Landslide runout no longer needs a solver at every step: a physics-constrained emulator runs about 10 ms per step and roughly 50 times faster over a 110-second rollout
Synopsis
Landslide runout forecasts usually come from numerical simulations that solve depth-averaged flow equations, and repeated runs for parameter sweeps are costly. A physics-constrained neural network was trained as a single-step surrogate that advances the thickness and velocity fields by a fixed time interval. On the test set, single-step root mean square errors were about 0.06 m for thickness and 0.11 m/s for velocity; after 110 autoregressive steps they were about 0.21 m and 0.09 m/s, with the full rollout taking about 1.1 s against about 16.7 s for the numerical solver.
Interpretation
Landslide runout prediction moves from solving partial differential equations step by step to a neural network that advances by a fixed time interval and directly returns the next thickness and velocity fields. Earlier neural surrogate models mapped landslide evolution straight to the final deposit and skipped intermediate states, while impact-area estimates in risk assessment need the temporal evolution. On 228,662 adjacent-time sample pairs generated over 12,214 digital elevation models from 19 mountain regions worldwide, the physics-constrained model reached an intersection over union of 0.989 for thickness and 0.904 for velocity on the test set, with Pearson correlations of 0.999 and 0.993.
The model takes the current thickness and velocity plus terrain features as input and outputs both physical fields for the next moment in one inference, and feeding outputs back as inputs reproduces the whole runout. It replaces a single numerical iteration of a depth-averaged flow model that treats the landslide as a continuum with uniform viscous-plastic properties and solves it with a staggered finite volume method on a 256 by 256 grid with 30 m spacing. On an unseen terrain in Alaska, the model rolled out 110 steps until the material stopped, with boundary intersection over union stabilizing near 0.94 after 30 s and thickness error rising from about 0.08 m at the first step to 0.20 m after 60 s before plateauing.
Adding mass conservation and kinetic energy loss to the training loss keeps predictions physically consistent at the aggregate level, not merely low in pointwise error. Without the physics constraints, single-step biases were small but accumulated during autoregression, and the final deposited volume fell to about one-third of its initial value while total kinetic energy reached zero about 20 s earlier than the reference. With the constraints, the single-step shift in total volume fell from -2130 to 36 cubic meters and the shift in maximum velocity from -0.043 to -0.025 m/s; a step takes about 10 ms and the 110-step rollout about 1.1 s, against 500 ms per step and 16.7 s for the numerical solver.
Perspective
The demonstrated scope is a two-dimensional depth-averaged flow model with one material parameter set and a uniform Gaussian initial thickness distribution, on terrain drawn from 12,214 elevation chips at 30 m resolution across 19 mountain regions; the model completed a 110-step autoregressive runout on an unseen Alaskan terrain.
Training data cover only one flow mobility regime and terrain-dependent friction is not explicitly represented, so the surrogate is not yet general enough for direct real-world use; the authors propose extending it with inputs such as material properties, initial source characteristics and spatially varying terrain friction.
