An Interpretable Metaheuristic-Optimized XGBoost Model for Plastic Zone Depth Prediction and Target-Oriented Parameter Screening in Underground Powerhouse Caverns

Yuxin Chen¹, Jian Zhou¹, Su Wang^2,*, Mohammad Rezaei³, Danial Jahed Armaghani^4,*
1 School of Resources and Safety Engineering, Central South University, Changsha, China
2 Kunming Prospecting Design Institute of China Nonferrous Metals Industry Co., Ltd., Kunming, China
3 Department of Mining Engineering, Faculty of Engineering, University of Kurdistan, Sanandaj, Iran
4 School of Civil, Environmental and Sociotechnical Engineering, University of Technology Sydney, Sydney, NSW, Australia
* Corresponding Author: Su Wang. Email: email ; Danial Jahed Armaghani. Email: email

Computer Modeling in Engineering & Sciences https://doi.org/10.32604/cmes.2026.087513

Received 17 June 2026; Accepted 27 August 2026; Published online 17 September 2026

Download PDF

Abstract

Rapid prediction of the plastic zone depth (PZD) at key points in the rock mass around underground powerhouse caverns is important for surrounding-rock stability assessment. However, existing machine-learning studies have primarily focused on forward prediction under prescribed conditions, with insufficient attention to model stability, prediction uncertainty, and the statistical significance of performance differences. Moreover, model interpretation has rarely been integrated with target-oriented cavern-layout parameter screening. To address these limitations, this study develops a metaheuristic-optimized XGBoost framework based on a numerical database containing 1920 two-dimensional finite-element results. The Runge-Kutta optimizer (RUN) and the weighted mean of vectors optimizer (INFO) are employed to search seven key XGBoost hyperparameters, while an unoptimized XGBoost model is established as a baseline within the same model family. Cross-validation, an independent test set, six evaluation metrics, bootstrap confidence intervals, residual diagnostics, and the Wilcoxon signed-rank test are combined to comprehensively evaluate model accuracy, stability, and statistical differences. The test results show that, compared with the unoptimized XGBoost model, RUN-XGBoost and INFO-XGBoost reduce RMSE by 27.63% and 28.94%, respectively, and MAE by 35.16% and 35.12%, respectively. Both optimized models achieve statistically significant reductions in absolute prediction errors, whereas the performance difference between them is not statistically significant. SHAP analysis indicates that, within the current database and parameter ranges, the coefficient of lateral stress, overburden depth, rock mass unit weight, tensile strength, and cavern-layout parameters are the main variables influencing the model predictions. The trained RUN-XGBoost model was further incorporated into a target-oriented parameter-screening framework. This framework rapidly identifies cavern-layout combinations that satisfy a prescribed PZD-control target and may reduce the need for repeated numerical simulations during preliminary scheme comparison. Finally, a graphical user interface integrating PZD prediction, local interpretation, parameter screening, result visualization, and data export is developed. The proposed framework can provide auxiliary support for rapid PZD prediction and preliminary cavern-layout scheme assessment within the parameter ranges covered by the current numerical database.