Machine learning (ML) for fluvial lithofacies identification from well logs: A hybrid classification model integrating lithofacies characteristics, logging data distributions, and ML models applicability

Identifying lithofacies plays a central role in studying sandbody architecture and reservoir quality in fluvial reservoirs. Logging data is widely considered the most effective method for identifying subsurface lithofacies. Many machine learning methods have been developed to automatically identify lithofacies by analyzing the value or patterns of well logs. However, poor generalization of many classification models has resulted from a lack of exploration into the intrinsic relationship between lithofacies characteristics, data distribution characteristics, and classification model applicability. To address this problem, we conducted research on core description, logging curve sampling processing for layer data, and lithofacies identification using gaussian mixture model (GMM) and back-propagation neural network (BPNN) for a tight sandstone reservoir in the northern part of the Sulige gas field. We investigated the relationship between lithofacies characteristics, logging data distribution, and the performances of machine learning classification models. Based on this relationship, we developed a gaussian mixture model-backpropagation neural network hybrid classification model (GMM-BPNN). The results indicate that the logging curve sampling method reduced deviation caused by adjacent lithofacies influence, and made the lithofacies characteristics constrain the distribution characteristics of logging data, thus improving the application of GMM and BPNN. We observe that the distribution of logging data becomes more centralized as the thickness of certain lithofacies increases, thus improving the performance of the GMM applicable to the classification of centrally distributed data. Conversely, the distribution of logging data becomes more discrete as the thickness of certain lithofacies decreases, thus improving the performance of BPNN applicable to the classification of discretely distributed data. Furthermore, the GMM-BPNN (with an F1-score of 0.95) outperformed individual GMM (F1-score of 0.76) and BPNN (F1-score of 0.77). The hybrid classification model also shows better outcomes in the identification of complex lithofacies in other areas.