Neural emulation of gravity-driven geohazard runout

TL;DR

Geohazard runout prediction is critical for risk reduction but existing methods trade realism for speed. The paper presents a neural emulator trained on tens of thousands of physics-based, depth-averaged simulations to predict runout footprint and deposit thickness directly from real topography and rheology. It achieves 100–10,000× speedups over numerical solvers while reproducing key behaviors such as avulsion and deposition and generalizing across varied terrains and flow types. This enables near-real-time, spatially resolved probabilistic forecasts and large ensembles, paving the way for scalable, impact-based early warning systems.

Abstract

Predicting geohazard runout is critical for protecting lives, infrastructure and ecosystems. Rapid mass flows, including landslides and avalanches, cause several thousand deaths across a wide range of environments, often travelling many kilometres from their source. The wide range of source conditions and material properties governing these flows makes their runout difficult to anticipate, particularly for downstream communities that may be suddenly exposed to severe impacts. Accurately predicting runout at scale requires models that are both physically realistic and computationally efficient, yet existing approaches face a fundamental speed-realism trade-off. Here we train a machine learning model to predict geohazard runout across representative real world terrains. The model predicts both flow extent and deposit thickness with high accuracy and 100 to 10,000 times faster computation than numerical solvers. It is trained on over 100,000 numerical simulations across over 10,000 real world digital elevation model chips and reproduces key physical behaviours, including avulsion and deposition patterns, while generalizing across different flow types, sizes and landscapes. Our results demonstrate that neural emulation enables rapid, spatially resolved runout prediction across diverse real world terrains, opening new opportunities for disaster risk reduction and impact-based forecasting. These results highlight neural emulation as a promising pathway for extending physically realistic geohazard modelling to spatial and temporal scales relevant for large scale early warning systems.

Figures (4)

• Figure 1: Comparison between the structure of a physics-based granular flow model (a) and our geophysical flow neural network emulator (b).
• Figure 2: Comparison between numerical simulations and neural emulator predictions for three representative test cases. Each row corresponds to a different flow scenario with varying topography and material properties. Columns show (from left to right): simulated deposit thickness, emulated deposit thickness, binary agreement between simulated and predicted runout footprints, and residuals in deposit thickness. The emulator reproduces the spatial extent and thickness patterns of the numerical solutions with high fidelity, capturing key flow features such as channelized propagation and lateral spreading while reducing computation time from tens of seconds to $\sim$0.04 s per case. (a) $V = 9.7\times10^{6}$ m$^{3}$, $c = 21.3$ kPa, $\rho = 1477$ kg m$^{-3}$; (b) $V = 5.7\times10^{6}$ m$^{3}$, $c = 19.7$ kPa, $\rho = 2092$ kg m$^{-3}$; (c) $V = 9.6\times10^{6}$ m$^{3}$, $c = 6.2$ kPa, $\rho = 1391$ kg m$^{-3}$.
• Figure 3: Comparison between numerical simulations (left panels) and neural network predictions (right panels) for systematic variations in material properties on a fixed topography. Rows correspond to increasing cohesion ($c = 5$, $25$, and $50$ kPa), and columns to increasing bulk density ($\rho = 917$, $1700$, and $2650$ kg m$^{-3}$). Both models reproduce the expected decrease in mobility with increasing cohesion and decreasing density, with consistent spatial patterns of deposition. Minor overprediction occurs for the highest densities, where nonlinear rheological effects are strongest, but overall the emulator captures the coupled influence of $\rho$ and $c$ on flow extent and deposit thickness with high fidelity.
• Figure 4: Probabilistic ensembles for three distinct types of rapid mass movements, showing pixelwise probability of reach and the 50th and 90th percentile deposit-thickness quantiles derived from 1,024 emulator runs. Each ensemble explores uncertainty across key flow parameters---volume, bulk density, and cohesion---within physically plausible ranges. (a) Zymoetz River landslide, Canada: rock avalanche transitioning into debris flow, with $V = 8\times10^{5}\text{--}3\times10^{6}\,\mathrm{m^{3}}$, $\rho = 1600\text{--}2200\,\mathrm{kg\,m^{-3}}$, and $c = 5\text{--}50\,\mathrm{kPa}$. (b) Swiss Alps avalanche: snow/ice avalanche with $\rho = 917\text{--}1100\,\mathrm{kg\,m^{-3}}$, $c = 5\text{--}15\,\mathrm{kPa}$, and $V = 8\times10^{5}\text{--}3\times10^{6}\,\mathrm{m^{3}}$. (c) Maoxian landslide, China: rock avalanche with $V = 5\times10^{6}\text{--}1\times10^{7}\,\mathrm{m^{3}}$, $\rho = 1600\text{--}2400\,\mathrm{kg\,m^{-3}}$, and $c = 5\text{--}50\,\mathrm{kPa}$. All ensembles were computed in $90\text{--}110$ s on a single GPU. The maps illustrate the emulator's ability to generate spatially resolved probabilistic forecasts across diverse flow types and rheological conditions. Source: Google Satellite 2025, Copernicus Sentinel-2.