Artificial Neural Network Architectures for Predicting Two-Phase and Three-Phase Relative Permeability Characteristics

Laboratory determination of relative permeability characteristics is labor intensive and can be complicated. Empirical models to predict relative permeabilities based on rock and fluid properties have experienced relatively limited success. Hence, alternate methodologies for accurate determination of relative permeability characteristics are always desirable. In this study, two-phase and three-phase relative permeability predictors are developed using backpropagation networks. In this category of networks, information is passed from input layer to output layer, and calculated errors are propagated back to adjust the connection weights in a sequential manner to improve the predictive capabilities of the models. In the development of the models, experimental relative permeability data along with some commonly reported rock and fluid properties obtained from the literature are used during the training stage, while some other data sets are preserved to test the prediction ability of the models. The two-phase relative permeability models are found to perform in a satisfactory manner within a wide spectrum of basic rock and fluid properties. Similarly, three-phase relative permeability models are observed to have good predictive capability in accurately producing the missing entries of three-phase data sets for a series of isoperms, and in constructing the missing isoperms for a system under consideration. Furthermore, they are found to be capable of effectively predicting the three-phase relative permeability values at various saturation combinations for systems with different rock and fluid properties.