Sunteți pe pagina 1din 5

Lithofacies discrimination using support vector machines, rock physics and simultaneous

seismic inversion in clastic reservoirs in the Orinoco Oil Belt, Venezuela


Atilio Torres*, Universidad Central de Venezuela, Jorge Reveron, PDVSA-Intevep and Juan Infante
Universidad Central de Venezuela

Summary properties volumes, and applying SVM as classification


algorithm. These results permit to identify reservoir pay
Lithologic discrimination was performed in clastic zones.
reservoirs in Orinoco Oil Belt, Venezuela, integrating rock
physics, simultaneous seismic inversion and support vector Reservoirs description
machines. The rock physics analysis allowed to obtain
relation between lithologic facies and elastic properties in The area of study is located in Caria Field in Orinoco Oil
reservoirs. Subsequently, through simultaneous seismic Belt, Venezuela. U sands of Oficina (Lower Miocene)
inversion was possible to generate P and S wave and Merecure (Oligocene) Formations, represent the main
impedances and density volumes. Finally, combining the reservoirs. Both Formations have a fluvial-deltaic origin,
previous two points was generated a lithofacies volume with more fluvial influence in Merecure Formation.
using the algorithm of support vector machines as
classification tool. Results of this study permitted to These sand reservoirs are very clean and unconsolidated,
identify reservoir pay zones. with a relative shallow depth (3500-4300 ft), porosities in a
range of 20-38% and permeabilities between 100-1800 mD.
Introduction Also, these reservoirs are saturated with heavy oil (10.3
API).
One of the biggest problems in the Orinoco Oil Belts
reservoirs is the low acoustic impedance contrast between Rock physics
sand and shale deposits. This circumstance forces to use
others elastic attributes besides the acoustic impedance, The goal of this phase is to find the elastic attributes that
which permit better lithological discrimination (Ovalles et provide a better lithological distinction between clean sand
al., 2011). and shaly deposits, this is important for SVM algorithm,
which requires to know the attributes and classes
Its well known that using prestack seismic inversion is (litofacies) to classify correctly the last ones. In addition,
possible to obtain elastic properties volumes (P (AI) and S Rock Physics Template (RPT) (Avseth et al., 2005) permits
(SI) wave impedances and density). The combination of to study the reservoirs properties at details, the idea is
these results with rock physics, which permits to link linking elastic properties and petrophysical and
lithologies and elastic properties, and the use of sedimentological data. The construction of the RPT in
probabilistic methods, derives in lithological volumes in unconsolidated sands considers three effective medium
order to define reservoir pay zones (Mukerji et al., 2001). models: Friable Sand Model (FSM) and Contact Cement
In this particular case, the low impedance contrast avoids Model (CCM) (Dvorkin and Nur, 1996), and Constant
using conventional probabilistic methods (discriminant Cement Model (CoCM) (Avseth et al., 2000). In this case,
analysis, Bayesian classification). To resolve this situation, the elastic properties of the fluids (brine and heavy oil) are
we use another classification algorithm, support vector very similar; consequently this study is focusing in
machines (SVM). sedimentological and facies conditions, and not in fluid
type in the pores. For this purpose, models were deployed
SVM is a statistics method with a huge success in in acoustic velocity versus total porosity crossplot, and
computation, generally is used in classification problems in considerating 100% of water saturation.
optical character recognition (OCR), data mining,
bioinformatics, etc. The application of this algorithm in Four wells were uses in this study, none had shear log,
lithological discrimination using elastic attributes therefore this was estimated applying Castagnas
represents a new point in comparison to other methodology (1985), with the estimation of lineal relations
investigations; this implies that the effectiveness of this Vp-Vs in surrounding fields. Furthermore, for better shear
method in the present study will be a relevant factor to velocity estimation, Oficina and Merecure Formations
point out in future works. required individual rock physics studies (Figure 1 and 2).

With all these antecedents mentioned, the main objective of Crossplot in Oficina Formation (Figure 1.a) shows an
this work is to generate a lithofacies volume through elastic overlap in the acoustic impedance values of sand and shaly
Lithofacies discrimination using support vector machines, rock physics and simultaneous seismic inversion

facies, in effect is the density the best lithological differentiable sands (black polygon) with low impedance
discriminator. Meanwhile, RPT in this case (Figure 1.b) and density values (Figure 2.a), being this last attribute that
point out a preferential direction in sands (black arrows), offers a better discrimination. Figures 2.b, and 2.d show
these directions was defined by Dvorkin and Gutierrezs separation between brine sands and oil sands. In RPT
model (2002) as a shale lamination direction. This is (Figure 2.d), oil sands respect FSM with high porosity,
validated with core information, which provides that sands while brine sands are fitted at CoCM with 1% or 2% of
usually have alternating shale lamination and dispersed cement and lower porosity. This difference is due to
shale. sorting, oil sand are cleaner and with higher porosity and
permeability than more shaly brine sands.

Simultaneous seismic inversion

In order to create a litofacies volume, its necessary to


generate elastic attributes volumes (AI, SI, RHOB), where
the solution of SVM classification will be propagated.
These volumes are generated by prestack simultaneous
seismic inversion. Besides of the seismic data required,
algorithm is based in a priori model information, provided
by interpreted horizons and well logs. CRP gathers had
incident angle up to 40 on the target that permits to obtain
a more confident density volumen. The correlation between
inversion and well was above 80%, ensuring a good quality
control of the project.

Support vector machines

The idea of SVM algorithm is to generate a surface, named


Figure 1: Oficina Formation: a) AI RHOB crossplot
hyperplane, which is able to separate two or more
(black polygon encloses differentiable sands); b) RPT
populations previously categorized. The hyperplane
(black arrows indicate shale lamination).
provides a mathematic equation (decision function),
capable to classify a new vector, after input data training
Elastic Elastic behavior in Merecure Formation is
and its crossvalidation to verify the SVM predictability.
completely different from the previous. Here, we observe

Figure 2: Merecure Formation: a) AI RHOB crossplot; b) AI RHOB crossplot (only sands); c) RPT ; d) RPT (only sands).
Black polygons encloses clean differentiable sands and adjusted to friable sand model.
Lithofacies discrimination using support vector machines, rock physics and simultaneous seismic inversion

The effectiveness of SVM resides in the fact to support


multidimensional input data (huge numbers of attributes).
In other hand, SVM decreases computation time in
comparison to other classification methods.

The mathematical formulation of SVM has been quite


studied (Cristianini et al., 2000; Schlkopf and Smola,
2000), this requires optimization and regularization of
objective function, also developing a complex vector and
matrix. In addition, SVM are divided into two main
branches: linear and nonlinear. In the lineal approach, the
classification is solved in the same input space, for
example, working with R3 data, the classifier hyperplane is
a plane, in the case of R2 data, a line represents the
separation of two populations.

For understanding graphically the linear SVM approach,


Figure 3 illustrates in R3 the classifier hyperplane in Figure 3: 3D visualization of elastic attributes and linear
combination with the populations categorized in Oficina classification SVM (Example in Oficina Formation).
Formation. Its clear to note that the central hyperplane
separates clean sands from shaly facies, being density the Results of SVM training are shown in Table 1. The linear
attribute with more influence in the differentiation. The two method is applied to seven (7) attributes in Oficina
planes parallel to the central are built by the closest vector Formation and it validates a greater population percentage.
(samples) of each class. These vectors are called support The increasing of the number of attributes validates more
vectors and are necessary for SVM mathematical elements, which does not happen in Merecure Formation.
formulation.
In the case of nonlinear solution its the opposite. The
The nonlinear approach is more complex; the classification solution with three (3) attributes is more efficient in
solution is done in higher dimension space than input one. Merecure Formation, however increasing the number of
In this spatial translation is necessary to introduce a kernel attributes in the PSVM nonlinear algorithm does not
function, which linearized the input function in the new increase the validation; in fact Oficina Formation solution
higher dimension space. For its mathematical simplicity, is unstable. Logically this instability is related to data
the most used kernel function is the Gaussian. characteristics, because Fung and Mangasarian (2001) had
Nevertheless, there are different types of kernel functions, success in different data sets with a high number of
adapted to the problem and data management. attributes, however there are not references to the
application of SVM in elastic attributes. For these reason,
In this research, we employed the algorithm Proximal we have not a solid argument to explain the instability of
Support Vector Machines (PSVM) (Fung and Mangasarian, SVM algorithm for this case.
2001), which resolves the classification problem in a linear
equation systems that has demonstrated to work quickly Despite these problems, optimal validations in both
and effectivly. Input data was divided in different sets, in Formations are quite favorable, with training and test
order to prove the combination of elastic attributes to get a validations above 86%. These results give a high level of
better classification solution. Additionally, classification confidence and credibility in SVM method, as lithofacies
solutions were estimated for each Formation. classifier using elastic attributes.
Table 1: SVM linear and nonlinear results. The red boxes highlight the optimum classification for each Formation.
Lithofacies discrimination using support vector machines, rock physics and simultaneous seismic inversion

Litofacies volumes sands above are not identified by SVM possible due to lost
of predictability towards the south of the field, where is
After getting SVM classification solutions, decision blind well. Nevertheless, correlations of lithofacies were
function are propagated in elastic properties volumes acceptable, getting values in order of 70% for different two
generated through simultaneous inversion. Figure 4 blind wells.
illustrate sand spatial distributions in Oficina Formation.
The sand greatest accumulation is in the west of the field.

Figure 4: 3D sands visualization in Oficina Formation.

In Merecure Formation case (Figure 5), which has more


fluvial characteristics, there is sand in almost whole field.
However, sand relative volume map identifies northwest
Figure 6: Lithofacies correlation in blind well. (Left: well
area with the biggest sand accumulation, also there is an
facies ; right: SVM facies classification)
orientation North-South that crosses wells F and D, which
coincides with sedimentation direction. This is an important
Conclusions
point to considerate, keep in mind that rock physics study
shows that these sands have a big possibility of oil
SVM algorithm classified lithofacies effectively using
accumulation.
elastic properties information. For this purpose, validations
were above 86%. These results proved that this method
could be an alternative to Bayesian classification, generally
employed in such studies.

By lithofacies volumes and sand relative volume maps,


after rock physics analysis, was possible to identify
reservoir pay zones. Oficina Formation presents higher
sand content in the west of the field, while in Merecure
Formation the pay zones are located in northwest. This last
Formation has a significant interest, because rock physics
analysis showed that sands are very clean and they have
high oil saturation.

Acknowledgments
Figure 5: 3D sands visualization in Merecure Formation.
The authors would like to thank PDVSA Intevep for their
Finally, Figure 6 deploys in a blind well the initial facies in cooperation in providing the data, support and permission
comparison to SVM facies in order to verify litofacies to publish this work.
volume predictability. Despite seismic resolution (70 ft), it
is possible to identify some thin sand intervals in Oficina
Formation. For Merecure Formation its possible to
correlate the big sand package around 3800 ft. The two
References

Avseth P., Dvorkin J. and Mavko G., (2000). Rocks physics diagnostic of North Sea sands: Link between microstructure and
seismic properties. Geophysical research letters, Vol. 27, NO. 17, pag 2761-2764.

Avseth, P., Mukerji, T. and Mavko, G. (2005). Quantitative Seismic Interpretation. Applying Rock Physics to Reduce
Interpretation Risk. Cambridge, New York, Melbourne: Cambridge University Press.

Castagna J., Batzle M. and Eastwood R., (1985), Relationships between compressional-wave and shear-wave velocities in elastic
silicate rocks. Geophysics Vol. 50, 571-581.

Cristianini N. and Hawe-Taylor J., (2000). An Introduction to Support Vector Machines and other kernel-based learnig metrhods.
Cambridge University Press.

Dvorkin J., and Nur A., (1996). Elasticity of high-porosity sandstones: Theory for two North Sea data sets. Geophysics, Vol. 61,
1363-1370.

Dvorkin, J., and Gutierrez, M., (2002) Grain-sorting, porosity, and elasticity: Petrophysics, 43.

Fung, G. and Mangasarian O., (2001). Proximal Support Vector Machine Classifiers. Computer Sciences Department Data
Mining Institut. University of Wisconsin.

Mukerji T., Jorstad A., Avseth P., Mavko G. and Granli J., (2001). Mapping lithofacies and pore-fluid probabilities in a North
Sea reservoir: Seismic inversions and statistical rock physics. Geophysics Vol 66. 988-1001.

Ovalles, A., Caldern, J. and Revern, J., (2011). Estimacin de la relacin Vp/Vs a partir de la interpretacin de datos ss micos
de onda convertida para la delineacin de yacimientos de petrleo pesado en el Campo Zuata, Faja Petrolfera del Orinoco,
Venezuela, XIV Congreso Latinoamericano de Geologa y XIII Congreso Colombiano de Geologa, Medelln, Colombia.

Schlkopf, B. and Smola, A.(2002) Learning with Kernels. MIT Press.

S-ar putea să vă placă și