Intelligent prediction of post-seismic debris flow source volumes: a physics-informed generative deep learning framework

Accurate prediction of postseismic debris flow source volumes is essential for geohazard risk management but remains challenging due to small event inventories, strongly nonlinear predictor interactions, and physically implausible extrapolations. This study proposes GAMMA, a framework integrating conditional diffusion based augmentation, cross attention based multimodal fusion, and an inventory consistent physics guided regulariser. A geomorphologically conditioned diffusion model generates 400 statistically consistent synthetic events conditioned on 60 measured events. A Transformer based predictor captures nonlinear couplings among landslide related and topographic attributes, while the physics guided regulariser suppresses nonphysical predictions in boundary conditions. In the Longmenshan fault zone, GAMMA achieves RMSE of 13.54, MAE of 8.25, and R2 of 0.9874, outperforming fifteen baseline regressors with RMSE and MAE reductions of 50.8 and 52.9 percent relative to XGBoost. Complementary fivefold cross validation under strict information isolation yields R2 of 0.912 ± 0.018 and confirms that the physics guided regulariser is the primary contributor enabling GAMMA to surpass baselines under small sample conditions, while diffusion generated events provide complementary coverage rather than substituting for real observations. The framework supports hazard zoning, threshold setting, and mitigation design in data scarce postseismic environments.