Modeling subsidence and building damage in central Gothenburg using machine learning

Other conference contribution, 2025

Subsidence is an important consideration affecting many cities around the world. It leads to an increased risk of flooding and may cause damage to buildings and other infrastructure, especially in low-lying urban areas on soft clays. Physics-based numerical models can provide estimates of ongoing and future subsidence, caused by the combined time-dependent processes of creep and consolidation, thereby increasing our understanding of when and where deformations will arise and at what magnitude. However, such models are computationally expensive and generally not feasible at a large scale with varying stratigraphy. To address these challenges, we apply a recently developed modeling approach that uses the data from a physics-based model to train a machine learning-based metamodel for subsidence prediction and a consequent building damage model. The approach was applied to an area within City of Gothenburg, Sweden, simulating ongoing background creep subsidence over a 30-year period. The predicted subsidence is validated against observed settlement rates from interferometry (InSAR) and point-based depth-integrated measurements from a bellow-hose. The building damage is calculated on a large scale using the limiting tensile strain method. The results demonstrate that the metamodel effectively represents subsidence as a function of time and space. The areas with discrepancies in comparison to InSAR are explained by ongoing construction and the effects of relative soil-foundation stiffness.