Multi-Modal Foundation · allenai
ACE2-ERA5
A global atmospheric emulator for subseasonal to decadal climate variability and multi-year projections.
What ACE2-ERA5 is
ACE2-ERA5 is the second iteration of the Ai2 Climate Emulator (ACE) developed by the Allen Institute for AI (allenai). It is designed to simulate global atmospheric variability across time scales spanning days to centuries, providing rapid, stable simulations for multi-year climate research.
The model emulates the Earth's atmosphere by learning from historical reanalysis data rather than integrating numerical fluid dynamical equations from first principles. It functions strictly as a research tool to investigate atmospheric dynamics and responses to external forcings, and is not designed for operational weather or climate forecasting.
How it works
The system is trained on 1-degree resolution ERA5 reanalysis data spanning from 1940 to 2022. It takes specified initial atmospheric conditions alongside forcing inputs—such as sea surface temperatures and greenhouse gas levels—to advance atmospheric states forward in time.
During execution via the fme software library, the model strictly enforces physical balance constraints, achieving exact conservation of global dry air mass and moisture.
Key facts
- Domain
- Climate emulation and atmospheric simulation
- Training data
- ERA5 reanalysis (1940–2022 at 1-degree resolution)
- Execution framework
- fme
- Digital Object Identifier
- 10.57967/hf/5377
What you can use it for
- Decadal Climate Response Simulation
- Simulating atmospheric warming trends driven by combined historical increases in sea surface temperatures and atmospheric carbon dioxide.
- Large-Scale Teleconnection Analysis
- Studying atmospheric circulation responses to El Niño sea surface temperature variability and the Madden–Julian Oscillation.
- Stratospheric Dynamics Research
- Evaluating the strength and variability of the stratospheric polar vortex across subseasonal to multi-year runs.
- Tropical Storm Distribution Studies
- Analysing geographic distributions and spatial frequencies of tropical cyclones under varying forced conditions.
Run it here
We list this model but cannot run it yet: no pipeline_tag — needs a human.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Demonstrates an accurate atmospheric warming response to the combined increase of sea surface temperatures and carbon dioxide over an 80-year span.
- Achieves exact physical conservation of global dry air mass and moisture.
- Accurately reproduces atmospheric responses to El Niño variations and Madden–Julian Oscillation variability.
- Captures realistic stratospheric polar vortex behaviour and geographical distributions of tropical cyclones.
Watch out for
- It must not be deployed for operational weather or climate predictions.
- Its medium-range weather forecasting capability across three to ten days is not state of the art.
- The model exhibits unrealistic responses when isolating individual sensitivities to sea surface temperature or carbon dioxide changes separately.
- It struggles to generalise reliably under large perturbations outside the training regime, such as an immediate doubling of carbon dioxide.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · apache-2.0
The apache-2.0 licence carries terms worth reading before you depend on it.
Questions
Can I use ACE2-ERA5 commercially?
Check the licence before commercial use. The apache-2.0 licence carries terms worth reading before you depend on it.
Can ACE2-ERA5 be used for real-time weather forecasting?
No. The model is intended strictly as a research tool for atmospheric and climate dynamics and lacks competitive medium-range forecasting skill.
What software environment is required to run inference?
Inference requires the fme Python library to load configuration files, atmospheric initial conditions, and specified boundary forcings.
What observational period was used to train the emulator?
It was trained using historical ERA5 reanalysis data covering the period from 1940 to 2022 on a 1-degree grid.
How does the model handle mass conservation during long runs?
The emulator maintains exact numerical conservation of both moisture and global dry air mass throughout simulations.
How to cite it
ACE2: Accurately learning subseasonal to decadal atmospheric variability and forced responses, Nature, https://www.nature.com/articles/s41612-025-01090-0
Written from ACE2-ERA5's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.