PINN for short-horizon potential-temperature forecasts with sparse observations (arXiv:2609.27290v1)
This arXiv preprint evaluates a physics-informed neural network (PINN) that enforces a pressure-coordinate thermodynamic advection–source equation and uses a diabatic-source closure fit from the preceding 12 hours (frozen during future-time training) to forecast potential temperature at 1–3 hour leads using hourly ERA5 data at three pressure levels. In tests, the PINN reduced RMSE versus the strongest baseline from 8.1% at 1 hour to 23.8% at 3 hours in an Oklahoma case, retained a 14.6–16.9% 3-hour advantage under observation-density down to 5%, transferred with 19.7–24.4% 3-hour improvement in an Alabama heat event, and degraded roughly 17.5% in a Montana complex-terrain stress test indicating a vertical-coordinate applicability limit.