A data-driven approach for classifying GCM errors and understanding their impact on climate projections.

Evaluating regional climate model performance often relies on simple error metrics that assume the largest mean errors matter most, sometimes combined with visual inspection of the large-scale atmosphere circulation fields, that introduces a priori assumptions. However, the influence of spatial error patterns in the large-scale circulation on other surface variables, such as precipitation, is complex, and their link to future projections remains uncertain.

We present a novel computer vision–based framework for regional climate model evaluation that emphasises interpretability and explainability. Rather than treating machine learning as a black box, our approach learns structural characteristics of seasonal mean sea-level pressure fields directly from CMIP6 model data. Using a convolutional variational autoencoder (CNN‑VAE), we construct an interpretable latent space in which large-scale atmospheric patterns cluster according to shared spatial structure.

These clusters enable the identification of systematic differences in how regional climate models represent key large-scale drivers of European climate. Deviations from reanalysis are quantified using simple distance metrics in latent space, allowing the magnitude of structural model errors to be directly compared in a physically meaningful way and without reliance on subjective visual assessment or pointwise error measures.

We further compare end‑of‑century precipitation and temperature projections from models occupying different regions of latent space to assess how distinct large‑scale circulation errors impact future surface climate projections. We present this methodology as a complementary approach to traditional error metrics that can be applied flexibly to any variable of interest in model evaluation.

By linking learned representations to physically interpretable circulation structures, this framework supports more trustworthy use of machine learning in climate model evaluation and provides new insights into how spatial error patterns may influence downstream variables and projections.