Identification of Laminar Structure in the Yingxiongling Shale Oil Sediment in China with Random Forests and SHAP Analysis

In this study, a diagnostic machine learning approach is proposed for automatic identification of laminar structures in sediments. The Yingxiongling shale oil in the Qaidam Basin of NorthWest China is used as a case study. First, various experimental methods were used to investigate the characterization of shale hydrocarbon potential, storage and seepage capacity, pore structure, and compressibility properties by sedimentary structures. Then, a particle swarm-optimized random forest algorithm (PSO-RF) is used to achieve continuous automatic identification of sedimentary formations, and the identification results are comprehensively interpreted using the Shapley Additive exPlanations (SHAP) method. The results show that laminated shale exhibits higher hydrocarbon generation, better seepage ability, and higher compressibility. In contrast, layered shale has a stronger storage capacity. The PSO-RF methodology achieves an approximate 90% accuracy in identifying sedimentary structures. Additionally, the SHAP method furnishes both global and local elucidations of the model's predictions. These interpretations enhance the model's trustworthiness and transparency. This provides a reference for further exploration and development of shale oil reservoirs in the study area.