Optimized prediction of tunnel stability using advanced machine learning and an ANN-based analytical expression

Abstract

This study introduces a hybrid machine learning framework for predicting the stability of circular and rectangular tunnels in cohesive-frictional soils under surcharge loading. Leveraging high-fidelity datasets generated via Isogeometric Analysis (IGA) and Upper-Bound Limit Analysis (UBLA), the model captures a wide range of geometric and geotechnical scenarios to ensure robust learning and generalization. A Real-Coded Genetic Algorithm (RCGA) and Tree-Structured Parzen Estimator (TPE) are employed for hyperparameter optimization, significantly improving model accuracy compared to traditional numerical and empirical methods. The RCGA-optimized Artificial Neural Networks (ANNs) and CatBoost models demonstrate strong adaptability to varying tunnel geometries and soil conditions. Interpretability is enhanced through SHAP (SHapley Additive exPlanations) values, Accumulated Local Effects (ALE), and Partial Dependence Plots (PDPs), offering valuable insights into the influence of key design parameters. An explicit analytical expression is further derived from the trained model by extracting its weights and biases, enabling rapid and efficient tunnel stability assessments without resorting to computationally intensive simulations. This work establishes a new standard for tunnel stability prediction, combining the strengths of advanced numerical modeling and explainable machine learning to deliver a scalable, interpretable, and practical solution for geotechnical engineering applications.