Physics-informed deep learning reveals climate-driven snowpack decline and threatens ecological water availability in a Californian snow-fed catchment

Mountain snowpack functions as a critical natural reservoir, and its gradual melt sustains streamflow for downstream ecosystems, particularly in arid and semi-arid regions. Climate warming is disrupting these historical patterns by shifting precipitation from snow to rain and accelerating melt, creating an urgent need for robust ecological forecasting tools that can anticipate threats to water availability. Physics-informed machine learning (PIML) offers a powerful approach to this challenge by integrating domain knowledge into flexible data-driven models. This study evaluates and compares the performance of three hydrological modeling approaches: (i) a calibrated process-based Soil and Water Assessment Tool (SWAT), (ii) a data-driven Long Short-Term Memory (LSTM) neural network, and (iii) a physics-informed LSTM (PIML) model that integrates a melt index and precipitation-phase constraints within the Upper West Walker River Watershed in California, USA. The models are assessed based on their ability to simulate historical daily snow water equivalent (SWE) and streamflow using performance metrics such as Nash–Sutcliffe Efficiency (NSE), Kling–Gupta Efficiency (KGE), root mean square error (RMSE), and peak-timing bias. Results show that the PIML model provides the most robust and balanced performance, particularly in minimizing bias, highlighting the benefit of embedding physically meaningful constraints into data-driven models in snow-dominated basins. Future simulations using bias-corrected CMIP6 high-emission scenarios project that peak SWE may decline by up to 60 % (52–73 %), and peak discharge by approximately 33 % (28–47 %). Moreover, warming-induced changes in precipitation phase are expected to shift snowmelt and runoff 10–19 days earlier, compressing the hydrologic season and inducing ecological stress. These findings emphasize the dual risk of increased springtime flows and diminished summer water availability and heightened ecological drought risk, highlighting the urgent need for adaptive water management strategies in the face of climate-driven hydrological shifts. • Physics-informed LSTM outperforms traditional hydrological and ML models. • Future warming may reduce peak snowpack by over 60 %. • Earlier, compressed runoff threatens summer ecosystem water security.