An innovative approach that combines the Upper-Bound finite-element method with Rigid Translatory Moving Elements (UBRTME) and Artificial Neural Networks (ANN) is presented in this paper to efficiently predict the critical load factors and failure characteristics of circular tunnels in heterogeneous clay. A high-fidelity dataset of primary tunnel failure characteristics is constructed by extracting meshlike failure mechanisms generated by the UBRTME analysis. ANN hyperparameters are systematically optimized through a combined strategy of K -fold cross-validation and early stopping. The optimized model exhibits robust performance across a range of tunnel burial depths and soil strength conditions, providing an accurate and computationally efficient tool for failure characteristic predictions of circular tunnels in heterogeneous undrained clay. Furthermore, the critical failure surface is reconstructed from the ANN outputs using cubic B-spline interpolation, and the Hausdorff distance is employed to quantify error transfer of the reconstruction process. By employing Permutation Feature Importance (PFI) analysis and SHapley Additive exPlanations (SHAP), this study quantitatively assesses the relative effects of input parameters on both tunnel stability and the morphology of the ultimate slip surfaces, thereby enhancing model interpretability and practical utility. This work effectively bridges limit analysis theory with interpretable machine learning, offering an explainable and practical framework for geotechnical failure prediction while preserving physical interpretability.
Yang et al. (Mon,) studied this question.