Speaker: Farshid Rahmani (Univ Louisville)
Topic: Integrating Process Knowledge and Machine Learning: Applications to Groundwater Recharge

Differentiable modeling offers a new paradigm in hydrology by integrating process-based representations with neural networks. In this talk, I will highlight the fundamental differences between conventional machine learning (e.g., LSTMs), process-based models, and differentiable frameworks, and discuss how these approaches can be applied to hydrologic prediction. I will then present an application of differentiable modeling to groundwater recharge, a critical but highly uncertain component of the hydrologic cycle. Groundwater recharge underpins ecosystem health and water sustainability, yet current projections are limited by coarse resolution, large model spread, and underuse of observational signals. This presentation introduces a differentiable (delta) modeling framework that integrates hydrologic and stream-temperature modules with a neural network. By leveraging large-scale datasets of streamflow and stream temperature, where thermal dynamics reveal groundwater contributions, the delta model infers recharge more accurately and at high spatial resolution (compared to an alternative available benchmarks), without direct reliance on recharge observations for training.