WeatherDreamer: Learning a World Model from Multimodal Earth Observations
WeatherDreamer beat NOAA’s ensemble for temperature, dew point and wind at held-out stations through 12 hours.
The researchers trained it directly on 18 global observing systems, without reanalysis data. It advances an internal atmospheric state hourly and corrects it with incoming observations, learning assimilation and forecasting together. Performance was uneven: departures from observations were close to NOAA’s GSI system for surface measurements but up to five times larger for temperature-sounding channels. An hourly global analysis takes about 1.5 minutes on one GPU.
ArXiv · AI/CL/LG's note
The researchers trained it directly on 18 global observing systems, without reanalysis data. It advances an internal atmospheric state hourly and corrects it with incoming observations, learning assimilation and forecasting together. Performance was uneven: departures from observations were close to NOAA’s GSI system for surface measurements but up to five times larger for temperature-sounding channels. An hourly global analysis takes about 1.5 minutes on one GPU.
ArXiv · AI/CL/LG's note
score 5