ClimX: a benchmark for extreme-aware climate model emulation

ClimX is a persistent benchmark for building fast and accurate machine learning emulators of the NorESM2-MM Earth System Model, with evaluation focused on climate extremes rather than mean climate alone.

An online Kaggle challenge will run from September 1 through November 30, 2026 as a launch event for the benchmark. The benchmark, data, code, and evaluation protocol remain available independently of the challenge. The organizers plan to submit an accompanying paper to the NeurIPS 2026 workshop Tackling Climate Change with Machine Learning in Sydney; if accepted, challenge results will be presented there.

Dataset summary

This dataset contains the full-resolution ClimX data in NetCDF-4 format (targets + forcings, depending on split) with a native grid of $192\times 288192\; \backslash times\; 288$ (about $1^{\circ }1^\backslash circ$) resolution. It also contains the lite-resolution version, with a native grid of $12\times 1812\; \backslash times\; 18$ (about ${16}^{\circ }16^\backslash circ$) resolution:

• Lite-resolution: <1GB, $16\times 16\backslash times$ spatially coarsened, meant for rapid prototyping.
• Full-resolution: ~200GB, full-resolution data for large-scale training.

What you will do (high level)

You train an emulator that predicts daily 2D fields for 7 surface variables:

• tas, tasmax, tasmin
• pr, huss, psl, sfcWind

However, the benchmark targets are 15 extreme indices derived from daily temperature and precipitation (ETCCDI-style indices). The daily fields are an intermediate output your emulator produces (useful for diagnostics and for computing the indices).

Benchmark users produce daily target variables from which the indices are derived. Challenge participants must follow the additional competition rule prohibiting direct prediction of leaderboard indices.

Conceptually:

$$x_t=g(f_t,f_{t-1},\dots ,f_{t-\alpha },x_{t-1},x_{t-2},\dots ,x_{t-\beta })x\_t\; =\; g(f\_t,\; f\_\{t-1\},\; \backslash dots,\; f\_\{t-\backslash alpha\},\; x\_\{t-1\},\; x\_\{t-2\},\; \backslash dots,\; x\_\{t-\backslash beta\})$$

where $f_{t}f\_t$ are forcings (greenhouse gases + aerosols) and $x_{t}x\_t$ is the climate state.

Dataset structure

Spatial and temporal shape

Full-resolution daily fields:

• Historical: lat: 192, lon: 288, time: 60224
• Projections: lat: 192, lon: 288, time: 31389

Splits and scenarios (official challenge setup)

Training uses historical + several SSP scenarios; testing is on the held-out SSP2-4.5 scenario:

• Train: historical (1850–2014) + ssp126, ssp370, ssp585 (2015–2100)
• Test (held-out): ssp245 (2015–2100)

To avoid leakage, targets for ssp245 are withheld in the official evaluation; only the forcings are provided for that scenario. The full outputs will be released after the competition.

Evaluation metric

The primary leaderboard metric is the region-wise normalized Nash–Sutcliffe efficiency (nNSE), averaged over 15 climate extreme indices.

For each index $vv$, grid cell $(i,j)(i,j)$, a validity mask $\mathcal{V}\backslash mathcal\{V\}$ excludes cells with negligible temporal variability. Cell-level $R^{2}R^2$ and nNSE are:

$$R_{ij}^2=1-\frac{{\mathrm{M}\mathrm{S}\mathrm{E}}_{ij}}{{\mathrm{V}\mathrm{a}\mathrm{r}}_t(g{t}_{ij})},{\mathrm{n}\mathrm{N}\mathrm{S}\mathrm{E}}_{ij}=\frac{R_{ij}^2}{2-R_{ij}^2}R^2\_\{ij\}\; =\; 1\; -\; \backslash frac\{\backslash mathrm\{MSE\}\_\{ij\}\}\{\backslash mathrm\{Var\}\_t(gt\_\{ij\})\},\; \backslash qquad\; \backslash mathrm\{nNSE\}\_\{ij\}\; =\; \backslash frac\{R^2\_\{ij\}\}\{2\; -\; R^2\_\{ij\}\}$$

For each AR6 land region $kk$, the area-weighted regional score is:

$${\mathrm{n}\mathrm{N}\mathrm{S}\mathrm{E}}_{kv}=\frac{\sum _{(i,j)\in k\cap \mathcal{V}}\cos {\varphi }_i{\mathrm{n}\mathrm{N}\mathrm{S}\mathrm{E}}_{ij}}{\sum _{(i,j)\in k\cap \mathcal{V}}\cos {\varphi }_i}\backslash mathrm\{nNSE\}\_\{kv\}\; =\; \backslash frac\{\backslash sum\_\{(i,j)\backslash in\; k\; \backslash cap\; \backslash mathcal\{V\}\}\; \backslash cos\backslash phi\_i\; \backslash ,\; \backslash mathrm\{nNSE\}\_\{ij\}\}\{\backslash sum\_\{(i,j)\backslash in\; k\; \backslash cap\; \backslash mathcal\{V\}\}\; \backslash cos\backslash phi\_i\}$$

The final score averages uniformly over valid regions and indices:

$$S=\frac{1}{\mathrm{\mid }V\mathrm{\mid }}\sum _{v\in V}\frac{1}{\mathrm{\mid }{K}_v\mathrm{\mid }}\sum _{k\in K_v}{\mathrm{n}\mathrm{N}\mathrm{S}\mathrm{E}}_{kv}S\; =\; \backslash frac\{1\}\{|V|\}\; \backslash sum\_\{v\; \backslash in\; V\}\; \backslash frac\{1\}\{|K\_v|\}\; \backslash sum\_\{k\; \backslash in\; K\_v\}\; \backslash mathrm\{nNSE\}\_\{kv\}$$ $S=1S=1$ is perfect agreement, $S=0S=0$ corresponds to a mean predictor, and $S\to -1S\; \backslash to\; -1$ is pathological.

How to load the data

This dataset is distributed as NetCDF-4 files. There are two common ways to load it.

Option 1 (recommended): clone the ClimX code and use the helper loader

The ClimX repository includes a helper module (src/utils/hugging_face_utils.py) that downloads the requested dataset variant from Hugging Face and opens it as three lazily-loaded “virtual” xarray datasets. Use Python 3.10 or newer:

git clone https://github.com/IPL-UV/ClimX.git
cd ClimX
pip install -U "huggingface-hub" xarray netcdf4 dask

from pathlib import Path
from src.utils.hugging_face_utils import get_dataset_from_hf, open_climx_virtual_datasets

# Download NetCDF artifacts from HF into a local cache directory.
root = Path("/path/to/hf_cache")
get_dataset_from_hf(root, variant="full")

# Open as three virtual datasets (lazy / dask-friendly).
ds = open_climx_virtual_datasets(root, variant="full")  # or "lite"

ds.hist # historical (targets + forcings)
ds.train # projections training scenarios (targets + forcings; excludes `ssp245` scenario)
ds.test_forcings # `ssp245` scenario forcings only (no targets)

Option 2: download NetCDFs and open with xarray directly

You can also download files from Hugging Face and open them with xarray.

Example:

from huggingface_hub import hf_hub_download
import xarray as xr

path = hf_hub_download(
    repo_id="isp-uv-es/ClimX",
    repo_type="dataset",
    filename="PATH/TO/A/FILE.nc",  # replace with an actual file in this dataset repo
)
ds = xr.open_dataset(path)
print(ds)

Links

• Kaggle main track
• Kaggle UQ track
• Full dataset (this page)
• Public benchmark repository
• Website

Challenge timeline and workshop plan

• September 1, 2026: challenge opens
• September 29, 2026: workshop paper acceptance notification
• November 30, 2026: submissions close, the leaderboard freezes, and finalist reproducibility checks begin
• December 11 or 12, 2026: results presented at Tackling Climate Change with Machine Learning in Sydney, if the paper is accepted

The website centralizes the challenge FAQ, tutorial resources, timeline updates, and submission guidance:

• Competition website
• Organizer contact for private questions: oscar.pellicer@uv.es

Prizes

The challenge, final rankings, and results paper will proceed whether or not the workshop submission is accepted.

Prize funding is conditional on acceptance of the accompanying workshop paper. If it is accepted, ESA Phi-lab funding is expected to provide:

• Main-track prizes: €1,000 for first place, €500 for second place, and €300 for third place.
• Travel support: up to €500 per winning team for the in-person results presentation in Sydney.

If the paper is not accepted, the challenge will still conclude normally and the organizers will write a paper reporting its results, but no prize or travel funding will be available.

License and usage

The ClimX software is released under the MIT License. Redistributed CMIP6 and input4MIPs data retain the licenses, attribution requirements, and provenance metadata of their original providers. Challenge participants must additionally follow the competition rules, including restrictions on external climate training data and redistribution.