The Role of Geology in Stochastic Reservoir Modelling: The Future Trends

Stochastic simulation is a popular approach for the quantification of reservoir heterogeneity. From the geological viewpoints, however, the role of reservoir geology in stochastic models is diminishing in recent years. This paper reviews the major characteristics of the stochastic models, and geological knowledge is identified as the major missing element in the current practices of the simulation techniques. A detailed discussion of the meaning and usefulness of geological knowledge is given. We have also provided a list of future research directions showing how we can make use of conceptual geological models to bridge the gap between reservoir geology and stochastic simulation practices. Introduction Reservoir characterisation plays a crucial role in reservoir management practices. It is an important joint between geology, geophysics and reservoir simulation. It aims to integrate information from various sources and generate numerical models from sparsely distributed reservoir properties, such as lithofacies, porosity, permeability and fluid saturations. Without a realistic geological framework incorporated into the reservoir study, no reservoir model can be used reliably as a predictive tool. Philosophical Issues in Reservoir Modelling. Obtaining realistic reservoir models requires fast and flexible processes to handle reservoir data which are, unfortunately, often incomplete. Because of the incompleteness of the available data, reservoir modelling becomes a very challenging problem as the establishment of truth is impossible (although the reservoir is of a deterministic nature in reality). This means that verification and validation of any reservoir models also becomes impossible. Many studies have used history matching as a tool to “verify” the accuracy of the reservoir models. This, in fact, is a result of committing a logical fallacy of “affirming the consequent” because there exists more than one numerical model which can produce the same outputs. This situation is referred to by scientists as nonuniqueness. Therefore, if a reservoir model fails to match the observed data (e.g. production data), then we know that the model is faulty in some way, but the reverse is never the case. This means that we may need to rely on other criteria in order to define a better model. Hence, reservoir characterisation studies are, in fact, not purely scientific works, but depend heavily on subjective modelling decisions. Data Integration. In order to reduce the subjectivity involved, the current practice is to incorporate as many inputs (e.g. well data, seismic attributes, well tests and production data) as possible into the modelling algorithms. The only advantage of data integration is to reduce the uncertainty of the model, but may not necessarilly produce a more “accurate” model because the truth is never available for verification. In order to integrate data from various sources into the reservoir models, geologists have experienced many difficulties, as it is difficult and nearly impossible for them to process these data quantitatively and to provide high resolution 3D models. With the rapid development of modern computer technology, stochastic simulation becomes a popular approach for simulating unknowable reservoir properties and quantifying reservoir heterogeneity with extensive risk analysis capability. Despite the usefulness of stochastic models, there is still much evidence that the predicted performance is below expectation. This is an indication of our failure to understand the process involved and to recognise the uncertainty inherent in the definition of important reservoir characteristics. There are many possibilities for poor predictions, and all of them are more or less associated with the input parameters to the simulation algorithms. We can certainly improve the predictions if we are able to identify the problem inputs and model their uncertainty appropriately. SPE 54307 The Role of Geology in Stochastic Reservoir Modelling: The Future Trends D. Tamhane, SPE, L. Wang and P.M. Wong, SPE, University of New South Wales, Sydney, Australia. 2 D. Tamhane, L. Wang and P. M. Wong SPE 54307 Model Uncertainty. According to Zimmermann, “certainty” implies that a person [model] has quantitatively and qualitatively the appropriate information to describe, prescribe or predict deterministically and numerically a system, its behaviour or other phenomena. The situations which are not described by this definition shall be called “uncertain.” The causes of uncertainty mainly include lack of information, abundance of information (complexity), conflicting evidence, ambiguity, engineering measurement, and subjective belief. From our experience, the major uncertainty in reservoir modelling is the subjective geological interpretation of the field. This strongly relates to the understanding of the reservoir geology. However, the commonly identified problem inputs in stochastic models are hardly related to reservoir geology. Hence, there is a great need to re-examine the characteristics of the conceptual geological models, which are known to be fully charged with valuable geological knowledge. Objective. The objective of this paper is to firstly revisit the current practices of stochastic simulation. This is followed by a discussion of the conceptual geological models. Lastly, we will provide a list of future research directions, showing how we can make use of geological knowledge in stochastic simulation for improved predictions. Stochastic Simulation Stochastic simulation is a very fast and flexible approach to generate reservoir models. The use of Monte Carlo methods is extremely convenient to simulate unknowable events. Many are able to generate multiple conditional realisations for local uncertainty analysis. With the increasing interest of stochastic modelling in reservoir characterisation, many stochastic models have been developed in the past several years. They can be classified into two broad types: pixel-based models and object-based models. In this section, we will briefly discuss the characteristics of these models. Pixel-based Models. Most of the pixel-based stochastic methods are based on kriging or cokriging in a sequential manner. Examples are sequential Gaussian simulation, sequential indicator simulation and truncated Gaussian simulation. They are all able to obtain estimates of the necessary conditional distributions by using simple kriging or ordinary kriging. In order to incorporate soft data (e.g. seismic impedance), cokriging technique can be used during the sequential simulation processes. They can directly analyse the quantitative data and utilise them for prediction purpose via the use of direct-variograms or cross-variograms. Apart from the sequential simulation algorithms, simulated annealing is also flexible to incorporate more data types into the model by adding new components to the objective function. Object-based Models. Object-based models present a promising concept to integrate stochastic geological objects. They treat each geological body (e.g. a sand body) as an object, and parameterise the body by constructing a series of probability distributions according to the known data, Then, they generate different bodies for the geological model by Monte Carlo sampling from the distributions of the parameters. During the process, geologists can also put specific sedimentary body with certain parameters into the model and match their geological knowledge. The most popular model is the marked point processes. The small-scale features are simulated mostly by the use of simulated annealing. The Major Difficulty. From the geological viewpoints, the role of reservoir geology in stochastic models is diminishing and is increasingly replaced by two-point statistics (e.g. variograms) used in pixel-based models that are far too simple to parameterise complex geology, and are often unrepresentative when the data are sparse, limited and biased. The object-based models present a more geologically intuitive concept, but it requires many parameters that are difficult to interpret geologically and does not consider information derived from sedimentary processes. From our experience, geological knowledge is extremely important and often represents the major uncertainty component in reservoir characterisation. Unfortunately, this is not what the current practices of stochastic simulation are trying to model. There also exists a general misunderstanding of what “geological knowledge” really means. The next section will give a detailed discussion of conceptual geological models and clarify the meaning of geological knowledge. Conceptual Geological Models Conceptual models have placed a central role in reservoir modelling for many decades, even before the birth of modern computers. Despite the boom of stochastic simulation in recent years, there are still many so-called “conservative geologists” who do not appreciate the value of stochastic simulation, partly because of the complex mathematics involved, but mostly because of the lack of geological soundness in the simulated models. We forecast that the merging of conceptual models and stochastic simulation will become an important breakthrough in reservoir characterisation. Construction of conceptual models is in fact a highly nonlinear and complex process, which is difficult to be described by precise mathematics. It involves a good understanding of physical and chemical reactions in earth sciences, and depends heavily on the knowledge of the geologist(s) involved. The next section will expound the meaning of geological knowledge. Geological Knowledge. The meaning of “knowledge” is abstractive and difficult to be quantified. It is a result of learning and understanding of a certain subject. In the formalism of knowledge, there is a set of facts contained in the knowledge. When the subject wants to speak about its knowledge, it can only use its concepts. In geology, the subject knowledge can give rise to conceptual models and is commonly shown as hand-drawings SPE 54307 The Role of Geology in Stochastic Reservoir Modelling: The Future Trends 3 for technical communication. The knowledge (conceptual model) contains two main groups of information: Geological rules. Geological rules are the result of the basic theories in petroleum geology. They include the scientific rules of structural geology, stratigraphy, sedimentology, paleontology, diagenesis, geochemistry and others. We can call them “objective knowledge,” as they are generally agreeable among geologists. The spatial distributions of lithofacies and reservoir properties are basically controlled by the geological rules for a given sedimentary environment. Geological experience. Geological experience from longterm geological practice is extremely valuable for reservoir prediction, especially when there is insufficient reservoir data to support the geological rules. As opposed to the geological rules, geological experience is generally considered as “subjective knowledge,” as they are almost disagreeable among geologists, especially those with different research experience. It is therefore important for geologists to work in various types of environments and subsequently improve the confidence of their own intuition. Because of the existence of subjectivity in geological intuition, it becomes the major cause of uncertainty in reservoir modelling. Model Construction. The construction of conceptual model involves a number of steps. The first step is to construct a structural model based on seismic, regional geology and well data. The next step is to identify a depositional model. This identification is based on the regional geological information (tectonics and stratigraphy), global sea level eustatic interpretation, seismic interpretation (structural and stratigraphic), sequence stratigraphy, well logs, core description, paleontological, geochemical and analog data. Based on the depositional model, sedimentary facies model comprising reservoir and non-reservoir facies along with reservoir geometry (external form) can be inferred. High resolution 3D seismic and statistical database regarding reservoir geometry in different environments, if available, can also be used to derive the reservoir geometry. Orientation of reservoirs can be inferred from 3D seismic (amplitude slices) and dip meter log interpretations. As the original reservoir characteristics are greatly affected by diagenesis, the next important step is to envisage diagenetic facies model. This model is constructed based on postdepositional, physio-chemical, biological and mechanical processes, detailed core description, logging, petrophysical and geochemical data. Based on the diagenetic facies model, reservoir flow units and barriers can be identified. The final model is a 3D conceptual geological model depicting structural picture, reservoir geometry, spatial distribution of flow units and barriers along with the qualitative knowledge regarding reservoir heterogeneities. It represents the geological knowledge and remains as an important tool for reservoir modelling. Model Limitations. There are a number of limitations in the use of conceptual models for reservoir simulation. Firstly, the conceptual model is mostly of a symbolic nature (e.g. sand distribution) rather than a numerical one (e.g. permeability). Secondly, it represents only the global trends of heterogeneities and is not able to provide any fine-scale features. The problems are worse if a full 3D model is being sought, as the traditional approach is based on 2D techniques (e.g. cross-section method) with limited interpolation choices. This is why stochastic models are useful as they have the ability to handle such problems efficiently. However, the simple statistical measures and distribution functions should never replace the contribution of the conceptual model, which provides a geological framework of the reservoir. More research works are required to develop methodologies for combining the usefulness of the conceptual models and the efficient stochastic models. Future Trends This final section outlines five major research directions showing how we can make use of geological knowledge and conceptual geological models to bridge the gap between reservoir geology and stochastic simulation practices. Formalisation of Geological Knowledge. A strong emphasis on how the geological knowledge is evolved is required. The formalisation of the knowledge allows the development of a flexible computer system for the generation of conceptual models with different scenarios. As the major uncertainty lies on the subjective interpretation of the reservoir data, it is necessary to consider multiple or alternative geological interpretations. This will be a dramatic improvement on the current methodologies (e.g. the use of different random seed) for simulating multiple, equally-probable realisations, which are only indirectly related to reservoir geology. The potential research tool to understand and model geological knowledge is artificial intelligence (AI). AI is a specialised field of computer science concerned with concepts and methods of symbolic inference by computer and symbolic knowledge representation for use in making inferences. AI can be seen as an attempt to model aspects of human thought on computers. The particular AI technologies are rule-based expert systems (or knowledge-based systems) with a strong focus on the incorporation of fuzzy-based information. These are extremely useful in modelling uncertain, qualitative and linguistic information, which contribute significantly to the derivation of the conceptual geological models. Use of Possibility Theory in Reservoir Modelling. So far, probability theories dominate the fundamentals in stochastic simulation, especially in facies simulation. From the geological viewpoints, however, some geological facies cannot simply be described by probability. For example, the diagenetic facies is generally identified in core plugs. There are a number of transitional types between the end facies types, and the definition of target diagenetic facies types is often imprecise, or fuzzy. Unlike simple depositional facies, which can be expressed by facies proportions in clastic sedimentary sequences, the definition of diagenetic facies is 4 D. Tamhane, L. Wang and P. M. Wong SPE 54307 not as straightforward. Possibility theory is a prominent methodology to handle such fuzzy information. There is a fundamental difference between probability and possibility. Probability expresses a binary decision (yes or no), but possibility is for multi-valued decision. Consider an attempt to describe a core plug using two minerals: type “A” and type “B.” Very often, we describe that the plug looks more like “A” than “B.” If we use probability to implement such a statement, the best we can do is to use a X% chance belonging to “A” and a (100-X)% chance belonging to “B,” where X>50%. This in fact does not reflect the true meaning of the statement. In the statement, the plug actually looks like both types of minerals, but is certainly not composed of X% of “A” and (100-X)% of “B.” In possibility theory, we are able to assume the coexistence of both types, and use X% to express the degree of membership belonging to “A” and Y% to express the degree of membership belonging to “B,” where X>Y. The major characteristic of possibility theory is that (X+Y)% is not 100%. From this example, we can clearly see the failure of probability in describing a core plug, which is visible to us. The problem is worse if we use probability concepts to describe the underground reservoirs. Therefore, research in the use of possibility theory in stochastic models will offer a completely new insight to reservoir modelling practices. Quantification of Reservoir Geology. The most commonly available quantitative geological model is the iso-porosity contours which are often hand-sketched by geologists. As discussed previously, the incorporation of as many inputs as possible into the modelling algorithm can reduce the uncertainty of the model. There is no reason why geological hand drawings cannot be used. Although they represent only global or large-scale features, they do contain valuable geological knowledge. If stochastic simulation is used, the small-scale features can be simulated, and the final model can be constrained by the large-scale features. Techniques such as neural network residual simulation are particular useful. As the quantitative information is so useful, it is a great challenge to provide more quantitative maps. Methods to convert the conceptual models into quantitative hand drawings (e.g. expressing the sedimentary facies model and the diagenetic model) will contribute to many stochastic models. Rule-based Ranking of Multiple Realisations. Ranking multiple realisations from stochastic models is important. Many criteria can be used to choose the “best” image. If geological knowledge could be converted into rules, we may develop rule-based systems to rank the multiple realisations and examine if any of the realisations are geologically interpretable or worth further investigation. This is strongly dependent on the success of the formalisation of geological knowledge discussed previously. Rule-based Reservoir Upscaling. Reservoir upscaling is a challenging field. It aims to convert fine-scale reservoir model to a coarser one, but to retain some characteristics according to some criteria. There is a potential research area for constraining the upscaled model to the conceptual model. It is because the conceptual model represents the large-scale geological trends of the reservoir features. When larger grid size is used, the upscaled model should reflect the same geological trends. Again, we may develop rule-based systems to evaluate the performance of various upscaling techniques. On the other hand, we may also use the conceptual model as the starting point and then downscale it using stochastic simulation. This is similar to the ideas presented in Wang et al. (1998), but further investigation is required. AcknowledgmentsThe authors would like to thank the Australian PetroleumCooperative Research Centre for funding the work publishedin this paper. References1. Simlote, V.N. and Nikolayenko, S.A.: “Dealing withUncertainty: Geostatistics and the New Era in ReservoirStudies,” paper SPE 39546 presented at the SPE Iindia Oiland Gas Conference and Exhibition, New Delhi, India(1998) 383-397.2. Oreskes, N., Shrader-Frechette, K. and Belitz, K.:“Verification, Validation, and Confirmation of NumericalModels in the Earth Sciences,” Science (1994) 263, 641-646.3. Foley, L., Ball, L., Hurst, A., Davis, J. and Blockley, D.:“Fuzziness, incompleteness and randomness: classificationof uncertainty in reservoir appraisal,” PetroleumGeoscience (1997) 3, 203-209.4. Zimmermann, H.-J.: “A fresh perspective on uncertaintymodeling: Uncertainty vs. Uncertainty modeling,”Uncertainty Analysis in Engineering and Sciences: FuzzyLogic, Statistics, and Neural Network Approach, B.M.Ayyub and M.M. Gupta (eds), Kluwer AcademicPublishers, Massachusetts, USA (1998) 353-363.5. Goovarts, P.: “Comparison of CoIK, IK, mIKperformances for modelling conditional probabilities ofcategorical variables,” Geostatistics for the next century,M. Armstrong and P.A. Dowd (eds), Kluwer AcademicPublishers, Netherlands (1994) 18-29.6. Gotway, C. A. and Rutherford, B. M.: “Stochasticsimulation for imaging spatial uncxertainty: Comparisonand evaluation of available algorithms,” Geostatistics forthe next century, M. Armstrong and P.A. Dowd (eds),Kluwer Academic Publishers, Netherlands (1994) 1-21.7. Deutsch, C. V. and Journel, A. G.: GSLIB: Geostatisticalsoftware library and user’s guide, Oxford UniversityPress, New York (1992).8. Lia, O., Tjielmeland, H. and Kjellesvik, L. E.: “Modellingof facies architecture by marked point processes,”Geostatistics Wollongong ’96 Volume 1, E.Y. Baafi andN.A. Schofield (eds), Kluwer Academic Publishers,Netherlands, (1997) 386-397.9. Syversveen, A. R. and Omre, H.: “Marked point modelsfor facies units conditioned on well data,” Geostatistics SPE 54307The Role of Geology in Stochastic Reservoir Modelling: The Future Trends5 Wollongong ’96 Volume 1, Kluwer Academic Publishers,Netherlands, (1997) 415-423.10. Holden, L., Hauge, R., Skare, and A. Skorstad, A.:“Modelling of Fluvial Reservoirs with Object Models,”Mathematical Geology, (1998) 30, no.5, 473-496.11. Phillips, D.C. “3D Geological Modeling,” Geotimes (July1993), 14-16.12. Umkehrer, E. and Schill, K.: “General perspective on theformalization of uncertain knowledge,” UncertaintyAnalysis in Engineering and Sciences: Fuzzy Logic,Statistics, and Neural Network Approach, B.M. Ayyub andM.M. Gupta (eds), Kluwer Academic Publishers,Massachusetts, USA (1998) 21-35.13. Chapman, R. E.: Petroleum Geology, Elsevier Publisher,Amsterdan, New York, (1983).14. Freulon, X. C., Dunderdale, I. D.: “Integrating FieldMeasurements with Conceptual Models to Produce aDetailed 3D Geological Model,” paper SPE 28877presented at the SPE European Petroleum Conference,London, U.K. (1994), 99-10815. Frodeman, R.: “The earth sciences and the public realm --rethinking geology's role,” Geotimes (March 1997), 24-26.16. Fang, J. H.: “Fuzzy Logic & Geology,” Geotimes (October1997) 23-26.17. Nordlund, U.: “Formalizing geological knowledge – withan example of modeling stratigraphy using fuzzy logic,”Journal of Sedimentary Research (1996) 66, no. 4, 689-698.18. Wong, P.M., Tamhane, D. and Wang, L.: “A neuralnetwork approach to knowledge-based well interpolation:A case study of a fluvial sandstone reservoir,” Journal ofPetroleum Geology (1997) 20, no. 3, 363-372.19. Wang, L., Wong, P.M. and Shibli, S.A.R.: “Modellingporosity distribution in the Anan Oilfield: Use ofgeological quantification, neural networks andgeostatistics” paper SPE 48884 presented at the 6thInternational Oil & Gas Conference and Exhibition,Beijing (1998) 509-515.20. Deustch, C.V.: “Fortran programs for calculatingconnectivitiy of three-dimensional numerical models andfor ranking multiple realizations,” Computes &Geosciences (1998) 24, no. 1, 69-76.