allenai/aimip-phase1-submissions

AIMIP Phase 1 model submissions

Summary

This dataset contains output from six model submission groups' ~46-year atmospheric simulations produced as contributions to the AI Model Intercomparison Project (AIMIP) Phase 1. AIMIP systematically evaluates AI weather/climate models trained on ERA5 reanalysis by running standardized AMIP-style simulations and comparing their climate statistics against the ERA5 reanalysis and a conventional climate model.

There are currently eight model submissions from the six model submission groups. For each submission, there are typically 15 simulations spanning 5 ensemble members (different initial atmospheric conditions) and 3 sea surface temperature (SST) forcing scenarios — historical baseline, +2 K, and +4 K uniform global SST perturbations — covering October 1978 through December 2024. Outputs are CMIP6-compliant NetCDF files at monthly and, for most submissions, daily mean temporal resolution, at approximately 1° global resolution, on each model's own grid (NeuralGCM is the exception, at 2.8°).

For the AIMIP Phase 1 specification that defines the model training and the required simulations that produced the output here, see https://github.com/ai2cm/AIMIP. For the initial evaluations conducted using a snapshot of this dataset, see https://arxiv.org/abs/2605.06944.

Contents

Group	Submission	Directory	Experiments	Frequency	Horizontal grid	Vertical grid
Ai2	ACE2.1-ERA5	Ai2/ACE2-1-ERA5	aimip, aimip-p2k, aimip-p4k	daily, monthly	1° × 1°	13 pressure levels (gr), regridded from native 8 model layers via ML; native layers also submitted (gn)
ArchesWeather	ArchesWeather	ArchesWeather/ArchesWeather-V2	aimip, aimip-p2k, aimip-p4k	daily, monthly	1° × 1°	7 pressure levels
ArchesWeather	ArchesWeatherGen	ArchesWeather/ArchesWeatherGen-V2	aimip, aimip-p2k, aimip-p4k	daily, monthly	1° × 1°	7 pressure levels
NVIDIA	cBottle-1.3	NVIDIA/CMIP6/AIMIP/NVIDIA/cBottle-1-3	aimip, aimip-p2k, aimip-p4k	daily, monthly	HEALPix order 6 (~0.9°)	8 pressure levels
University of Washington and NVIDIA	DLESyM	DLESyM/DLESyM	aimip, aimip-p2k, aimip-p4k	daily, monthly	HEALPix order 6 (~0.9°)	some surface, 850 hPa and 500 hPa variables
University of Maryland (PARETO group)	MD-1.5 v0.9	UMD-PARETO/MD-1p5	aimip, aimip-2k, aimip-4k	monthly	1.5° × 1.5° native; 1° × 1° submitted (gr)	7 pressure levels
Google Research	NeuralGCM	Google/NeuralGCM	aimip, aimip-p2k, aimip-p4k	daily, monthly	2.8° × 2.8°	32 sigma levels native; 7 pressure levels submitted
Google Research	NeuralGCM-HRD	Google/NeuralGCM-HRD	aimip, aimip-p2k, aimip-p4k	daily, monthly	1° × 1°	32 sigma levels native; 7 pressure levels submitted

References

Model and code citations follow Table 1 of the AIMIP Phase 1 evaluation paper.

Submission	Model reference	Code
ACE2.1-ERA5	Watt-Meyer et al. (2025), doi:10.1038/s41612-025-01090-0	Henn et al. (2026), doi:10.5281/zenodo.19831256
ArchesWeather, ArchesWeatherGen	Couairon et al. (2026), doi:10.1126/sciadv.adx2372; Singh et al. (2026), arXiv:2605.29976	Couairon et al. (2026), doi:10.5281/zenodo.20784771
cBottle-1.3	Brenowitz et al. (2025), arXiv:2505.06474	Manshausen et al. (2026), doi:10.5281/zenodo.20832634
DLESyM	Cresswell-Clay et al. (2025), doi:10.1029/2025AV001706	Cresswell-Clay (2026), doi:10.5281/zenodo.21270137
MD-1.5 v0.9	Hall and Molina (2026), arXiv:2604.13481	Hall and Molina (2026), doi:10.5281/zenodo.21430292
NeuralGCM, NeuralGCM-HRD	Kochkov et al. (2024), doi:10.1038/s41586-024-07744-y; Yuval et al. (2026), doi:10.1126/sciadv.adv6891	Kochkov et al. (2026), doi:10.5281/zenodo.21303721

Notes on the table:

• DLESyM and cBottle-1.3 are on HEALPix grids rather than latitude/longitude. The resolution is comparable to the 1° submissions, but the data must be remapped before gridpoint comparison with them. In the files, DLESyM carries face, height and width dimensions and cBottle-1.3 a single i dimension of 49152 cells.
• MD-1.5 v0.9 uses aimip-2k and aimip-4k as experiment IDs, without the p the other submissions use. Code that globs across submissions has to allow for both spellings.
• NeuralGCM-HRD is the same model as NeuralGCM with outputs downscaled to a 1° grid. The submitting group recommends NeuralGCM for comparisons involving pressure-level wind and precipitation.
• DLESyM was trained on 1983–2016 and so saw some of the AIMIP holdout period during training. Only a small subset of variables is available; pressure-level output, as submitted, is ta at 850 hPa and zg at 250, 500 and 1000 hPa.
• Per-submission documentation, where the group provided it, is published alongside the data (for example Ai2/DATASET_CARD.md, Ai2/LICENSE-DATASET, NVIDIA/README.md, NVIDIA/data-card.md).

Reference data

ERA5/mon_1deg_dkrz_v3/ holds monthly-mean ERA5 for 1978–2024 on a regular 1° grid, one file per variable (hus, pr, ps, psl, ta, tas, tdps, ts, ua, uas, va, vas, zg). They were prepared for AIMIP evaluation with ESMValTool by Bettina Gier (University of Bremen), conservatively regridded in one step from 0.25° ERA5, and are version 3 of the files on the DKRZ store, published unchanged. README.txt and the recipe recipe_era5_aimip_mon.yml in the same folder record the sources and settings. ts is skin temperature everywhere.

ERA5/day_1deg_arco/ holds daily-mean ERA5 on the same grid for 1 October 1978 – 31 December 1979 and for 2024, the two periods of the daily submissions. Each day is the mean of its 24 hourly values from the public ARCO-ERA5 mirror, conservatively regridded, built by evaluations/notebooks/data_prep/ERA5-monthly-1deg-ARCO.ipynb in the AIMIP repository.

Experiment design

AIMIP Phase 1 compares AI weather/climate models through standardized multidecadal AMIP-style simulations. Its goals are to (1) systematically compare time-mean climate, trends, and variability across AI models trained on ERA5; (2) produce CMIP7-compatible outputs so the broader climate science community can evaluate them; and (3) strengthen the credibility of AI climate models through structured intercomparison. See the AIMIP specification for the full protocol.

Parameter	Value
Simulation period	Oct 1, 1978 – Dec 31, 2024
Spinup period	Oct–Dec 1978 (3 months, excluded from analysis)
Analysis period	Jan 1, 1979 – Dec 31, 2024
Ensemble members	5 per experiment (typical)
SST scenarios	3 (baseline, +2 K, +4 K)
Total simulations	15 per submission (typical)

Forcing

Models are forced with prescribed monthly sea surface temperature and sea-ice concentration derived from ERA5, spatially interpolated by each group to its own native grid. AIMIP supplies a common 1979–2024 monthly AMIP-like SST and sea-ice forcing dataset built from daily 0.25° ERA5 output.

Experiment ID	Forcing
aimip	Historical observed SST and sea-ice concentration (baseline)
aimip-p2k (aimip-2k for MD-1.5)	Historical SST with a uniform global +2 K perturbation
aimip-p4k (aimip-4k for MD-1.5)	Historical SST with a uniform global +4 K perturbation

Ensemble members

Each scenario is run five times, from different atmospheric states on or around October 1, 1978, labelled r1i1p1f1 through r5i1p1f1. cBottle-1.3 is the exception: its five members differ by training-checkpoint combination rather than initial condition, and are labelled r1i1p1f1 through r1i1p5f1 (the "physics index"); see NVIDIA/README.md.

Output frequencies

• Amon (monthly means): the full simulation period, October 1978 – December 2024. ArchesWeather submissions extend one month further, through January 2025.
• day (daily means): two sub-periods only — October 1978 – December 1979 (spinup assessment) and January – December 2024 (out-of-sample testing). DLESyM instead submitted continuous daily output from 1978 to 1984.

File organization

Files are NetCDF4, CF-compliant, and follow CMIP6 naming conventions for variables, units, and coordinates. They are laid out in a CMIP6 Data Reference Syntax (DRS) hierarchy below each submission directory:

<experiment_id>/<variant_label>/<table_id>/<variable>/<grid_label>/<version>/<filename>.nc

Field	Values
experiment_id	aimip, aimip-p2k, aimip-p4k (aimip-2k / aimip-4k for MD-1.5)
variant_label	r1i1p1f1 – r5i1p1f1 (cBottle-1.3: r1i1p1f1 – r1i1p5f1)
table_id	Amon (monthly mean) or day (daily mean)
grid_label	gn (model's native grid) or gr (regridded, e.g. to standard pressure levels)
version	e.g. v20260304; ArchesWeather submissions have no version directory

The filename repeats the same fields:

<variable>_<table_id>_<source_id>_<experiment_id>_<variant_label>_<grid_label>_<start>-<end>.nc

Examples:

Ai2/ACE2-1-ERA5/aimip/r1i1p1f1/Amon/ta/gr/v20251130/ta_Amon_ACE2-ERA5_aimip_r1i1p1f1_gr_197810-202412.nc
Google/NeuralGCM/aimip-p4k/r2i1p1f1/day/tas/gn/v20260306/tas_day_NeuralGCM_aimip-p4k_r2i1p1f1_gn_20240101-20241231.nc
ArchesWeather/ArchesWeather-V2/aimip/r1i1p1f1/Amon/tas/gn/tas_Amon_ArchesWeather_aimip_r1i1p1f1_gn_197810-202501.nc

Variable coverage differs by submission. All eight provide ta, zg, and tas; most provide hus, ua, va, uas, vas, pr, ps, and ts. DLESyM provides only ta, zg, tas, and pr. cBottle-1.3 additionally provides radiative and column-integrated fields (rlut, rsds, rsut, prw, clivi). ArchesWeather submissions provide tos and psl, and no pr; NeuralGCM provides tdas (2 m dewpoint temperature) instead of huss. The AIMIP protocol asked for tdas by mistake; the CMIP6 name is tdps.

Usage

Download the whole dataset, or a subset, with the Hugging Face CLI:

# everything (~1.4 TB)
hf download allenai/aimip-phase1-submissions --type dataset --local-dir local_data

# one submission plus the ERA5 reference data
hf download allenai/aimip-phase1-submissions --type dataset --local-dir local_data \
    --include "DLESyM/*" "ERA5/*"

The layout matches what the AIMIP evaluation notebooks expect: place it at local_data/ in a clone of that repository, or point AIMIP_DATA_ROOT at it. Individual files open with any NetCDF tool:

import xarray as xr

ds = xr.open_dataset(
    "local_data/Ai2/ACE2-1-ERA5/aimip/r1i1p1f1/Amon/ta/gr/v20251130/"
    "ta_Amon_ACE2-ERA5_aimip_r1i1p1f1_gr_197810-202412.nc"
)
ta = ds["ta"]  # (time, plev, lat, lon); plev in Pa

Known minor gaps

These do not affect the overall characteristics of the dataset, but are worth knowing before computing ensemble statistics.

• NeuralGCM-HRD, daily v20260306 (hus, ta, ua, va): some realization/period files were removed after a corrupted upload and were not replaced. Seven of the 1978–79 files are absent, so those variables have three or four members rather than five for that period.
• NeuralGCM: three files were truncated in the source archive and are left out, so these have four members rather than five: aimip-p2k monthly va (r1i1p1f1 absent), and aimip-p4k daily hus for 1978–79 and daily va for 2024 (r4i1p1f1 absent).
• ArchesWeather-V2, aimip-p4k: four realizations; r3i1p1f1 is absent.
• cBottle-1.3, aimip-p2k and aimip-p4k: a single member (r1i1p4f1) rather than five.

The AIMIP evaluation code fills missing realizations with NaN and continues, so ensemble means and medians are taken over the members that exist.

Provenance

Snapshot of the AIMIP archive hosted by DKRZ at s3://ai-mip, taken 2026-03-31, plus the DLESyM v20260406 resubmission from April 2026 and the license and documentation files. It is published here as a citable, immutable revision of the Phase 1 submissions for the evaluation paper. Each submission is included at the version the evaluations read; the ERA5 reference data were replaced in October 2026 (see below).

Changes since phase1-2026-09

The first revision, tagged phase1-2026-09, remains available unchanged. Since then:

• ERA5 monthly reference replaced. Apart from pr, ERA5/mon_1deg/ was interpolated bilinearly from ERA5's native grid, so even a perfect model regridded conservatively showed a bias near coastlines and steep terrain. It also had a spurious column at longitude 359.5° near the poles, and its pr file lacked January 1978 – September 1979. ERA5/mon_1deg_dkrz_v3/ is conservatively regridded throughout and complete.
• ERA5 daily reference rebuilt. ERA5/day_1deg/ (a Gaussian grid, each day the mean of four 6-hourly values) is replaced by ERA5/day_1deg_arco/ (the regular 1° grid, 24-hour means, built from public data, and adding 2024).
• Superseded submission versions removed: NeuralGCM and NeuralGCM-HRD v202512*, DLESyM v20250825, MD-1.5 v20251214, and cBottle-1.3 v20260120 and v20260217. The evaluations read none of them.
• Five truncated files removed that could not be opened: two were among the superseded versions, and the other three are listed under the known gaps above.

License

All submission files are copyright of their respective institutions, organisations, or companies and are licensed under CC BY 4.0; see LICENSE.txt at the root. Individual submission folders may add guidelines consistent with CC BY 4.0, for example Ai2/LICENSE-DATASET and NVIDIA/data-card.md.

The ERA5 reference data is derived from ECMWF ERA5 reanalysis and is subject to the Copernicus licence; credit ECMWF and the Copernicus Climate Change Service.

This dataset consists entirely of weather and climate model output and reanalysis-derived reference data. It contains no private, personal, or otherwise sensitive data.

Citation

Please cite this dataset by its DOI, together with the AIMIP Phase 1 evaluation paper (https://arxiv.org/abs/2605.06944) and the papers for the individual model submissions you use, listed in the table above.