 1), the single Gaussian prior distribution in Fig. 2A is used to
compute the conditional distribution in Equation (17), to obtain the MAP solution for impedance and porosity (Figs. 5A
and B).
Based on the three Gaussian components in Fig. 2B, we apply the Gibbs sampling (Algorithm 1) to obtain several realiza-
tions of the subsurface model (Method 2). The model realizations of impedance and porosity after 100 iterations are shown
in Fig. 6A and B. After all the iterations of the algorithm are completed, the means and standard deviations of impedance
and porosity are calculated (Figs. 5C and D and Figs. 6C and D respectively).
In the inversion results, the acoustic impedance and porosity well log are shown in the same color scale of the inversion
for comparison. We observe that the estimated properties are conditioned to the well-log data at the well locations and
also honor seismic data in the regions between the wells and the background model. By comparing the results of the two
methodologies, we point out that the inversion results using a single Gaussian prior distribution have a smaller range of
properties values than the inversion results obtained using the Gaussian mixture prior. We speculate that it might be due
to the low probability of extremes values of a single Gaussian distribution.
Using the impedance and porosity results shown in Figs. 5A and B and their variances, we then use Equation (23) of
Method 1 to calculate the probability of each facies given the seismic data. The point-wise probability of the reservoir facies
is shown in Fig. 7A. From the point-wise probability of each facies, we can predict the most likely facies for all subsurface
points. Fig. 7B shows the facies model along the arbitrary line (each facies is represented by the color of their corresponding
distribution in Fig. 2). The point-wise probability of the reservoir facies calculated from the multiple facies congurations
obtained from the Gibbs algorithm (Method 2) is shown in Fig. 7C and the corresponding most likely facies model is shown
in Fig. 7D.
We then display the isosurface of reservoir facies probability to show the regions with more than 70% of probability of
belonging to the reservoir facies in the 3D reservoir grid. Fig. 8A shows the isosurface probability calculated with Method 1
using a single Gaussian prior distribution, while Fig. 8B shows the isosurface probability obtained through Method 2 using
a Gaussian mixture prior.
136 L.P. de Figueiredo et al. / Journal of Computational Physics 336 (2017) 128142
Fig. 3. The interpolated acoustic impedance and porosity from well data (A and B respectively), and their variances (C and D respectively) used to dene
the background subsurface model f and the covariance matrix f .
Comparing the two results in Fig. 8, we notice that the rst approach with a single Gaussian distribution underestimates
the reservoir volume compared to the second approach. Indeed the Gaussian mixture approach better allows capturing
values close to the boundaries of the property ranges. Typically, the reservoir facies shows high porosity and low impedance.
These values are better predicted by the Gaussian mixture case, leading to a better prediction of the reservoir volume.
To validate the inversion results, we perform a blind well test. The results of the proposed methods are compared
to the results obtained by applying a standard Bayesian linearized inversion [6]. In the standard approach, the posterior
distribution of the properties of interest is analytically computed through Bayes rule, assuming the same linearized forward
model used in the proposed methods. For a fair comparison, in the standard approach the prior mean is dened according
to a low-frequency model [11]. Fig. 9 shows the impedance and porosity well logs in black, the MAP solution of Method 1
in light blue, the posterior distribution of Method 2 in color scale (mean in blue and MAP solution in dashed blue), and
the MAP of the standard Bayesian linearized inversion in green. All the predictors show a satisfactory match with the well
log data. In general, the results of Method 2 better capture the multimodal behavior of the rock and elastic properties. The
mismatch in the time range from 3615 ms to 3635 ms is probably due to the noise and limited resolution of seismic data.
In Fig. 9, we also compare the actual facies classication at the well location with the most likely facies model obtained
in Methods 1 and 2, as well as the standard Bayesian facies classication of the impedance and porosity estimated by the
standard Bayesian seismic inversion. The predicted models of the proposed methods show similar results including some
misclassications due to the limited bandwidth of seismic data and overall provide a better prediction than the standard
approach. To analyze the quality of the inversion, we calculate the mean absolute percentage error (Table 2). From this
quantitative analysis, we conclude that the results of Method 2 better match the well logs. These results corroborate the
L.P. de Figueiredo et al. / Journal of Computational Physics 336 (2017) 128142 137
Fig. 5. Maximum a posteriori of acoustic impedance and porosity (in A and B respectively) for the single Gaussian prior distribution, and point-wise means
of impedance and porosity of the multiple congurations of the Gibbs algorithm (in C and D respectively). The well log data ltered using a maximum
frequency of 100 Hz are shown for comparison.
Fig. 6. One realization of impedance and porosity (in A and B respectively) after 100 iterations of the Gibbs algorithm of Method 2, and point-wise standard
deviation models calculated from all the iterations for impedance and porosity (in C and D respectively).
observation that Method 2 better describes the multimodal behavior of the reservoir properties, which leads to a better
prediction of the reservoir volume (Fig. 8).
Using a computer with Intel Core (TM) i7-3930k 3.2 GHz processor and 50% of processing capacity, Method 1, using
a single Gaussian prior distribution, takes 10 minutes for the calculation of Equation (17) for all the traces in the entire
seismic volume. If we simplify the variance model by assuming a homogeneous variance, the computational cost decreases
to only 1 minute. For a non-homogeneous variance model, the covariance matrix f must be recomputed for each trace,
which leads to a higher computational cost. The Gibbs algorithm of Method 2, assuming a Gaussian mixture for the prior
138 L.P. de Figueiredo et al. / Journal of Computational Physics 336 (2017) 128142
Fig. 7. Probability of reservoir facies and most likely facies model obtained using the analytical approach and a single Gaussian prior distribution (plots A
and B) and obtained using a Gaussian mixture prior and the Gibbs algorithm (plots C and D).
Fig. 8. The isosurface with 70% of the reservoir facies probability by applying the Method 1 in A and Method 2 in B.
Table 2
Mean absolute percentage error of the estimated impedance and porosity at the well location used for the blind test.
distribution, takes 90 minutes for the generation of 115 models of the rock properties, using the same computer and the
same processing capacity.
4. Conclusions
We introduced an analytical formulation and a Monte Carlo algorithm for the inversion of seismic data to jointly es-
timate acoustic impedance, porosity and facies conditioned to well and seismic data. The input data are the post-stack
seismic amplitudes and traveltimes, a background subsurface model, the seismic wavelet and the variance parameters. The
innovative component of the proposed research is the denition of the joint prior distribution for impedance and porosity
based on a linearized rock-physics model for the analytical computation of the joint distribution in the Bayesian framework.
The advantage of the rst method is that the solution of the Bayesian inverse problem is analytical, since the prior dis-
tribution is Gaussian and the model is linear. The analytical approach allows a fast assessment of the model parameters.
However, this approach cannot describe multimodal data and facies-dependent rock-physics relations. The advantage of the
L.P. de Figueiredo et al. / Journal of Computational Physics 336 (2017) 128142 139
Fig. 9. Blind well test. The rst two plots show impedance and porosity (well log in black, MAP of Method 1 in light blue, posterior distribution of the
Method 2 in color scale with mean in blue and MAP solution in dashed blue, and MAP of standard Bayesian inversion in green). The other plots show
the actual facies classication at the well, most likely facies model obtained by Method 1, by Method 2 and by the standard Bayesian approach. (For
interpretation of the colors in this gure, the reader is referred to the web version of this article.)
second method is the multimodal prior distribution and the facies-dependent rock-physics linearization, which leads to a
more accurate approximation of the relation between the model parameters. However, this approach requires a numerical
evaluation of the posterior distribution through a Gibbs sampling algorithm, resulting in larger computational cost.
Acknowledgements
The authors would like to acknowledge Petrobras for the support and availability during the research project and CGG
Jason for providing the academic licensed software used in this research. Figueiredo L.P. acknowledges CAPES for the schol-
arship (process No. 3224/15-5). Figueiredo W. acknowledges the nancial support from the Brazilian agency CNPq (Grant
No. 303253/2013-4) and INCT-FCX (FAPESP-CNPq Grant No. 573560/2008-0).
A rock-physics model is a set of equation that links rock and uid properties of a porous rock to their elastic response.
In our application we focus on rock physics relations between porosity and P-impedance (or its logarithm). These relations
can be estimated, in poroelasticity and effective-medium theory, in terms of the elastic behavior of spherical or ellipsoidal
cavities and inclusions. Several inclusion models are available in literature and are generally suitable in carbonate rocks [25].
[21] derived expressions for bulk and shear moduli, K s and G s of a saturated rock by using a long-wavelength rst-order
scattering theory
K m + 43 G m
(K s Km) = K f Km P , (A.1)
4
Ks + G
3 m
(G m + )
(G s G m ) = G m Q , (A.2)
(G s + )
where
G m 9K m + 8G m
= , (A.3)
6 K m + 2G m
140 L.P. de Figueiredo et al. / Journal of Computational Physics 336 (2017) 128142
K m and G m are the bulk and shear moduli of the solid phase (matrix), K f is the bulk modulus of the uid phase, is
the porosity of the porous rock, and P and Q are coecients that describe the effect of the inclusion of the uid in the
background matrix. These expressions are written for a variety of inclusion shapes. In this work, we assume that the pores
are spherical; therefore, according to KusterToksoz equations:
K m + 43 G m
P= , (A.4)
K f + 43 G m
Gm +
Q = . (A.5)
If we explicitly solve for the bulk and shear moduli, we obtain:
4K m G m + 3K m K f + 4G m K f 4K m G m
Ks = , (A.6)
4G m + 3K f 3K f + 3K m
G m (9K m + 8G m )(1 )
Gs = . (A.7)
9K m + 8G m + 6( K m + 2G m )
From the elastic moduli, we can estimate the P-wave velocity of the saturated rock as:
K s + 43 G s
Vp = , (A.8)
where is the density of the saturated rock:
= m (1 ) + f , (A.9)
where m and f are the densities of the solid and uid phases respectively. Finally P-impedance is dened as:
z = V p = f K T (). (A.10)
Kuster Toksoz model f K T is non-linear respect to porosity. In order to derive a linearization of the following rock-physics
model:
g () = g (G ) + g (G )( G ) = g (G ) + g (G ) g (G )G = A + B . (A.12)
(b a)( f m ) + (c d + e + f )
g () = , (A.13)
2(b a)
where
4G m (9K m + 8G m )( 1)
a= ,
3(9K m + 8G m + 6( K m + 2G m ))
3K f K m + 4G m ( K m + K f K m )
b= ,
4G m + 3( K f K f + K m )
8G m ( K m + 2G m )(9K m + 8G m )( 1)
c= ,
(9K m + 8G m + 6( K m + 2G m ))2
4G m (9K m + 8G m )
d= ,
3(9K m + 8m + 6( K m + 2G m ))
3( K f K m )(3K f K m + 4G m ( K m + K f K m ))
e= 2 ,
4G m + 3( K f K f + K m )
4G m ( K f K m )
f = .
4G m + 3( K f K f + K m )
L.P. de Figueiredo et al. / Journal of Computational Physics 336 (2017) 128142 141
The following theorems [1,2] allow computing the joint distribution of the properties of interest conditioned by the
observed data.
Theorem 1. Let x Rn be distributed according to N n (x; , ). Then y = T x, is distributed according to N n ( y ; T , T T T ), for T
nonsingular.
Theorem 2. Let the components of x Rn , x N n (x; , ) be divided into two sub-vectors x1 e x2 . Suppose the mean is similarly
divided into 1 e 2 , and suppose the covariance matrix is divided into 11 , 12 e 22 , the covariance matrices of x1 , of x1 and x2 ,
1
and of x2 , respectively. Then, the conditional distribution of x1 given x2 is Gaussian with mean 1|2 = 1 + 12 22 x2 2 and
1
covariance matrix 1|2 = 11 12 22 21 .
The multivariate Gaussian distribution in the canonical notation can be written as:
T 1 T 1
p (x) = exp + x x x , (B.1)
2
where
1
= 1 , = d ln(2 ) + ln|| + T .
2
Using this notation, the product of n distributions is
T
n
n
n
1
n
p i (x) = exp i + i x xT 1
i x . (B.2)
2
i =1 i =1 i =1 i =1
n
n
1
s =
i
1
, s = i , (B.3)
i =1 i =1
1
s = d ln(2 ) + ln| s | + sT s s . (B.4)
2
The Equation (B.2) can be written as
n
n
1
T T 1
p i (x) = exp i s exp s + s x x s x , (B.5)
2
i =1 i =1
or
n
n
p i (x) = exp i s N (x; s , s ). (B.6)
i =1 i =1
Therefore, the product of n Gaussian distributions results in another Gaussian distribution multiplied by a scale factor.
References
[1] T. Anderson, An Introduction to Multivariate Statistical Analysis, 3rd edition, Wiley Series in Probability and Mathematical Statistics, Wiley, 2003.
[2] M. Bilodeau, D. Brenner, Theory of Multivariate Statistics, Springer Texts in Statistics, Springer, New York, 2008.
[3] M. Bosch, C. Carvajal, J. Rodrigues, A. Torres, M. Aldana, J. Sierra, Petrophysical seismic inversion conditioned to well-log data: methods and application
to a gas reservoir, Geophysics 74 (2) (2009) O1O15.
[4] M. Bosch, T. Mukerji, E. Gonzalez, Seismic inversion for reservoir properties combining statistical rock physics and geostatistics: a review, Geophysics
75 (5) (2010) 75A165.
[5] S. Brooks, A. Gelman, G. Jones, X. Meng, Handbook of Markov Chain Monte Carlo, Chapman & Hall/CRC Handbooks of Modern Statistical Methods,
Taylor & Francis, 2011.
[6] A. Buland, H. Omre, Bayesian linearized AVO inversion, Geophysics 68 (1) (2003) 185198.
[7] A. Buland, H. Omre, Joint AVO inversion, wavelet estimation and noise-level estimation using a spatially coupled hierarchical Bayesian model, Geophys.
Prospect. 51 (6) (2003) 531550.
142 L.P. de Figueiredo et al. / Journal of Computational Physics 336 (2017) 128142