Physics-Informed Deep Learning for Revealing the Evolutionary Characteristics of Landslides Induced by Rainfall Process
Citation
Li, S., Feng, X., Xu, Q., Guo, C., Zhao, K., Zheng, G. (2025). Physics-Informed Deep Learning for Revealing the Evolutionary Characteristics of Landslides Induced by Rainfall Process. Geophysical Research Letters, 52: e2025GL117356. Link to paper
Abstract
Rainfall-induced landslides involve complex interactions of multi-physical fields between seepage and stress fields. This study proposes a Physical Information-Driven and Guided Multimodal Multi-task Neural Network (PMNN), which explicitly integrates fundamental geomechanical equations, including the Van Genuchten seepage model and Bishop's effective stress equation, while incorporating hydromechanical governing laws within the loss function to enforce physical constraints. Experiments on 800 numerically simulated and 55 model tested rainfall scenarios demonstrate that the internal physics embedding leads to over a twofold improvement in multi-physical fields prediction accuracy, while external physical constraints effectively reduce stability prediction errors. High-intensity rainfall causes rapid surface saturation, leading to quick accumulation of pore pressure in shallow layers, thereby forming a shallow critical zone that is more prone to trigger shallow landslides.