In-situ stress inversion method in reservoirs: An efficient framework based on coupled numerical simulation and reinforcement learning

The in-situ stress field is a key factor influencing reservoir geological stability. High-efficiency, high-accuracy in-situ stress inversion is crucial for reservoir stability assessment and exploitation strategy optimization. To improve efficiency and accuracy, this study proposes an in-situ stress inversion method integrating numerical manifold models, efficient surrogate models, and deep reinforcement learning. It addresses traditional numerical simulation limitations in handling discontinuous interfaces and overcomes high computational costs and strong sensitivity to input conditions. The method's effectiveness is validated through an engineering application for a shale gas reservoir in the Sichuan Basin. Results show that the surrogate model constructed from extensive simulation data significantly accelerates the inversion process. While maintaining accuracy, its computational efficiency is 14 times higher than the numerical simulation model. Compared to the single-step scheme, the multi-step inversion optimization scheme, which dynamically adjusts the inversion path through staged, iterative steps, provides more accurate results under the same parameters. The maximum error rates of normal stress and shear stress are reduced by 6.78% and 198.23%, respectively, while corresponding average error rates decrease by 1.72% and 23.53%. The proposed method provides an effective tool for mitigating geological hazards and optimizing exploitation, demonstrating strong potential for practical engineering applications.