DOE highlights groundwater modeling gains—and limits—from machine learning
A DOE research highlight reports better simulated groundwater depths, but mixed results for related water processes.
What the record says
A machine-learning approach improved one part of a major Earth system model without consistently improving the rest of its water simulations, according to a November 3 Department of Energy research highlight. Researchers used a surrogate computer model and observation-constrained groundwater data to adjust parameters in the E3SM Land Model. Simulated groundwater depths improved, especially in the one-to-five-meter range where groundwater can affect surface processes. The changes did not consistently improve soil moisture, runoff or evapotranspiration. Some regional runoff and soil-moisture results became less accurate. DOE says the findings point toward calibrating several connected water processes together and testing interactions between land and atmosphere. The work addresses how scientists model the water cycle, rather than providing a local water-supply forecast.
Source details and supporting locations.
Practical next step
Read DOE’s research highlight for the study’s methods, limitations and linked journal publication.
Limits and what is not established
Improvement in one simulated variable does not establish better overall forecasts. This report does not demonstrate a new operational flood or groundwater forecast.
Sources checked November 4, 2026. Later notices may change instructions or arrangements.
Original sources
U.S. Department of Energy: original information