Researchers develop STAINet, an attention-based deep learning model that predicts groundwater levels more flexibly and reliably than traditional physics-based approaches.

STAINet uses a pure deep learning approach with attention mechanisms to forecast weekly groundwater levels at multiple variable locations

The model combines sparse groundwater measurements with dense weather data to make predictions at arbitrary locations

Physics-guided strategies are incorporated to improve the model's trustworthiness and ability to generalize beyond training data

The approach addresses limitations of traditional theory-based models, which require significant computational resources, simplifying assumptions, and extensive calibration