license: apache-2.0
ACE-climSST
Ai2 Climate Emulator (ACE) is a family of models designed to simulate atmospheric variability from the time scale of days to centuries.
Disclaimer: ACE models are research tools and should not be used for operational climate predictions.
ACE-climSST is the original ACE model, as described in ACE: A fast, skillful learned global atmospheric model for climate prediction. It is trained on output from the FV3GFS atmospheric model forced with annually-repeating climatological sea surface temperature.
Code for doing inference with ACE models can be found here: https://github.com/ai2cm/ace with corresponding documentation here: https://ai2-climate-emulator.readthedocs.io/en/stable/.
Briefly, the strengths of ACE-climSST are:
- long-term stability
- highly accurate time-mean climate compared to its target dataset
- very fast inference compared to typical physics-based atmospheric models
Some known weaknesses are:
- responses to El Niño-like sea surface temperature and long-term warming trends are not accurately captured
- small but non-zero drifts in total dry air mass of the atmosphere
Refer to published manuscript for more details.
Note the checkpoint provided here is the same as the one in this Zenodo repository, just with the optimizer state removed to decrease the checkpoint size.