ClimX: Extreme-aware climate model emulation is a persistent benchmark designed for fast and accurate machine learning emulators of the NorESM2-MM Earth System Model. Unlike traditional benchmarks that focus on mean climate, ClimX evaluates models based on their ability to predict climate extremes. This Kaggle challenge serves as a time-bounded launch event to seed benchmark use and establish an initial community leaderboard.
What Participants Will Do
Participants will train models to emulate daily climate fields for seven surface variables (tas, tasmax, tasmin, pr, huss, psl, sfcWind) on the native NorESM2-MM grid. The primary target of the benchmark is not the daily fields themselves, but 15 climate extreme indices derived from these fields (ETCCDI-style indices). Models must predict the daily target variables first; direct prediction of the leaderboard indices is not allowed.
Models can utilize: * Forcings: Greenhouse gases and aerosols for the target period. * Past climate state: Optional autoregressive emulation.
Training data includes historical periods and several SSP scenarios, while testing is conducted on the held-out SSP2-4.5 scenario.
Tracks and Metrics ClimX utilizes two competition pages: * Main Track: Deterministic submissions scored with region-wise normalized Nash–Sutcliffe efficiency (nNSE). * UQ Track: Probabilistic submissions scored with the analogous CRPS-based metric.
The primary leaderboard metric averages the nNSE over 15 climate extreme indices across valid AR6 land regions. This approach prioritizes reproducing physically meaningful extremes—such as heat intensity, drought persistence, and flood intensity—rather than average climate conditions.
Timeline and Prizes The challenge runs from September 1 through November 30, 2026. Prize funding is conditional on the acceptance of an accompanying paper submitted to the NeurIPS 2026 workshop Tackling Climate Change with Machine Learning in Sydney. If accepted, ESA Phi-lab funding provides: * €1,000 for first place, €500 for second place, and €300 for third place in the main track. * Travel support of up to €500 per winning team for the in-person results presentation in Sydney.
If the paper is not accepted, the challenge concludes normally, but no prize or travel funding will be available.
Data Access To balance accessibility with large-scale training needs, data is provided via a two-tier strategy: * Kaggle (Lite/Debug): A spatially coarsened version (~GB) for rapid prototyping in Kaggle Notebooks. * Hugging Face (Full Dataset): ~200GB of full-resolution daily NorESM2-MM data required for competitive entries and native-resolution modeling.