Abstract: In-situ stress inversion is a crucial method for determining the initial stress state in underground engineering. To improve the inversion accuracy, this study proposes a multi-model fusion method for in-situ stress inversion correction based on the stacking algorithm, demonstrating the limitations in single intelligent algorithms, particularly their insufficient generalization capabilities. We develop a stacking ensemble model with input data from the errors of finite element analysis results and multiple linear regression values against the in-situ measurements obtained at borehole monitoring points. This model adopts LightGBM, XGBoost and Linear Regression as base learners and RidgeCV as the meta-learner, and determines error correction by learning the underlying error patterns, with the optimal parameters selected through cross-validation and grid search. This method preserves structural and physical significance inherent in the finite element model while overcoming the nonlinear fitting deficiencies of traditional methods. Its application to an engineering case study demonstrates that compared to conventional multiple linear regression methods, it reduces the Root Mean Square Error (RMSE) by roughly 32% and the Mean Absolute Relative Error (MARE) by 45%. We use the SHapley Additive exPlanations (SHAP) model to interpret the contributions of various geological factors on the predicted and corrected errors, thereby enhancing the interpretability of this stacking-based inversion correction model. This study shows our new method exhibits strong applicability and significant potential for generalization in in-situ stress inversion correction, and would help engineering design and safety assessment.

Key words: in-situ stress inversion, stacking algorithm, finite element analysis, machine learning, error correction, SHAP interpretability model