A Bayesian Approach in Machine Learning for Lithofacies Classification and Its Uncertainty Analysis

As a qualitative process, classification of subsurface lithofacies is very important for the characterization of hydrocarbon reservoirs. Machine learning has been a potential method to automate the prediction of this parameter based on the well-logging data. In order to incorporate the geological trend into the classification process, a Bayesian approach has been designed in this letter, in which the Markov transition matrix is added. Compared with the Artificial Neural Networks (ANN), a typical representative of machine learning, the proposed Bayesian-ANN (B-ANN), can honor the lithofacies orderings in the vertical direction. A better performance of B-ANN can be observed in the precision-recall curves, as well as indicated by larger values of the average precision. Because of the dominance in the internal transitions within the lithofacies themselves, the posterior probabilities of each lithofacies in B-ANN are less varied than those in ANN.