An integrated simulation–learning framework for rapid prediction of regional snow avalanche runout and hazard metrics

Abstract

Snow avalanches pose significant and growing risks in the southeastern Tibetan Plateau, where steep terrain and limited data availability challenge effective hazard assessment. This study proposes a hybrid modeling framework (ADS-DNN), which integrates a limited set of physically based avalanche simulations with deep neural networks to enable rapid prediction of key hazard metrics at a regional scale. A total of 206 simulations were conducted using real terrain and snow parameters across diverse avalanche-prone basins. The simulation results serve as training data for the neural network, which uses terrain and snow features, such as slope, elevation, depth, and density, to predict four key indicators: the maximum runout distance, velocity, flow depth, and deposition area. Field photos were used to validate the simulation reliability and support model calibration. Applied to the Nyingchi region, the ADS-DNN model achieves high predictive performance while reducing computation time from four days to a few minutes. This framework provides a scalable and transferable solution for avalanche hazard mapping and early warning in mountainous regions with limited monitoring data. While the approach demonstrates good performance, its accuracy depends on the representativeness of the simulated scenarios and is constrained by the limited availability of detailed field observations.