rampaln/REMS-MR-LargeClimateEnsemble

NIWA-REMS generative-AI CMIP6 large ensemble

These AI outputs are entirely experimental. They are shared so that internal variability and uncertainty in fine-scale climate projections can be investigated. They are not official climate projections for Aotearoa New Zealand. They have no output bias correction. Do not use them on their own for adaptation decisions, engineering design, planning rules, or regulatory assessments.

For those purposes, use the official national products:

• Updated national climate projections for New Zealand (NIWA)
• Ministry for the Environment climate-data portal
• Climate change projections (Ministry for the Environment)

The official NIWA / Ministry for the Environment CMIP6 products are a different dataset. They are based on a small set of global models, regional climate modelling, and bias correction. This archive is the separate generative-AI large ensemble described below.

Paper

Rampal, N., Gibson, P. B., Sherwood, S. C., Queen, L. E., Lewis, H., & Abramowitz, G. (2026). Downscaling with AI reveals the large role of internal variability in fine-scale projections of climate extremes. Climate Dynamics, 64(8). https://doi.org/10.1007/s00382-026-08269-y

The paper is published under CC BY 4.0. That licence applies to the article, not automatically to this data archive. The files are the public NIWA-REMS release (CMOR_v = v1.0 Public). Questions about reuse go to Neelesh Rampal (neelesh.rampal@niwa.co.nz).

The paper shows why an ensemble of this size exists. Dynamical downscaling is too expensive to run at this membership. A generative-AI downscaler was used to produce daily fields over New Zealand from many CMIP6 realisations, including large initial-condition ensembles. At the end of the century, the internal variability of daily precipitation extremes grows with warming faster than the mean change, and can dominate other sources of uncertainty. This archive is the daily pr and tasmax output for that kind of investigation. It is not a bias-corrected national projection product.

@article{Rampal2026,
  author  = {Rampal, Neelesh and Gibson, Peter B. and Sherwood, Steven C. and Queen, Laura E. and Lewis, Hamish and Abramowitz, Gab},
  title   = {Downscaling with {AI} reveals the large role of internal variability in fine-scale projections of climate extremes},
  journal = {Climate Dynamics},
  year    = {2026},
  volume  = {64},
  number  = {8},
  doi     = {10.1007/s00382-026-08269-y}
}

What is in this repository

Original daily NetCDF files for historical and ssp370 only, uploaded without rewriting, recompression, or member deletion. Only pr and tasmax.

Experiment	pr files	tasmax files	Size
historical (1950–2014 in the files inspected)	77	77	1.250 TiB (1.343 TB)
ssp370 (SSP3-7.0)	77	77	1.624 TiB (1.786 TB)
Total	154	154	2.846 TiB (3.129 TB, 308 files)

Sizes are measured from the source files. Per-file sizes are in CATALOG.csv. ssp126, ssp245, and ssp585 are not in this release: pr and tasmax for all five experiments is 4.53 TB, which does not fit. Historical plus SSP3-7.0 is the subset that stays near 2.7 TiB.

Not included

The source tree also contains other variables and an alternate precipitation version. Those files were not uploaded:

Left out	Why
ssp126, ssp245, ssp585	Not in this release. All-scenario pr and tasmax is 4.53 TB.
pr_v280125	Alternate precipitation version (the pr_… directory to ignore). Present only for ssp370 and ssp585.
sfcwind, sfcwindmax, tasmin	Not part of this release.
historical/daily/test.nc	Scratch file, not a member of the ensemble.
Shell scripts and notebook checkpoints	Not data.

The uploaded precipitation is only the pr directory (NIWA-REMS_v110425_pr), not pr_v280125.

How the files are named and stored

On disk, and in this repository, the layout is:

{experiment}/daily/{variable}/{filename}.nc

experiment: historical | ssp370
variable:   pr | tasmax

Example:

historical/daily/pr/pr_CMIP_AS-RCEC_TaiESM1_historical_r1i1p1f1_day_NZ12km_NIWA-REMS_v110425_pr.nc
ssp370/daily/tasmax/tasmax_ScenarioMIP_CSIRO_ACCESS-ESM1-5_ssp370_r1i1p1f1_day_NZ12km_NIWA-REMS_tasmax_v280125.nc

Filename fields, split on _, with the fixed marker _day_NZ12km_NIWA-REMS_ before the version token:

{variable}_{activity}_{institution}_{source_id}_{experiment}_{variant}_day_NZ12km_NIWA-REMS_{version}.nc

Field	Meaning	Examples in this archive
variable	Daily field	pr, tasmax
activity	CMIP6 activity	CMIP for historical, ScenarioMIP for the SSPs
institution	Institution token as written in the filename	CSIRO, CCCma, MPI-M, DKRZ, EC-Earth-Consortium
source_id	CMIP6 source model	ACCESS-ESM1-5, CanESM5, TaiESM1
experiment	Experiment id	historical, ssp370
variant	CMIP6 variant label r{k}i{l}p{m}f{n}	r1i1p1f1, r40i1p1f1, r4i1p1f1
version	NIWA-REMS product token	v110425_pr or tasmax_v280125

The version token is not in the same order for the two variables. Take everything after NIWA-REMS_ rather than assuming one pattern. NZ12km is the nominal grid spacing, not a claim that every cell is exactly 12.000 km.

Do not join historical and future files by string-matching the whole filename. One known inconsistency, in both pr and tasmax: MPI-ESM1-2-HR is labelled MPI-M in the historical filenames and DKRZ in the scenario filenames. The member lists otherwise match between the two variables. Join on source_id and variant, then check CATALOG.csv.

Ensemble map

historical and ssp370 are the large ensembles (77 members each variable):

• ACCESS-ESM1-5: 40 initial-condition members, r1i1p1f1 through r40i1p1f1
• CanESM5: 19 members, r1–r8 and r10 at both p1 and p2, plus r9i1p1f1. r9i1p2f1 is not in the archive
• 18 other source models: one variant each

ssp126, ssp245, and ssp585 are not uploaded. In the source archive those experiments have one variant of each of 20 source models, not the 40-member ACCESS-ESM1-5 or 19-member CanESM5 ensembles.

Use the large historical and ssp370 membership to study internal variability. Use the spread across source models to study model uncertainty. A single member is not "the" projection for that model.

Daily fields and grid

Every file inspected is a daily time series on a regular longitude–latitude grid.

Item	Value
Frequency	Daily (frequency = day). Consecutive samples in the opened files are 1 day apart.
Grid	172 latitude × 179 longitude
Latitude	−51.213°N to −32.860°N, spacing about 0.10733°, degrees_north
Longitude	164.860°E to 183.965°E, spacing about 0.10733°, degrees_east
Domain	New Zealand and the surrounding ocean. Longitudes above 180° are west of the dateline (longitude - 360).
Coordinates	One-dimensional lat and lon. This is not the native curvilinear CCAM cube.
Missing data	_FillValue = NaN on the data variable
Calendar samples	proleptic_gregorian

Time coverage is stored on each file and is not identical across models.

Sample opened	Decoded time	Length	time units
TaiESM1 historical pr and tasmax	1950-01-01 to 2014-12-31	23725 days	days since 1950-01-01 00:00:00
TaiESM1 ssp245 pr	2015-01-01 to 2100-12-31	31390 days	days since 2015-01-01 00:00:00
ACCESS-ESM1-5 ssp370 pr and tasmax	2015-01-01 to 2100-12-31	31411 days	days since 2015-01-01 00:00:00

Read time, its units, and its calendar from the file you open. Do not assume every model has the same length or the same units origin, even when the decoded end date is 2100-12-31.

lat and lon advertise bounds = lat_bnds and lon_bnds, but those bound variables are not in the files.

Variables

Name	Role	In which files
pr	Daily precipitation field from the generative downscaler	pr files
tasmax	Daily maximum near-surface air temperature	tasmax files
pr_GAN_mask	Integer companion array stored with precipitation. No standard_name, units, or flag meanings are set.	pr files only

pr and tasmax also have no units or standard_name attribute in the files that were opened. Do not convert units until that gap is resolved. Inspected single time slices had non-negative pr and tasmax values near 280. That is the scale expected for millimetres per day and kelvin, but it is a sample of slices, not a declared unit and not a file-wide check.

pr_GAN_mask is retained only because the NetCDF files were uploaded unchanged. It is not a climate variable to analyse, and it has no documented flag values.

File metadata

Global attributes are written by CDO from the NIWA-REMS processing chain. They vary slightly by variable and member. In the TaiESM1 historical files they include:

Attribute	Content
title	Generative AI Downscaled data
source	NIWA-REMS generative-AI downscaling. Training data are from the CSIRO conformal-cubic atmospheric model (CCAM).
Conventions	CF-1.7
frequency	day
modeling_realm	atmos
project	NIWA CMIP6 Downscaling
CMOR_v	v1.0 Public
BC_version	no output bias correction
CMIP6_nudging	GAN downscaled with u, v, q, and t fields at 1.5° from the CMIP6 GCM
CMIP6_model, GCM	Source model, for example TaiESM1
CMIP6_scenario, scenario	Experiment, for example historical
version	NIWA-REMS_v110425_pr or NIWA-REMS_tasmax_v280125, plus CCAM r5251:5254M
contact	Neelesh Rampal (neelesh.rampal@niwa.co.nz)
training_data	https://zenodo.org/records/13755688
rlat0, rlong0, schmidt	Parent CCAM grid parameters (−41.5, 173.5, 0.3428572). The stored field itself is regular lat–lon.

Method code linked from source: On the Extrapolation of Generative Adversarial Networks for downscaling precipitation extremes.

How to use the files

The Hub dataset viewer cannot preview these NetCDF files. Download a file, or a subset, and open it with xarray, netCDF4, or CDO.

One member:

from huggingface_hub import hf_hub_download
import xarray as xr

repo_id = "rampaln/REMS-MR-LargeClimateEnsemble"
filename = (
    "historical/daily/tasmax/"
    "tasmax_CMIP_AS-RCEC_TaiESM1_historical_r1i1p1f1_"
    "day_NZ12km_NIWA-REMS_tasmax_v280125.nc"
)
path = hf_hub_download(repo_id=repo_id, filename=filename, repo_type="dataset")
ds = xr.open_dataset(path)
print(ds.time.values[[0, -1]])
print(ds.tasmax.isel(time=0).mean().item())

List what has actually been uploaded, then parse the catalog rather than guessing names:

from huggingface_hub import hf_hub_download, list_repo_files
import pandas as pd

repo_id = "rampaln/REMS-MR-LargeClimateEnsemble"
print([p for p in list_repo_files(repo_id, repo_type="dataset") if p.endswith(".nc")][:5])

catalog = pd.read_csv(
    hf_hub_download(repo_id=repo_id, filename="CATALOG.csv", repo_type="dataset")
)
members = catalog.query("scenario == 'ssp370' and variable == 'pr' and source_id == 'ACCESS-ESM1-5'")
print(members[["variant", "path_in_repo"]].head())

A whole experiment is hundreds of gigabytes to more than a terabyte. Prefer allow_patterns over a full snapshot:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="rampaln/REMS-MR-LargeClimateEnsemble",
    repo_type="dataset",
    allow_patterns=["historical/daily/tasmax/*r1i1p1f1*.nc"],
    local_dir="rems_tasmax_r1",
)

From a shell, after pip install huggingface_hub:

huggingface-cli download rampaln/REMS-MR-LargeClimateEnsemble \
  historical/daily/pr/pr_CMIP_AS-RCEC_TaiESM1_historical_r1i1p1f1_day_NZ12km_NIWA-REMS_v110425_pr.nc \
  --repo-type dataset \
  --local-dir rems_example

Suggested checks before an analysis:

1. Confirm the file is the experiment, source model, and variant you intended. Use CATALOG.csv, not a partial filename match.
2. Read time.units and time.calendar. Do not assume 1950 or 2015 as the origin, and do not assume a common length.
3. Do not apply a unit conversion until units is confirmed. It is missing on the inspected data variables.
4. Remember BC_version = no output bias correction.
5. For internal variability, use the ACCESS-ESM1-5 or CanESM5 members inside historical and ssp370. For a national climate projection, use the official datasets linked at the top.

Source tree

The files were read from:

/esi/project/niwa03712/ML_Downscaled_CMIP6/NIWA-REMS_CCAM_public

Only historical/daily/{pr,tasmax}/*.nc and ssp370/daily/{pr,tasmax}/*.nc were selected. The NetCDF contents were not modified for this upload.