Connectionist and mutual information tools to determine water saturation and rank input log variables

Characterization of petroleum reservoirs plays an important role to effectively manage and forecast the recovery performance. A number of subset log variables such as gamma-ray, resistivity, density, neutron, and sonic porosity logs are generally used to characterize/predict the reservoir properties. The data attributes selection and ranking in reservoir characterization are vital to determine the output variables with the best performance and cost-effective manner during exploration and production operations. The objectives of this research work are to estimate the water saturation in the reservoir with an acceptable accuracy and to rank the log variables according to their importance. To achieve the objectives, the mutual information (MI) and artificial neural network (ANN) techniques are implemented with the non-linear predictors using log variables. The feed-forward ANN model is employed and optimized to predict the water saturation, where the Levenberg-Marquardt algorithm is used for the network training. There is a good match between the real data and predictions so that the regression coefficient and the maximum error is 99.98% and 5.55%, respectively. In addition, both ANN and MI approaches lead to the same ranking levels of log variables, implying high accuracy and reliability of the introduced strategies. It is found that the primary (or most important) log variables are the true resistivity and bulk density to obtain the pore fluid saturation. The approach suggested in this study (connectionist and MI strategies) can assist engineers/operators to run a few numbers of logging tools for prediction of water saturation, resulting in saving the exploration costs through an efficient manner. In addition, further understanding is attained to conduct proper data selection for determination of reservoir petrophysical properties.