Physics-Informed Seismic Inference of Rockfall Sources and Motion Regimes

Rockfalls radiate complex seismic signals that encode both evolving source dynamics and path-dependent interactions, yet these signals are rarely exploited to support real-time trajectory inference within a unified framework.Here we develop a physics-constrained framework that integrates deep learning with forward motion modeling to jointly infer rockfall source location and motion mode from multi-station seismic observations, and to translate these inferences into early trajectory prediction.A spatiotemporal network combining temporal convolution and graph convolution exploits inter-station waveform variability to estimate three-dimensional source locations and discriminate motion regimes in near real time.Field experiments in China and the French Alps demonstrate meter-scale localization accuracy and enable early estimates of subsequent impact points before terminal deposition, providing actionable lead time for dynamic hazard response.Guided by rockfall source–path mechanisms, we further introduce spatial information as a physically meaningful proxy for propagation effects, which substantially improves motion-mode discrimination and yields spectrally coherent attention patterns consistent with observed impact and rolling processes.Finally, we show that localization accuracy is jointly controlled by dataset size and spatial scale, revealing that site-scale effects can outweigh gains from simply increasing sample numbers.Together, these results demonstrate that embedding physical cognition into deep learning enables seismic wavefields to be translated into real-time, interpretable constraints on rockfall dynamics, outlining a pathway toward physics-informed monitoring of gravity-driven hazards.