AI Training Datasets
The E3SM project and Allen Institute for AI (Ai2) have developed several datasets for AI and machine learning applications. These datasets have been postprocessed for ingestion by the ACE/FourCastNet emulator.
Dataset Details
EAMv2: 73-year EAMv2 simulation (F2010, perpetual 2010 forcing, repeating annual SST cycle from 2005-2014 average). 6-hourly outputs. More details see: Duncan et al. 2024
EAMv3: 51-year EAMv3 AMIP-style simulation (1970-2020, F2010 with AMIP SSTs, constant 2010 CO2). Includes multiple ENSO cycles and global warming trend. More details see: Wu et al. 2025
E3SMv3: 105-year fully coupled E3SMv3 pre-industrial control (piControl) simulation, regridded to a 180x360 grid and vertically coarsened to 8 atmosphere layers and 19 ocean depth levels. Includes 6-hourly atmosphere output and 5-day mean ocean and sea ice output. Used to train SamudrACE, which couples the ACE2 atmosphere emulator with the Samudra full-depth ocean emulator. See the E3SM newsletter article for an overview. More details see: Wu et al. 2026
SCREAMv1: Simple Cloud-Resolving E3SM Atmosphere Model version 1 training data (coming soon)
Tip
Check the archive_contents text file to see files included in each tar archive. You can selectively download the files you need.
Note
The EAMv2 and EAMv3 datasets are distributed as netCDF files. The E3SMv3 coupled dataset is
distributed as Zarr v3 stores
(using the sharding codec), which require zarr-python >= 3.0 to read. In both cases the
vertical dimension is split across separate variables (T_0, T_1, …), with the vertical
grid defined by the ak_*/bk_* coefficients as pressure = ak * P0 + bk * PS.
Data access: the datasets are archived on NERSC HPSS. See Accessing E3SM Data for download instructions.
Run scripts: the run scripts for the source simulations are not yet archived in this repository.
Simulations
Note
Data on Disk paths are provided for convenient short-term access on NERSC. These disk copies are not guaranteed to persist – the HPSS archive is the authoritative long-term copy. Please use the HPSS path or URL for durable access. The Globus URL opens the on-disk copy in the Globus web app; see Globus sharing.
Dataset |
Status |
Data Size |
Data on Disk (NERSC) |
Globus URL |
HPSS Path |
HPSS URL |
|---|---|---|---|---|---|---|
EAMv2 |
||||||
EAMv2 AI Training Dataset |
Available |
1.2T |
/global/cfs/cdirs/m4492/gsharing/e3sm-v2-climsst-180x360-gaussian |
/home/projects/e3sm/www/AI_training_data/e3sm-v2-climsst-180x360-gaussian |
||
EAMv3 |
||||||
EAMv3 AI Training Dataset |
Available |
1.0T |
/global/cfs/cdirs/m4492/gsharing/e3sm-v3-amip-180x360-gaussian |
/home/projects/e3sm/www/AI_training_data/e3sm-v3-amip-180x360-gaussian |
||
E3SMv3 |
||||||
E3SMv3 Coupled AI Training Dataset |
Available |
2.2T |
/global/cfs/cdirs/e3sm/gsharing/SamudrACE-E3SMv3 |
/home/projects/e3sm/www/AI_training_data/samudrace-e3smv3-piControl-180x360 |
||
SCREAMv1 |
||||||
SCREAMv1 AI Training Dataset |
Coming Soon |
TBD |
TBD |
TBD |
TBD |
TBD |