New Neural Network Framework Models Snow-Water Hydrology with Physical Constraints

Traditional hydrologic models are often complex and require significant domain expertise to develop and interpret. Neural networks, while powerful for prediction, typically lack the physical interpretability and mass-conservation properties crucial for scientific applications like water resource management. The Mass-Conserving Perceptron (MCP) framework bridges this gap by transforming conceptual hydrologic models into neural networks that inherently respect physical laws, such as mass conservation. Applied to snow-water dynamics across numerous US basins, the MCP network demonstrates predictive accuracy on par with established hydrologic models, but with the added benefits of neural network flexibility and reduced parameter count. This approach allows for the discovery of compact, interpretable representations of precipitation-storage-runoff processes.