Earth2Studio: An Open Inference Toolkit for AI Weather Forecasting

Earth2Studio is an open-source Python toolkit that turns state-of-the-art AI weather and climate models into composable, reproducible workflows that researchers and operators can run and adapt on their own infrastructure.  It targets a key bottleneck in AI-for-weather: the difficulty of moving from standalone model checkpoints to fully integrated forecasting systems that span data, models, uncertainty, and verification.

Earth2Studio provides a unified API for prognostic and diagnostic AI models, heterogeneous data sources, perturbation methods, metrics, and I/O backends, enabling users to assemble end-to-end inference pipelines with only a few lines of code.  The model zoo includes leading global and regional AI forecast models such as Altas, StormScope, GraphCast, Pangu, Aurora, FourCastNet 3, CorrDiff and more. Standardized data interfaces expose operational initial conditions and reanalyses
(e.g. GFS, HRRR, ERA5, IFS) through a shared Xarray-based vocabulary and coordinate system.

Building on the broader Earth-2 initiative, Earth2Studio is designed to cover the entire weather forecasting value chain including AI data assimilation for initial conditions, global medium-range prediction, generative downscaling, and kilometer-scale severe weather nowcsating.  Ensemble-ready perturbation schemes and built-in statistics (RMSE, ACC, CRPS, rank histograms, spread–skill diagnostics) allow consistent quantification of forecast skill and uncertainty across models, lead times, and regions, supporting methodologically robust intercomparison studies.

Released as OSS, Earth2Studio emphasizes openness and sovereignty: all core components are optimised to run on NVIDIA local or cloud platforms, enabling national meteorological services, research institutions, and industry users to integrate proprietary data and maintain ownership over operational chains.

Presented here, are the design principles of Earth2Studio, illustrative exemplar workflows, and a discussion of how this shared software infrastructure can help the EGU community accelerate AI weather research and bridge the gap between experimental models and operationally relevant forecasting systems.