The Middle Permian Maokou Formation carbonate rocks in the southern Sichuan Basin are import targets for hydrocarbon exploration, with numerous gas fields discovered in structural traps. However, as exploration extends into slope and syncline zones, the limestone reservoirs become denser, and fluid distribution becomes increasingly complex, limiting efficient exploration and development. Identifying the key factors controlling natural gas accumulation is therefore critical. This study is the first to apply deep learning techniques to fault detection in the southern Sichuan Basin, identifying previously undetected WE-trending subtle strike-slip faults (vertical displacement < 20 m). By integrating well logging, seismic, and production data, we highlight the primary factors influencing natural gas accumulation in the Maokou Formation. The results demonstrate that 80% of production comes from less than 30% of the well, and that high-yield wells are strongly associated with faults, particularly in slope and syncline zones where such wells are located within 200 m of fault zones. The faults can increase the drilling leakage of the Maokou wells by (7–10) times, raise the reservoir thickness to 30 m, and more than double the production. Furthermore, 73% of high-yield wells are concentrated in areas of fault intersection with high vertical continuity. Based on these insights, we propose four hydrocarbon enrichment models for anticline and syncline zones. Key factors controlling gas accumulation and high production include fault intersections, high vertical fault continuity, and local structural highs. This research demonstrates the effectiveness of deep learning for fault detection in complex geological settings and enhances our understanding of fault systems and carbonate gas reservoir exploration.