Turn Your Latest Observations Into Timely Weather Decisions With NVIDIA Earth-2

NVIDIA released a tutorial describing AI data assimilation tools in its Earth-2 platform, aimed at incorporating proprietary or third-party observations into weather forecasting pipelines.

The tutorial covers two techniques: constraining diffusion models with point observations, typically for regional models, and assimilating disparate datasets into a consistent state, typically for global models. Prerequisites listed are an Earth2Studio installation, an NVIDIA RTX PRO or data center GPU, basic Python knowledge, and approximately 30 minutes.

Score-Based Data Assimilation (SDA) constrains diffusion models such as StormCast and CorrDiff with point observations. According to the reported figures, it reduces wind-speed RMSE by 54% in a CorrDiff-COSMO downscaling example and by an average of 7.2% across six StormCast-CONUS forecast steps. SDA does not require retraining the model, and its output is probabilistic, with less uncertainty near observation locations. Users define an observation operator mapping model output to expected measurements at each location.

HealDA estimates the global atmospheric state in seconds by mapping remote-sensing and in situ observations within a time window onto a 1° HEALPix grid (HPX64), using an observation encoder and a vision transformer backbone. Earth2Studio provides a pretrained global data assimilation model for research purposes that integrates microwave sounder, radio occultation, surface station, aircraft, and buoy data, among other sources. A HealDA training pipeline is available in the open-source PhysicsNeMo library.

Earth2Studio data sources provide unified access to gridded satellite and radar data, conventional observation archives, and operational observation streams for developing and validating weather models. Named sources include GOES, Himawari, Meteosat, MRMS, OPERA, GHCN/ISD, NNJA, UFS, GDAS, ASOS, MetOp, and JPSS. NVIDIA pointed to an end-to-end example library, a custom data source example, and the Earth2Studio user guide as next steps.