A new forecasting model doesn't just predict groundwater levels. It also shows its work, generating the physical variables a hydrologist would want to see.
Researchers describe a Physics Informed Recurrent Neural Network, or PIRNN, built to predict unobservable intermediate variables alongside the target forecast, both on historical data and into the forecast itself. The model borrows its equations from Gardenia, an existing physical model that simulates groundwater levels using transfer equations between reservoirs, tuned through data assimilation. The team tested PIRNN against several established neural network models and against Gardenia itself across twelve real-world datasets, and it won on five of them. An ablation study found that stripping out the physical grounding hurt performance, underlining that the physics content is doing real work, not just adding complexity.
The gap this fills is real. Physics Informed Neural Networks have been around for a while, but most treat the underlying physical process as a black box, producing a forecast without exposing the intermediate variables domain experts actually reason about. PIRNN's pitch is that a model useful to a hydrologist needs to speak the language of reservoirs and transfer equations, not just spit out a number.
A domain expert reviewed the coherence of those predicted physical variables, though the paper does not report what that review concluded. Winning five of twelve benchmarks is a real result, not a rout, and the interpretability gains matter regardless of who eventually claims the forecasting crown.