Landslide-generated impulse waves (LGIWs) are highly destructive hydrodynamic phenomena in mountainous reservoirs, characterized by complex near-field evolution and pronounced nonlinear-dispersive wave effects. To investigate the generation and propagation mechanisms of granular LGIWs and to establish a physically consistent predictive framework, this study integrates Froude-scaled physical model experiments with physics-informed deep learning. A representative granular reservoir-bank landslide in a deep V-shaped reservoir was selected as the prototype, and 27 experiments were conducted with varying landslide volumes, sliding velocities, and still-water depths. The results show that the generated wave trains exhibit weakly nonlinear oscillatory behavior, with the maximum amplitude generally occurring at the third wave crest. The maximum wave amplitude increases with landslide volume and sliding velocity, and decreases with still-water depth, reflecting the influence of momentum input and hydrodynamic confinement on energy transfer. An empirical formula for predicting the maximum amplitude is established based on these relationships. A coupled transformer-physics-informed neural network model is further developed to predict the spatiotemporal evolution of LGIWs. A modified Korteweg–de Vries equation is incorporated as a governing physical constraint during training. The predictions show excellent agreement with experimental data, with coefficients of determination consistently exceeding 0.97 across all 27 cases. Model analysis indicates that the governing-equation constraint improves the physical consistency of wave propagation, while boundary-condition constraints enhance waveform stability. This study advances the understanding of granular LGIW dynamics and provides a physically consistent framework for impulse-wave prediction and hazard assessment.
Wang et al. (Mon,) studied this question.