Characterizing irregularly shaped rockfall dynamics using seismic signals: Geometric source effects combined with deep learning-driven strategy

Seismic signals generated by rockfall-ground impacts encode critical information about source dynamics and medium response. However, establishing quantitative, physics-based relationships remains challenging because of complex contact mechanics and distortion of high-frequency waveforms. To address nonlinear variations in source behavior and wave propagation under different contact conditions, we develop a hybrid physics-AI framework that integrates a visco-elastoplastic impact simulator with a conditional variational autoencoder (CVAE)-based seismic propagation model. The source module resolves impact dynamics for three representative contact-nose geometries, while the CVAE module leverages empirical Green’s functions to reproduce frequency-dependent wavefield evolution. Validation using 126 controlled single-block experiments and multi-sensor seismic recordings demonstrates that the framework reliably captures seismic signatures associated with different impact geometries. The results reveal that rock–ground contact geometry exerts a dominant control on signal characteristics, underscoring the need for physics-informed constraints on nose morphology to achieve robust dynamic parameter inversion. This study establishes a mechanics-guided inversion strategy that directly links recorded waveforms to impact kinematics, providing a new pathway for quantifying rockfall hazards from seismic observations.