Skip to content
AI.info

The Pulse

Ai2 releases three ACE2.5-ERA5 climate research checkpoints

Ai2 has published three configurations of its ACE2.5-ERA5 atmospheric emulator on Hugging Face. The research models are not intended for operational climate predictions.

Ai2 releases three ACE2.5-ERA5 climate research checkpoints

Three configurations of Ai2’s ACE2.5-ERA5 atmospheric emulator are now available on Hugging Face, with daily and six-hourly versions trained on ERA5 data. The Nov. 19 repository release is for research, and Ai2 cautions that ACE models should not be used for operational climate predictions.

Source huggingface.co

The model is designed to simulate atmospheric variability over periods ranging from days to centuries. It succeeds ACE2-ERA5 and uses a stochastic approach to produce ensemble predictions.

Three configurations

The primary daily model operates at 1-degree resolution and has 46 million parameters. A second daily version has 4-degree resolution and 14 million parameters; Ai2 says it has comparable climate skill at its target resolution, five times the throughput and one-tenth the training cost.

A third configuration uses a six-hourly time step at 1-degree resolution, with 46 million parameters. Ai2 describes its short-range weather skill as competitive and its ensemble spread as calibrated.

Each configuration is released as seed 0. The daily models also have seeds 1 through 3 in a separate directory, and Ai2 says climate skill is consistent across those seeds. The repository also contains ablation-model checkpoints for reproducing figures from the paper, along with documentation of checkpoint provenance.

Changes from ACE2

Ai2 says spectral bottlenecking reduces the parameter count from 456 million in ACE2-ERA5 to 46 million. It reports that the daily time step and bottleneck together make inference 3.2 times faster than in the earlier model.

For precipitation structure, the model card says ACE2.5-ERA5 retains 86% of ERA5’s spectral power at wavenumbers 100–179, compared with 36% for ACE2-ERA5. Ai2 says the model’s ensemble spread matches error to within 4% from day 3.

Among its reported strengths are the atmospheric warming response to combined sea-surface-temperature and carbon-dioxide forcing over 1940–2020, and the response to El Niño sea-surface-temperature variability. The card also describes the six-hourly model’s tropical cyclone geographic distribution as realistic and reports accurate Madden-Julian Oscillation variability.

Limits and data

The daily model’s medium-range weather skill is below that of ACE2-ERA5, while the six-hourly version matches the earlier model, according to Ai2. The model’s individual sensitivities to sea-surface temperature and carbon dioxide are not entirely realistic, and its interannual variability of global mean surface temperature is not calibrated.

Ai2 does not expect the model to generalize accurately to large perturbations, such as a doubling of carbon dioxide. It lists exact conservation of global dry-air mass and moisture among the model’s strengths.

The training data is available in a Hugging Face storage bucket, with 1-degree and 4-degree regriddings of ERA5 from 1940–2025 at daily and six-hourly time steps. The bucket also holds normalization statistics and climatological forcing data; its datasets include temperature, water and wind fields at eight pressure levels, as well as surface and diagnostic fields.

To run inference, the model card instructs users to download a checkpoint, install the fme dependencies and prepare a configuration specifying the checkpoint, initial condition and forcing data.

Sources

  1. Original source Ai2 · page 1 huggingface.co

Other pages read

  1. Ai2 · page 2 huggingface.co
  2. Ai2 · Ace2.5 ERA5 reference data huggingface.co
  3. The Deploy Log deployedbyai.com

Related stories

Explore

More articles