Item: A MACHINE LEARNING SURROGATE FOR ALPINE3D SNOWPACK SIMULATION: PERSISTENCE-BASED EVALUATION AND A PATH TOWARD SNOWDRIFT
-
-
Title: A MACHINE LEARNING SURROGATE FOR ALPINE3D SNOWPACK SIMULATION: PERSISTENCE-BASED EVALUATION AND A PATH TOWARD SNOWDRIFT
Proceedings: Proceedings, International Snow Science Workshop, Whistler, BC, Canada, 2026
Authors:
- Michael Brandon Hurd [ AutoRoboto, San Francisco, CA, USA ]
Date: 2026-09-28
Abstract: Distributed snow-cover models such as Alpine3D and SNOWPACK provide spatially detailed information on snowpack evolution, surface energy balance, and terrain-driven variability that is relevant to avalanche operations. The drawback is the computational cost of high-resolution, multi-season simulations, which limits repeated scenario testing, ensemble analysis, and rapid visualization. This paper presents a physics-supervised graph neural network surrogate for selected gridded Alpine3D/SNOWPACK outputs over a Kananaskis Country case-study domain in the Canadian Rockies. The foundation model was trained on multiple winter seasons and evaluated on a fully held-out season. The model represents the distributed domain as a graph and uses terrain, previous snow state, meteorological forcing, and time to predict selected snow-state and process variables. The first evaluation compared absolute snow height (HS) and snow water equivalent (SWE) predictions with a copy-previous-hour persistence baseline. Although the model achieved R² values of 0.99886 for HS and 0.99939 for SWE, persistence performed better, with R² values of 0.99992 and 0.99998. For this reason, hourly changes were reconstructed from predicted process variables. The reconstructed changes achieved positive skill against persistence of 0.373 for ΔHS and 0.795 for ΔSWE, with the strongest performance during melt and runoff. A separate frozen-checkpoint test on 290 corrected-2022 timesteps and approximately 312 million valid cell observations retained positive flux-reconstructed skill of 0.479 for ΔHS and 0.461 for ΔSWE. This provides evidence of robustness to the corrected forcing lineage within the same domain. End-to-end inference required 23.8 s per predicted hour on the 1.12-million-cell domain, compared with 432 s per simulated hour for the measured Alpine3D workflow. This represents an approximately 18× system-level wall-clock speedup across different GPU and CPU hardware. The foundation teacher uses simple radiation and does not include wind-driven snow transport. Higher-fidelity physics extensions are being evaluated, but they remain preliminary and are not reported as validated results. The objective is not to replace Alpine3D, SNOWPACK, or field observations. The objective is to determine which teacher-model processes can be reproduced by a graph surrogate, how that reproduction should be evaluated, and what validation is required before operational use.
Object ID: ISSW2026_P1.58.pdf
DOI: https://doi.org/10.15788/1790098857
Language of Article: English
Presenter(s): Brandon Hurd
Keywords: graph neural network, Alpine3D, SNOWPACK, snowpack modelling, avalanche forecasting, persistence baseline
Page Number(s): 2255 - 2262
-