Physics-informed neural networks for predicting the coupled hydro-mechanical behaviour of partially saturated soils

Accurately modelling coupled hydro-mechanical behaviours of soils is critical for ensuring geoengineering safety. Physics-informed neural networks (PINNs) provide a promising approach to solving these coupled problems. However, traditional PINNs perform poorly in partially saturated soils under surface loading. This difficulty arises from the varying coupling degrees between the water and solid phases in the unsaturated and saturated zones, which leads to sharp changes in excess pore water pressure near the water table during the initial loading stage, thus complicating network training. To address this, this study develops a sequentially trained PINNs framework based on the fixed-stress split method, along with two novel enhancement strategies: initial condition-guided training and domain decomposition. These strategies help the sequentially trained PINNs effectively capture sharp pressure changes and improve prediction accuracy. The developed PINNs are subjected to comprehensive inverse analysis and compared with a finite element-based surrogate model under a rainfall scenario. PINNs display advantages in noise resistance when predicting hydro-mechanical parameters and avoiding costly data generation. However, PINNs face optimisation challenges with high computational cost and susceptibility to local minima. These results highlight the benefits of the developed PINNs and underscore optimisation challenges that need to be overcome for more practical applications in geoengineering.