Detection and Spatial Delineation of

Thin-Sand Sedimentary Sequences

With Joint Stochastic Inversion of
Well Logs and 3D Prestack Seismic
Amplitude Data
Germán D. Merletti* and Carlos Torres-Verdín, The University of Texas at Austin

Summary properties (Castagna and Backus 1993). More sophisticated appli-

We describe the successful application of a new prestack stochastic cations consist of directly determining spatial distributions of rock
inversion algorithm to the spatial delineation of thin reservoir units properties that best reproduce recorded seismic amplitudes (Sheriff
otherwise poorly defined with deterministic inversion procedures. and Geldart 1982). These procedures make use of seismic inver-
The inversion algorithm effectively combines the high vertical sion algorithms implemented in fully or partially stacked normal-
resolution of wireline logs with the relatively dense horizontal moveout (NMO) corrected gathers. The latter approach, referred
coverage of 3D prestack seismic amplitude data. Multiple partial- to in this paper as prestack inversion, attempts to capture enhanced
angle stacks of seismic amplitude data provide the degrees of degrees of freedom in the measurements to yield independent
freedom necessary to estimate spatial distributions of lithotype and estimates of density and P- and S-wave velocities (Contreras and
compressional-wave (P-wave) and shear-wave (S-wave) velocities Torres-Verdín 2005), which can be used to estimate lithology-
in a high-resolution stratigraphic/sedimentary grid. In turn, the dependent petrophysical properties (Goodway et al. 1997).
estimated volumes of P- and S-wave velocity permit the statistical Despite the fact that many variants of seismic inversion proce-
cosimulation of lithotype-dependent spatial distributions of poros- dures have been reported in the open technical literature, in this
ity and permeability. paper we classify them into two main categories: deterministic
The new stochastic inversion algorithm maximizes a Bayesian and stochastic methods. Geoscientists commonly use determin-
selection criterion to populate values of lithotype and P- and S- istic inversion methods in their first attempt to interpret seismic
wave velocities in the 3D simulation grid between wells. Property amplitude reflection data in terms of quantitative rock properties.
values are accepted by the Bayesian selection criterion only when To accomplish this, deterministic inversion algorithms minimize an
they increase the statistical correlation between the simulated and objective function that contains at least two additive terms (Debeye
recorded seismic amplitudes of all partial-angle stacks. Further- and van Riel 1990; Pendrel 2001): One of these terms controls the
more, inversion results are conditioned by the predefined measures misfit between recorded and synthetic seismic amplitudes, and a
of spatial correlation (variograms) of the unknown properties, second term controls the energy of the inverted reflectivities. The
their statistical cross correlation, and the assumed global lithotype inversion includes a stabilization factor that places a differential
proportions. weight between the metric of data misfit and the energy of the
Using field data acquired in a fluvial-deltaic sedimentary-rock inverted reflectivities. Practical applications of prestack inversion
sequence, we show that deterministic prestack seismic-inversion indicate that S-wave velocity and bulk density are more difficult
techniques fail to delineate thin reservoir units (10–15 m) pen- to estimate than P-wave velocity. Consequently, additional terms
etrated by wells because of insufficient vertical resolution and low are often included in the objective function to enforce soft nonlin-
contrast of elastic properties. By comparison, the new stochastic ear relationships between elastic properties (Fowler et al. 2000).
inversion yields spatial distributions of lithotype and elastic prop- Deterministic inversion procedures do not honor well logs in a
erties with a vertical resolution between 10–15 m that accurately direct manner, and, hence, the frequency bandwidth of the inverted
describe spatial trends of clinoform sedimentary sequences and products is the same as that of the recorded seismic amplitudes.
their associated reservoir units. Stochastic (or geostatistical) inversion techniques, on the other
Blind-well tests and cross validation of inversion results con- hand, use a completely different approach to estimate elastic and/or
firm the reliability of the estimated distributions of lithotype and petrophysical properties (Pendrel 2001). In their simplest form,
P- and S-wave velocities. Inversion results provide new insight to stochastic inversion methods consist of two sequential procedures:
the spatial and petrophysical character of existing flow units and The first one populates the interwell space with rock properties.
enable the efficient planning of primary and secondary hydrocar- This interpolation is performed in a 3D stratigraphic/structural
bon recovery operations. simulation grid with time sampling interval two- or four-fold the
original sampling interval of the recorded seismic amplitudes.
Introduction Subsequently, interpolated samples are entered into an algorithm
3D seismic amplitude data are commonly used to identify hydro- that numerically simulates seismic amplitudes to quantify the
carbon reservoirs and to assess their spatial continuity. Modern corresponding data misfit (Haas and Dubrule 1994). The stochas-
seismic processing techniques enable the interpretation of time- tic inversion algorithm will accept or reject the geostatistically
migrated common-midpoint (CMP) gathers in terms of amplitude- simulated property values depending on their agreement with the
vs.-offset (AVO) variations to estimate subsurface petrophysical recorded seismic amplitudes.
Because they explicitly combine well logs and seismic amplitude
variations, stochastic inversion procedures are particularly useful
* Now with BP Americas. to detect and delineate thin reservoirs otherwise poorly defined by
deterministic inversion techniques. In this paper, we refer to as “thin
Copyright © 2010 Society of Petroleum Engineers
reservoir” one whose thickness is below the vertical resolution of
This paper (SPE 102444) was accepted for presentation at the SPE Annual Technical the available seismic amplitude data. Some authors have success-
Conference and Exhibition, San Antonio, Texas, USA, 24–27 September 2006, and revised
for publication. Original manuscript received for review 30 June 2006. Revised manuscript fully implemented post-stack stochastic inversion techniques to
received for review 9 January 2009. Paper peer approved 29 July 2009. improve the detection and spatial delineation of individual fluvial

246 April 2010 SPE Reservoir Evaluation & Engineering

hydrocarbon reservoirs within amalgamated sandstone complexes Data Set and Local Geology. The data set used for the study
(Torres-Verdín et al. 1999; Merletti et al. 2003). reported in this paper was acquired in an active hydrocarbon
The objective of this paper is to describe the successful imple- field located in Venezuela’s Barinas-Apure basin. Fig. 1 shows an
mentation of a new prestack stochastic inversion algorithm recently approximately 6.5×5.5-km2 subset of the 14.2×25.4-km2 seismic
developed by Fugro-Jason. The inversion algorithm effectively data used to implement the stochastic inversion procedure. The
combines the high vertical resolution of well logs with the dense seismic data were acquired with a time sampling interval of 4 ms
horizontally coverage of seismic amplitude data. We integrate these and a square bin size of 75 m. Spectral analysis within the hydro-
two complementary measurements with a stratigraphic geological carbon-producing units indicates a frequency content between 5
model constructed with quantitative knowledge of reservoir archi- and 65 Hz and a dominant frequency of 28 Hz. Wavelet extractions
tecture. Inverted products exhibit a vertical resolution intermediate evidenced an average phase spectrum of 90°. For a value of P-wave
between the vertical resolution of well logs and seismic amplitude velocity of 4500 m/s, we estimated a tuning wavelength of 30 m.
data. The relative increase of vertical resolution, compared to that This calculation indicates that the available seismic data do not
of seismic amplitude data, improves the interpretation of complex enable the direct detection of individual hydrocarbon-producing
delta progradations (clinoform geometries) of the fluvial-deltaic units in the area.
sedimentary rock sequences. In addition, we make use of analog The data subset includes three vertical and two deviated wells.
synthetic reservoir models constructed with the edited logs to Fig. 2 displays the well logs acquired in Well 37. Wireline mea-
verify the reliability and vertical resolution of inverted products surements usually consist of gamma ray, resistivity, neutron, bulk-
under controlled conditions of seismic data quality, well proximity, density, and P-sonic logs; S-wave sonic logs were available only
and number of wells. in Well 37. Well logs comprise several depth intervals in which
Previously documented experience with prestack stochastic the measurements are strongly affected by washouts and borehole
inversion considered the spatial delineation of relatively thick rugosity. We edited such intervals before computing petrophysical
sedimentary units, as well as the estimation of their porosity, per- properties. Two seismic-time horizons were available throughout the
meability, and water saturation (Contreras et al. 2005; Contreras entire seismic cube. These horizons coincide with the tops of the two
et al. 2006). However, in this paper we describe the first applica- main hydrocarbon-producing units of the basin. For simplicity, we
tion of prestack stochastic inversion to the delineation of laterally refer to these formation units as Formation E and Formation C.
varying thinly bedded reservoirs. Formation E is a 25- to 35-m-thick mixed carbonate and silici-
In the following sections, we first introduce the data set used clastic interval composed of limestone, dolostone, arenitic dolos-
to implement the stochastic inversion procedure. Secondly, we tone, and calcareous sandstone intercalated with shales and calcare-
describe the method used to approach the petrophysical evaluation ous shales. Fig. 2 describes the location of Formation E in the depth
of wireline logs. We perform deterministic inversion to verify its range of 10,700–10,800 ft. The depositional environment interpreted
limitations in detecting reservoir-scale heterogeneities. Subse- for this formation is a shallow marine carbonate platform with vari-
quently, we describe the new stochastic inversion as implemented able periods of exposure (Aquino et al. 1997). On the other hand,
with three partial-angle stacks. Later, and after applying rigorous Formation C, which is the target of our studies, is a 250- to 300
quality control to the inverted elastic properties, we cosimulate ft-thick medium- to coarse-grained sandstone (Fig. 2) with slight
petrophysical properties statistically related to elastic properties. amounts of shale. The base of this formation is interpreted as depos-
We further appraise determinist and stochastic inversion products ited by large bedload-dominated braided rivers filling incised valleys
with analog synthetic reservoir models and noise-free numerically (Bejarano 2000). Such deposits were transgressed and overlain by
simulated seismic amplitude data. Finally, we discuss relevant an extensive sheet of shoreface and braided delta sediments (center
observations concerning the interpretation of inversion results and Formation C). There were some episodes in which the rate of sedi-
summarize our conclusions. mentation was greater than the rate of accommodation (Bejarano

Fig. 1—Seismic-time horizon displaying the embedding geological structure and the location of vertical and deviated wells. The
right-hand side panel is an enlarged view of the 6.5×5.5 km2 seismic data subset used to implement stochastic inversion.

April 2010 SPE Reservoir Evaluation & Engineering 247

Fig. 2—Wireline logs acquired in Well 37. The far-right panel describes the lithology proportions (sandstone in yellow, carbon-
ate in red, and shale in dark grey) and porosity (cyan). The depth interval of interest considered in this paper corresponds to
Formation C (9,800–10,200 ft).

2000). This explains the extensive progradational stacking patterns Fig. 3 also displays an enlarged view of typical well-log
displayed in wireline logs and prograding clinoform geometries responses. The gamma ray log (shown in green) does not evidence
evidenced in seismic-amplitude sections. important lithology variations through the interval, whereas the
Fig. 3 is a seismic cross section displaying high-angle clino- log-derived porosity (shown in yellow) does display grain-size
forms. Dip angles of 4–5° measured at well locations outside the trends with better reservoir conditions toward the top. Either one or
study area support the interpretation of fluvial-dominated deltaic two upward-coarsening units are found in the interval represented
deposits rather than shoreface deposits (Walker and Plint 1992). by deltaic deposits; these trends are interpreted as the advance of

Fig. 3—Seismic cross section displaying some prograding clinoforms within Formation C. For reference, the gamma ray and
computed-porosity logs are posted along the trajectory of Well 55.

248 April 2010 SPE Reservoir Evaluation & Engineering

delta-front sands over prodelta and platform muds (Galloway and Fluid/Lithology Sensitivity Analysis. Three independent exer-
Hobday 1990). Fluvial deposits are also found in the upper sec- cises were performed to determine the sensitivity of elastic proper-
tion and are overlain by thin bioclastic carbonates, referred in this ties to variations of lithology, porosity, and pore fluid. We applied
paper to as Formation B. Given the small thickness of the basal these techniques in the following order: (1) numerical simulation
fluvial deposits and the strong seismic interference produced by of synthetic seismic gathers (together with analysis of recorded
carbonates in the upper Formation C, we focus our efforts on seismic gathers), (2) Biot-Gassmann fluid substitution, and (3)
improving the spatial delineation of deltaic deposits in areas in crossplot analysis. Before the application of these techniques, we
which clinoform configurations are poorly defined or not defined reconstructed several synthetic S-wave velocity logs because the
at all. The average thickness of individual reservoirs is 10–15 m, original data set included only one measured log. Log reconstruc-
although in some cases reservoir units appear as amalgamated lay- tion included: (1) estimation of elastic constants (bulk and shear
ers whose thickness is close to the vertical resolution of seismic moduli) for dry rock given initial values of P-wave velocity,
amplitude data (30 m). Both the high degree of compaction of porosity, and water saturation (Hilterman 1983) and (2) calcula-
reservoir units and the absence of thicker shaly intervals between tion of S-wave velocity using the estimated shear modulus and the
them make it very difficult to delineate individual reservoir units measured bulk density.
with conventional seismic interpretation techniques. First, we simulated NMO corrected synthetic gathers using well
logs. Synthetic gathers are the result of the convolution of a previ-
Petrophysical Analysis. We synthesized volume-of-shale, poros- ously estimated wavelet with well-log angle-dependent reflectivi-
ity, water saturation, and permeability logs using standard petro- ties. Reflectivities were computed by means of Knott-Zoeppritz
physical evaluation techniques. Volumetric shale concentration equations (Aki and Richards 1980) using recorded and synthetic
(Csh) was calculated using a linear shale index (Ish) (Dewan P- and S-wave velocities, and density logs. The simulated seismic
1983). Second, we computed porosity from the density log using amplitude gathers are displayed in Fig. 5a, and the measured
a quartz-mineral matrix and a single-fluid (brine) model. Given gathers in Fig. 5b. Differences between simulated and measured
that neither abundant shaliness nor gas is found within producing amplitudes can be attributed to the way in which amplitudes were
reservoirs, an arithmetic average of shale-corrected neutron and generated. Specifically, simulated gathers were computed from
density porosities is a good approximation of nonshale porosity. log-derived elastic properties, which are representative of a volume
We simultaneously computed water saturation and fluid-corrected of investigation a few inches from the borehole wall. On the other
porosity using Archie’s model with an iterative minimization hand, measured gathers represent the average of much larger rock
procedure that corrected the error introduced by the assumption volumes of sensitivity. Both synthetic and recorded seismic gath-
of brine saturation during the previous estimation of porosity. ers exhibit a monotonic decrease of amplitude with an increase of
Calculation of irreducible water saturation logs was the first angle of incidence. This behavior corresponds to the classic AVO
step in computing permeability. We then used linear multivariate response of Class-I sands (Rutherford and Williams 1989), which
regression with well-log-derived porosity and irreducible water is typical of sediments that underwent a gradual burial history.
saturation to estimate permeability. Fig. 4 shows a good agreement Secondly, we performed fluid-substitution analysis to determine
between core- and log-derived permeabilities within hydrocarbon the influence of pore-fluid saturation on P- and S-wave velocities.
producing units. In so doing, we applied Biot-Gassmann’s equations assuming a

Fig. 4—Results from standard petrophysical evaluation of wireline logs acquired in Well 67. There is a good agreement between
the available rock-core data and well-log-derived porosity and permeability.

April 2010 SPE Reservoir Evaluation & Engineering 249

 (a) (b) (c)

Fig. 5—Synthetic (a) and recorded (b) angle gathers in Well 84 exhibit a decrease of amplitude with an increase of angle of
incidence (typical behavior of Class-I reservoirs). Synthetic angle gathers were computed using angle-dependent reflectivities
and extracted wavelets. For reference, Fig. 5c shows the measured P-wave velocity log.

constant porosity of 20% (this is the largest value of porosity Crossplot analysis of well log-derived elastic properties (P-
computed from well logs). Fig. 6 shows that P-wave velocity (red and S-wave velocities, and density) and petrophysical properties
curve) is nearly insensitive to mixtures of oil and water filling the (porosity and permeability) shows that the cleanest sandstones are
pore space; this figure also displays the P-wave velocity for the associated with values of P- and S-wave velocity smaller than 4300
case of gas saturation instead of liquid hydrocarbon saturation and 2600 m/s, respectively (Fig. 7). On the other hand, high values
(green curve). The analysis indicates that S-wave velocity (blue of elastic properties are associated with sandstones containing
curve) does not change appreciably over the entire range of fluid larger volumes of fine materials or smaller grain sizes. In addition,
saturation. As predicted by Biot-Gassmann’s equations applied to we computed the fluid- and lithology-sensitive Lamè petrophysi-
consolidated sediments, both rigidity and bulk moduli increase cal parameters, LambdaRho and MuRho (Goodway et al. 1997).
with compaction, thereby minimizing the effect of fluid saturation Values of MuRho below 30 GPa·g/cm3 indicate good-quality rocks
on the measured rock velocities (Tatham et al. 1991). (Fig. 8). Fig. 9 displays the crossplot of these parameters. As

Vp modeled
Vs modeled
Vp gas
Vp wells
Vs wells

Fig. 6—Biot-Gassmann fluid substitution indicates that P-wave velocity (red curve) does not change significantly with an increase
of oil saturation. There is a larger variation of P-wave velocity for the hypothetical case of gas saturating the pore space (green
curve), but this occurrence has not been reported in the study area. S-wave velocity (blue curve) remains almost constant with
changes of oil/water ratio. Numbered symbols identify average P- and S-wave velocities (red and blue, respectively) selected
from well-log readings of 20% porosity.

250 April 2010 SPE Reservoir Evaluation & Engineering

 shaly sands

clean sands

Fig. 7—Crossplot of P- and S-wave velocities constructed with wireline logs acquired in Well 37. The color code identifies the
corresponding value of porosity. Porous sands exhibit low values of P- and S-wave velocity, whereas shaly sands are associated
with high values of P- and S-wave velocity. Histograms evidence the existence of two distinct populations of samples clearly
differentiated by P- and S-wave velocities.

expected, MuRho is an excellent discriminator of facies, whereas Because of the lack of sensitivity of elastic parameters to
LambdaRho shows the same dynamic range for all the facies. Such variations of oil/water saturation, in what follows we focus our
behavior supported our previous observations stemming from the attention only on estimating spatial distributions of porosity and
Biot-Gassmann behavior of oil/water mixtures: Elastic parameters permeability from inverted elastic properties. Crossplot and histo-
remain almost insensitive to variations of oil/water saturation. gram analysis indicated that P- and S-wave velocities exhibited a

shaly sands

clean sands

Fig. 8—Crossplot of Vp /Vs ratio and density constructed with wireline logs acquired in Well 37. The color code identifies the
corresponding value of porosity. Density is an excellent discriminator of porous and shaly sands whereas the Vp /Vs ratio is not
because the two sand groups are associated with the same dynamic range (1.55–1.75).

April 2010 SPE Reservoir Evaluation & Engineering 251

 shaly sands

clean sands

LambdaRho [GPa⋅g/cm3]

Fig. 9—Crossplot of LambdaRho and MuRho constructed with wireline logs acquired in Well 37. The color code identifies the
corresponding value of porosity. Porous sands are associated with low values of MuRho; the vertical histogram indicates that
this parameter is an excellent discriminator of rock quality. LambdaRho, in turn, does not differentiate petrophysical properties
(all the samples fall within the range of 23 and 35 GPa·g/cm3).

bimodal distribution of samples corresponding to an equal num- not enforce value-range constraints when minimizing the objec-
ber of lithotypes: (1) clean sands, including porous and perme- tive function.
able intervals, and (2) shaly sands, including finely grained and Fig. 10 is a three-panel seismic-time cross section that shows
shale-laminated sands. Such an approach for the differentiation of the deterministically inverted volumes of P-wave and S-wave
lithotype based on the combined use of P- and S-wave velocities impedances (impedance being the product of velocity and density)
is attractive because it permits the determination of lithotype from and density, together with the gamma ray and computed porosity
seismic inversion products. logs posted along two well trajectories. Despite the low seismic
residuals yielded by the inversion (low data misfit), the computed
Deterministic Inversion. NMO-corrected partial-angle stacks porosity log clearly indicates that inverted parameters are unable to
were entered to inversion on the basis of comparisons between syn- reproduce reservoir-scale heterogeneities. From direct comparison
thetic (numerically simulated) and recorded angle gathers. Angu- of well logs and inverted products along well trajectories, we found
lar ranges were available from seismic amplitude data between that the vertical resolution of deterministically inverted volumes
5 and 25°. For the study described in this paper, we chose the of P- and S-wave impedance was approximately 25 m. As noted
following angle ranges: 5–14°, 10–19°, and 16–25°. In an effort earlier, the relatively high degree of compaction of reservoir units
to improve signal-to-noise ratios, a small overlap between angle causes a small contrast of elastic properties between reservoir
ranges was included in the separation of angle ranges, given the units and their embedding rocks, and this makes it very difficult to
limited number of traces available for stacking. Angle-dependent delineate individual reservoir units from deterministically inverted
wavelets were estimated for each partial-angle stack in two steps: elastic properties.
(a) calculation of angle-dependent reflectivities (for the angle
ranges listed above) using density and P- and S-wave velocity Stochastic Inversion. In most cases, low vertical resolution of
logs and (b) wavelet estimation by minimizing the least-squares seismic amplitude data is the main motivation for implementing
difference between simulated and recorded partial-angle stacks in stochastic inversion procedures. However, in this paper, the choice
the vicinity of wells. of stochastic inversion is equally motivated by the geometrical
The deterministic inversion used in this paper is an algorithm properties of the reservoir units under consideration. Specifically,
developed by Fugro-Jason as an extension to nonzero offsets of and as noted in Fig. 3, there are several reservoir sections in which
their constrained sparse-spike inversion (CSSI) (Debeye and van clinoform geometries are well defined by seismic amplitude data
Riel 1990). This algorithm minimizes an objective function that (left-hand side of the section), whereas there are other places where
combines the L1-norm of reflectivity series and the L2-norm of the such geometries are poorly defined, whereby their existence had to
seismic misfit (expressed as the difference between simulated and be corroborated with wireline logs and well-derived petrophysical
recorded seismic amplitudes). The interplay of these two additive logs (right-hand side of the section). The integration of seismic
terms in the inversion is adjusted with a stabilization factor (or amplitude data, well logs, and geological knowledge of reservoir
Lagrange multiplier). A third additive term uses the L1-norm met- architecture allowed us to construct a high-resolution stratigraphic/
ric to bias low-frequency components of inverted products toward structural property volume with one-fourth of the original time
low-frequency models constructed with the interpolation of well sampling rate of seismic amplitude data.
logs. We emphasize that, for the deterministic inversion exercises Given that not only low-frequency components of wireline logs
described in this paper, we did not enforce an explicit relationship are used by the stochastic inversion but also some of their high-fre-
between density and P- and S-wave velocities. Likewise, we did quency components, we performed a strict well-log quality control

252 April 2010 SPE Reservoir Evaluation & Engineering

 [g/cm3⋅m/s]

(a)

[g/cm3⋅m/s]

(b)

[g/cm3]

(c)

Fig. 10—Seismic-time cross section showing the results of deterministic inversion of prestack seismic amplitude data: P-impedance,
S-impedance, and density. The well-log-derived porosity (the curve posted on the right-hand side of each well track) indicates
that inverted property volumes are unable to resolve reservoir-scale heterogeneities (some of these heterogeneities are indicated
with white arrows).

to identify intervals during which wireline logs did not respond to residual) between simulated and recorded seismic traces. The final
formation properties. In those sections, we computed synthetic logs point of the random walk is the location in model space in which
(P- and S-wave velocities and density) following three steps: (a) the posterior PDF exhibits a local maximum.
assessment of lithology fractions by means of multilinear regres- In this project, a priori models (volumes of lithotype, P-wave
sion, (b) calculation of synthetic logs using the product of lithology velocity, S-wave velocity, and density) were constructed with
fractions and theoretical readings for pure-mineral compositions, sequential Gaussian simulation (SGS) (Chilès and Delfiner 1999)
and (c) replacement of measured logs with synthetic logs at places of well-log properties using predefined PDFs and assumed mea-
where caliper readings exceeded 10% of the nominal bit size. sures of spatial correlation (variograms). We selected the variogram
Partial-angle stacks, angle-dependent wavelets, high-resolution (1 model and its properties (lateral and vertical ranges) with adher-
ms) stratigraphic/structural framework, lithotype (clean and shaly ence to outcrop geological analogs. In addition, the construction of
sands) logs, and corrected-by-caliper P- and S-wave velocities and a priori models enforced predefined lithotype-dependent statistical
density logs formed the input data set used for prestack stochastic cross-correlations (joint PDFs) between P-wave velocity, S-wave
inversion. velocity, and density constructed with well logs. We estimated joint
The stochastic inversion algorithm used in this paper is based PDFs of elastic properties (density, and P- and S-wave velocities)
on a Bayesian strategy with Markov-Chain Monte-Carlo (MCMC) based on two possibilities of lithotype: porous sands and shaly
updates (Gilks et al. 1996; Chen et al. 2000) to sample the poste- sands. Furthermore, simulations of lithotype were made to honor
rior (model space) probability density function (PDF). Beginning a predefined global measure of facies proportion.
with a prior PDF of the unknown model properties, the objective Figs. 11 and 12 show samples of properties collected from
of Bayesian inversion is to sample the posterior PDF to find all well logs of “shaly sands” and “clean sands,” respectively, for the
possible model realizations that honor the measurements within complete interval of interest; the same figures display projections
the predefined signal-to-noise ratio. MCMC implements a biased over 2D Cartesian planes of 3D joint PDFs constructed with well-
random walk to sample the posterior PDF. The trajectory of the log data. Fig. 13 displays a 3D joint PDF of elastic properties
walk is modified continuously to target points in model space confirming that each lithotype is associated with nonoverlapping
which honor the measurements and their uncertainties (Tarantola 3D subdomains of elastic properties.
2005). This procedure requires that a data likelihood function be Fig. 14 compares the high vertical resolution of stochastically
evaluated at each step of the random walk; in other words, such a inverted P-wave velocity (lower panel) against the deterministi-
procedure requires the repeated solution of the forward problem. cally inverted P-wave impedance (center panel) and the near
The numerical simulation consists of (a) computing angle-depen- partial-angle stack (upper panel). We note that both the expression
dent reflectivities from the simulated values of elastic properties, of clinoform geometries and the P-wave velocity contrast between
(b) convolving angle-dependent wavelets with angle-dependent embedding rocks and reservoir units have been enhanced by the
reflectivities to numerically simulate angle-dependent seismic stochastic inversion when compared to those of deterministically
traces, and (c) computing the metric of the difference (seismic inverted products. Fig. 15 shows seismic-time cross sections of the

April 2010 SPE Reservoir Evaluation & Engineering 253

 [g/cm3]

Fig. 11—Example of a 3D lithotype-dependent joint PDF of P-wave velocity, S-wave velocity, and density constructed with
wireline logs acquired in Well 37. Green cubes identify well-log samples for “shaly sands” gathered from the entire interval of
interest. Color-coded 2D Cartesian planes describe projections of the sampled 3D joint PDF, with the colors indicating relative
concentration of samples.

stochastically inverted elastic properties (P- and S-wave velocities non-uniqueness of inversion results. We observed this behavior
and density) in the 1-ms high-resolution stratigraphic framework. in the central and southern areas of the seismic data subset that,
These figures confirm the increase of vertical resolution of sto- incidentally, included the location of the five available wells.
chastically inverted products compared to that of the corresponding The best inversion results were obtained with a priori models
deterministically inverted products. Such a result comes as a direct constructed with variogram vertical and lateral ranges of 6 ms
consequence of the combined quantitative use of well logs, 3D and 900 m, respectively. Assumed lithotype proportions for such
prestack seismic amplitude data, and the geological/stratigraphic models were 30% for “clean sands” and 70% for “shaly sands.”
framework. Moreover, we found that the choice of variogram type did not exert
relevant influence on inversion results. However, for normalization
Analysis of Results. We found that the stochastically inverted purposes, all the inversion examples described in this paper were
property distributions remained sensitive to some of the parameters implemented with exponential variograms.
used to construct a priori models and to parameters associated Fig. 16 displays the normalized sum of quadratic differences
with the inversion process. To shed light on such dependence, we between angle-dependent synthetic and recorded seismic data as
analyzed several combinations of inversion parameters to deter- cross-correlation maps. The cross correlation is greater than 95%
mine their impact on the cross correlation between measured and in most places and decreases to 70% in the proximity of faults.
synthetic seismic amplitudes. In this context, synthetic amplitudes In addition to the appraisal of data fit (cross correlation), we
correspond to seismic amplitudes numerically simulated from the performed blind-well tests to appraise the choice of inversion
elastic-property models yielded by the inversion. Inversion param- parameters. Fig. 17 shows the stochastically inverted density, and
eters analyzed included: (1) assumed global lithotype proportions, P- and S-wave velocities at the location of a well not included in
(2) vertical and lateral ranges of the assumed variograms, (3) sig- the construction of the a priori model. The agreement between
nal-to-noise ratio of the seismic amplitude data, and (4) variogram measured and stochastically inverted P- and S-wave velocities is
type. Sensitivity analyses near available wells indicated that the remarkable (two first panels). Moreover, the vertical resolution
first two parameters listed above exerted the greatest influence on gained is consistent with the variability of well-log measurements
inversion results. We interpreted such dependence as caused by and helps to delineate the most important reservoir heterogene-
(1) the relatively low vertical resolution of the recorded seismic ities. For instance, the thin bed located between 2075 and 2085
amplitudes or (2) the large degree of amalgamation of coarse ms is successfully resolved by stochastic inversion (black curve),
deposits between successive clinoforms (possibly associated with whereas the same reservoir feature is vertically averaged with
a decrease in accommodation space). Any of these two adverse shouldering beds by deterministic inversion (red smooth curve).
conditions can obscure the imprint of dipping layer boundaries For the case of density, the inverted distribution does not match
on seismic amplitude data, thereby considerably increasing the the measured well-log density (far-right panel). We believe that

254 April 2010 SPE Reservoir Evaluation & Engineering

 Density
[g/cm3]

Fig. 12—Example of a 3D lithotype-dependent joint PDF of P-wave velocity, S-wave velocity, and density, constructed with
wireline logs acquired in Well 37. Red cubes identify well-log samples for “clean sands” gathered from the entire interval of
interest. Color-coded 2D Cartesian planes describe projections of the sampled 3D joint PDF, with the colors indicating relative
concentration of samples.

Density
[g/cm3]

Fig. 13—3D joint PDF of P-wave velocity, S-wave velocity, and density constructed with wireline logs acquired in Well 37 and
displaying the non-overlapping model space occupied by the two lithotypes: clean sands (red) and shaly sands (green).

April 2010 SPE Reservoir Evaluation & Engineering 255

 (a)

[g/cm3⋅m/s]

(b)

(c)

Fig. 14—Seismic-time cross section emphasizing the enhancement of vertical resolution of the P-wave velocity volume (c) derived
from the stochastic inversion of prestack seismic amplitude data, (a) near partial-angle seismic stack, and (b) deterministically
inverted P-wave impedance panels are included for comparison.

the limited angular coverage available from the recorded seismic between P- and S-wave velocities and permeability, respectively,
gathers (25°) does not allow the reliable and accurate estimation for the lithotype “clean sands.” The same figures display the pro-
of density. Fig. 17 also compares the vertical resolution associated jection over 2D Cartesian planes of the 3D joint PDFs constructed
with deterministic and stochastic inversion products. On the basis with well-log samples.
of the comparison of the two results, we estimate the vertical reso- We cosimulated porosity and permeability from inverted elastic
lution of stochastic inversion products to be approximately 10–15 properties by constructing 3D joint PDFs of P-wave velocity, S-
m, whereas the resolution of deterministic inversion products is wave velocity, and each of these two petrophysical properties. In
estimated to be 25–30 m. We emphasize that the difference in verti- so doing, we decided not to include density in the cosimulations
cal resolution is because of the fact that the stochastic inversion because the stochastically inverted distributions of this property did
enforces a tight connection between seismic amplitude data and not agree with measured density logs at blind-well locations (Fig.
well logs, whereas the deterministic inversion does not. 17). The final petrophysical volumes were obtained by averaging
30 independent cosimulations of porosity and permeability.
Cosimulation of Petrophysical Properties. A common procedure After performing the cosimulation, we mapped eight porous
used to estimate petrophysical properties from inverted elastic and permeable units within the area of study, all of which exhibited
properties is to establish a statistical correlation between one elas- a north-northeast/south-southwest strike. Fig. 20 shows one such
tic property (for instance, P- or S-wave impedance) and one pet- reservoir unit, 1500–2000 m long and less than 1000 m wide; it
rophysical property, such as porosity (Pendrel and van Riel 1997). also describes the spatial distribution of the cosimulated porosity
Prestack stochastic inversion yields spatial distributions of density, within the reservoir unit.
and P- and S-wave velocities that, when properly cross validated,
provide the degrees of freedom necessary to reliably cosimulate Generation and Inversion of Analog Synthetic Reservoir
more than one petrophysical property. Our approach consists of Models. One of the central objectives of our study was to quantify
constructing 3D joint PDFs of elastic and petrophysical properties; the accuracy and reliability of the new stochastic inversion algorithm
namely, we quantify the degree of statistical correlation between to delineate thin, dipping reservoir geometries. As emphasized in
elastic and petrophysical properties with a lithotype-dependent previous sections, there are not well-defined clinoform structures
joint multidimensional PDF. Such joint PDFs are constructed from at existing borehole locations. On the other hand, clinoform geom-
lithotype-dependent well-log samples of elastic properties and etries are evidenced in the northern areas of the study that, coinci-
well-log derived petrophysical properties. Figs. 18 and 19 show dentally, are currently devoid of wells. In order to assess the ability
3D joint PDFs between P- and S-wave velocities and porosity, and of prestack seismic amplitude data to detect and spatially delineate

256 April 2010 SPE Reservoir Evaluation & Engineering

 (a)

(b)

Density, g/cm3

(c)

Fig. 15—Seismic-time cross sections of the elastic properties derived from the stochastic inversion of prestack seismic amplitude
data: P-wave velocity (a), S-wave velocity (b), and density (c).

Fig. 16—Maps of cross correlation between recorded and numerically simulated seismic amplitudes for three partial-angle seis-
mic stacks. The cross correlation is greater than 0.95 for most CMPs and decreases to 0.7 in the proximity of faults.

April 2010 SPE Reservoir Evaluation & Engineering 257

 (g/cm3⋅m/s) (g/cm3⋅m/s)

Density, g/cm3

(a) (b) (c)

Fig. 17—Blind-well inversion test performed at Well 17 between the depths of 2,040 and 2,100 ft. Stochastically inverted (black
curves) and measured (blue curves) P- and S- wave velocity logs [(a) and (b), respectively] agree very well. The inverted density
(c) does not agree well with well-log measurements. Red curves identify deterministically inverted density, P-wave impedance,
and S-wave impedance included in each panel for comparison.

Fig. 18—Example of a 3D lithotype-dependent joint PDF of P-wave velocity, S-wave velocity, and porosity, constructed with
wireline logs acquired in Well 37. Red cubes identify well-log samples for the lithotype “clean sands” gathered from the entire
interval of interest. Color-coded 2D Cartesian planes describe projections of the sampled 3D joint PDF, with the colors indicat-
ing relative concentration of samples.

258 April 2010 SPE Reservoir Evaluation & Engineering

Fig. 19—Example of a 3D lithotype-dependent joint PDF of P-wave velocity, S-wave velocity, and permeability, constructed with
wireline logs acquired in Well 37. Red cubes identify well-log samples for “clean sands” gathered from the entire interval of
interest. Color-coded 2D Cartesian planes describe projections of the sampled 3D joint PDF, with the colors indicating relative
concentration of samples.

Fig. 20—Spatial distribution of the average value of porosity obtained from 30 independent cosimulations of porosity from sto-
chastically inverted distributions of lithotype and P- and S-wave velocity within one of the hydrocarbon-producing reservoirs.
The reservoir under consideration strikes in the north-northeast/south-southwest direction and was penetrated by Well 13 within
one of the best porosity zones in the area of study.

April 2010 SPE Reservoir Evaluation & Engineering 259

such clinoform reservoir configurations, we experimented with a inverted trace of the same property. This comparison shows that
controlled appraisal environment that included perfectly known the upper reservoir unit (identified with the number “1”) is not
elastic-property volumes and noise-free seismic amplitude data. resolved with deterministic inversion.
The objective of this exercise was to analyze the dependence of The synthetic models used for simulation included reservoir
inversion products on a priori models when the seismic amplitude units with thickness above and below the vertical resolution of
data were sensitive to the actual reservoir geometry. In addition, the simulated seismic amplitude data (Fig. 22). We first analyzed
we tested the reliability of the new stochastic inversion to delineate the inversion parameter that exerted the strongest influence on
thin dipping layers when the a priori model was constructed with inversion products obtained from field data, namely the assumed
incorrect horizontal layers. lithotype proportion. The upper panels of Fig. 23 display initial
The construction of synthetic models and the simulation of the models constructed with assumed lithotype proportions between
corresponding seismic amplitudes used for appraisal were made clean sand and shaly sands of 0.2/0.8 (Fig. 23a) and 0.4/0.6 (Fig.
consistent with reservoir-unit features and acquisition parameters 23b); remaining inversion parameters were kept the same as in
of the field data set discussed earlier. These choices helped us to previous sections of this paper. The impact of the difference of
further cross validate the reliability of inversion results obtained initial models on inversion results is evident in the cross sections
with field data. of Fig. 23a. Lower panels display inversion results derived from the
Prestack seismic amplitude data were simulated using noise- synthetic models. We note that lithotype distributions are almost
free convolution that used log-derived elastic properties, angle- identical, thereby indicating a negligible dependence of inversion
dependent wavelets, and a stratigraphic framework, which repro- products on the assumed global lithotype proportion.
duced clinoform geometries observed in the northern sectors of the Fig. 24 describes the sensitivity of inversion results (P-wave
study area. We simulated three partial-angle stacks using the fol- velocity) to the choice of lateral variogram range. Blue colors
lowing steps: (a) interpolation of elastic properties from synthetic identify low-velocity values typical of reservoir units in the actual
well logs following the stratigraphic microlayering imposed by the field data. The upper panels of Fig. 24 describe a priori models
geological model; (b) generation of angle-dependent reflectivities constructed with lateral variogram ranges of 900 (Fig. 24a) and
using Knott-Zoeppritz’s equations for angular ranges of 5–14°, 1500 m (Fig. 24b). We confirm that variogram lateral range exerts a
10–19°, 16–25°; and (c) convolution of modeled reflectivity vol- significant control over the inverted spatial distribution of reservoir
umes with their corresponding angle-dependent wavelet. Wavelets units. For purposes of cross validation of inversion results, Well
used in the convolution had a maximum frequency of 70 Hz and S3 (see Fig. 24) was not included in the construction of a priori
a dominant frequency of approximately 32 Hz. Fig. 21 displays property models. We note that the upper reservoir unit in Well S3,
the models of elastic properties (density, and P-wave and S-wave indicated with the white arrows, is poorly defined by the synthetic
impedances) constructed to generate synthetic partial-angle stacks, seismic amplitude data. The lower panels of Fig. 24 show spatial
together with the simulated near, middle, and far partial-angle distributions of P-wave velocity inverted with different values of
stacks. Figs. 22a and 22b show two different seismic cross sec- lateral variogram range. In both cases, the thin layer in the upper
tions, one evidencing clinoform reflections on the measured near section of Well S3 is delineated accurately regardless of the var-
partial-angle stack (Fig. 22b) and the other describing our attempt iogram lateral range used to construct the a priori model. Fig. 25
to reproduce the influence of such a geometry on prestack seismic displays the results of a blind-well test in the location of Well S3.
amplitudes (Fig. 22a). Fig. 22c describes the P-wave impedance We note that the same thin reservoir unit, which was not resolved
volume obtained from the deterministic inversion of synthetic with deterministic inversion (Fig. 22), is now resolved accurately
seismic amplitude data. We note that thick reservoir units coincide with stochastic inversion.
with relatively low values of P-wave impedance (blue colors). Finally, we performed a sensitivity analysis of inversion prod-
The lower panel of the same figure compares the original P-wave ucts to the selection of stratigraphic framework. Fig 26 describes
impedance log in Well S3 against the nearest deterministically the results from this exercise. The upper panels in that figure

Density, g/cm3

(a) (b)

Fig. 21—Synthetic model of elastic properties (P-wave velocity, S-wave velocity, and density) interpolated from well logs and
used to generate the synthetic prestack seismic amplitude data (a). The three synthetic partial-angle stacks (PAS) numerically
simulated with the convolution of angle-dependent wavelets and the corresponding angle-dependent reflectivity volumes (b).

260 April 2010 SPE Reservoir Evaluation & Engineering

 (a)
(b)

[m/s⋅g/cm3]

(c)

(d)

Fig. 22—Seismic-time cross sections of synthetic (a) and recorded (b) near partial-angle seismic stacks displaying the spatial
distribution of reservoir units included in Formation C. The intermediate panel shows a cross section of the P-wave impedance
volume deterministically inverted from synthetic data. We note that thick reservoir units coincide with low values of P-wave
impedance (blue colors). Fig. 22d shows the P-wave impedance log in Well S3 (blue curve) and the nearest inverted trace of the
same property (red curve); it indicates that the upper reservoir (indicated with the number 1) is not resolved by the deterministic
inversion because its thickness is below the vertical resolution of the input prestack seismic amplitude data.

(a) (b)

Fig. 23—The upper panels describe a priori distributions of lithotype (clean sands shown in yellow and shaly sands shown in
green) for assumed global lithotype proportion ratios of 0.2/0.8 (a) and 0.4/0.6 (b). Remaining inversion parameters were kept
constant when performing the two stochastic inversions. The lower panels describe the inverted spatial distributions of lithotype.
White arrows indicate places where the inverted lithotype distributions give rise to the same stratigraphic features even though
the corresponding a priori models are not the same.

describe models constructed with the interpolation of lithotype Discussion

fractions (Fig. 26a) and P-wave velocity (Fig. 26b) using an The synthetic inversion exercises described above indicate that the
incorrect horizontal stratigraphic framework. Moreover, we did dependence of inversion products on a priori models significantly
not include Wells S3 and S4 in the construction of a priori models decreases with an increase in the quality and vertical resolution
to further bias the initial premisses of the inversion. The central of the input prestack seismic amplitude data. Partial-angle stacks
panels of Fig. 26 show inversion results obtained with the use of effectively constrain the stochastic inversion when they are sensitive
incorrect a priori models. Inverted spatial distributions of lithotype to angle-dependent reflectivities along clinoform surfaces. Inverted
and P-wave velocity do evidence the actual dipping geometry near properties shown in the lower panels of Figs. 23 and 24 compare
wells used to construct the a priori models (S1, S2, and S5), despite well even though they originate from substantially different a priori
the use of an incorrect horizontal framework to guide the inver- models (upper panels). Blind-well tests indicate that, regardless of
sion. Likewise, inversion results agree very well with the spatial the a priori model, stochastic inversion accurately estimates the
distributions of lithotype and P-wave velocity near cross-validation thickness and lateral extent of the thin-layer reservoir intersecting
wells (S3 and S4). Well S3. In addition, when assuming an incorrect stratigraphic a

April 2010 SPE Reservoir Evaluation & Engineering 261

 (a) (b)

Fig. 24—Sensitivity of inversion products to the assumed variogram lateral range. The upper panels describe a priori models
of P-wave velocity constructed with a 900- (a) and 1500-m (b) lateral variogram range, and without the use of Well S3 (which
was kept aside as a blind well). The lower panels describe inverted products. In both cases, the thin layer included in the upper
section of Well S3 (indicated with a white arrow) is delineated accurately by the inversion regardless of the parameters used to
construct the a priori model.

Density [g/cm3]

Fig. 25—Blind-well test of stochastic inversion products performed at the location of Well S3. The reservoir unit identified with
the number 1 is not resolved with deterministic inversion (see Fig. 22). The same reservoir unit was resolved consistently with
stochastic inversion regardless of the a priori model or the specific choice of inversion parameters (e.g., variogram type, vari-
ogram range, lithotype proportion).

priori model, stochastic inversion properly detected the thin-layer On the basis of the above observations with synthetic models,
reservoir but the geometrical reconstruction was poor. and fully aware of the limitations of signal-to-noise ratio, angular
When seismic amplitude data do not evidence the actual res- range, and vertical resolution of the recorded seismic amplitude
ervoir geometry, either because of their low vertical resolution or data in the study area, we directed a considerable portion of our
because of the high degree of facies amalgamation, the reliability efforts to the construction of a priori models that reproduced the
of inversion products becomes increasingly dependent on param- assumed reservoir architecture. In addition, we conducted a large
eters used to construct a priori models and on the choice of inver- number of cross-validation and blind-well tests to appraise the non-
sion parameters. In such cases, the stochastic inversion loses its uniqueness and vertical resolution of inversion results.
ability to narrow down the likely range of solutions in model space On the basis of our experience, we strongly recommend that
that simultaneously honor both the seismic amplitude data and the the stochastic inversion be preceded by a systematic analysis of the
well logs, thereby increasing the degree of non-uniqueness. influence of a priori models and inversion parameters on inversion

262 April 2010 SPE Reservoir Evaluation & Engineering

 (a) (b)

(c)

Fig. 26—Sensitivity of inversion products to the stratigraphic framework used to construct a priori models. The upper panels
describe the spatial distributions of lithotype (a) and P-wave velocity (b) interpolated with an incorrect horizontal stratigraphic
model (stratigraphic layering parallel to top and base). Note that the stratigraphic framework used to simulate the synthetic
seismic amplitude data is similar to the one shown in Fig. 26c. The middle panels describe the inversion products. The inverted
lithotype (a) and P-wave velocity (b) volumes indicate that the dipping-layer geometry is reconstructed properly by the inversion,
despite the use of an incorrect a priori model.

products. Numerical simulation of synthetic reservoir models quality logs, presence of several wells, and proximity of wells.
constructed with available wells and knowledge of stratigraphic Blind-well tests confirmed that absence of wells at key locations in
architecture provides sufficient guidance to quantify non-unique- the stratigraphic model decreased the resolution of inversion prod-
ness, reliability, and vertical resolution of inversion products. ucts to that of properties obtained with deterministic inversion.

Conclusions Acknowledgments
Joint stochastic inversion of well logs and prestack seismic ampli- The authors would like to thank PDVSA for providing the data set
tude data enabled the accurate detection and spatial delineation used for this research study. We remain grateful to Fugro-Jason for
of reservoir units otherwise poorly resolved with deterministic their software support and to Elizabeth Fisher for her technical
seismic inversion. The stochastic inversion of field data pursued in advice during the inversion phase of the study. Our gratitude goes
this paper yielded P- and S-wave velocity volumes with a vertical to Jeff Kane (Bureau of Economic Geology of The University of
resolution of 10–15 m, compared to the vertical resolution of 25– Texas at Austin) for his assistance during the well-log petrophysi-
30 m associated with deterministic inversion. This enhancement cal evaluation. The work reported in this paper was funded by The
in vertical resolution allowed us to delineate most reservoir-scale University of Texas at Austin’s Research Consortium on Forma-
heterogeneities. Moreover, cosimulation of porosity from inverted tion Evaluation, jointly sponsored by Anadarko, Aramco, Baker-
elastic properties enabled the spatial delineation of clean sands Hughes, BHP-Billiton, BP, BG, ConocoPhillips, Chevron, ENI,
with the best storage conditions. ExxonMobil, Halliburton, Hess, Marathon, Mexican Institute for
In cases in which the available seismic amplitude data were not Petroleum, Nexen, Petrobras, RWE, Schlumberger, StatoilHydro,
sensitive to reservoir geometry, we found that inversion products TOTAL, and Weatherford.
were biased by both the a priori models and the inversion param-
eters themselves. The only possible way to appraise non-unique- References
ness and reliability of inversion products derived from similar data Aki, K. and Richards, P.G. 1980. Quantitative Seismology, Theory and
sets is to perform a systematic study of the sensitivity of inversion Methods, 535. New York City: W.H. Freeman & Co.
products to a priori models and inversion parameters. Non-unique- Aquino, R., Figueroa, L., Prieto, M., Salazar, R., Garcia, E., Kupecz, J., and
ness of inversion products can be appraised with: (a) blind-well Hernandez, E. 1997. Sedimentological study of cores and correlation
tests at each well location and (b) checks of the cross correlation with well-logs, “O” Limestone, Barinas Basin, Venezuela. Abstract P
between recorded and simulated prestack seismic amplitudes. A5 presented at the Annual AAPG-SEPM-EDM-DPA-DEG Conven-
Exercises with synthetic analog models indicated that stochastic tion, Dallas, 6–9 April.
inversion successfully reproduced clinoform configurations with Bejarano, C. 2000. Sequence analysis and sedimentology of the Gobernador
noise-free seismic amplitude data. Formation, Barina Basin, Venezuela. Proc., VII Simposio Bolivariano de
The use of a priori models in close agreement with the under- Exploración Petrolera en las Cuencas Subandinas, Venezuela, 1–26.
lying stratigraphic framework effectively decreases the degree of Castagna, J.P. and Backus, M.M. 1993. Offset-Dependent Reflectivity—
non-uniqueness of inversion products. We obtained practically the Theory and Practice of AVO Analysis, No. 8, 348. Tulsa: Investigations
same spatial distributions of lithotype and elastic properties with in Geophysics, Society of Exploration Geophysicists.
different a priori models and inversion parameters. All stochastic Chen, M.-H., Shao, Q.-M., and Ibrahim, J.G. 2000. Monte Carlo Methods
inversion tests performed over synthetic models using the correct in Bayesian Computation, 375. New York City: Springer Series in
stratigraphic framework successfully detected and delineated thin- Statistics, Springer-Verlag.
bed reservoirs. Chilès, J.-P. and Delfiner, P. 1999. Geostatistics: Modeling Spatial Uncer-
The gain in vertical resolution of stochastic inversion, compared tainty, 695. New York City: Wylie Series in Probability and Statistics,
to that of deterministic inversion, is a direct consequence of good- John Wiley & Sons.

April 2010 SPE Reservoir Evaluation & Engineering 263

Contreras, A. and Torres-Verdín, C. 2005. Sensitivity analysis of fac- Rutherford, S. and Williams, R. 1989. Amplitude-versus-offset variations
tors controlling AVA simultaneous inversion of 3D partially stacked in gas sands. Geophysics 54 (6): 680–688.
seismic data: application to deepwater hydrocarbon reservoirs in the Sheriff, R.E. and Geldart, L.P. 1982. Exploration Seismology, 575. Cam-
central Gulf of Mexico. SEG Expanded Abstracts 24 (CH3): 464. doi: bridge, UK: Cambridge University Press.
10.1190/1.2144355. Tarantola, A. 2005. Inverse Problem Theory and Methods for Model Param-
Contreras, A., Torres-Verdín, C., Chesters, W., Kvien, K., and Fasnacht, eter Estimation, 333. Philadelphia, Pennsylvania, USA: SIAM.
T. 2005. Joint stochastic inversion of 3D pre-stack seismic data and Tatham, R.H., McCormack, M.D., Neitzel, E.B., and Winterstein, D.F.
well logs for high-resolution reservoir characterization and petrophysi- 1991. Multicomponent Seismology in Petroleum Exploration, No. 6,
cal modeling: application to deepwater hydrocarbon reservoirs in the 248. Tulsa, Oklahoma: Investigations in Geophysics, Society of Explo-
central Gulf of Mexico. SEG Expanded Abstracts 24 (RC2): 1343. doi: ration Geophysicists (SEG).
10.1190/1.2147935. Torres-Verdín, C., Victoria, M., Merletti, G., and Pendrel, J. 1999. Trace-
Contreras, A., Torres-Verdín, C., Chesters, W., Kvien, K., and Fasnacht, T. based and geostatistical inversion of 3-D seismic data for thin-sand
2006. Extrapolation of Flow Units Away From Wells With 3D Pre-Stack delineation: an application in San Jorge Basin, Argentina. The Leading
Seismic Amplitude Data: Field Example. Petrophysics 47 (3): 223–238. Edge 18 (9): 1070–1077.
Debeye, H.W.J. and van Riel, P. 1990. Lp-Norm Deconvolution. Geo- Walker, R.G. and Plint, A.G. 1992. Wave- and storm-dominated shallow
physical Prospecting 38 (4): 381–404. doi: 10.1111/j.1365-2478.1990. marine systems. In Facies Models: Response to Sea-Level Change,
tb01852.x. ed. R.G. Walker and N.P. James, 219–238. St. John’s, Newfoundland,
Dewan, J. 1983. Essentials of Modern Open-Hole Log Interpretation, 351. Canada: Geological Association of Canada.
Tulsa: PennWell Publishing Company.
Fowler, J., Bogaards, M., and Jenkins, G. 2000. Simultaneous Inversion
of the Ladybug prospect and derivation of a lithotype volume. SEG
Germán Merletti holds a degree in geology from the National
Expanded Abstracts 19 (RC4): 1517. doi: 10.1190/1.1815696.
U. of Cordoba, Argentina (1995). E-mail: German.Merletti@
Galloway, W. and Hobday, D. 1990. Terrigenous Clastic Depositional bp.com. He worked for YPF S.A. as a petrophysicist and res-
Systems: Applications to Fossil Fuel and Groundwater Resources, 485. ervoir geologist. In March 2000, Merletti began working with
Berlin: Springer-Verlag. integrated projects that included well-log correlation, petro-
Gilks, W.R., Richardson, S., and Spiegelhalter, D.J. 1996. Markov Chain physics and seismic inversion for Repsol-YPF. In August 2004,
Monte Carlo in Practice: Interdisciplinary Statistics, 486. Boca Raton, he jointed the Masters program at the Jackson School of
Florida, USA: Chapman & Hall/CRC Press. Geosciences of The U. of Texas at Austin where he worked
Goodway, W., Chen, T., and Downton, J. 1997. Improved AVO fluid detec- for the Research Consortium on Formation Evaluation as
tion and lithology discrimination using Lamè petrophysical parameters; Graduate Research Assistant on Seismic Inversion. Merletti
now works at BP as a petrophysicist and reservoir modeler.
“␭ ␳ “,”␮␳”,&”␭/␮ fluid stack”, from P and S inversions. SEG Expanded
Carlos Torres-Verdín holds a PhD degree in engineering geo-
Abstracts 16 (AVO2): 183. doi: 10.1190/1.1885795. science from the U of California at Berkeley (1991). E-mail:
Haas, A. and Dubrule, O. 1994. Geostatistical inversion—a sequential [email protected]. During 1991–1997, he held the posi-
method for stochastic reservoir modeling constrained by seismic data. tion of research scientist with Schlumberger-Doll Research.
First Break 12 (11): 561–569. From 1997–1999, Torres-Verdín was reservoir specialist and
Hilterman, F.J. 1983. Stratigraphic interpretation of seismic data. Continu- technology champion with YPF (Buenos Aires). Since 1999, he
ing Education Lecture Series, AAPG School, Houston. has been affiliated with the Department of Petroleum and
Merletti, G., Hlebszevitsch, J.C., and Torres-Verdín, C. 2003. Geostatis- Geosystems Engineering of the U. of Texas at Austin, where
tical inversion for the lateral delineation of thin-layer hydrocarbon he currently holds the position of Zarrow Centennial Professor
in Petroleum Engineering and conducts research on bore-
reservoirs: a case study in San Jorge Basin, Argentina. SEG Expanded
hole geophysics, formation evaluation, well logging, and inte-
Abstracts 22 (IT2): 662. doi: 10.1190/1.1818017. grated reservoir characterization. Torres-Verdín is the founder
Pendrel, J. 2001. Seismic Inversion—The Best Tool for Reservoir Charac- and director of the Research Consortium on Formation
terization. CSEG Recorder 26 (1): Feature Article. Evaluation at the U. of Texas at Austin. He is recipient of the
Pendrel, J. and van Riel, P. 1997. Methodology for seismic inversion and mod- 2006 Distinguished Technical Achievement Award from the
eling: a Western Canadian reef example. CSEG Recorder 12 (5): 5–15. SPWLA and the 2008 SPE Formation Evaluation.

264 April 2010 SPE Reservoir Evaluation & Engineering