Comparing Snow Water Equivalent Estimations From Long Short‐Term Memory Networks and Physics‐Based Models in the Western United States

Abstract In snowmelt‐dominated regions like the western U.S., mountain snowpacks supply 50%–70% of the total runoff. Accurate estimation of snow water equivalent (SWE) is critical for informed management of water resources, including reservoir operations, flood risk assessments, and drought mitigation. However, the spatial variability and complexity of snow accumulation and dispersion processes are challenging to accurately model. Recently, machine learning (ML) techniques, particularly Long Short‐Term Memory (LSTM) neural networks, have demonstrated success in modeling complex hydrological processes; however, their application to SWE estimation remains relatively underexplored. In this paper, we develop and train an LSTM model to estimate SWE at point locations in the western U.S. and assess its sensitivity, transferability, and robustness to biased forcing in comparison with two representative physics‐based models, ParFlow‐CLM and the University of Arizona SWE product. We identify strengths of the LSTM model, including superior performance on metrics of magnitude and temporal accuracy and higher adaptability across snowpack regimes and erroneous forcing conditions, as well as weaknesses, including a lack of physical constraints on estimations. We also show the LSTM model relies on physically relevant characteristics—both static and meteorological—to predict snowpack. Overall, this paper represents a novel investigation into the behavior and characteristics of an LSTM model of SWE compared to a range of physics‐based models, extending beyond traditional performance‐focused assessments. Developing a deeper understanding of LSTM models in comparison to physics‐based models can pinpoint areas for improvement, representing an important step toward the development of operational ML models of SWE.