Yiou et al. (2025) Using artificial intelligence to identify CMIP6 models from daily SLP maps

Identification

• Journal: npj Climate and Atmospheric Science
• Year: 2025
• Date: 2025-11-17
• Authors: Pascal Yiou, Soulivanh Thao
• DOI: 10.1038/s41612-025-01246-y

Research Groups

• Laboratoire des Sciences du Climat et de l’Environnement (UMR 8212 CEA-CNRS-UVSQ, Université Paris-Saclay, Gif-sur-Yvette, France)
• Institut Pierre-Simon Laplace (Sorbonne Université, Paris, France)

Short Summary

This study employs a neural network to classify CMIP6 climate models from single daily sea-level pressure (SLP) maps over the North Atlantic, revealing high model identifiability in summer and enabling the identification of model families and the assessment of climate change impacts on SLP patterns.

Objective

• To determine whether it is possible to recognize a climate model from a single daily sea-level pressure (SLP) map over the North Atlantic.
• To investigate how a warmer climate, according to Shared Socioeconomic Pathway (SSP) scenarios, affects the daily weather patterns of climate models.

Study Configuration

• Spatial Scale: North Atlantic region [50° W–20° E; 30° N–65° N], with all SLP fields interpolated onto 1° × 1° maps.
• Temporal Scale:
  • Historical period: 1970–2000 (approximately 22 years for training, 8 years for validation).
  • Future period: 2070–2100 (under SSP5-8.5 scenario).
  • Daily time steps, analyzed seasonally (June–July–August (JJA), September–October–November (SON), December–January–February (DJF), March–April–May (MAM)).

Methodology and Data

• Models used:
  • Artificial Intelligence (AI) model: A simple dense neural network (multi-layer perceptron) with one hidden layer of 256 neurons (ReLu activation) and an output layer of 17 neurons (softmax activation).
  • Climate Models: 16 selected CMIP6 Global Climate Models (GCMs) (e.g., BCC-CSM2-MR, FGOALS-g3, CanESM5, CNRM-CM6-1, ACCESS-ESM1-5, EC-Earth3, INM-CM5-0, IPSL-CM6A-LR, MIROC6, HadGEM3-GC31-LL, MPI-ESM1-2-LR, MRI-ESM2-0, CESM2, NorCPM1, KACE-1-0-G, NESM3) and their "sister" models.
  • Reanalysis Models: ERA5 and NCEP reanalyses.
• Data sources:
  • Daily sea-level pressure (SLP) fields from the CMIP6 archive (historical and SSP simulations).
  • European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 reanalysis (0.25° horizontal resolution).
  • National Centers for Environmental Prediction (NCEP) reanalysis (2.5° horizontal resolution).
  • Data primarily sourced from the IPSL database server (a subset of ESGF data).

Main Results

• CMIP6 models are highly identifiable from single daily SLP maps over the North Atlantic, particularly in summer (JJA), with classification success probabilities often exceeding 0.6 for several models.
• Model identifiability is significantly lower in other seasons (SON, MAM, DJF), with winter (DJF) showing the poorest classification scores, making GCMs and reanalyses largely indistinguishable.
• The NorCPM1 model consistently exhibits high classification rates across all seasons, while the EC-Earth3 model consistently shows low classification rates and is frequently confused with ERA5 due to shared atmospheric model components.
• Internal variability within different runs of the same model does not substantially affect SLP identification in summer.
• Family ties among GCMs (from the same research group with comparable horizontal resolutions) can be identified using daily summer SLP maps.
• Under the SSP5-8.5 climate change scenario (2070-2100), classification probabilities in summer slightly decrease, suggesting some models develop new SLP patterns, but GCMs generally remain identifiable based on historical training.
• The NESM3 model is an exception, exhibiting a notable change in winter atmospheric circulation patterns under SSP5-8.5, rendering its future SLP patterns unrecognizable from its historical simulations.
• The AI model's classification decisions are based on small, localized SLP patterns (e.g., over the Sahara, Mediterranean, Alps, Greenland) and, in some cases, larger-scale wave-like patterns over the North Atlantic.

Contributions

• Demonstrates the feasibility of identifying individual climate models from single daily SLP maps using a relatively simple neural network, particularly highlighting strong identifiability in the summer season.
• Provides a nuanced perspective on the interchangeability of climate models, suggesting that pooling models is more appropriate in seasons where identifiability is low (e.g., winter).
• Establishes a method for identifying family ties among GCMs based on daily SLP patterns, underscoring the robustness of the atmospheric component of coupled models when horizontal resolutions are similar.
• Investigates the influence of climate change (SSP5-8.5) on model identifiability, revealing general stability but also identifying specific models (e.g., NESM3 in winter) that undergo significant pattern shifts.
• Offers valuable insights for AI-based weather forecasting, by identifying potential candidate GCMs suitable for transfer learning or pooling, and emphasizing the need for complex bias correction otherwise.
• Extends previous research by focusing on daily timescales and a climate variable (SLP) that exhibits less obvious biases than surface temperature, thereby addressing a more challenging classification problem.
• Proposes a framework that functions as a "Turing test" for climate models, particularly relevant for their use and trustworthiness at daily timescales.

Funding

• ANR-20-CE01-0008-01 (SAMPRACE)
• Agence Nationale de la Recherche—France 2030 (PEPR TRACCS program, ANR-22-EXTR-0002)
• European Union’s Horizon 2020 research and innovation program (grant agreement No. 101003469, XAIDA)

Citation

@article{Yiou2025Using,
  author = {Yiou, Pascal and Thao, Soulivanh},
  title = {Using artificial intelligence to identify CMIP6 models from daily SLP maps},
  journal = {npj Climate and Atmospheric Science},
  year = {2025},
  doi = {10.1038/s41612-025-01246-y},
  url = {https://doi.org/10.1038/s41612-025-01246-y}
}