Distilling Hydrological and Land Surface Model Parameters from Physio-Geographical Properties Using Text-Generating AI

Abstract The estimation of parameters for distributed hydrological and land-surface models is inherently difficult, particularly in data-scarce regions. We present a novel approach that employs variational autoencoders (VAEs) that have been trained as text-generating models to automatically generate optimal, interpretable parameter transfer functions. This AI-driven methodology transforms equation discovery into continuous latent-space optimization, thereby improving the interpretability of the model and the efficiency of the process. When evaluating this method in a prediction-in-ungauged-basins setting with the mesoscale Hydrological Model (mHM) across 162 German basins, it showed significant improvements in runoff predictions compared to both traditional regionalization techniques and state-of-the-art regional Long Short-Term Memory (LSTM) networks. The generated transfer functions are robust, scalable, and physically interpretable, which is a substantial improvement in the field of large-scale process-based environmental modeling.