q 

8 m
m
01 Petroleum Engineers, Inc Tachmcal Conference and Exhlbmon of the Society of Pelroleum Engineers held m Houslon. Texas, 3-6 Oclohar 1993 for presentation at the 66th Annual

SooIety of PetroleumEnglneere

SITE 26419 A New, Fast Parallel Simulated Annealing Algorithm for Reservoir Characterization
Ahmed Ouenes, New Mexico Petroleum Recovery Research Center, and Naji Saad, Mobil E&P Technical Center SPE Members
Copyright 1993, Society

This paper was prepared

This pepar was selectad for presantatlon by an SPE Program Commdtee following review 01 Intormatlon contemed in an abstract aubmnted by the eulhor($). Contents of the peper. as presented, have not been reviewed by Ihe Society of Pelroleum Engineers and sue Subjecl to correction by the aulhor(a) The material. aa presented, does no! necessarily reflacl any poshon of the Soclely of Petroleum Enginaars. ds officere, or mambers. PaPars presented at SPE maelinga are subject to pubbcallon rewaw by Editorial Commmeea of Iha Society of Petrofewm Enginaere. Perm@amn to copy ie restricted to an abatract of not more than 300 words. Illustretlona may not be copied. Tha abstrect snould wntain conspicuous acknowledgment of whare and by whom the peper la presented. Write Lfbranan. SPE, P.O. Box 833836, Richardson, TX 75083.3836, USA. Telex, 163245 SPEUT

ABSTRACT

This paper presents a new parallel simulated annealing algorithm for computational intensive problems. The new algorithm enables us to reduce the overall time required to solve reservoir engineering problems by using the simulated annealing method (SAM). A simple geostatistical optimization problem (variogram matching) applied to two fields is used for illustration purposes. The reduction of computation time starts by optimizing the sequential simulated annealing algorithm. This task is achieved by an efficient coding and an appropriate choice of topology. Three different topologies are used and their effects on the overall run time and the quality of the generated image are discussed. After optimizing the sequential algorithm, the problem of high rejection rate at low annealing temperature is solved by using parallelizaticm. The new algorithm uses, in an optimal manner concurrently many sequential algorithms. The number of concurrent algorithms is adjusted throughout the optimization to increase the acceptance rate with the optimal use of a CPU. The new algorithm was implemented on a CRAY Y-M P with 4 processors. A 50,400 (280x180) gridblock field was used to test the parallel optimization method. The overall run (clock) time was reduced by approximately the number of concurrent calls of the sequential algorithm.
INTRODUCTION

During the last three years. the interest in global optimization methods in the oil and gas industry has increased. Farmer introduced simulated annealing method (SAM ) in geostat istical optimization. This global optimization

References and illustrations

at end of paper 19

method, developed simultaneously by Kirkpatrick et al.z at 113M, Siarry3 in Paris, and Cernj4 at 13ratislava, has been appiied successfully in more than twenty industrials applications. The resulting computer codes has been used as mature engineering tools and allowed an unexpected gain of productivity in various industrial processes. However, the application of simulated annealing to geostatistical optimization has yet to reach this maturity. This new technique is now added to the increasing number of stochastic models developed these recent years hoping to provide some additional capabilities that will improve the current status of reservoir description. Deutsch and Journe15 added this technique to the GSLIB library which includes more than twenty stochastic modeling tools. Ouenes et al.6 and Ghori et al.7 compared SAM to other geostatistical methods, not only by looking at the generated images, but by using 3-D reservoir simulation and tracer t.est.7. From these comparisons based on fluid flow, they found that SAM w? probably better than the smooth kriging but not necessarily better than other stochastic models tested for that particular problem. However, one fact was certain- SAM was slower than other stochastic models. From the early use of SAM, the major complaint has been the computation time. One solution to this problem is the use of parallel algorithms and computers. Sen et al.8 proposed genetic algorithms instead of SAM, because of the higher potential of parallelization of genetic algorithm, H~wever, the computation time problem requires global consideration, and all the factors contributing to the high run time requires careful investigation. Actually, the computation time is not at all a problem for the application of simulated annealing to geostatistical optimization. Ouenes et aL9- 12 and Deutsch and Journe15 used simulated annealing to match experimental g- 12 and n~ode15 variograms. The advantage of choosing an objective function based on variograms is the considerable reduct]on of

e
.

A New Fast Paralfel Simulated Annealing Algorithm for ReservoirCharacterization

SPE 26419

computation time because of the updating procedure. In other words, the new simulated variogram is obtained from the old one with some simple operations as described by Deutsch and Journe15. Moreover, with the vectorization of this updating procedure the cost of computing the objective function becomes insignificant compared to the computation time required in other problems. The most important problems in oil and gas reservoirs are those which involves fluid flow, which is the major difference between a mine and an oil or gas reservoir. Unfortunately, extensive use of traditional geostatistical methods have not dramatically improved reservoir forecasting*3. This major characteristic (fluid flow) of oil and gas reservoir was not taken into account when assessing reservoir properties. Recently, other authors 14-15 combined geostatistical methods and some localized flow performances by using simulated annealing. The obtained realizations are conditioned to well test and honor the effective absolute permeability around the wel114, However, the simulated annealing is set in a geostatistical framework where the only focus is the permeability distribution. The simulated annealing as a global optimization has a wide range of applicability in the oil and gas industry, and the real potential of this algorithm, and other global optimization methods, are not fully used. The most promising applications of SAM in reservoir engineering are for inverse modeling problems. In such problems, the major focus is on the dynamic data such as production or pressure history, which is the most important information in a reservoir. For many decades, various deterministic methods have been used for inverse modeling and automatic history matching. None of the algorithms developed has yet found widespread use. On the other hand, the use of SAM in inverse problems open new horizons, The first inverse problem solved by SAM was at core scalel 6 where relative permeability and dynamic capillary pressure curves were estimated simultaneously from laboratory coreflood. The second and third applications were at reservoir scale, where the upscaling problem is avoided. By using a simplified reservoir model, permeability and porosity distributions as well as other reservoir engineering parameters were estimated by matching pressure history 17 in gas reservoirs. In the third application, the previous automatic history matching algorithm was extended to oil reservoirsls and infill drilling problems. The application of SAM to inverse model in lead to useful engineering tools routinely used by industry, 1~-]7 where engineers are freed from the time consuming history matching procedure. Because of these practical benefits, our major use and interest in global optimization and SAM is for inverse modeling, which implicitly includes geostatistical optinlization. For such applications, a reservoir simulator is used at each trial solution and makes the cost of computing the objective function a real concern. At the same time, the geostat.istical optimization problem is also solved at each iteration. Therefore, the computation time for the geostatistical optimization component must be reduced as well. Our objective is to use all possible means to reduce the overall computation time: the hunt for unnecessary computation, and optimal use of current and future parallel computers hw begun. In this paper, we address the reduction of computation for the geostatistical component. The first concern is to reduce the computation time required for the objective function. The ideal solution for this problem is to update the objective function rather than re-

compute it at each iteration. However, this ideal solution is not easy to implement in inverse modeling where a system of partial differential equations need to be solved. The second concern is to reduce the number of iterations required to reach a certain low value of the selected objective function. There are two major ways to reduce the overall number of iterations, which is proportional to the execution time. The first way is to use appropriate topologies to transform the current state of the estimation paratneters, to obtain a new state to be tested. The second way is to use parallel computers, so many trials solutions can be tested concurrently and at least one of them will be successful. These two concepts will be illustrated with the application of SAM to geostatistical optimization.
SAM TOPOLOGY

In the past three years at least fifteen papers have been produced on the application of SAM to geostatistical optimization. However, many aspects of the algorithm were not addressed. In geostatistical optimization, a simple and peculiar application of SAM, the algorithm is used to generate a given realization of a geological or petrophysical parameter, for example rock permeability (or its logarithm), by minimizing the objective function J. J = ~
U/r

w.
L-I

~(ml)
h.

- 7e(L))2

(1)

where ~S(ha) and -ye(ho) are the simulated and experimerltal values of the variogram at the lagged distance ha in the specific direction a, and W= is a weight factor. The constraints imposed on the optimization problem are the known locations where field data are available. The algorithm starts by generating an initial random image which satisfies all the statistics of the available field data. This image is transformed during each iteration by using a given topology that gives the rules to change a given image. Once the new image is generated, the new objective function is computed and the change in the objective function, AJ, is deduced for the Metropolislg acceptance rule. The rule consists of keeping the obtained configuration if AJ < 0 and continuing the process. If AJ > 0, the new configuration is kept with a probability equal to eap(-AJ/f3) where O is the annealing temperature. The algorithm starts with an initial temperature, L90, which is reduced after a given number of iterations according to the simple rule : O= AXOO (2)

where A is the reduction factor. Both initial temperature, 00, and J can be chosen automatically; additionally} A, can be adjusted during the optimization procedure20. In a geostatistical optimization problem where the objective function varies smoothly, constant reduction factor such as ~ = 0.9 can still allow convergence. However, in inverse modeling where the objective function may have chaotic variations* 7, it is better to adjust the reduction factor A according to the behavior of the objective function. From the early use of SAM in various industrials applications, we have noticed that the rate of convergence, or the total number of iterations required to achieve a certain low value of the objective function, depends on the topology used, This important point was rarely reported in the literature and usually was considered as a know-how of the user. 20

9 ,.

SPE 26419

AHMED

OUENES,

and NAJI

SAAD

However, some authors21 22 discussed this aspect briefly in their publications. In the application of SAM to geostatistical optimization this issue is crucial since it is directly linked to the computation time. In this section, the effect of various topologies on the computation time and total number of iterations is discussed. There are various ways to perform an elementary transformation on a given image. One simple topology is to select two points ( 1 pair ) randomly in the image and swap their values. The same idea can be extended to 2 or more pairs. Another topology commonly used in Markov fields are neighborhood points. A point is selected randomly in the field, and the swapping is performed on 4 or more of its close neighbors. However, as mentioned earlier, the application of SAM to geostatistical optimization is peculiar. The reason is that the cost for computing the new variogram depends on the number of points used in the topology. If the cost for updating the variogram after 1 pair swapping is 2 computing units, then the cost for 2 pairs (random or neighbors) swapping will be 4 computing units. The cost of an objective function depending on the number of parameters disturbed, is rarely found in other applications. When swapping the values from one location to another, the Markovian formalism guarantees that the mathematical convergence of SAM is respected. However, because of the computation time constraint, and the peculiarity of this problem, Sen et al.s reduced the computation to its minimum by disturbing only one value at each iteration. Therefore the cost of updating the variogram is reduced to 1 computing unit. However, this transformation do not respect the Markovia~ formalism and the mathematical convergence needs to be justified. Additionally, the moments of the distribution are not necessarily preserved. We are limiting this paper to numerical experimentation, as it is very often the case with SAM, and use a practical example to illustrate the difference between 3 topologies. For the first case, a training image of 2500 gridblock (50 x 50) shown in Fig. 4a and its variogram (Fig. 5) is used for generating simulated images by using three different topologies. The horizontal and vertical experimental variograms are matched and 100 randomly distributed conditioning points are used as a constraint for generating the simulated images. The stopping criterion of the algorithm for all the topologies is the same. Automatic matching is stopped when the error on both horizontal and vertical variograms for all the lags considered is less than 0.01, which means that the simulated variogram values match the experimental ones up to the second digit (Fig. 5b). Notice that we are matching experimental variograms rather than a model variogram. Because SAM includes matching the given variogram, lag by lag, it is not necessary to use model variograms of the experimental data, and the experimental variogral,~ itself can be used for the matching. In this new framework, the more valid lag distances (lags that have a significant number of pairs) the user includes in his matching, the better is the generated image. For the considered example all the lags are used for the matching. The first topology, called the 1pt topology, consist of randomly choosing one point in the image and replacing its current value with a new value drawn randomly from the cdf of the training image. After 99,430 iterations, the convergence with this topology was achieved for the realization shown in Fig. 4b. The major problem with this topology is the statis21

tics of the final realization. Because, the initial statistics of the random image are disturbed when drawing from the calf, the a~erage value and the variance of the simulated image are shghtly different from the values of the training image. This will also be true for the higher moments of the distribution. However, this difference does not exceed 1 % if an appropriate cdf is used for this example. To minimize this difference, the number of points in the cdf need to be close to the number of gridblock in the training image. In other words, it is not recommended to use a limited cdf and interpolation. Another problem with this topology is the risk of substantial changes in the simulated images caused during the optimization process. As mentioned earlier, this topology allows the best run time since it uses only one point in the updating procedure. The total run time for this examp, s will be considered as the base line (see table 1). The second topology, called 2pts topology, consists of swapping two values chosen randomly in the image. The convergence with this topology was obtained after 139,780 iterations, and the simulated image is shown in Fig. 4c. The final realization satisfies all the statistical attributes of the training image including the shape of the calf. Finally, the third topology, called 4 pts topology, consist of swapping 2 pairs at the same iteration. With this topology the algorithm needed 758,931 iterations to converge and produce the image shown in Fig. 4d. Before discussing these results it is important to emphasize that these num-ber of iterations are given for one realization. Another run will use another sequence of random numbers and will lead to different numbers of iterations as well as different images, However, the difference in the number of iterations will rarely exceed 10 %, and the main features of the images will be represented in a similar way. When comparing the total computation time of these three examples (see table 1) as expected the Ipt topology is the best, followed by the 2pts topo!ogy which is 2 times slower and finally the 4pts topology which is 18 times slower than the 1 pt topology. Based on the computation time, the 1 pt topology is the fastest, but the obtained image has slightly different statistics from the training image statistics (see. Fig. 3). Although all the generated images (Fig. 4b, 4c, and 4d) seem to contain the main features of the training image (fig. 4a), it is better to compare tbe produced realizations in a quantitative manner. A simple way to achieve this goal is to use the correlation coefficient, CC, which ranges from-1 to 1. CC = 2X21 -

3(Y -Y)

(3)

~~(Zi - 7)2 ~i(Yi ~)2 where Zi is the value of the stochastic variable in the gridblock i, and Yi is the value of the training image at the same location i. ~ and ~ are the average values of Z~ and Yi for all the N gridblock i, When CC is close to 1, the generated image is completely correlated to the training image, or in other words, contains all its features. At CC = O, there is no correlation between the generated image and the training image. Based on the correlation coefficient. the best image is the one that has the highest CC (see table 1). In this example, the best image is the one generated with the 2 pts topology. Based on both criteria used, the 4 pts topology is not recommended for this example. However, for other fields

q  

A New Fast Parallel

Simulated

Annealing

Algorithm

for Reservoir

Characterization

SPE 26419

with different sizes and different variogram structures, the performances of these topologies may be reversed, and a 4 pts topology (2 pairs random or 4 neighbors) may give better computation time than a 2 pts topology by simp ; requiring 2 times to 4 times less iterations23. It seems that when heterogeneities are very localized, higher number of perturbation points in the topology will lead to fewer number of iterations. However, when the variations are more spatially apart, the 2pts topology will result in fewer number of iterations. Finally, this example shows clearly how the overall run time can change drastically when using different topologies. But, there is no general rule for choosing the best topology. That choice depends on the field in consideration. For other applications of SAM, the same problem arises, and considerable saving in run time can be obtained if appropriate topologies are used. Once the ap >ropriate topology is found, the sequential algorithm can be rurther optimized by appropriate coding. For example, the update procedure used for the variogram can be highly vectorized when using a CRAY. These reductions in computation time are designed for a sequential simulated annealing algorithm which still have a major problem when an image that decreases the objective function is not found. At low annealing temperature, the probability of accepting images which increases the objective function become very low. Therefore, man y trials images are rejected, and the time spent for computing the objective function is wasted. This problem can be solved by using parallel simulated annealing.
PARALLELIZATION
OF

on the first chain and performs the first LC/NP iterations (the first sub-chain); from these computation, the algorithm computes the temperature for the second Markov chain. The second processor uses this temperature, and the optimal configuration obtained by the first processor to start the second Markov chain. During this time, the first processor starts to work on the second sub-chain of the first chain. This procedure is extended to all the processors, which are updating both temperature and optimal configuration according to the computation performed by each processor by the one before it. The computation time for such an algorithm is TCpa, = where TCe is the tion, and NC the the computation. tion time for the me x 1X x (fvc-t-NP
NP

- 1)

(4)

SAM

Before addressing the parallelization of SAM, it is important to remind the reader that other global optimization methods exist and are currently used in various applications. A brief description and the parallelization potential of each method can be found in 0uenes20. As mentioned earlier, genetic algorithm is better suited for parallelization than simulated annealing. However, before considering the parallelization, it is important to reduce the computation time by considering the topology aspect and the cost of computing the objective function. Because of the possibility of only updating the variogram when using SAM, it is may be unnecessary to consider genetic algorithms for geostatistical optimization. The reason for this is that implementing the update procedure for computing the new variogram is not practically possible with genetic algorithm. Although the number of necessary generations required to obtain a good image with genetic algorithm is considerably less than the number of iterations required by SAM, still, the computation time with genetic algorithm is 3 to 5 times more than the computation time with SAM. Therefore, in geostatistical optimization we will focus on SAM in this paper. The idea of a parallel simulated annealing algorithm may seem to contradict the philosophy of the method which is undoubtedly sequential. Different types of alternatives are proposed in the literature, and the Azencotts book24 is the complete reference in this area. Many parallel simulated annealing algorithms have been developed since the mid 1980s. One approach proposed by Aarts et af.25 consists of running the algorithm in parallel on N P Markov chains. Each Markov chain of constant length LC is assigned to a given microprocessor. In addition, each Markov chain is divided into sub-chains of length LC/N P. The first processor starts 22

time required for one elementary perturbatotal number of Markov chains used during We remind the ,eader that the computasequential algorithm is LCx NC Tc,eq =TCex (5) which shows that the computation time is divided by a factor NP (if N P is small compared to NC). The main disadvantage of such an algorithm is that the convergence is not guaranteed, and using it may lead to a waste of time instead of reducing computation time. Another approach to parallelization was proposed by Kravitz and Rutenbar26 where parallel computation are performed on the same Markov chain. The idea is based on the fact that at low temperature, many trials are rejected. Therefore, instead of trying different configurations in a sequential manner, it is possible to try them in parallel. Roussel-Ragot27- 2s proposed the following strategy to implement such an algorithm in a way that guarantees that the result obtained will be the same as the sequential one. There are N P processors which are used to perform one or more elementary perturbation and Metropolis acceptance rule is applied after each transformation. The N P processors are used at the same time during the same process, and there are two modes: 1. At high temperature, each processor performs only one elementary perturbation. After all the N P processors are done, one of the optimal configurations obtained by the processors is chosen randomly, and used to update all the processors. 2. At low temperature, the NP processors may not provide a new optimal configuration during one process. Therefore, many process are required to find an acceptable transformation. Once one processor finds this new solution, all the processors are updated. In contrast to the previous parallel algorithm, this approach converges exactly like the sequential algorithm. Finally, a third parallel algorithm proposed by Casotto et U1.29tested on a placement problem, gave comparable results to those obtained by a sequential algorithm. The new algorithm presented in this paper uses the Roussel-Ragot2s approach with major improvements in the use of the parallel processes and annealing schedule. Our objective in using a parali-4 simulated annealing algorithm is to reduce the overall computation time for a given problem. However, we want this time reduction at minimum cost in terms of real dollars. When using a parallel computer. the CPU time increases with the increase of parallel or concurrent runs of a specific sequential algorithm. For example if the cost of running one sequential algorithm is 1 unit

SPE 26419

AHMEDOUENES,and

NAJ[

SAAD IMPLEMENTATION

time or $1, the cost or running two of the same algorithm h parallel will be 2 unit time, or $2. Therefore, the goal in developing a parallel algorithm is to increase the number of parallel runs only when needed . The parallel algorithm will be composed of a master and parallel slaves. The master or optimization manager, is responsible for the following tasks: 1. Determining the number of necessary concurrent calls of the sequential algorithm (optimal number of slaves), and 2. managing and adjusting the annealing schedule. The number of concurrent calls required by the master depends on the acceptance rate, AR, at the previous temperature level. ~R _ M (6) LC where LS is the total number of successful iterations obtained during the previous level of temperature that has a constant Markov chain length LC (maximum number of iterations allowed for a given temperature level). The genera] trend of the AR for a sequential algorithm is shown in Fig. 2. Notice two major points: 1. Half of the number of iterations required to converge are performed with an acceptance rate less than 0.1 2. The average acceptance rate decreases, but AR have a chaotic behavior where the increase and decrease alternate. The new parallel algorithm uses in an optimal way the variation of the acceptance rate. [n contrast to Roussel-Ragot where only two or three modes of temperature are used, we use more than 6 modes corresponding to different ranges of acceptance rate. Moreover, the number of processors used is not the same for all the modes. For example, the highest temperature mode is defined by AR = [1., 0.7]. In other words, when the acceptance rate is in this range, the master will use only one sequential SAM ( 1 slave), and unnecessary parallel runs are avoided to save CPU time. When the acceptance rate drop below AR = 0.7 another temperature mode starts, where the master will require the use of two parallel or concurrent calls of the sequential SAM (2 slaves). If both sequential calls are successful (the two images are accepted by the Metropolis rule), then the master will choose randomly one of them. If only one of the concurrent calls satisfies the Metropolis rule, then the master will select this image as the winner, and update all its parameters. The new image is then passed to the two parallel sequential algorithms for another trial. The number of parallel calls is adjusted throughout the optimization according to the acceptance rate. If the acceptance rate drop into a lower temperature mode, the master will add one more parallel call. However, if the acceptance rate increases to a higher temperature mode, the master will reduce the number of parallel calls. In this manner, the CPU time and the cost of running a parallel machine will stay in a reasonable range. At the end of each Markov chain, the master uses the lowest value and the average value of the objective function provided by all t.~ winner slaves. to reduce the temperature as described in uuenes20. This new algorithm is illustrated with a second geostatistical optimization problem where a 50,400 (280x180) gridblock field will be used (Fig. 8). The experimental variogram of this field shown in Fig. 6 will be matched by using a 2 pts topology and 720 conditioning points. The parallel SAM is tested on a CRAY Y-MP. 23

CRAY Y-MP

The parallel architecture of the CRAY Y-M P is convenient because with a CRAY Y-MP we can use two or more parallel slaves and only one processor. The main program will represent the master, and the sequential SAMS to be called in parallel are simply subroutines seql, seq2. etc. of the main program. The slaves can be called in parallel by the CMID command used in FORTIIAN: CMID CMID call CMID call CMID CMID CMID $ parallel $ case seql $ case seq2 $ case $ end case $ end parallel

With these simple commands the program will run seq 1 and seg2 in parallel. The subroutines will not run in parallel on two processors with the CM 1 D command. However, another command can force seql and seq2 to run in parallel on two processors. This is not necessarily a more efficient way of parallelization because the overhead time increases. The maximum number of slaves allowed during the optimization and its effect on the convergence was tested on the 50,400 gridblock field. First, the algorithm was run without any parallel ization, and convergence was achieved after N, = 1,487,012 iterations. When using at most two slaves, or two parallel calls of the sequential algorithm, the total number of iterations dropped to iV2 = 761,729 iterations which represents a speed factor SU F = 0.51, The speed-up factor SU F, is defined as the ratio of the number of iterations when using the parallel SAM, divided by the number of iterations obtained with the sequential SAM. When the maximum number of slaves is increased to 3, the speed-up factor drops to SUF = 0.41. The speed up factor on the total computation time is very ciose to the one obtained based on the total number of iterations. However, the CPU time increases by a factor close to the inverse of the speed-up factor. The reduction of total number of iterations reaches an asymptotic behavior after a maximum number of 6 slaves where the speed-up factor dces not exceed SU F = 0.38 for this example. When using the parallel SAM, the acceptance rate increases significantly and the need to use more concurrent calls, because of a low value of AR, almost vanishes. Therefore, it is expected to see little effect on the number of iterations when doubling the number of concurrent calls from 3 to 6. The main reason is that the acceptance rate, AR, never drops below AR = 0.1 when using 3 concurrent calls. In addition to the parallelization of the optimization method, each sequential subroutine was highly vectorized in this application.

CONCLUSIONS

Based on the results presented in this paper, the folIowing conclusions can be drawn. !. The topology has a considerable effect on the number of iterations and computation time required for convergence.

A New Fast ParaIlel

Simulated

Annealing

Algorithm

for Remrvoir

Characterization

SPE 26419

2. The lpt topology has the advantage of being the fastest for the example considered in this paper. However, the statistics of the obtained realization is slightly different from the original statistics of the training image. The degree of variation in the statistics depends on the problem and cannot be known a priori. 3. The quality of the images measured by the correlation coefficient are different for the three topologies considered. The 2 pts topology provided the best image for the example considered. 4. A new parallel simulated annealing algorithm was proposed that uses in an optimal way the parallel CPU. 5. A substantial reduction in computation time was obtained when using the developed parallel simulated annealing algorithm on a CRAY Y-MP.
NOMENCLATURE

REFERENCES

AR h J LC LS N NC NP TC Y

= = = = =

cl 1 A 0

acceptance rate lag distance objective function (dimensionless) Markov chain length number of successful iterations during one temperature level = total number of gridblock = total number of Markov chains during the optimization = total number of processors = computation time = stochastic variable for training image = stochastic variable for simulated image = anisotropic direction = variogram value = reduction factor for f? = annealing temperature (dimensionless)

Superscript e = experimental s = simulated Subscript e = i = par = seq = elementary perturbation index for gridblock i parallel sequential = index for initial d

ACKNOWLEDGMENTS

The authors wish to thank Patrick Siarry (Ecole Centrale de Paris) for his valuable comments and suggestions. Appreciation is extended to Akhtarul 13assan (University of New Mexico), for his help in the parallelization work. The authors would like to thank CRAY Research Inc. for their support and for providing machine time. Ahmed Ouenes would like to thank Mobil R&D, Gaz de France DETN, Marathon Oil Company, the New Mexico Petroleum Recovery Research Center, and New Mexico Techs Petrcdeurm Engineering Department for their active support. The authors would like to thank Scott Richardson for the graphics, Steve Whitlach, and K. Stanley for reviewing the paper.

1, Farmer, C. L.: Numerical Rocks, The Mathematics of Oil Recovery, P. R.. King (cd. ) Clarrmdon Press, Oxford (1992), 437. 2. Kirkpatrick, S., Gelatt Jr, C. D., Vechi, M. P,: Optimization by Simulated Annealing, Science ( 1983) 220, 671. 3. Siarry, P. and Dreyfus, G.: An Application of Physical Methods to the Computer Aided Design of Electronic Circuits, Journal de Physique, LeUres ( 1984] 45, L-39. 4. dernj.,V.: A Thermodynamical Approach to the Traveling Salesman Problem, J. OSOptimization Theory and Applications (1985) 45, 41. 5. Deutsch, C. and Journel, A.: GSLIB; Geostatistical Sofiware Library, Oxford University Press, New York (1992). 6. Ouenes, A., Meunier, G., Pelc6 V., and Lhote, 1.: Enhancing Gas Reservoir Characterization by Simulated Annealing Method (SAM), paper SPE 25023 presented at the 1992 European Petroleum Conference, Cannes. France, Nov. 16-18. 7. Ghori, S., Ouenes, A., Pope, G., Sepehrnoori, K., and Heller, J.: The Effect of Four Geostatistical Methods on Reservoir Description and F1OWMechanism , paper SPE 24755 presented at the 1992 SPE Annual Technical Conference and Exhibition, Washington, D. C., Oct. 4-7. 8. Sen, M., Datta Gupta, A., Stoffa, P., Lake, L., and Pope, G.: Stochastic Reservoir Modeling Using Simulated Annealing and Genetic Algorithm, paper SPE 24754 presented at the 1992 SPE Annual Technical Conference and Exhibition, Washington, D. C., Oct. 4-7. 9. Ouenes, A., Bahralolom, 1., Gutjahr, A., and Lee, R.: A New Method for Predicting Field Permeability Distribution, Proc, 1992 Mediterranean Petroleum Conference and Exhibition, Tripoli, Libya, Jan. 19-22., 468-477. 10. Ouenes, A., Bahralolom, I., Gutjahr, A., and Lee, R.: Application of Simulated Annealing Method (SAM) to Laboratory and Field Anisotropic Porosity? Proc., 1992 Lerkendal Petroleum Engineering Workshop, Norwegian Institute of Technology, Trondheim, Norway, Feb. 5-6, 107-118. 11. Ouenes, A., Meunier, G., and Moegen, H. de: Application of Simulated Annealing Method (SAM) to Gas Storage Reservoir Characterization , paper 96e presented at the 1992 Annual AICh E National Spring h!eeting, NewOrleans, March 29 -April 3. 12. Ouenes, A., Bahralolom, I., Gutjahr, A,, and Lee, R.: Conditioning Permeability Fields by Simulated Annealing, Proc., Third European Conference on the Mathematics of Oil Recovery, DeIft, Netherlands, June 17-19, 41-50. 13. Haldorsen, H. and Damsleth, E.: Challenges in Reservoir Characterization, The American Association oj Petroleum Geologists Bulletin (1993), 77, No 4,541-551, 14. Deutsch, C.: Annealing Techniques Applied to Reservoir Modeling and the Integration of Geological and Engineering (Wel/ Tesi) Data, Ph D Dissertation, Stanford [University, Stanford CA (1992). 15. Hird, K.B. and Kelkar, M.: Conditional Simulation for Reservoir Description Using Spatial and Well Performance Constraints, paper SPE 24750 presented at the 1992 SPE Annual Technical Conference and Exhibition, Washington, D. C., Oct. 4-7. 24

SPE 26419

A}IMED

OUENES,

and NAJI SAAD

16. Ouenes, A., Fasanino, G., and Lee, R.: Simulated Annealing for Interpreting Gas/Water Laboratory Coreflood , paper SPE 24870 presented at the 1992 SPE Annual Technical Conference and Exhibition, Washington, D. C., Oct. 4-7. 17. Ouenes, A., Br+fort, B., Meunier, G., and Dup&6, S.: A New Algorithm for Automatic History Matching : Application of Simulated Annealing Method (SAM) to Reservoir Inverse Modeling, Unsolicited paper SPE 26297 (1993). 18. Sultan, J .A., Ouenes, A., and Weiss, W.: Reservoir Description by Inverse Modeling: Application to EVGSA U Field , paper SPE 26478 to be presented at the 1993 SPE Annual Technical Conference and Exhibition, Houston, Oct. 4-6. 190 Metropolis, N., Rosenbluth, A. W., Rosenbluth, M. N., Teller, A. H., and Teller, E.: Equations of State Calculations by Fast Computing Machines, J. Chernicai Physzcs (1953) 21, 1087. 20, Ouenes, A.: Application of Simulated Annealing to Reservoir Characterization and Petrophysics Inverse Problems, PhD dissertation, New Mexico Tech, Socorro, NM (1992). 21. Lundy, M. and Liees, A.: Convergence of an Annealing Algo~thm, Mathernuficaf Progran;mzng, ( 1986), 34, 11~

22. Otten, R. H.J.M and Van Ginneken, L.P. P. P.: The Annealing Algorithm, Kluwer Academic Publishers, Dordrecht (198!3). 23. Ouenes, A., Gutjahr, A.: Conclitioning Random Fields by Simulated Annealing Method (SAM) Using Multiple Swapping, accepted for publication in Mathematical Geology. 24. Azencott, R.: Simulated A nnealing ; Parallelization Techniques, John Wiley, New York (1992). 25. Aarts, E., de Bent, F., Habers, J,, and Van Laarhoven. P.: Parallel Implementations of the Statistical Cooling Algorithm, Integration 4, 209-238. 26. Kravitz, S. and Rutenhar, R.: Placement by Simulated Annealing on a Multiprocessor, IEEE Trans. CAD (1987), 6, No 4,534-549. P., Siarry, P. and Dreyfus, G. La 27 Roussel-Ragot, M6thode du Recuit Simu16: Principe et Parall$lisation, paper presented at the second national colloquim for electronic circuit design, Grenoble, France (1986). 28. Roussel-Ragot, P. and Dreyfus, G.: Parallel Annealing by Multiple Trials: An Experimental Study on a Transputer Network, Simulated Annealing: Parallelization ?echnigaes, R. Azencott (cd), John-Wiley, New York (1992). 29. Casso~o, A., Romeo, F. and Sangiovanni-Santelli, A.: A Parallel Simulated Annealing Algorithm for the Placement of Macro Cells, IEEE Trans. CAD (1987) 6, 838847.

Table 1: Topology Comparison TODO

1

I Iterations I 99,430

I Run Time I
1

CC

I I [

0.351 0.427 0.324

2 Pts 4
pts

I 139,780
!

I I

2 18.7

758,931

25

Fig. I: Rate ofconvergence

for fieldl
I

SPE26i
1

I b

102

10

1 ---

.......

pt topo 2 pts topo 4 pk tapo

al > .-

102

g10
e z

q f10-
q

5 .glo

-2

81()-2
.. 4...,

G a -

8
0
. .

................ 10 5 Iterations (xl O*4)

....

15
.....

1o-t
\ [ \ I

,........ ... ... .... .. ...

. . . . .

.....

D O..*..,,,..

. . .. . . . . . . . . . . .......? . .. .. . . . .

4 3 Iterations (xl 05)

Fig. 2: Acceptance

rate, AR, for field 1

Fig. 3: Histogram

of lpt topo reaUzation

0.25 0.2 % :0.15 3 ~ 0.1 Training Image Pt. Topo.

0001

Iv

10 5 Iterations (xl 04)

5 Log Permeability

10

26

SPE26419
Fig. 4 Grayscale maps for field 1

a. Training Image

b, 1 pt topo

c. 2 pts topo

d. 4 pts topo

27

SIPE~6419

a. Initial Variogram 2 Experimental ........ Simulated

2 E fgl.5 . m 9 ~ 1

a. Initial Variogram

1.5 EJ gl .% > 0.5

)
q

Experimental ........ Simulated

20

40

oo~

100

200

300

Lag Distance

Lag Distance

b. Optimal Variogram 2 Experimental ........ Simulated

2.5

b. Optimal Variogram

1.5 E ~ !?1 .% > 0.5

20

40

100

200

300

Lag Distance
Fig. 5: bIatching experimental variograms for field 1 Fig. 6: Matching

Lag Distance
experimental variograms for field 2

28

..

Fig. 7: Effect of CRA.Y

parallel

SAM

on the rate of convergence

sPE26& 19 *,
q

1.5 G t a E ~1
0 .~

1487012

Sequential -2 parallel - - 3 parallel o Ho 6 parallel

;. .. O*5.:.O ..*, N6 = 562934 --~. 0.

1Oz

a ~

SW= Ni/Ns Ns: SW = 1 N2: S(.JF= 0.51 N3: SUF = 0.41 N6: SUF = 0.38

~.~, n
1o 10

.
10-

Objective Function

Fig. 8: Grayscak map of 50,400 gridblock

for field 2

--

29