From wetting-front coalescence to urban road collapse: interpretable machine learning of hydro-mechanical failure under coupled rainfall and pipeline leakage

Urban ground collapse poses a severe threat to public safety, road operation and underground infrastructure in densely developed cities. This study investigates the hydro-mechanical failure mechanism of road collapse induced by coupled rainfall infiltration and pipeline leakage through physical model testing, finite element simulation and interpretable surrogate modelling. A validated numerical model incorporating saturation-dependent strength and stiffness degradation was used to generate 120 cases with different pipe burial depths, rainfall intensities and leakage pressures.The results indicate that collapse develops through shallow-soil pre-wetting, leakage-induced wetting-zone expansion, convergence of the upper and lower wetting fronts, and upward propagation of deformation. Leakage pressure is the dominant factor, contributing 86.42%, 87.79% and 92.77% to the variations in collapse depth, near-saturated zone area and wetting-front depth, respectively. Rainfall mainly dissipates matric suction and lowers the failure threshold, whereas burial depth regulates the wetting-front convergence and deformation-transfer path. GPR/Kriging achieved test-set (R 2 ) values of 0.997–0.998, while XGBoost-SHAP confirmed the mechanism of leakage-pressure domination, rainfall amplification and burial-depth regulation. Risk zoning further shows that increasing rainfall shifts high-risk conditions towards lower leakage pressures, particularly at intermediate burial depths. The proposed framework provides an efficient tool for identifying vulnerable pipeline sections and supporting targeted inspection and collapse prevention.