Skillful high-resolution weather forecasting independent of physical models

Microsoft’s ObsCast paper introduces a regional weather forecasting system that is explicitly independent of numerical weather prediction from training through inference, a notable shift from prior machine-learning weather models that still depend on NWP reanalyses. ObsCast is a two-stage architecture: an ObsCast-Analysis model reconstructs 0.05° gridded near-surface fields from sparse surface stations plus satellite and radar inputs, and an ObsCast-Forecast model then learns autoregressive evolution from those learned analyses using only raw observations, with no NWP-derived labels. Over CONUS, the analysis component reduces station RMSE versus NOAA’s RTMA by 11.2% to 34.5% on training stations and 2.9% to 31.5% on held-out stations for 2-m temperature, dew-point temperature, 10-m wind speed, and specific humidity, while precipitation threat scores improve by 11.1% to 25.0% over MRMS. In forecasting, ObsCast beats HRRR, ECMWF IFS-HRES, and NOAA GFS on most near-surface variables across 1–18 hour lead times, with especially strong gains for moisture and wind; it also produces significantly better precipitation skill during the first 4 to 8 hours, extending beyond classic nowcasting windows. The same pipeline transfers to Europe with Meteosat and OPERA data, again outperforming HRES and GFS and generalizing to held-out stations. The model combines Vision Transformers, Swin Transformers, and adaptive Fourier neural operators, showing that high-resolution, operationally relevant weather prediction can be learned directly from heterogeneous observations rather than inherited from physical-model reanalyses.