Armor Pro Research · Data and code
Data and Code for the ACE2 Reproduction
Every table behind the note, the fields, and the scripts that made them, released under CC BY 4.0.
This page lists what the reproduction note is made of. The tables are small and open in a spreadsheet; the netCDF fields open in any climate toolkit; the code is the exact set of scripts that turned the public inputs into the runs and the runs into the numbers. All of it is CC BY 4.0. The model output itself (about 190 GB of six-hourly fields for 139 runs) is not hosted; the scripts regenerate it from the public inputs in about a day on a single GPU.
Tables
| File | Contents |
|---|---|
| hindcast_by_year.csv | September 1991–2020, one row per year: ACE2 (observed-ocean run) MDR zonal and vector shear, ERA5 MDR zonal shear, both as absolute values and anomalies, ASO RONI and phase |
| members_2026-07-01.csv · members_2026-08-01.csv | The 51 ACE2 members per ECMWF initialization: September MDR zonal and vector shear anomaly against the emulator's own 1991–2020 climatology |
| seas5_vs_ace2_sep_members_2026-07-01.csv | Member by member, 1 July initialization: ECMWF SEAS5's own September MDR anomaly and ACE2's, both against ERA5 1991–2020 |
| clim_sep_shear_by_year.csv · era5_sep_shear_by_year.csv | The climatology runs and ERA5, September, MDR and western-Atlantic boxes, zonal and vector shear (ERA5 through 2025) |
| clim_sep_sst_indices.csv · fcst_2026-07-01_sep_sst_indices.csv | September sea-surface-temperature indices from the forcing files (Niño 3.4, MDR, tropical mean, global; relative Niño 3.4 and relative MDR) for the 30 climatology years and the 51 SEAS5 members |
| ace2_aug1-12_members.csv | The 51 members' MDR anomaly over 1–12 August, the window compared with ERA5's preliminary August |
| note_numbers.json | Every number quoted in the note, as written by note_data.py |
Fields (netCDF)
| File | Contents |
|---|---|
| shear_2026-07-01_sep.nc · shear_2026-08-01_sep.nc | September 2026 ensemble-mean zonal and vector shear, the emulator's own baseline, and the anomaly, on the model's 180 × 360 grid |
| clim_sep_shear_baseline.nc | The emulator's 1991–2020 September shear climatology and its interannual standard deviation |
| seas5_sep2026_zonal_anom_ensmean_vs_era5.nc | ECMWF SEAS5's September 2026 ensemble-mean zonal shear anomaly against ERA5, 1° grid |
Code
ace2-reproduction-code.zip holds the pipeline as run, with its working README: download_seas5.py (Copernicus requests), make_forcing.py (SEAS5 GRIB → the emulator's forcing file, following Ai2's own out-of-sample recipe), make_ic.py (initial state), make_configs.py (one run directory per member or year, plus a resumable runner), shear.py (monthly layer winds → zonal and vector shear, ensemble mean, anomaly against either baseline), sst_indices.py, era5_check.py, note_data.py (every number in the note), and the plotting scripts. The emulator itself is Ai2's fme package with the public ACE2-ERA5 checkpoint; the environment is Python 3.11 with PyTorch built for CUDA 12.6.
Sources and attributions
ECMWF SEAS5 ocean forecasts (system 51, daily SST and sea-ice cover; monthly 200 and 850 hPa winds, forecasts and 1993–2016 hindcasts) and ERA5 / ERA5T reanalysis: contains modified Copernicus Climate Change Service information 2026; neither the European Commission nor ECMWF is responsible for any use of it. ACE2-ERA5 checkpoint, initial-condition and forcing files: Allen Institute for AI, Apache 2.0 (Watt-Meyer et al. 2025). Global-mean CO₂: NOAA Global Monitoring Laboratory. ENSO phase and RONI: NOAA Climate Prediction Center, as tabulated in the hurricane occurrence study's data. Colorado State University forecast reports of 9 July and 5 August 2026 are cited, not reproduced.