Pascal Yiou and Soulivanh Thao, LSCE, ESTIMR :

Motivation

One often treats different global climate models (GCMs) as interchangeable when analyzing simulations or training machine-learning systems. But what if the daily-scale atmospheric patterns produced by each CMIP6 model actually carry a distinct “fingerprint”? In this study we explore whether a neural-network classifier can recognize which GCM from the CMIP6 archive (and reanalysis datasets) generated a given daily map of sea-level pressure (SLP) over the North Atlantic region.

Key findings

The neural network achieved good identification performance in summer (June-July-August), correctly classifying more than ~60 % of daily maps to the right model. For other seasons (spring, autumn, winter) the performance drops significantly, especially in winter where the models and the reanalysis become harder to tell apart (Figure 1).
The neural network can identify “sister” models (i.e., with different atmospheric or ocean configurations) in the summer season.
The patterns contributing to the classification reveal that even subtle spatial details in the SLP fields—such as pressure features near the Sahara, Mediterranean, Greenland or wave-like patches in the North Atlantic—help distinguish between models.

Broader implications

The fact that models are identifiable from a single daily SLP map means that they are not truly interchangeable at this temporal and spatial resolution—pooling their outputs without caution may mislead machine-learning or statistical analyses.
On the other hand, we suggest that grouping models by family (rather than treating all models as independent) may be a robust strategy when generating large ensembles for attribution or extreme-event studies.
From the perspective of AI in weather/climate forecasting: if one wishes to train AI systems on large model-derived archives to improve predictive skill, one must be aware that each model’s patterns of spatiotemporal variability might differ. Bias-correction steps may be necessary.

In short: even a simple neural network can pick out which model produced a daily North-Atlantic SLP map—especially in summer—indicating that models carry distinguishable internal structures of variability. This challenges the common assumption that multi-model ensembles are freely interchangeable, and opens up new avenues for how we might group models, correct biases, or design AI forecasting systems based on climate-model outputs.

Figure: Classification test. Empirical probabilities of classifying CMIP6 models or ERA5 onto those models, for the four seasons (panels a to d). CMIP6 model simulations in the vertical axes are distinct from the training simulations in the horizontal axes. Darker colors indicate higher probabilities. Model simulations listed on the vertical axis are classified onto the list of models on the horizontal axis.

Yiou, P., and S. Thao, 2025: Using artificial intelligence to identify CMIP6 models from daily SLP maps. npj Clim Atmos Sci, 8, 366, https://doi.org/10.1038/s41612-025-01246-y.