A metamodel for estimating time-dependent groundwater-induced subsidence at large scales

Construction of large underground infrastructure facilities routinely leads to leakage of groundwater and reduction of pore water pressures, causing time-dependent deformation of overburden soft soil. Coupled hydro-geomechanical numerical models can provide estimates of subsidence, caused by the complex time-dependent processes of creep and consolidation, thereby increasing our understanding of when and where deformations will arise and at what magnitude. However, such hydro-mechanical models are computationally expensive and generally not feasible at larger scales, where decisions are made on design and mitigation. Therefore, a computationally efficient Machine Learning-based metamodel is implemented, which emulates 2D finite element scenario-based simulations of ground deformations with the advanced Creep-SCLAY-1S-model. The metamodel employs decision tree-based ensemble learners random forest (RF) and extreme gradient boosting (XGB), with spatially explicit hydrostratigraphic data as features. In a case study in Central Gothenburg, Sweden, the metamodel shows high predictive skill (Pearson's r of 0.9–0.98) on 25 % of unseen data and good agreement with the numerical model on unseen cross-sections. Through interpretable Machine Learning, Shapley analysis provides insights into the workings of the metamodel, which alignes with process understanding. The approach provides a novel tool for efficient, scenario-based decision support on large scales based on an advanced soil model emulated by a physically plausible metamodel. • A ML-based metamodel emulates a hydro-geomechanical model accurately. • Subsidence due to pore-pressure reductions in soft soil is estimated at large scale. • Predictions possible at high resolution with high computational efficiency. • Interpretable ML confirms that the metamodel matches physics. • Metamodels are a reliable basis for large scale infrastructure decision support.