Prediction and Controlling Factors of Macrolithotypes in Deep Coal Reservoirs.

Deep coalbed methane (CBM) in the Ordos Basin is a critical resource for China's unconventional gas supply. Accurate identification of coal macrolithotypes is essential for reservoir characterization and guiding efficient exploration and development. However, the deep coal seams in the Ordos Basin feature complex geological conditions and strong spatial heterogeneity. Traditional prediction models exhibit low accuracy at the basin-wide scale, thereby constraining the fine-scale identification of macrolithotypes. This study proposes a machine learning prediction workflow driven by dual geological constraints: roof lithology and burial depth. Based on core samples and well-logging data, multiple machine learning models were constructed to evaluate and compare the performance of global models versus geologically partitioned models. The findings reveal that (1) Geologically partitioned modeling significantly outperforms global modeling, and the optimal prediction models vary distinctly under different geological backgrounds. (2) Within the depth and regional scope of this study, the distribution of macrolithotypes is dominated by sedimentary microfacies. Bright coal is most extensively developed beneath limestone roofs (accounting for 53.3%), corresponding to a low-lying mire facies with stable hydrodynamics, whereas semibright coal dominates beneath sandstone roofs. (3) A "geology-engineering" dual sweet-spot evaluation paradigm is proposed: the "limestone roof + bright coal" combination, characterized by exceptionally high gas content, serves as the premier geological resource sweet spot; conversely, the "sandstone roof + semibright coal" combination, featuring a higher elastic modulus and stress stability, acts as an engineering stable-production sweet spot that ensures long-term fracture conductivity. This study provides a robust scientific basis for sweet-spot evaluation and tailored fracturing designs in deep CBM exploration.