Beyond Local Models: Generalizing Deep Learning for Groundwater Simulation Across Hydrogeological Diversity and Mixed Influences

Overview & Synopsis

Groundwater level dynamics vary widely across aquifer types, from highly reactive to inertial systems, and are further complicated by human influences such as pumping and land-use changes. We investigate the ability of a single global LSTM model to capture this full diversity of responses, including both natural and anthropogenically influenced systems, using precipitation and temperature as inputs, combined with station embeddings to differentiate site-specific behaviors. The model is evaluated across 1150+ monitoring stations in metropolitan France with 20+ years of records. Results highlight the potential of a unified deep learning approach to generalize across contrasted hydrogeological contexts and mixed human-natural influences, while identifying directions for further improvement on complex site-specific dynamics. This work is a contribution to the France 2030 OneWater "DEESAC" project.