Field-Informed Hydro-Mechanical Identification and Data-Driven Inversion of an Equivalent Groundwater Leakage Pathway in a Deep Ship-Lock Excavation Along the Xiangjiang River, China

Abstract

Localized groundwater leakage through retaining and cutoff systems can reorganize hydraulic gradients and trigger deformation that is difficult to distinguish from the normal response to excavation and dewatering. This study tests whether spatial deformation residuals, interpreted with transient hydro-mechanical modeling, can identify and characterize an equivalent leakage pathway at a 31 m-deep ship-lock excavation along the Xiangjiang River, China. A three-dimensional fully coupled model using the Hardening Soil model with small-strain stiffness was validated against inclinometer and surface-settlement measurements. Relative error in maximum wall displacement was 7.1% with transient seepage, compared with 18.7% under a steady-state assumption. At the final excavation stage, residual screening localized anomalous deformation to monitoring sections SC08–SC09 and vertical coordinates Z = −22 to −19 m, within the rounded-gravel layer. A local permeable zone was subsequently introduced into the equivalent wall, and 200 orthogonal Latin hypercube simulations sampled its hydraulic conductivity, equivalent area and position. Within the investigated ranges, hydraulic conductivity and equivalent area caused the largest deformation changes. The CNN–BiGRU surrogate yielded an independent-test coefficient of determination R2 of 0.93 and yielded k = 6.5 × 10−4 cm s−1, A = 16.2 cm2, Z = −20.8 m and X = 32.7 m. To quantify the reported k–A non-uniqueness, a dense sweep of 10,201 k–A combinations was performed at the inverted X and Z coordinates, producing a low-misfit feasible valley in the k–A plane. For a normalized response misfit not exceeding 10%, the feasible combinations were approximately k = 2.8 × 10−4 to 1.6 × 10−3 cm s−1 and A = 10.0 to 22.5 cm2, while the minimum-misfit core (J ≤ 8%) clustered around k = 4.8 × 10−4 to 8.8 × 10−4 cm s−1 and A = 13.5 to 18.8 cm2. Reintroduction of the inverted parameters reproduced wall displacement and ground settlement within 6.9–8.1% of field measurements. The inferred parameters represent an equivalent seepage channel rather than a unique defect geometry, and the feasible-region analysis shows that hydraulic conductance is better constrained than k or A individually.