• New limit state solutions of tunnel stability in Bolton sand under blowout and collapse conditions using finite element limit analysis (FELA) • Developing a hybrid CatBoost Algorithm by coupling with Dwarf Mongoose Optimization • Employing SHAP analysis, performance metrics, and residual distribution to evaluate the performance of proposed machine learning models This study presents a hybrid framework combining finite element limit analysis (FELA) and machine learning (ML) to predict the stability factor of circular tunnels in dense sand. The Bolton model is incorporated into the FELA formulation to simulate non-linear strength behavior under blowout and collapse conditions. A parametric study reveals that relative density ( D R ), particle crushing strength ( Q ), and critical-state friction angle ( ϕ cv ) are the most influential factors governing tunnel stability. While the embedment ratio ( H/D ) significantly enhances blowout resistance, its effect on collapse capacity is less pronounced. Unit weight ( γ ) exhibits minimal influence in the tested range. To enhance predictive efficiency, a CatBoost algorithm optimized with Dwarf Mongoose Optimization (DMO) is developed and trained on FELA-generated data. The model demonstrates strong predictive performance (R² > 0.999), with low RMSE and MAE values observed across the training, validation, and testing phases. SHAP analysis confirms that the ML model identifies dominant features consistent with physical interpretations from FELA, ensuring both accuracy and interpretability. The integration of physics-based modelling and data-driven learning offers a reliable and computationally efficient tool for geotechnical analysis. This approach facilitates rapid assessment of tunnel stability in dense sand and supports informed decision-making in design and risk evaluation.
Suesaming et al. (Sun,) studied this question.