Lithofacies prediction driven by logging‐based Bayesian‐optimized ensemble learning: A case study of lacustrine carbonate reservoirs

Although lithofacies routinely is featured by distinct logging responses from each other, many types of lithofacies in practical cases show similar measuring characteristics on logs, and then to achieve a desirable solution from logging-based lithofacies prediction actually is challengeable. Since the mathematical essence of lithofacies prediction can be explained as an issue of pattern recognition, a light gradient boosting machine, a state-of-the-art ensemble learning, specifically developed to address supervised classification, could be a potential solver. Nonetheless, due to an incompatibility of inherent exclusive feature bundling algorithm for logs and usage of a great deal of hyper-parameters, a raw light gradient boosting machine might not be suitable or functional to predict lithofacies. Thus, continuous restricted Boltzmann machine and Bayesian optimization, respectively, employed to process original logging data and optimize hyper-parameters, are adopted as technical assistants for the light gradient boosting machine, and accordingly a new ensemble learning-based predictor called continuous restricted Boltzmann machine - Bayesian optimization - the light gradient boosting machine, is proposed for lithofacies. To validate classifying capability and robust nature of new predictor, three experiments are designed purposefully based on the application of a dataset collected from the wells located within pre-salt lacustrine carbonate reservoirs of the Santos Basin. Simultaneously, to highlight validating effect, other three sophisticated classifiers are introduced as competitors in all experiments, including supper vector machine, random forest and extreme gradient boosting. According to the analysis and comparison of all experimental results, overall four points are verified: (1) the integration of continuous restricted Boltzmann machine and Bayesian optimization indeed is beneficial for the light gradient boosting machine in data preprocess and modelling stage; (2) the light gradient boosting machine cored predictor shows more capability of producing reliable predicted lithofacies information compared to other three competitors; (3) training more learning samples is a simple and effective approach to enhance computing capability of any validated predictor, and under this circumstance the light gradient boosting machine cored predictor still performs relatively better; (4) the light gradient boosting machine cored predictor is characterized with a better robustness as its predicting outcomes, even derived from a sparse-condition dataset, are also proved more qualified. Consequently, the new proposed predictor is demonstrated as a high-efficient and robust solver for the prediction of lithofacies and deserves a widespread application in the geological, geophysical and petrophysical research.