An application of coupling RAFELA and CatBoost model for tunnel stability prediction by considering the nonstationary random field of undrained shear strength

This study presents a novel machine learning framework for predicting the probability of failure (PoF) of tunnels constructed in spatially random clays considering a soil strength gradient that increases with soil depth. By using random adaptive finite element limit analysis (RAFELA) and an advanced machine learning algorithm, namely Categorical Gradient Boosting or CatBoost, this research offers a robust alternative model to predict the probabilistic tunnel stability number, which is more effective than traditional deterministic methods. The proposed approach considers four critical factors such as the cover depth ratio (H/D), strength gradient factor (ρH/μ_Su0), dimensionless vertical correlation length (CL_y/D), and coefficient of variation (COV_Su), along with probabilistic simulations using Monte Carlo (MC) techniques. A dataset of 864 samples was used to train and validate the models, with 70% allocated for training and 30% allocated for testing. The CatBoost algorithm demonstrated exceptional predictive accuracy, with high R² values and minimal error metrics, making it a powerful tool for estimating tunnel stability in the context of soil spatial variability. This work contributes to more reliable and risk-informed tunnel design practices by incorporating stochastic variability and machine learning into geotechnical assessments, providing engineers with advanced methods for predicting tunnel failure probability under uncertain conditions.