Enhancing debris-flow mobility and hazard forecasting using machine-learned spatial–temporal parameters in a two-phase flow model

Debris flows are destructive multi-phase events composed of soil, water, and rock, characterized by complex interactions between granular solids and interstitial fluids. While the Two-Phase Flow Model provides a physical foundation for these events, its predictive accuracy is often constrained by a reliance on fixed parameters that fail to account for significant spatiotemporal variations. This study bridges this critical gap by developing a novel physics–AI hybrid framework that replaces static assumptions with dynamic, machine-learned parameter fields. We integrate a Convolutional Neural Network (CNN) to extract geomorphological patterns from terrain data while utilizing a Gated Recurrent Unit (GRU) to capture temporal dynamics from rainfall forcing and vibration signals. A primary highlight of this integrated system is its superior predictive accuracy; validation across 72 events in the Xiuguluan River Basin demonstrates that this approach reduces Mean Absolute Error (MAE) by approximately 84% (from 0.225–0.338 m to 0.031–0.059 m). This framework maintains strict physical interpretability by using neural networks to optimize internal parameter fields rather than approximating the final output. While tested within two specific regional sub-basins, this adaptive framework advances debris-flow prediction by clarifying how dynamic parameters evolve across both space and time.