Rapid stability assessment of straight-wall arch tunnel via a high-fidelity data-driven surrogate

• The highlights of this manuscript are as follows: • A Python-FLAC3D automated framework was developed to generate a dataset of 10,000 tunnel stability samples, providing sufficient data for machine learning model training in geotechnical engineering applications. • An ensemble learning model was developed that combines multiple algorithms through voting strategies, achieving R² = 0.9979 for regression and 98.95% accuracy for classification in tunnel deformation prediction. • Feature importance and SHAP analysis identified elastic modulus as the primary controlling factor for tunnel deformation, with an importance score of 0.608, providing guidance for parameter selection in tunnel engineering. Accurate prediction of excavation-induced peak deformation is critical for tunnel stability assessment. While a single FLAC 3D run is not computationally prohibitive, orchestrating and executing large numbers of simulations for broad design-space studies remains operationally cumbersome. We therefore construct a dataset from high-fidelity simulations via a FLAC3D–Python pipeline for vertical-wall, arched-roof tunnels, generating 10,000 simulations across diverse rock properties, and train a data-driven surrogate on top of it. The proposed approach couples HF simulation with machine learning: an ensemble of eight regressors forms a voting regressor to emulate peak deformation with low-latency inference, and six classifiers are aggregated by soft voting to provide a five-level stability rating defined by deformation thresholds, yielding probabilistic outputs for enhanced robustness and basic uncertainty awareness. Feature engineering is physics-informed: five fundamental mechanical parameters ( E, ν, φ, c, σ t ) are augmented with four quantities derived from elastic constitutive relations—bulk modulus ( K ), shear modulus ( G ), tan(φ) , and sin(φ) —resulting in a nine-dimensional feature space. Extensive experiments show R 2 =0.9979 and RMSE=0.00575 for regression and 99.35% accuracy for classification. The surrogate maintains fidelity to the HF simulator while substantially reducing the operational burden of large-scale parameter exploration, supporting rapid scenario screening, sensitivity analysis, and preliminary design evaluation.