ClimX is a persistent benchmark for emulating high-resolution daily climate outputs from the NorESM2-MM Earth System Model, with special emphasis on accurately reproducing climate extremes (e.g., heatwaves, droughts, and extreme precipitation), not just mean climate behavior. A September--November 2026 Kaggle challenge serves as a launch event to seed community use and establish an initial leaderboard.
- Core task: Predict daily, 1° resolution climate variables from greenhouse gas and aerosol forcings.
- Submission rule: Models must predict the daily target variables first; leaderboard indices must be computed from those outputs.
- Data: Full dataset (~200GB, Hugging Face) and lightweight prototype dataset (<1GB, Kaggle).
- Evaluation: Region-wise nNSE averaged across 15 extreme climate indices.
- Test setting: Held-out SSP2-4.5 scenario.
- Optional track: Probabilistic predictions are evaluated with CRPS on a separate Kaggle UQ track.
For this repository, using mamba is recommended for faster and more reliable environment solves.
# one-time: install mamba into base conda
conda install -n base -c conda-forge mamba
# create the environment from this repo
mamba env create -f environment.yml
conda activate clima_emu_newThen launch Jupyter and open playground.ipynb:
jupyter notebookplayground.ipynb: End-to-end notebook for data loading, preprocessing, training baseline models, evaluation, Kaggle submission formatting, and result visualization.environment.yml: Full Python environment specification used by the notebook and training scripts.train.py: Script-based baseline training workflow (useful when you prefer Python scripts over notebooks).src/: Core code for preprocessing, models, metrics, utilities, and evaluation helpers.rnNSE.py: Exact reference implementation of the Kaggle main-track leaderboard metric.
The main track uses region-wise normalized Nash–Sutcliffe efficiency (nNSE). For each climate index, a grid cell is valid when the population variance of its ground-truth time series is greater than 1e-6 (ddof=0). Only cells in the embedded AR6 land-region raster are scored. Cell nNSE values are weighted by cosine latitude within each region; valid regions are then weighted equally, followed by equal weighting across indices.
The scorer expects complete 192×288 maps with rows ordered as time, latitude, then longitude. NaN and infinite predictions are invalid. See rnNSE.py for the exact implementation used by the competition.
- Kaggle main track: https://www.kaggle.com/competitions/climx-main-track
- Kaggle UQ track: https://www.kaggle.com/competitions/climx-uq-track
- Hugging Face dataset: https://huggingface.co/datasets/isp-uv-es/ClimX
- Website: https://ipl-uv.github.io/ClimX/
The launch challenge runs independently from September 1 through November 30, 2026; the benchmark remains available afterward. The organizers plan to submit an accompanying paper to Tackling Climate Change with Machine Learning at NeurIPS 2026 in Sydney. If accepted, additional ESA Phi-lab prize and travel support is expected to become available.