Large pretrained models for the atmosphere or Earth system that are fine-tuned for many tasks (forecasting, downscaling, other variables).
Search and filter this topic Plain list on GitHub
Aurora is a large Earth-system foundation model trained on over a million hours of data that beats operational systems for air quality, waves, cyclone tracks and high-resolution weather.
The Alan Turing InstituteMicrosoft Research (AI for Science)Silurian AIUniversity of Cambridge (Aardvark)
ClimaX is a transformer foundation model pretrained on CMIP6 that fine-tunes to many weather and climate tasks, outperforming data-driven baselines.
Microsoft Research (AI for Science)University of California, Los Angeles (UCLA)code
ORBIT-2 is a 10-billion-parameter ViT downscaling foundation model reaching 0.9 km global resolution on 65,536 GPUs with R² of 0.98–0.99 on 7 km benchmarks.
WeatherGFM is a generalist weather foundation model using in-context visual prompting to handle up to ten weather tasks, including forecasting, super-resolution and post-processing.
Prithvi WxC is a 2.3-billion-parameter MERRA-2 foundation model that transfers to forecasting, downscaling, gravity-wave parameterization and extreme-event tasks.
Aurora is a large Earth-system foundation model trained on over a million hours of data that beats operational systems for air quality, waves, cyclone tracks and high-resolution weather.
The Alan Turing InstituteMicrosoft Research (AI for Science)Silurian AIUniversity of Cambridge (Aardvark)
ORBIT scales a vision transformer to 113 billion parameters for Earth system prediction, sustaining 1.6 exaFLOPS on 49,152 GPUs of the Frontier supercomputer.
Review and perspective arguing that transformer-based weather and climate foundation models are mature enough to build, outlining criteria for downscaling, wildfire and hurricane tasks.
Stochastic, task-independent atmosphere model learned by self-supervised representation learning; skillful for nowcasting, interpolation, model correction and downscaling without task-specific training.
Federated prompt tuning of a frozen weather foundation model for on-device forecasting, using dynamic graphs to handle heterogeneity and cut communication cost.
W-MAE pretrains a masked autoencoder on ERA5 and fine-tunes it for forecasting, lifting FourCastNet's anomaly correlation by about 20 percent.
ClimaX is a transformer foundation model pretrained on CMIP6 that fine-tunes to many weather and climate tasks, outperforming data-driven baselines.
Microsoft Research (AI for Science)University of California, Los Angeles (UCLA)code
Federated prompt learning trains a spatio-temporal Transformer foundation model across regions with heterogeneous meteorological sensors while protecting data exposure.
Proposes a climate-change-oriented benchmark to develop foundation models for Earth observation imagery, with a call for community collaboration on evaluation.
cBottle is a generative diffusion emulator of global 5 km climate simulations and reanalysis on HEALPix, avoiding autoregressive rollout and reaching 12.5M-pixel resolution.
Introduces an operator-algebra framework for sequential data assimilation with kernel-based data-driven schemes, tested on Lorenz 96 and ENSO in a climate model.
Convolutional autoencoder reconstructs missing observation fields before EnSRF assimilation, cutting analysis error in a barotropic vorticity testbed with sparse data.
Perspective framing data assimilation and machine learning as equivalent inverse methods in a Bayesian network, e.g. linking 4D-Var to recurrent neural networks, to learn earth system models from observations.
