ClimX Challenge

ClimX is an extreme-aware climate emulation challenge that tasks machine learning models with predicting daily weather variables and 15 extreme climate indices.
1 September 2026
Challenges & Competitions | ELIAS Alliance | Opportunities
7
Read more

Competition Details

About the Challenge

The ClimX Challenge is an ML competition on Kaggle and Hugging Face focused on extreme-aware climate model emulation. Organized by the Universitat de València and supported by ESA Phi-lab, it tasks participants with building fast surrogate models for the computationally heavy NorESM2-MM Earth System Model. Driven by greenhouse gas and aerosol inputs, models predict daily 2D maps across seven surface variables. Submissions are uniquely evaluated on 15 derived extreme climate indices—such as heatwaves, droughts, and severe rainfall—to ensure emulators capture high-impact tail risks rather than simple averages.

Z
Scope & Objectives

  • Problem Tackled: Earth System Models (ESMs) like NorESM2-MM are computationally expensive, restricting dense exploration of climate forcing scenarios, initial condition ensembles, and extreme events. Traditional climate emulators often optimize for average climate state, failing to capture severe, policy-relevant tail events.

  • Objective: Build fast, accurate machine learning surrogate models (emulators) driven by greenhouse gas and aerosol forcing trajectories (and optionally past climate states) that output daily 2D climate variable maps and faithfully capture high-impact climate extremes.

  • Impact: Enables rapid climate scenario testing, robust quantification of tail risks, and decision support for climate adaptation without requiring computationally prohibitive supercomputer ESM runs.

Dataset & Resources

  • Source Simulation: High-resolution daily outputs from the NorESM2-MM Earth System Model.

  • Target Daily Variables (7 total):

    • Near-surface air temperature (tas), Daily max temperature (tasmax), Daily min temperature (tasmin)

    • Precipitation (pr), Near-surface specific humidity (huss), Sea level pressure (psl), Near-surface wind speed (sfcWind).

  • Derived Extremes (15 Leaderboard Indices): ETCCDI-style extreme climate indices derived from daily temperature and precipitation fields (e.g., TXx, SU, TR, CDD, Rx5day, R95pTOT).

  • Data Splits & Scenarios:

    • Training: Historical period (1850–2014) + SSP1-2.6, SSP3-7.0, and SSP5-8.5 scenarios (2015–2100).

    • Testing: Held-out SSP2-4.5 scenario (2015–2100).

  • Data Access Tiers:

    • Hugging Face (Full Dataset): ~172GB–200GB full-resolution dataset in Zarr/NetCDF format ($192 \times 288$ native grid, ~1° resolution) for full model training.

    • Kaggle (Lite Dataset): <1GB spatially coarsened dataset ($12 \times 18$ grid) for rapid pipeline prototyping and debugging.

  • Codebase & Baseline Resources: Official GitHub repository (IPL-UV/ClimX) containing end-to-end notebooks (playground.ipynb), environment specifications, helper utilities, and baseline benchmarks (Climatology, Linear Surrogates, Neural Networks, Graph Neural Networks).

Eligibility

  • Sign Up: Register on Kaggle for either the Main Track (deterministic) or UQ Track (probabilistic/uncertainty quantification).

  • Develop Emulator: Build an emulator that predicts the 7 daily 2D surface variables for the held-out test forcing scenario (SSP2-4.5).

  • Compute Indices & Format: Deriving the 15 extreme climate indices from your daily target predictions.

  • Submit Entry: Upload the derived extreme index maps (86 annual time steps for SSP2-4.5 at the full $192 \times 288$ grid resolution, totaling 4,755,456 rows per index) to Kaggle.

    • Rule: Models must predict daily variables first; direct prediction of leaderboard indices without generating daily intermediate fields is strictly prohibited.

Z
Expected Outcomes
  • Model Innovation: Development of state-of-the-art spatio-temporal ML emulators capable of preserving fine-scale physical tail distributions.
  • Scientific Insights: Rigorous benchmark comparison across architectures (CNNs, GNNs, Transformers, probabilistic emulators) evaluated specifically on climate extremes rather than spatial/temporal means.
  • Publication & Dissemination: Winning entries and challenge results will be highlighted and presented at the Tackling Climate Change with Machine Learning workshop at NeurIPS 2026 in Sydney.
Key Dates

  • Running Period: September 1, 2026 – November 30, 2026 (Launch Challenge).

  • Key Milestones:

    • September 1, 2026: Competition officially opens on Kaggle.

    • September 29, 2026: NeurIPS workshop paper acceptance notification.

    • November 30, 2026: Submissions close, leaderboard freezes, and finalist code reproducibility checks begin.

    • December 11–12, 2026: Results presentation at the NeurIPS 2026 Tackling Climate Change with Machine Learning workshop in Sydney (upon paper acceptance).

  • Persistence: ClimX is a persistent benchmark; all datasets, metrics, and starter code remain open and publicly available after the launch challenge concludes.

Please make sure to check out the detailed description:
GET STARTED
Evaluation Criteria

How Entries Will Be Evaluated

  • Main Track (Deterministic): Evaluated using the region-wise normalized Nash–Sutcliffe efficiency (nNSE), averaged across all 15 derived climate extreme indices over IPCC AR6 land regions:
    nNSEij = R²ij ⁄ (2 − R²ij)
    This metric maps cell-level R² to the range (-1, 1] (where 1 = perfect match, 0 = climatology/mean predictor), computes area-weighted regional scores, and averages them across indices.
  • UQ Track (Probabilistic): Evaluated using an analogous regional Continuous Ranked Probability Score (CRPS)-based metric.
  • Masking Filter: Grid cells with negligible temporal variability (Var(y) ≤ ε²) are masked out to avoid instability in ratios.

Judging Panel

  • Review Panel: Challenge chairs and research leaders from the Image Processing Laboratory (IPL / ISP) at Universitat de València.
  • Responsibilities: Leaderboard validation, code review for top entries, and verification of rule compliance (confirming models predict daily fields prior to index calculation).
Prizes

  • Sponsoring Body: ESA Phi-lab (European Space Agency).

  • Prize Pool (conditional on NeurIPS workshop paper acceptance):

    • 1st Place: €1,000

    • 2nd Place: €500

    • 3rd Place: €300

  • Travel Support: Up to €500 per winning team to support in-person presentation at NeurIPS 2026 in Sydney.

Have Questions?

For further information or queries, please contact:
📧Oscar José Pellicer Valero (oscar.pellicer@uv.es)

Repository / Technical Support: Open an issue or discussion thread on the ClimX GitHub Repository.

Registration Links & Resources
    Academic Partners

    Upcoming Events

    View Calendar