Geologically Constrained Deep Learning for Lithofacies Identification of Mixed Terrestrial Shale Reservoirs: Permian Fengcheng Formation, Mahu Sag, Junggar Basin, Western China

Summary The Permian Fengcheng Formation within the Mahu Oilfield is predominantly characterized by mixed terrestrial shale reservoirs, which exhibit a profound influence on reservoir quality, particularly in their pivotal role in governing hydrocarbon enrichment. However, these shale lithofacies present substantial variations in rock composition, posing notable challenges for precise identification. Differentiating their small log responses in contrast to conventional laminated and interlayer-type shales further complicates their identification, underscoring the need for refined analytical techniques to accurately discern the nuances within this complex lithological framework. In this study, we introduce an interpretable and geologically constrained deep learning model, which is designated as the geological constrained convolution-gated recurrent unit (GCConv-GRU). The GCConv-GRU ingeniously integrates geological expertise, specifically the precession signal extracted from gamma ray (GR) curve, as training input for the purpose of lithofacies classification. In addition, we use purely data-driven machine learning methodologies, including gradient boosting decision tree (GBDT), long short-term memory (LSTM), and Conv-GRU, to benchmark the performance of GCConv-GRU, enabling a comprehensive comparison of classification accuracy. It was revealed that the lithofacies identification outcomes produced by the GCConv-GRU model exhibited a remarkable congruency with the logging profile, and it is capable of identifying thin layers of felsic shale within mixed shale formations. Meanwhile, pure data-driven machine learning methods, such as GBDT and LSTM, struggled to discern different shale lithofacies. Consequently, we also discuss the paleoenvironment and vertical distribution of shale lithofacies intricately regulated by astronomical cycles. This study underscores the advantages of dual knowledge- and data-driven models, which combine the powerful fitting capabilities of deep learning algorithms (DLAs) with consistent geological principles.