The authors trained Astra for 2.3 days on eight H200 GPUs and evaluated it in WeatherBench 2 against ERA5, a reconstruction of historical atmospheric conditions.
Astra matches leading weather ensemble accuracy with substantially less compute
Choi and Hong report quarter-degree forecasts matching GenCast’s average probabilistic accuracy using 9.8% of its hardware-normalised training compute.
Top University
Junu Choi · Youngjoon Hong
Seoul National University
Research Digest··3 min read
Choi and Hong present Astra, a global weather model that generates ensembles, sets of plausible forecasts representing uncertainty, at 0.25° resolution.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.