Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

 

SPE 135911

Methods of Multiple-Point Statistics in Geological Simulation Practice:


Prospects for Application
A.V. Bezrukov, A.R. Davletova, RN-UfaNIPIneft Ltd.

Copyright 2010, Society of Petroleum Engineers

This paper was prepared for presentation at the 2010 SPE Russian Oil & Gas Technical Conference and Exhibition held in Moscow, Russia, 26–28 October 2010.

This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents of the paper have not been
reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect any position of the Society of Petroleum Engineers, its
officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written consent of the Society of Petroleum Engineers is prohibited. Permission to
reproduce in print is restricted to an abstract of not more than 300 words; illustrations may not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright. Write
Librarian, SPE, P.O.Box 833836, Richardson, TX 75083-3836 U.S.A., fax 01-972-952-9435.
 

 
Abstract
The conventional methods of two-point statistics (TPS) applied in geological simulation have preserved histograms and
variograms. However, they have serious limitations in creating complex reliable distributions of geological bodies. Multiple-
point statistics (MPS) methods enable solving a wider spectrum of geological simulation problems while also preserving the
common statistical parameters. However, the application of these methods in practice is complicated by the requirement of
reliable litho-facial models and the corresponding training images. A possible solution could be the development of a library of
training images (TI) to cover the variety of sedimentation environments and a wide distribution of statistical parameters in geo
models. The selection of training images and their scales for specific simulation tasks is based on matching statistical
parameters; the simulation result in this case inherits these parameters. This approach ensures the match in general statistical
parameters of the models created using TPS and MPS. However, the difference in higher-order distributions may affect flow
model properties.
In this paper, we investigate some issues of implementing MPS methods, developing a training image library, as well as
selecting and scaling TIs. Some comparison results between TPS and MPS models are presented.
 
Introduction
As of now, geostatistics offers the standard inventory for simulating spatial distribution of geological model properties. When
modeling sedimentary geological objects (bodies), a special classification is used based on facies – the original rock matter
types. The spatial variables describing the distribution of facies in space are discrete. Specific indicator modeling techniques
are used to model discrete parameters. In two-point statistics, the Sequential Indicator Simulation algorithm (SIS) constitutes
the basis of the conventional modeling approach. Here are some important issues to be borne in mind when speaking about this
kind of model:
• The model of spatial continuity is described by a single covariance function which implies minimum effort to obtain a
geological model variant/scenario. Practical examples can, however, be given to show that entirely different facial distribution
patterns might match this covariance function1. This effect can be defines as degeneracy.
• Among all degenerated models, variogram models are characterized by ‘maximum entropy’2. That is, if the initial
(input) model defines some explicit geological patterns/textures attributed to the geometry of facial bodies, such patterns will a
priori be misrepresented in the variogram models.
• Variogram facial modeling suggests using only a certain class of variograms that guarantee discontinuity of
parameters at small distances3. This discontinuity property makes standard models increasingly more ‘disjointed’ when scaled
down, bringing about fractal facial bodies. In fact, discontinuity is linked to the maximum entropy characteristic. Justifying
these peculiar characteristics in modeling geological bodies is yet to be agreed upon. Here it is seen as a direct sequence from
the underlying theoretical assumptions.
Degeneracy is the most conspicuous of the mentioned characteristics. Geostatistics experts agree on the fact that the
absence of a defined heterogeneous structure model is equivalent to a ‘maximum entropy’ model. This may have been put up
with only before the advent of algorithms which, on the one hand, have inherited the main characteristic of variogram models
– its keeping the exact well properties, and on the other hand, allow incorporating and conceptualizing real geological data
pertaining to different depositional environments, hence different facial models. Such set of algorithms is termed
Multiple/Multi-Point Statistics4 to reflect the salient feature of using higher than second-order correlations in the new modeling
paradigm.
2 SPE 135911

In general, it should be pointed out that models created using MPS can stand very close to traditional variogram models in
the liquid holdup and integrated reserve density parameters, and also have quite similar variogram functions. Still, the two
model types compared together will show stunningly different hydrocarbon flow distributions during the production phase5.
The hidden optimization potential of different MPS model realizations and degenerated variogram realizations bespeaks
considerable financial gains on the development/drilling design stage.
 
Methods of Multiple-Point Statistics
All geo modeling problems share the same objective – investigation of spatial (regionalized) variables, i.e. those whose values are
dependent on the position in space. Such variables may include the depth geological horizons, isopach maps, reservoir (porosity
& permeability) properties in different space points, or discrete space-point categorization properties, e.g. according to the
capacity for holding reservoir fluids or as belonging to a depositional stage (lithofacies). When modeling inter-well space, various
uncertainties inevitably present themselves, since the wells are usually spaced too sparsely for an unambiguous estimate of the
reservoir behavior in-between them. For the task of estimating reserves, it is normally sufficient to know just the average value of
the expected pore volume, which can be easily obtained by interpolation. But for modeling reservoir flow patterns, this
information is insufficient; these models require reviewing one or several reservoir property distribution scenarios accounting, at
the same time, for geological heterogeneity (discontinuity) and all the input data. Interpolation methods become inapplicable here
for one reason: they would treat geological heterogeneity as dependent on well density, but this is not correct. To model
heterogeneity, a special mathematic inventory is employed – that is the theory of random processes.
Stochastic methods assume the investigated parameter to be a random field, i.e. a set of interdependent random values
corresponding to each point of the space analyzed (a cell in the geo model grid). The objective of stochastic simulation is
obtaining numeric realizations and the values of the random fields in all the grid cells accounting for their mutual dependence,
the input data, and any other a priori assumptions. A resulting model of interdependent random field values in the grid cells is
a spatial continuity model of the reservoir that largely defines the final appearance of the geological model, the reservoir’s
flow parameters, and its sweep efficiency.
The MPS methods are such that the simulation result honors well data and indirect seismic data, and inherits the
heterogeneity pattern of the training image, as well as its basic statistical properties, such as mean values, variograms,
compartmentalization parameters, etc. This latter again brings the methods of MPS and TPS together, inasmuch as the
variograms underlying the TI are carried through into the final realizations. Strictly speaking, the MPS methods are a direct
generalization of the TPS methods and can therefore produce equivalent results if unconditional TPS simulation results are
used for the TIs. The methods of MPS allow explicit definition of the conceptual model, whereas the conventional (TPS)
methods ‘impose’ an a priori artificial special continuity model. 
 
Training Images
In MPS, a training image (TI) is a 3D conceptual model or pattern that defines the basic laws of property alteration across
space. TI have the following requirements:
• three-dimensional space;
• stationarity, i.e. invariability of the statistical parameters of the TI throughout its volume;
• recurrence, i.e. repeated re-use of the same structure elements;
• aperiodicity, i.e. no part of the TI may be an identical copy of another part of this TI; the structure elements must vary
in different combinations to cover all the possible variants;
• relative simplicity, i.e. the TI must not abound in complex structures that may not be reproduced in the realizations;
• the scale and orientation of the TI measured in grid cells are to be set according to the field being simulated;
• statistical parameters, such as means, variograms, unit compartmentalization (per number of cells) and dimensions of
geological bodies are to be matched against well data and the target values.
The quality of a TI and the correctness of algorithm settings are determined by a visual comparison of the statistical
parameters of the realizations with those of the TI.
The largest stumbling block that stymies rapid spreading of MPS in reservoir simulation tasks is the difficulty of creating
TIs for each definite modeling case. A good solution may be found in preparing a training image library classified according
to:
• depositional environments;
• statistical distribution parameters.
When creating TIs, the geologist is not limited in means and algorithms; however, special programs have to be written for
more complicated modeling scenarios. A TI library may thus act as a source of:
1. ready-to-use TI for ‘standard’ cases;
2. templates for custom-built TIs;
3. algorithm templates for making up non-standard TIs.
There are roughly two types of workflows to generate TIs:
• object-oriented modeling methods;
• imitation models.
SPE 135911 3

The object-oriented generation methods imply a random distribution of objects of different types, shapes and sizes in the
model. An example of applying such methods is the FLUVSIM software6.
Imitation models imitate sedimentation processes, i.e. erosion and deposition. Examples are TSSIM7 that simulates deep-
water cones and ALLUVSIM8 simulating meandering bed deposits. Imitation algorithms can be based on mathematical
sedimentation models.
The TI library sections are filled by iterative generation of TIs with different input data. Each section must contain a
sufficient number of TIs with a sufficient range in the parameters. This can be achieved, for example, using a Monte-Carlo
process that would randomly select parameters from a predefined distribution function. The module (grid-block) sizes must be
sufficiently large to allow multiple contractions along any axis.
 
Statistical Principles of Scaling and Selecting Training Images
Regarding the methods of multiple-point statistics as generalized TPS methods naturally brings about the principle of
equivalent basic statistical parameters (hereinafter we imply discrete training images for lithofacial models), viz.:
• the average number of facies in a volume unit (TI);
• indicator variograms for each facies in at least three directions – horizontal along and athwart the direction of
sedimentation, and vertical.
Apart from these, the methods of MPS allow controlling a number of other parameters, including compartmentalization per
grid unit.
Provided that TIs do not have a definite scale, they must be scaled to match the model variogram radiuses. Scaling means
vertical and/or horizontal extension or contraction of the model. This process changes the dimensions of the TI in grid cells.
The scaling factors are calculated by:
hd
rvar
k = TI , (1)
rvar
hd TI
where rvar is the radius of a variogram obtained from well data and rvar is the radius of a variogram obtained from a
training image. The scaling factors are to be calculated for each of the three directions (horizontal along and athwart the
direction of sedimentation and vertical); the variogram radiuses are selected accordingly. It must be highlighted that for
discrete algorithms (working with grid cell values), variogram radiuses also have to be defined in grid cell units.
Compartmentalization is the average number of homogeneous intervals in the reservoir. Provided that a TI may have
arbitrary dimensions, using an absolute count of intervals is not justified. What does matter is compartmentalization given per
some unit of reservoir thickness, which in our case is the radius of the vertically oriented variogram. Let us remark that the
average number of intervals per model thickness unit equal to the vertical variogram radius does not change during scaling (if
the segmentation and merging effects of intervals during smaller scaling procedures are ignored). Thus, matching unit
compartmentalization (the average number of intervals per vertical variogram radius) is one of the criteria for selecting the
required TI.
Another transformation is necessary to match the depositional orientation of the reservoir with that of the training image;
for this purpose, the TI is to be rotated by an angle equal to the difference between the depositional orientation azimuth of the
reservoir and that of the TI:
α = α reservoir − α TI .
(2)
The standard workflow for the statistical processing of TIs can thus be summed up as follows:
1. Calculation of parameters from well data, including the average share of each facies; compartmentalization factor;
variogram radiuses in the vertical and horizontal directions (along and athwart sedimentation).
2. Calculation of unit compartmentalization matched with the vertical variogram radius:
R
R= rvar , (3)
Nz
where R is compartmentalization; N z is the vertical cell count; and rvar is the vertical variogram radius.
3. Collation of the parameters obtained from well data with those taken from the TI library; selection of one or several
‘best’ TIs characterized by the best fitting average sand content and unit compartmentalization R .
4. Determining the scaling factors for the selected training image(s) from equation (1).
5. Determining the rotation angle of the selected training image(s) from equation (2).
6. Scaling and rotation of the training image(s).
Under otherwise equal conditions, scaling factors close to or less than 1 are preferable. This will help reduce losses while
rescaling the TI(s).
The above steps can be rather easily automated by using an electronic database containing the basic TI statistics and a
simple TI search tool; this would dramatically boost the geologist’s work efficiency.
 
4 SPE 135911

Hierarchical Facies Simulation


Creating a model with a large number of lithological facies necessitates a sequential hierarchical approach to simulation, in
particular when the reservoir structural cross-section presents a spatial superposition of several basic facies deposition
conditions.
The following hierarchical simulation scheme is proposed9:
1. 2 or 3 facies groups are singled out by the genetic and geometric similarity criteria. An ideal classification would be
such where each group represents the simplest geometric shapes.
2. The facies data obtained from well data, the training image and auxiliary data corresponding to the probabilities of
occurrence of each facies are transformed into facies group data.
3. Simulation is performed, in which each grid cell is attributed to one of the facies groups.
4. For each facies group, the internal lithofacial content is simulated. This has three variants:
a. facies within a group are simulated independently of the group boundaries; in this case, a training image is
used which contains only the facies to be simulated;
b. facies within a group are simulated with account for the group boundaries; in this case, a training image is
used which contains both the facies to be simulated and the other facies groups;
c. facies within a group are simulated with account for the group boundaries and the boundaries of the other
facies groups; in this case, the entire training image is used containing all the facies from all the groups.
In case of simulating very complex scenarios, the simulation workflow can have more than two levels of hierarchy. It
should be remembered that variants (a) and (b) in this workflow neglect part of the information, which causes some loss in
simulation accuracy. This, however, may be acceptable when the geological cross-section is a superposition of several
independent deposition processes.
 
Simulation Example
Below is an example of applying the methods of multiple-point statistics to Reservoir YAK III-IV of the Vankor oil field. The
reservoir is classified as a meandering river deposition. Well logging data suggested the presence of four facies:
• flood clays (20.4%);
• straightened river channels (36.1%);
• meandering channel point bars (38.8%);
• flooding sands (4.7%).
The simulation process was divided into two steps. At first, the straightened channels were simulated; then the flood clays,
point bars and sands were simulated in the sections of the model free from the simulated channels.
The first application of MPS is obtaining the training image. The object-oriented TI generation method was used for the
straightened channels facies – the model was randomly filled with straightened channel-like bodies (Fig. 1a). The channels
have an East-West orientation. Training images for the point bars, flood clays and sands were generated using the imitation
algorithm (Fig. 1b). The basic simulation parameters (the shares of the facies and the geometric properties of the sand bodies –
width and thickness) were retrieved from well data.
 

   
a  b 
Fig. 1 a) TI for straightened river channels, b) TI for meandering channels
 
SPE 135911 5

The MPS simulation was performed in Petrel v. 2009.1. Fig. 2a shows the result of a TPS (SISIM) simulation; Fig. 2b
shows the result of a MPS (SNESIM) simulation. A visual comparison of these results indicates that the MPS model is
characterized by a more complex facial structure, which directly affects the model’s continuity/discontinuity.
 

   
a  b 
Fig. 2 а) model created by TPS, b) model created by MPS
 
Let us compare the statistical parameters of the two resulting models. The real compartmentalization factor is 1.2 (observed
from well data), 5.6 for the TPS model and 2.4 for the MPS model. Table 1 also compares the average thickness of the facies
for the wells (target values) and the two models. It may be seen in the table that the MPS model is significantly closer to the
target values than the TPS model, albeit the coincidence is not ideal. The prediction inaccuracy is due to the numerical
constraints – the search radius, the TI size, and the number of nearest-neighbor points used in probability calculation.

Table 1: Average facies thickness, in grid cells


  Well Data TPS Model MPS Model
Flood clays 14.3 2.9 11.2
Straightened river 16.1 3.4 21.6
channels
Meandering point bars 31.1 2.9 20
Flooding sands 14.5 2.1 7

Conclusion
A training image library is part and parcel of the analogous field simulation approach using the MPS methods. It enables
amassing and classifying geological knowledge and experience obtained from different fields and crucial to modern numeric
simulations. A standard tool for searching / adding / manipulating training images will render the practice of MPS simulation
more efficient. Compared to stochastic variogram methods, the methods of multiple-point statistics turn a new leaf in geological
simulation by providing the possibility to explicitly formulate a conceptual model with an underlying geological meaning.

References
1. Krishnan, S. and Journel, A. G.: “Spatial connectivity: from variograms to multiple-point measures”: Math. Geol. (2003) v. 35 no.
8 p. 915–925
2. Journel A.G., Deutsch C.: “Entropy and Spatial Disorder”, Mathematical Geology (1993) v. 25 no. 3
3. Steven F. Carle, Graham E. Fogg: “Transition Probability-Based Indicator Geostatistics”, Mathematical Geology (1996) v. 28 no. 4
4. Strebelle S.: “Conditional Simulation of Complex Geological Structures Using Multiple-Point Statistics”, Mathematical Geology
(January 2002) v. 34 no. 1
5. Andre Journel, Tuanfeng Zhang: “The Necessity of a Multiple-Point Prior Model”, Mathematical Geology (July 2006) v. 38 no. 5
6. C.V. Deutsch, T.T. Tran: "FLUVSIM: a program for object-based stochastic modeling of fluvial depositional systems", Computers
& Geosciences (2002) no. 28 p. 525–535
7. X. Zhang, M.J. Pyrcz, C.V. Deutsch: "Stochastic surface modeling of deepwater depositional systems for improved reservoir
models", Journal of Petroleum Science and Engineering (2009) no. 68 p. 118–134
8. M.J. Pyrcz, J.B. Boisvert, C.V. Deutsch: "ALLUVSIM: A program for event-based stochastic modeling of fluvial depositional
systems", Computers & Geosciences (2009) no.35 p. 1671–1685
9. A. Maharaja and A. Journel: "Hierarchical Simulation of Multiple-Facies Reservoirs Using Multiple-Point Geostatistics" (2005),
SPE 95574

You might also like