Mass flow runout prediction using neural network emulators

Geohazard mass flow runout prediction is critical for protecting lives, infrastructure, and ecosystems. Rapid mass flows such as landslides, and avalanches are among the most destructive geohazards, often travelling many kilometres from their source. Uncertain initial conditions and strong sensitivity to topography make these events difficult to anticipate, particularly for downstream communities that may be exposed to severe impacts with little warning. In this context, computational speed is essential for enabling timely forecasting and scenario-based risk assessment.

Accurately predicting runout requires models that are both physically realistic and computationally efficient. However, existing approaches face a fundamental trade-off between realism and speed, limiting their use for large-scale forecasting, ensemble analysis, and operational early warning.

Here we demonstrate that neural networks can emulate the final outcomes of mass flow runouts across diverse real-world terrains. Our model is trained on approximately 90,000 high-fidelity simulations spanning more than 5,000 globally representative topographies. The model predicts both flow extent and deposit thickness with high spatial accuracy while achieving computation speeds orders of magnitude faster than numerical solvers. Importantly, the emulator reproduces key emergent physical behaviours, including avulsion and heterogeneous deposition patterns, and generalizes across a wide range of rheologies, volumes, and terrain types. Probabilistic outputs further enable scalable uncertainty quantification.

These results show that data-driven emulation can shift geohazard runout forecasting from site-specific analysis towards rapid prediction frameworks, supporting impact-based early warning and regional-scale hazard assessment. We anticipate that this approach will form a foundation for next-generation forecasting models that integrate physical simulation and machine learning to address transient dynamics, multi-hazard interactions, and cascading effects relevant to landslide hazard forecasting in space and time.