.

,,

SPE -of~mme
SPE 20750 Geostatistics for Reservoir Characterization
A.G . Journel, Stanford
SPE Member

Copyright 19S0, Socie!y of Petroleum Engineers Inc. This paper was prepared for presentation at the 65th Annual Technical Conference and Exhibition of the Society of Petroleum Engineers held in NewOrleans, LA, September23-2S, 1S90. Thle paper was selected for presentation by an SPE Program Committee following review of information contained in an abatracl eubmitted by the aulhor(e). Conteme of the paper, aa preeenled, have not been reviewed by the Society of Petroleum Engineera and are subject to correction by the author(s), The malerial, as presented, does not necessarily rellecl any poaltion of the Society of Petroleum Engineers, its officers, or members. Papere presented at SPE meetings are subject to publication review by Editorial Committees O! the Society of Petroleum Engineere. Permieelon 10copy is restricted to an abetract of not more than S&3words. Illuelrations may not be copied. The abstract should contain conspicuous acknowledgment Of where and by whom the paper is presented. Write Publications Manager, SPE, PO. Box 83383S, Richardson, TX 75083.383S. Telex, 730989 SPEDAL.

Abstract
Geostatistics and, more specifically, stochastic mod-

elirig of reservoir heterogeneities are being increasingly considered by reservoir analysts and eagirteers for their potential in generating more accurate reservoir models together with usable measures of spatial uncertainty. Geostatistics provides a probabilistic framework and a toolbox for data analysis with early integration of infer. mation. The uncertainty about the epatial distribution of critical reservoir parameters is modeled and transferred all the way to a risk-conscious reservoir management, The stochastic imaging (modeling) algorithms al. low the generation of multiple, equiprobable, unsoothed reservoir models yet all honoring the data available, In trod uctiol~l Numerical models of reservoirs often fails to capture the heterogeneities that are critically imPortantl-4 for reservoir performance. With production always being the primary functioq of w well, the available data cue typically biased toward the more productive regions in a resevvoir and are regrettably sparse, The interpolation and griclding algorithms commonly used by industry further cxaccrbatc the problem since they are low-pass filters that tend to smcwth out the little spatird variability that the sparse data reveal. While core plugs and well logs are not the only sourcee of information, other data, such as geophysical information, arc often difficult to integrate since they have different levels of reliability4$ and are reprcsentativc of very different voh.rtnes of rock, History matching on historical production information and well test data does not guarantee reliable forcca.sts of a reservoirs future performance. All of M is not NCWS;these problems have come to the forefrrmt as industry focuses on enhancing rccovcry from known rwmrvoirs. With performance prediction for EOR processes I]tcferellces and [Ihtstratioas at ctd of paper

calling for better numerical models of a reservoir rock and fluid propcrties, geostatistics is receiving renewed attention, Until recently, geostatistics was often associated with only one of its important contributions and was used as a synonym for kriging, a multiple regression technique6-s that has been most commonly used as a gridding procedure, By reducing geostatistics to another canned gridding softwsreg, early users failed to realize its full potential as a set of spatial data analysis tools, as a probabilistic language to be shared by geologist slO, geophysicists and reservoir engineers23, and as a vehicle for integrating various sources of uncertain informat ion2, Oversold as canned software with w regtird for prior education, geostatistics generated overcxpectations and the ensuing disrtppointment. Fortunately, rwccnt experience has brought forward more reasonable expectations and has broadened the scope of applications, .Spatial Data Analysis Geostatistics begins with an emphasis on describing and modelling the spatial variability of reservoir properties rmd the spatial correlation between related properties such as porosity and scisrnic velocity, These models can then be used in the construction of numerical mod& for a variety of purposes interpolations for a property W11OSC average is critically impor. tant, stochastic simulations for a property whose extremes arc critically important, Whether onc needs to transfer information from one rescrvoir to rtnother, or between different units within the same reservoir, or whether one needs to transfer information from one discipline to another, a quantitative vchic]c is necessary, Gcostutistical models of spatial variability and dependence pro. vide a quantitative summary of geological observatiotd, and can therefore serve as a such a vehicle. Gcostatistical nlodCIS make it eaaicr to compare data from different sedimentary basins, from different formations, and from different horizons, and are therefore valuable aids to any attempt at buildh~g a

t
GEOSTATISTICS FOR RESERVOIRCHARACTERIZATION
arc

..

en-

,fi,~~
-ra w

classification system. They aiso enable geologists, for example, o put their vaiuable information in a format that can be used ~y reservoir engineers, Geologica.i interpretation of the direction of maximum continuity can be corroborated through the use of directional ccmelogrtuns and the plot of correlat ion ranges, see figure 1. In the ibsence of sufficient data in a particular reservoir, correlograms md other statistical characteristics can be borrowed from other ;elevant data sets such as originating from more mature fields13 n similar geological environments, outcrop studies14 or geologcai cross-sections and maps. By capturing the criticai heterogeneities in a quantitative form, directional correlograms mable the use of this important information that often goes gnored since it commonly has only a quaikative expression, Geostatisticians commonly use variograms to describe spa;ial continuity. While a correlogram, such as that shown in figlre 1, shows the decay in the correlation between sample vaiues w a function of increasing separation distance, the variogram6 ;hows the increase in dissimilarity between sample vaiues verms increasing separation distance. Though variograms and correlograms are commonly used ;O study spatiai continuity of a rvticular variable, the same ;OOIScan be applied to the study J the cross-continuity of difkrent variables at different locations. One could, for example, :ompare porosity at a particular location to travel time at a ocation nearby. Once modelled, this spatiai cross-correlation :an be used in a muitivariate regression procedure known as :okrigingll for building a porosit y map not only from the availible porosity data but aiso from the more abundant seismic information, Non-iinear transformations of the data vaiues can also be very useful in spatiai data anaiysis. The spatiai continuity of Very skewed data, such as permeability vaiues, is usually better understood through anaiysis of the logarithms of these data, Another example is the non-iinear binary indicator transform, which can be defi;ed as 1 if the permeability sample vaiue is greater than a given threshold and O otherwise, The vari~grarn of such an indicator transformation provides a measure of spatial connectivity of the high permeability values that are responsible for flow pat hs and the low permeability y values that serve as barriers to flow16, In addition to variograms and correlograms, a geostatisticians toolkit contains a wide variety of plots and graphs for detecting, corroborating and modelling patterns of spatial variability and interdependence.

variables at d~fferent locations. These generalized regression ~. gonthms are collectively known as kriging and cokrigi~. They generdlze tradtionai regression as appiied in weii log analysis in two senses: (i) The data used (the independent variables in tradhionai statistical jargon) need not be independent one from another, This allows redundancy to be taken into account, an important factor when using several related well logs simultaneously. (ii) The sample correlation is modelled before being appiied to the regression anaiysis. This ailows for filtering certain aspects (frequencies) of the sample correlation. For example, white noise and more generaily the high frequency components of the data can be filtered out, leaving an in. terpolated map that reflects only the large scaie trends of these datas. Conversely, a particular trend known frcxn indirect information can be built into the map in such a way that it honors as closely as possible the available data17.

Integrating

Data of Different

Types

Multivcuiate regression, or coknging as a geostatistician would tail it, is usuaily not a convenient framework for the integration of too widely different types of data such as qualitative geological information, which is usuaily oniy indicative in nature, and direct laboratory measurements. At a particular location where one does not yet have a porosity measurement, a consideration of the iithofacies information might provide a reasonable intervai within which the unknown value should faills, If we are certain that we am in a particular type of sandstone, for example, we might know that the porosity must be somewhere in the intervai from 10% to 30%. If, in addition, we also have enough core plug measurements within that type of sandstone to build a histogram we could go further than simply stating the previous cm-istrained interval. We could use that histogram to provide a probability distribution that might, for example, tell us that the unknown porosity is more likely to be on the low end of our 109$ to 30% range than on the high end, The indicator framework of geostatistics and the soft kriging algorithm ailow an updating of such prior distribution using nearby data that may be either soft or hard, To continue with our previous example, we could locaily update the previ. ous probability y distribution obtained from lithofacies consider. ations with more specific local information, This could be soft information, such as the fact that the location in question is i] the upper haif of a fining-upwards sequence, or hard informs tion, such as the fact that a full core measurement taken in [ well only 50 feet away had a porosity of 15,670, The result of this updating is a posterior probability dis tribution that provides the probability for the unsampled vaiu( to belong within any given class of values, say, between 109 and 1570 porosity. From such a dist ribut ion, any optimai estimate for the unmrnplcd vaiue can be derived once an opti maiity criterion has been specifiedie, The optimaiity criterio] preferred by most statisticians is the least square error trite rion; for this definition of optimaiity, the mean of the posterio

Generalized

Regression

(Kriging - CokrigingO-6J1 )

Any unsampled vaiue, a porosity for example, can be estimated by generalized regression from surrounding measurements of the same value once the statistical relationship between the unknown being estimated and the available sample values hus been defined, This is exactly what the correlograxn provides: a prior model of the statistical similarity between data vaiues, As was pointed out in the previous section, this generalized regression can also include nearby measurements travel time, for example, or of .somc different variable-seismic a facies code, When using such secondary information, one aiso necd6 a prior model of the cross-corrclogram, which pro. vides information on the statistical similarity between different

-1

SPE 20750

A. G. Journel

probability distribution provides the best estimate, One could :hose instead to use the absolute er:or, rather th~ the squ~ed :rror, es the yardstick by which to measure the goodness of an :stimate; for this optimtdity criterion, the median of the posterior probability distribution turns out to be the best estimate, Rather than chosing some convenient, but arbitrary, definition of optimality, one could also analyze the intended use of my estimate and come up with a specific {loss function lg120 that describes the penalty associated to any particular level of error, Such a loss function would likely be asymmetric since the impact of an underestimation oi a particular magnitude is rarely the same as the i:npact of an overestimation of the same magnitude, Thk project-specific 10SSfunction can also be minimized, yielding a best estimate that aims specifically at minimizing the impact of error on each specific project. Indicator geostatistics encodes all amenable information into a series of indicator (binary) dat a718. Hard data result in a coding that consists entirely of Os and 1s; soft data resuit in a coding that consists of intermediate values between D and 1. These indicator data are then interpreted as prior probability for any unsampled variable to take upon any partitular value. No matter what their origin, be it well logs, seismic data or geological interpretation, all indicator data are pooled together. Experimental indicator variograms are then calculated to reveal the patterns of spatial continuity in these indicator data, Once the pattern of spatial continuity is well understood, it can be modelled and used in a multiple indicator regression to yield the posterior probability distribution of any unsampled value, Traditionally, data of different types have been processed separately, leading to several different modelsa geological model, a geophysical model, a production model, ...which are difficult, if not impossible, to merge. Indicator geostat istics takes a quite different route by first merging all of the relevant information through a common coding of that information and, then, producing reservoir models consistent with that information.

Just as one can draw a series of outcomes from a univariate probability distributional, one can also draw a series of outcomes from a multivariate probability distribution47~ 13. With a (K . JV)-variate probability distribution representing the uncertainties in K rock and fluid properties at N locations, each drawing from such a multivariate probability distribution represents a stochastic image cjf the reservoir. If IV is very large, several million nodes on a, 3-dimensioned grid, each outcome provides a high resolution, image of the reservoir rock and fluid properties, Such stochastic images honor all of the available information, both hard and soft, at all locations and all of the structural information entered through the correlograms and cross-correlograms, yet they are different one from each other, see figure 2. The differences between these stochastic images provide a directly usable visualization of the uncertainty about the reservoir rock and fluid properties. Where all the different outcomes agree, there is little or no uncertainty; where they differ most there is maximum uncertainty. For example, a streak of high permeability values that persists on 90% of the images may be considered as reliable, whereas a streak that is present on only 50% of the images is less likely to exist. If the uncertain streaks happen to be in areas that are highly consequential to reservoir performance, the need for additional data has been established together with the locations at which these data are seeded. Stochastic imaging might, for example, reveal that there is a lot of uncertainty about whether or not a particular EOR injector/producer pattern contains a high permeability streak near the top of the producing unit, !,f the nttture of the EOR process makes gravity override a concern, the exercise of stochastic imaging will have identified the neeu for more accurate information on the upper portion of the pattern in question. Stochastic images represent alternative, equiprobable numerical models of the reservoir rock acd fluid properties and, as such, can be used as input to flow simulators for sensitivity analysls 2i122. Since a complex flow simulator could be very ex pensive to run, it is often necessary to consider only a limited number of reservoir models representative of the range of uncertainty: some unfavorable ones, some intermediate ones and some favorable ones. This selection can be done by running a stripped-down simulator that is faat and efficient yet also captures the relevant feature of the flow problem under consideration, A simple particle tracking algorithm, for example, might be useful in identifying the main flow paths23, thereby allowing a trained engineer quickly to assess whether a particular reservoir model will produce favora$le or unfavorable results. ~eference~ 1, Weber, K, J,: How heterogcneit y affects oil recovery, in (1986), Lake and Caroll cd,, Rest rvoir Characterizntio~ Academic Press, 487-584, 2. Hewett, T. A,: Fractal distributions of reservoir heterogeneity and their influence on fluid transport, SPE papcr 15386 presented at the 1986 SPE Annual Conference, New Orleans (Oct. 5-8, 1986),

Stochastic

Imaging of Reservoir Heterogeneities

The thrust of modern geostatistics is not least-squares spatial regression but the building of probability distributions that characterize the present uncertainty about a reservoir rock and fluid properties, These probability distributions should account for all relevant information through models of the spatial dependence between each piece of information and the unknown variable of interest. As more information is collected, the uncertainty about the unknown variable of interest lessens and the spread of the posterior probability distribution d> creases. The unknown can be a single particular unsarnple~ value, say porosity, at a single location, or it can be the unsampled values of a particular variable at many locations (the nodes of a regular grid, for example). It can even be the unsampled values of all relevant variables-porosity, permeability, saturations, pressures,... -at many locations. With a single variable at a single location, one hss a l-variate problem; with a single variable at JV locations, the problem becomes an .V-variate problmn; with K interdependent variables at N locations, the problem becoms a (K oN)-variate problem.

1-

GEOSTATISTICS FOR RESERVOIRCHARACTERIZATION

SPE 20750

3. Tang, R.W,, Behrens, R. A,and Emanuel, A.S; Reservoir studies using geostatistics to forecast performance, SPE paper 18432 presented at the 1989 SPE Symposium on Reservoir Simulation, Houston (Feb. 6-8, 1989). 4, Haldorsen, H.H and Damsleth, ing, JPT (Apr. 1990), E.: Stochastic model-

14, Go@n, D. J., Chandler, M. A., Lake, L.W. and Kocurek, G.: Permeability transects in eolian sands arid their use in generating random permeability fields: SPE paper 19586 presented at the 1989 SPE Annual Conference, San Antonio (Oct. 8-11, 1989). 15. Journel, A.G. and Alabert, F. G.: Non-Gaussian data expansion in the Earth Sciences, Terra Nova, 1(2), 123134 (1989). 16. Sandjivy, L,: The factorial kriging analysis of regionalized data: Its application to geochemical prospecting, . . in Geostatistics for Natural Resources Charact~ 1 Part 1, 571-579, (1984), Verly et al. cd., Reidel publ. 17, Henning, O. and Halvorsen, K. B.: The Bayesian bridge between simple and universal kriging, Math Geology, 21(7), 767-786, (1989). 18. Kostov, C, and Journel, A. G.: Coding and extrap~ lating expert information for reservoir description, in ~ Character izatio~ (1986), Lake and Caroll cd,, Academic Press, 249-264. 19. Berger, J. O.: Statistical Decision Theory Analysis (1980), Springer Verlag, 617p, and Bayesian

5. Krause, F.F. and Collins, H. N.: Pembina Cardium, recovery efficiency study: A geological and engineering synthesis, vol. I and II, Petroleum Recovery Institute, Calgary, 1984. 6. Hohn, M. E.: Geostatistics and Petroleum Van Nostrand Reinhold, 264 p. Geology, (1988),

7. Journel, A. G,: Fundamentals of Ge ostatistics in Five Les~, Short Course in Geology 8, (1989), AGU publ. 40 P. 8. Isaaks, E.H. and Srivastava, R. M.: An Introduction lied Geostatistics, (1989), Oxford Press, 561 p. 9. Bluepack-3D Users Manual, Fontainebleau, l?ra~;e, 209p. Centre to App-

de Geostatistique,

0. Fogg, G.E. . ucia, J. F.: Reservoir modelling in the Dune field, Crane County, TX: Geologic/geostatistical characterization and analysis of heterogeneity, recovery efficiency and in-fill drilling: a Report of the Bureau of Economic Geology, U. of Texas at Austin, 64 p. (1989). 1. Doyen, P.: Porosity from seismic data A geostatis~ical approach, Geophysics, 53(10), 1263-1275 (1988). 12, Journel, A.G. and Alabert, F.: New method \oir mapping, JPT (Feb. 1990). for reser-

20. Srivastava, R, M,: Minimum variance or maximum profitability, CIM Bulletin (May 1987), 80(901), 63-68. 21. Genrich, J,)7mand ~ommer, sitivity 985. analysis, JPT
Fos,: L~Novel approach to sen-

(Sept,

1989), 930-932 and 983-

22. Matthews, J. L., Emanuel, A,S, and Edwards, K, A.: Erat tal methods improve Mitsue miscible predictions, JPT (NOV. 1989) 1136-1142. 23. Giordano, R., Salter, S, and Mohanty, K.: The effect! of permeability y variations on flow in porous media, SPE paper 14365, 1985.

l?,. Journel, A.G. and Gomez-Hernandez, J. J.: Stochastic imaging of the Wilmington clastic seq~ence, SPE paper 19857 presented at the 1989 SPE Annual Conference, San Antonio (Oct. 8-11, 1989).

,.,

SPE 20750

model Correlograms

fww)
1 1

Experimental value

Rose of correlation ranges

a2

al
( direction Ml)

Ihl

(direction X2)

Figure

1: Typical spatial dependence analysis

The variable is porosity as sampled (after calibration) from sonic logs. A directional correlogram p(l~l, a) is a plot versus 1~1of the linear correlation data O(Z), c#(g+k) direction 1~1 between coefficient between all pairs of porosity

separated by approximately the same distance l~i in approximately the same

p(l~, a), measures the loss of correlation aa the separation vanishes distance is called values increases, The distance at which the correlation

a. The correlogram two sample

the range aa, The polar plot, aa vs. al, of these range values provides a rneamre of anisotropy.

with the largest range corresponding to the direction of best spatial continuity (correlation),

357

... ,,
i.

WR

r
X$n .!

b
ma

I L
2

W*

r
K9n

Figure 2: Four equiprobakle realizations of the sand-shale sequence over a vertical section, (The vertical exaggeration is 10:1)

358