Testing the realism of interannual to centennial climate variability in a generative coupled atmosphere-ocean deep learning model

Application of deep learning has proved useful in many scientific domains and has also gained increased interest as a tool for weather and climate modeling in recent years. Deep Learning weather models have already demonstrated competitive prediction performance to state-of-the-art methods while hybrid models and emulators have shown some promise for climate simulation. However, the realism of simulated climate variability, and climate modes of pure deep learning models trained only on observational or reanalysis data, has not received as much attention.
As one example of these models, we investigate DLESyM, an autoregressive deep learning model based on the U-Net architecture and originally trained on ERA5 reanalysis data from 1981 to 2017 (REF1). Unlike many weather-generating deep learning models, DLESyM does not draw on sea-surface temperatures as boundary conditions, but learns to generate ocean surface patterns. Its applications could, therefore, extend to free-running simulations. The original authors showed its ability to generate stable climate simulations for time-spans up to three millenia, with the absence of spurious drifts and unphysical smoothing in the annual cycle. Here we test how realistic the simulated climate variability of DLESyM is, focusing on interannual to centennial spatio-temporal modes of internal climate variability. We seek to identify whether it is able to generalize to the underlying physical processes of the climate system, or if it is only capable of reproducing spatio-temporal statistical patterns of its training data. We compare the unforced variability of the deep learning model to that in equilibrium simulations out of General Circulation Models out of the Coupled Model Intercomparison Project phase 6 (CMIP6 GCMs), and palaeoclimate reconstructions (REF2). We focus on regional and global power spectra of surface temperatures, and gradients between land and ocean, tropics and extratropics, as well as the high latitudes. To assess the model’s ability to generalize outside the distribution of the training data we perform simulations from varying initial conditions, and comparing them with the output of CMIP6 GCMs. Based on this we discuss potentials and limitations of such a purely data-driven model for climate simulations and future climate risk assessment, where characteristics beyond mean state and slow changes become relevant.

 

REF1 Cresswell-Clay, N., Liu, B., Durran, D. R., Liu, Z., Espinosa, Z. I., Moreno, R. A., & Karlbauer, M. (2025). A deep learning Earth system model for efficient simulation of the observed climate. AGU Advances, 6, e2025AV001706. https://doi.org/10.1029/2025AV001706

REF2 Laepple, T., Ziegler, E., Weitzel, N. et al. (2023) Regional but not global temperature variability underestimated by climate models at supradecadal timescales. Nat. Geosci., 16, 958–966. https://doi.org/10.1038/s41561-023-01299-9