A physics-constrained neural network is beating standard machine-learning weather models at short-term temperature forecasts, and the gap widens the further out you look.
Researchers built a neural network for forecasting potential temperature (a measure tied to how air behaves at different altitudes) that bakes in a thermodynamic equation governing how heat moves through the atmosphere, with that physical closure calibrated on the prior 12 hours of data and then locked before testing. Using hourly ERA5 reanalysis data at three pressure levels, they compared it against persistence and trend baselines plus two ordinary neural networks, one of which received the same future weather data as the physics-informed model, a control meant to separate the value of the physics constraint from simple data access. In an Oklahoma test, the physics-informed model's RMSE advantage over the strongest baseline grew from 8.1% at a one-hour forecast to 23.8% at three hours, a gap that held up even when observation coverage was thinned to 5% of candidate sites. Applied to an Alabama heat event, three-hour improvements ranged from 19.7% to 24.4%, but over Montana's mountainous terrain, where the model's fixed pressure levels cut across complex topography, three-hour accuracy dropped by about 17.5%.
Climate-aware digital twins, systems meant to simulate real regions or infrastructure in near real time, live or die by the sensor networks feeding them, and most of the world's weather-observation grid is thin. This study's main claim is that physics constraints can substitute for missing sensors rather than just adding computational overhead, which matters for utilities, agriculture, or disaster planners who can't blanket a region in weather stations. The Montana result is the more interesting data point, though: it shows the same physics that helps on flat ground can actively hurt once terrain breaks the model's assumptions.
Worth remembering this is an unreviewed preprint tested against historical reanalysis data, not live forecasts, so the real test is whether these gains hold once fed real-time, noisy sensor data instead of curated hindcasts.