A Physics-Constrained Neural Differential Equation Framework for Data-Driven Snowpack Simulation
Andrew Charbonneau \aff a, Katherine Deck, \aff a Tapio Schneider, \aff a · 2024
Read A Physics-Constrained Neural Differential Equation Framework for Data-Driven Snowpack Simulation by Andrew Charbonneau \aff a, Katherine Deck, \aff a Tapio Schneider, \aff a (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-06819.txt.
Introduction
Andrew Charbonneau, Katherine Deck, and Tapio Schneider build a snow pack mock that puts physics in a neural math rule. The set traits tell how depth of snow by time of year shifts under force from the sky. Trained on more than one SNOTEL watch site, it calls day depth with less than nine per cent mid flaw. That holds through a wide range of snow zones.
The model also works at sites it did not see in train, which set snow models often fail to do. When it must call snow water weight as well as depth, flaw rises just a bit. Since rules of physics are safe by how it is built, the traits work at new time steps with no new train. The team sees scope for climate models and for more bound move sets.
Catalog record
| CREATOR | Andrew Charbonneau \aff a, Katherine Deck, \aff a Tapio Schneider, \aff a |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 84 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-06819
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A Physics-Constrained Neural Differential Equation Framework for Data-Driven Snowpack Simulation Andrew Charbonneau \aff a, Katherine Deck, \aff a Tapio Schneider, \aff a Abstract This paper presents a physics-constrained neural differential equation framework for parameterization, and employs it to model the time evolution of seasonal snow depth given hydrometeorological forcings. When trained on data from multiple SNOTEL sites, the parameterization predicts daily snow depth with under 9% median error and Nash Sutcliffe Efficiencies over 0.94 across a wide variety of snow climates. The parameterization also generalizes to new sites not seen during training, which is not often true for calibrated snow models. Requiring the parameterization to predict snow water equivalent in addition to snow depth only increases error to ∼ \sim 12%. The structure of the approach guarantees the satisfaction of physical constraints, enables these constraints during model training, and allows modeling at different temporal resolutions without additional retraining of the parameterization. These benefits hold potential in climate modeling, and could extend to other dynamical systems with physical constraints. 1 Introduction Seasonal snowpacks help regulate the Earth’s energy balance, provide freshwater storage, and are crucial for understanding Earth’s climate. They hold economic and ecological significance, supplying a majority of the western United States’ (and a sixth of the world’s) water supply and influencing agriculture, flood, drought, and avalanche hazards ( De Michele et al., 2013a ; Gao et al., 2021 ) . Their seasonal importance and susceptibility to climate change emphasize the need for ongoing modeling and monitoring on both seasonal and multi-decadal timescales. Modeling the evolution of seasonal snow for regional or global climate applications offers a challenging problem of scales; it is the bulk properties of the snow (albedo, snow cover fraction, snow temperature, and snow water content) that are critical, yet microphysical and location-specific processes control these properties and must be taken into account. The most detailed models represent vertically-resolved snowpacks, including liquid percolation, phase changes, metamorphism effects, and other types of compaction; they are often calibrated and used on the site-level (e.g., De Michele et al. 2013b ). Models used in global climate simulations range in complexity from single-layer/bulk models to multi-layer models with parameterizations for bulk properties that are calibrated with observational data ( Menard et al., 2021 ) ; the horizontal resolution of these models is typically {\sim}10 – 100 km where microscopic processes cannot be tractably resolved. While the laws of physics ultimately govern the evolution of snowpacks, uncertainty in how to relate essential but unresolved small-scale processes to snowpack bulk properties on the spatial scales of global models makes developing snow models a challenging task ( Kapnick et al., 2018 ; Bair et al., 2018 ) . This challenge is exacerbated by data availability ( Menard et al., 2021 ; Kouki et al., 2022 ) . Among bulk variables in global snow models, the snow water equivalent, \mathrm{SWE} (units of \mathrm{m} ), represents the total water storage in snow. It relates to the snow depth z (units of \mathrm{m} ) through the bulk snow density \rho_{\mathrm{snow}} (units of \mathrm{kg~m^{-3}} ) and the density of liquid water \rho_{\mathrm{water}} (1000 \mathrm{kg~m^{-3}} ) as \mathrm{SWE}=(\rho_{\mathrm{snow}}/\rho_{\mathrm{water}})\;z; (1) it is typically used as a prognostic variable in bulk models for determining snowpack mass balance. Density and depth are critical in climate models for determining the snowpack energy balance, as they influence thermal, mechanical, and optical properties, impacting mass/energy fluxes, water retention, and spring thaw ( Kouki et al., 2022 ; Bormann et al., 2013 ) . However, for a given \mathrm{SWE} , the snow depth and bulk density can vary considerably over time at a single location, or between locations under similar forcings, due to compaction, melt/refreeze cycles, and changes in the density of falling snow. These variations give rise to ongoing challenges in snow modeling. Many prevalent snow models rely on the seminal parameterizations of Kojima (1967) and Anderson (1976) to derive snow density and depth from modeled \mathrm{SWE} . These works modeled compaction and microstructure metamorphism by assuming a linear relationship between the strain rate of compaction and the weight of the overlying snow and suggested empirical parameterizations that were calibrated from a select number of observational sites and laboratory experiments and analytically extended. Snow evolution in models such as Snow17, iSnobal, Noah/NoahMP, CLM5, HTESSEL, and others ( Anderson, 2006 ; Marks et al., 2018 ; Niu et al., 2011 ; Lawrence et al., 2019 ; Dutra et al., 2010 ; Menard et al., 2021 ) employ this simplified formulation based on small-scale representations; thus, it is integral to most current US and European global snow predictions. However, these parameterizations were primarily developed for the hydrological community instead of earth system models, and their calibration to prioritize accurate \mathrm{SWE} and subsequent local runoff induces a trade-off in density/depth errors. Such errors impact the snowpack energy balance, which is as crucial as mass balance in global climate simulations ( Diro and Sushama, 2018 ; Xu and Dirmeyer, 2011 ) , and snow depth errors remain problematic in climate predictions ( Menard et al., 2021 ) . The inadequacy for energy tracking in these formulations has been recognized for decades and spurred development of more detailed and realistic alternatives such as CROCUS or SNOWPACK, which evolve the metamorphism of snow microstructure ( Vionnet et al., 2012 ; Lehning et al., 2002 ; Brun et al., 1989 ) . While these complex models address the limitations of Anderson’s and Kojima’s original formulations, they require site-specific calibration and struggle with computational scalability for global applications. This necessitates more efficient representations for global models that can accurately predict snow depth/density. With the push toward even finer localized (1–10 km) land modeling, simple yet accurate models are increasingly essential for computational efficiency ( Clark et al., 2015 ; Schär et al., 2020 ; Ban et al., 2021 ) . Advances in computing and sensing technology have led to initiatives in data assimilation and machine learning (ML) aimed at disrupting longstanding parameterizations. Today, extensive snow depth observations surpass the resolution and precision of \mathrm{SWE} and density data/estimates ( Fontrodona-Bach et al., 2023 ; Bruland et al., 2015 ) . Improving snow models with these resources has become a dominant avenue of hydrology research, with a bias towards SWE modeling, like incorporating remotely sensed depth into iSnobal to infer \mathrm{SWE} ( Hedrick et al., 2018 ) . Several ML models (e.g., Bair et al. 2018 ; Meloche et al. 2022 ; Duan et al. 2024 ; Steele et al. 2024 ) predict \mathrm{SWE} or depth from meteorological and topographical inputs in specific regions. Such models show satisfactory snowpack estimation but frequently yield errors over 15% when tested, especially beyond their training or calibration locations ( Meloche et al., 2022 ; Ebner et al., 2021 ; Viallon-Galinier et al., 2020 ) . The ability of these empirical models to generalize to new locations or future climates and act as a universal model is limited, and their statistical or black-box nature does not inherently respect physical constraints, impeding their integration into land models ( De Michele et al., 2013a ; Gao et al., 2021 ) . Combining depth observations and ML techniques to improve depth/density parameterizations has the potential to lead to improved simulation of key