Selecting Observationally Constrained Global Climate Model Ensembles Using Autoencoders and Transfer Learning

Abstract Climate modes of variability are recurring patterns that influence climate phenomena across spatial scales. Accurately representing these modes in Global Climate Models (GCMs) is crucial for assessing model performance and reducing uncertainty in future climate projections. In this study, we present a novel approach utilizing autoencoder neural networks (AEs) combined with transfer learning to evaluate the representation of monthly sea level pressure (SLP) modes over North America across five GCMs: the Geophysical Fluid Dynamics Laboratory Climate Model (GFDL‐CM4), Centro Euro‐Mediterraneo sui Cambiamenti Climatici Climate Model (CMCC‐CM2‐SR5), Canadian Earth System Model Version 5 (CanESM5), Institut Pierre‐Simon Laplace Climate Model (IPSL‐CM6A‐LR), and Hadley Center Global Environment Model Version 3 (HadGEM3‐GC31‐LL). We derived the reference regional SLP modes using autoencoders (AE) from the European Center for Medium‐Range Weather Forecasts Reanalysis (ERA5), capturing more physically consistent SLP patterns. Transfer learning was employed to adapt the pre‐trained AE, from ERA5 to the GCM outputs, enabling a direct and robust evaluation of each model’s ability to produce the observationally constrained SLP modes. This approach allowed us to rank the GCMs based on how well they replicated the reference SLP modes, providing an observationally constrained assessment of model performance. The congruence coefficients between the modeled and reference modes exceeded 0.91 for all GCMs, demonstrating strong performance in simulating regional SLP modes over North America. Among the models, HadGEM3‐GC31‐LL achieved the highest performance with an average congruence coefficient of 0.94. These results highlight the effectiveness of neural network techniques in evaluating and ranking GCMs for model intercomparison projects.