Neural Network Inversion Algorithm for Geostress Field Based on Physics-Informed Constraints

Traditional methods for geostressfield inversion face issues such as weak physical interpretability and insufficient generalization ability. This study pioneers the application of Physics-Informed Neural Network (PINN) to this problem, developing a data- and physics-driven inversion algorithm. The framework incorporates a constitutive-equation-based regularized loss function as a hard constraint during training to ensure physical consistency. To address boundary load uncertainty, two quantification approaches—Bayesian linear regression and surrogate model optimization—are proposed to establish 95% confidence intervals for boundary coefficients. Verification based on simple three-dimensional models and actual geological models of mines shows that PINN inversion achieves a mean absolute relative error as low as 0.0772%, with an error of 15.67% under sparse sampling conditions—significantly lower than the 31.07% error of the traditional Back propagation neural network. This demonstrates excellent robustness and data efficiency. In the practical engineering application of complex geological bodies, the average error of principal stress inversion is 9.35% with a minimum error of 0.137%. All inversion results fall within the permissible accuracy range of engineering, and the stress distribution conforms to basic laws, with an average error of 0.453 in the constitutive relation. Compared with BP neural network and multiple linear regression methods, it shows obvious accuracy advantages. This method provides a new solution for intelligent ground stress prediction with high accuracy, high efficiency, and strong physical interpretability, and also lays the foundation for early identification of geological disasters.