Fast Geostatistical Stochastic Inversion in a Stratigraphic Grid

I. Escobar∗ , P. Williamson and A. Cherrett - Total UK

P.M. Doyen, R. Bornard, R. Moyen and T. Crozat - Compagnie Générale de Géophysique

SUMMARY Following the work of Buland and Omre (2003), we use a

Bayesian framework and a linearized AVA model to calculate
We have developed an efficient stochastic AVA inversion tech-
a log-Gaussian joint posterior distribution for the elastic pa-
nique that works directly in a fine-scale stratigraphic grid, and
rameters. We use an SGS procedure to decompose the joint
is conditioned by well data and multiple seismic angle stacks.
posterior into a series of local posteriors for each column of
We use a Bayesian framework and a linearized, weak con-
cells that are sampled sequentially to generate multiple real-
trast approximation of the Zoeppritz equation to construct a
izations of elastic properties. Our fast, layer-based stochastic
joint log-Gaussian posterior distribution for P- and S-wave
inversion procedure supports non-stationary spatial statistics
impedances. We apply a Sequential Gaussian Simulation al-
via vertically or laterally variable variogram models, incorpo-
gorithm to sample the posterior PDF. We perform a trace-by-
rated in the prior model. Non-stationary noise statistics can
trace decomposition of the global posterior into local poste-
also be handled by inputting trace-dependent S/N information.
rior distributions, conditioned by previously simulated traces.
Trace-by-trace sampling of the local PDFs generates multiple,
high-resolution realizations of the elastic properties. The new METHODOLOGY
sequential algorithm has been implemented to take full advan-
tage of parallel architectures and scales approximately linearly The proposed GeoSI algorithm relies on a linear forward
with the number of CPUs. The technique has been success- model linking elastic parameters of to seismic reflectivi-
fully tested using real data and a large layered model contain- ties. When considering for example the acoustic and shear
ing more than 30 × 106 grid-cells. impedances IP and IS , the angle-variant reflectivity for a single
interface can be approximated by (Fatti et al. (1994)):

INTRODUCTION 1 I¯S 2
rθ ≈ ∆ ln(IP ) − 4 sin2 θ ∆ ln(IS ) (1)
In order to overcome the band-limited nature of determinis- 2 cos2 θ I¯ 2
P
tic inversion methods (Oldenburg et al. (1983)), stochastic ap-
proaches have been proposed as a way of generating multi- In a stratigraphic grid, it is straightforward to extend this equa-
ple high-frequency realizations of elastic properties from seis- tion to a full vertical column i of layers. If we denote by mi
mic data (Francis (2002)). By construction, the realizations the discrete vector containing the logarithms of the IP and IS ,
are constrained to reproduce the observed seismic data, within the reflectivity series can be written as:
some noise-dependent tolerance limits and to honour condi-
tioning well data. The multiplicity of models reflects the in- ri,θ ≈ Ai,θ mi (2)
herent non-uniqueness of the inverse problem.
where Ai,θ differences ln(IP ) and ln(IS ) and multiplies them
Haas and Dubrule (1994) have developed a rejection-based with the coefficients of equation (1). Strictly speaking, this
stochastic method using trace-by-trace Sequential Gaussian equation is linear in log-impedances only if the squared
Simulation (SGS), where a simulated acoustic impedance trace impedance ratios from equation (1) are assumed to be known.
is accepted only if the corresponding match between synthetic Following Buland and Omre (2003), we express the inverse
and real seismic is good. More recently, Debeye et al. (1996) problem in a Bayesian framework with Gaussian PDFs. If the
and Contreras et al. (2005) have extended this work using sim- seismic noise is assumed to be uncorrelated across traces and
ulated annealing and MCMC techniques, respectively. These angles, the likelihood of seismic data s can be written as:
methods, which apply sample-by-sample model perturbations,
remain very computer-intensive. Buland et al. (2003) have p(s|m) ∝
developed an elegant Bayesian linearized AVA inversion ap- !
1
proach that works in the Fourier domain and allows fast calcu- exp − ∑ ∑(si,θ − Gi,θ mi ) Cs i,θ (si,θ − Gi,θ mi )
T −1
2 θ i
lation and sampling of a joint Gaussian posterior distribution
for the elastic parameters. However, their approach assumes (3)
that the data and model are sampled regularly in time and that
where Cs i,θ are the noise covariance matrices. The rectangu-
spatial covariance functions and noise covariance are strictly
lar matrices Gi,θ group together Ai,θ from equation (2) with
stationary.
the wavelet convolution and transform a model trace with ir-
To facilitate integration with reservoir modelling workflows, regular layers into a seismic trace with regular samples. On the
the Geostatistical Stochastic Inversion (GeoSI) presented here other hand, considering the IP and IS as a log-Gaussian random
operates in a stratigraphic grid defined in the time domain, with field, a Gaussian prior PDF for m can be defined as:
horizontal sampling fixed by the seismic bin size and verti-  
1
cal columns of cells with variable thickness, typically much p(m) ∝ exp − (m − µm )T C−1 m (m − µm ) (4)
smaller than the seismic resolution. 2
 Fast Geostatistical Stochastic Inversion in a Stratigraphic Grid

Figure 1: Sections showing inverted IP and IS . Top panels show the mean impedance of 25 realizations. Bottom panels show the prior mean models.
Measured IP at two wells is shown as black curve. Time axis is in milliseconds, impedances are in g/cc × m/s.

(a) Near angle (10 degrees) (b) Far angle (30 degrees)

Figure 2: QC sections across the center of the area. Top: real seismic data. Middle: difference between real and synthetic seismic data using mean
of 25 realizations. Bottom: standard deviation of 25 realizations. Measured P-wave impedance at two wells is shown as black curve. Time axis is
in milliseconds.

where µm and Cm are respectively the prior mean and covari-
ance matrix, which are not required to be stationary. Cm is
typically modelled using vertical and lateral variograms and
cross-correlations. The Gaussian posterior PDF p(m|s) is con-
structed by combining equations (3) and (4). When well mea-
surements are available, they can easily be integrated to the
posterior PDF with their uncertainties as an additional likeli-
hood term. We have developed an SGS-type algorithm to sam-
ple the posterior PDF: it is possible to perform a trace-by-trace
decomposition of the global PDF into a number of approxi-
mate local PDFs, which are conditional to previously visited
traces. These local Gaussian PDFs can be sampled easily and
 Fast Geostatistical Stochastic Inversion in a Stratigraphic Grid

multiple realizations can therefore be efficiently generated. obtained by averaging the 25 realizations.
As most of the computations can be shared between realiza- Prior mean models for IP and IS were obtained by 3-D krig-
tions, extra simulations can be generated at little computational ing interpolation of the well log data after Backus averaging
expense. The algorithm has been implemented to take full ad- in the layered framework. Spatially constant relative stan-
vantage of parallel architectures because a significant part of dard deviations of about 10% were used in the prior model
the computation does not depend on the previously generated for both elastic attributes. A prior linear correlation between
traces. Use of a local trace search neighbourhood allows a IP and IS log residuals of about 0.7 was estimated from the
very efficient parallelisation of the algorithm with near-linear well data and used as a constraint in the inversion. A ver-
scalability with the number of CPUs. tical variogram model with exponential structure and correla-
tion length of about 20 ms was estimated by statistical analysis
The restrictive assumption that the impedance ratios in equa-
of the log residuals after blocking onto the grid and subtract-
tion (1) are known can be overcome by a post-processing step.
ing the prior model values. Lateral variograms were estimated
The consequence of using erroneous impedance ratios is that
by cross-correlation between the near and far seismic data af-
there is usually a mismatch between input seismic and synthet-
ter resampling into the stratigraphic grid. The analysis was
ics computed from the stochastically simulated impedance ra-
averaged over several micro-layers in each interval. Similar
tios. We have implemented post-inversion corrections, which
isotropic exponential models were fitted to both macro-layers
are applied independently to each trace and each realization.
with a range of 200 meters.
Small impedance perturbations are calculated so that the cor-
rected simulated reflectivities match the reflectivities com- Synthetic and residual seismic cubes were generated using the
puted using the initial fixed impedance ratios. These pertur- mean from the 25 realizations (Figure 2). Good agreement is
bations give the correct fit to the seismic data, at the expense obtained for both angle stacks with RMS errors below 10% and
of a slight degradation of the realization geostatistics. 15% for the near and far angle stacks respectively. A better re-
construction is achieved for the near angle stack because of its
higher signal-to-noise ratio. Some remaining signal is visible
TEST ON REAL DATA in the far angle residuals, probably due to a wavelet scaling
problem and/or non-stationarity. The impact of well condi-
The GeoSI method was tested using data from a deep-water
tioning is demonstrated by the reduced variance of simulated
field. The reservoir is a faulted anticline comprising turbidite
impedance values in the neighbourhood of the well shown by
channels and lobe complexes. The main challenge in field de-
the black vertical line on the figure.
velopment has been the assessment of lateral and vertical con-
nectivity between channels, which is a hard task given the rel- A blind well test was performed to assess the quality of the
atively low resolution of the seismic in the area. inversion results. Figure 3 shows a comparison of measured
and predicted IP and IS and confirms the good level of match
The data include two angle stacks (10 and 30 degrees), dipole
achieved with GeoSI. Individual realizations were also inter-
sonic and density logs at 4 wells and 3 seismic horizons in-
terpreted over the 8.1km × 8.7km project area. Zero-phase
wavelets were estimated independently for the two partial
stacks using a simultaneous multi-well extraction procedure.
The near angle wavelet has a dominant frequency around
25Hz. Although similar in shape, the extracted far angle
wavelet is consistent with a reduction of about 10% in fre-
quency content compared to the near angle data.
S/N map estimation was performed separately on near and far
angle cubes using a standard coherence-type analysis, result-
ing in average S/N values of 6 and 4 respectively. Very low
values of S/N were observed in the Southwest region, partly
due to strong migration artifacts. These maps were used to
define trace-by-trace seismic noise statistics in the GeoSI in-
version.
A stratigraphic grid framework was constructed in time by
proportional interpolation from the three interpreted horizons
to define micro-layers in each macro-interval with maximum
time-thickness of 2 msec. The inverted volume corresponds
to 325 × 350 seismic traces at 25 m CDP-spacing and spans a
400 msec time window. This results in a layered model con-
taining about 35 × 106 cells. It took 70 hours to generate 25
realizations of both IP and IS on a dual-Opteron 64-bit com- Figure 3: Comparison of measured (green) and inverted (black)
puter. Figure 1 is a vertical section showing the prior model impedance values at blind well with predicted confidence margins (2.5
for IP and IS , together with the corresponding inversion results standard deviation) in light blue.
 Fast Geostatistical Stochastic Inversion in a Stratigraphic Grid
Figure 4: 3D view of a single high-resolution realization of inverted impedances clearly showing one of the main channel system in the lower
reservoir section. Left: colourmap is between 3000 (black) and 9000 (red) g/cc × m/s. Right: colourmap is between 2000 (black) and 5000 (red)
g/cc × m/s
preted and used to evaluate uncertainty in the spatial distribu-
tion of the channelized sand deposits. Figure 4 shows 3D vi-
sualizations of one high-resolution realization for both P-wave
and S-wave impedance.
CONCLUSIONS
We have illustrated a new Bayesian linearized stochastic in-
version approach that operates directly in a fine-scale strati-
graphic grid, simultaneously inverts multiple angle stacks, can
be conditioned to well data and controls spatial continuity via
layer-dependent variogram models. Use of a linearized for-
ward model allowing analytical calculation of the joint pos-
terior distribution, combined with a trace-by-trace sequential
simulation strategy leads to an efficient parallel inversion im-
plementation with fast turnaround time, even for large data vol-
umes. Sampling from the joint posterior distribution delivers
multiple, high-resolution images of elastic attributes in the lay-
ered framework. The multiple realizations of IP and IS can be
used for uncertainty analysis and cascaded stochastic simula-
tion of petrophysical reservoir properties.
ACKNOWLEDGMENTS
The authors would like to acknowledge TOTAL E&P UK PLC
for permission to publish this paper.
REFERENCES
Buland, A., O. Kolbjornsen, and H. Omre, 2003, Rapid Spa-
tially Coupled AVO Inversion in the Fourier Domain: Geo-
physics, 68, 824–836.
Buland, A. and H. Omre, 2003, Bayesian Linearized AVO In-
version: Geophysics, 68, 185–198.
Contreras, A., C. Torres-Verdin, K. Kvien, T. Fasnacht, and
W. Chesters, 2005, AVA Stochastic Inversion of Pre-Stack
Seismic Data and Well Logs for 3D Reservoir Modeling:
EAGE 67th Conference and Exhibition.
Debeye, H., E. Sabbah, and P. van der Made, 1996, Stochastic
Inversion: SEG 66th Annual Meeting.
Fatti, J. L., G. C. Smith, P. J. Vail, P. J. Strauss, and P. R. Levitt,
1994, Detection of gas in sandstone reservoirs using AVO
analysis: A 3-D seismic case history using the Geostack
technique: Geophysics, 59, 1362–1376.
Francis, A. M., 2002, Deterministic Inversion: Overdue for
Retirement?: PETEX 2002 Conference and Exhibition,
London, UK.
Haas, A. and O. Dubrule, 1994, Geostatistical inversion: a
sequential method of stochastic reservoir modelling con-
strained by seismic data: First Break, 12, 561–569.
Oldenburg, D., T. Scheuer, and S. Levy, 1983, Recovery of
the acoustic impedance from reflection seismograms: Geo-
physics, 48, 1318–1337.