Physics-Informed Residual Convolutional Network Model for Depth-Averaged Landslide Dynamics

Rapid landslide motions control impact area, flow velocity, deposition pattern, and, in extreme cases, are a river-blocking hazard; therefore, reliable dynamic simulations are of direct importance to engineering–geological hazard assessments. Depth-averaged models provide an efficient framework for simulating large-scale mass movements, but conventional physics-informed neural networks (PINNs) remain challenged with regard to nonlinear flows, which can limit their applicability in landslide analysis. To address these limitations, this study develops a physics-informed residual convolutional network model (PI-RCN) for depth-averaged landslide dynamics. The proposed framework combines sequential residual learning with depth-wise separable convolutions (DSCs) and incorporates physics-based residuals, mass conservation, and hard constraints to preserve physical consistency during time marching. The model is evaluated using a 1+1D frictionless dam-break benchmark, a Hong Kong landslide, and the Yigong rock avalanche. Results show that PI-RCN accurately reproduces the benchmark flow evolution with substantially fewer trainable parameters than a baseline fully connected PINN. In the Hong Kong case, the model demonstrates improved convergence stability and optimization efficiency. In the Yigong case, PI-RCN reproduces the main spatiotemporal evolution and multi-stage velocity variation of a long-runout rock avalanche. These results suggest that PI-RCN provides a useful physics-informed framework for efficient and consistent landslide dynamic simulation.