Physical-Informed Neural Network for Predicting Spatiotemporal Variation of Pore Water Pressure in Soils Due to Consolidation

Rapid urbanization has caused numerous construction solid waste landfills. The historical stepwise landfilling process may generate excess porewater pressure (PWP) within the slope, dissipatinggradually due to the consolidation. Accurately modeling the spatiotemporal variation of PWP is essential for keeping slope stability. The physical-informed neural networks (PINNs) provide a promising approach for predicting soil responses in geoengineering. However, PINNs' training is often unstable and inefficient. This study develops a novel PINN framework with an enhanced network architecture to model soil consolidation behaviors. The enhanced architecture modifies the traditional fully connected neural networks (FNNs) by adding two additional transformer networks. Compared to the vanilla PINNs with traditional FNNs, the novel PINNs demonstrate enhanced training stability inforward simulation and convergence speed in the inverse analysis through a benchmark test of Terzaghi's problem. The novel PINNs are then used to model a consolidation problem close to real-world scenarios with continuous drainage boundary conditions and noisy monitoring data. The results demonstrate that the novel PINNs can accurately identify the coefficient of consolidation and interface parameters simultaneously in a noise dataset. Furthermore, the model not only reconstructs the excess PWP variation at the sampled locations but also provides reliable prediction in the unsampled areas, demonstrating the physical interpretability of PINNs. This methodology offers a new idea for developing a geo-digital twin model by incorporating real monitoring data and respecting the physical laws.

Keywords: Data-driven, Physical-informed, PINNs, Consolidation, Digital twin.