Investigation of Artificial Neural Networks, Alternating Conditional Expectation, and Bayesian Methods for Reservoir Characterization

The objective of reservoir characterization is to describe the complex distribution of properties of a reservoir based on available geological, petrophysical, and engineering data. Some of the reasons for the complexity are randomness or nonlinearity among petrophysical parameters and the subjective nature of geological interpretation. The nonlinear relationships among the reservoir properties can be quantified using artificial neural networks (ANN),· alternating conditional expectation (ACE), and Bayesian methods. First,· an approach is developed to correlate oil recovery efficiency with the petrophysical, engineering, and volumetric parameters reported in the Atlas of Major Texas Oil Reservoirs database compiled by the Bureau of Economic Geology at The University of Texas at Austin. Results are obtained by using the alternating conditional expectation (ACE) method on the database and dividing reservoirs according to drive mechanisms and/or reservoir classes. The categorical classification according to drive mechanism gives better predictions than classification by lithologies. This approach can be applied for prediction of oil recovery efficiency in a new reservoir. Second, an approach is developed for facies classification in a reservoir from wireline logs and core data using back-propagation artificial neural networks (BP-ANN) and Bayesian methods. The example facies selected from a sandstone reservoir are turbidities, debris flow, shallow marine, shoreface, and lower shoreface. Core and wireline logs (gamma ray, density, neutron porosity, and resistivity) are used for facies and facies pay prediction. The accuracy of the facies predicted from these methods usually ranges from 75 to 93%. Gamma ray and density logs are the most crucial for some types of facies while neutron porosity logs are most important for others. These results can be used where quantitative classification of a large number of logs by visual observation can be time-consuming and tedious. The approach can also be used to deteimine which logs are the most crucial for determining different types of facies. Third, the Bayesian approach is further extended for the prediction of facies pay and net pay using wireline log ·and core data. The facies pay is predicted based on the results from facies classification using Bayes theorem. The net pay is. predicted by Classifying the core data into permeability classes. Neutron porosity and density logs are usually important for prediction of facies pay. Sonic logs are usually important for net pay prediction.