Extratropical Atmospheric Circulation Response to ENSO in Deep Learning Pacific Pacemaker Experiments

TL;DR

This study investigates whether a deep learning climate emulator (DLESyM) can reproduce ENSO-driven teleconnections when forced to observed tropical Pacific variability. It employs the Pacific Pacemaker (PACE) framework to isolate the atmospheric response to tropical forcing by nudging Pacific SSTs toward observations, and contrasts two background climatologies to identify mean-state influences. The results reveal that while DLESyM captures the spatial pattern of the extratropical response, it exhibits a significantly amplified teleconnection amplitude and biases in blocking statistics compared to CESM2, with internal variability largely realistic. The findings highlight the necessity for physically grounded validation of deep learning emulators and show how background climatology can strongly shape teleconnections, informing future improvements and risk assessment for AI-based climate tools.

Abstract

Coupled atmosphere-ocean deep learning (DL) climate emulators are a new frontier but are known to exhibit weak ENSO variability, raising questions about their ability to simulate teleconnections. Here, we present the first Pacific pacemaker (PACE) experiments using a coupled DL emulator (DLESyM) to bypass this weak variability and isolate the atmospheric response to observed ENSO forcing. We find that while the emulator realistically captures internal atmospheric variability, it produces a significantly amplified forced teleconnection response to ENSO. This amplified response leads to biases in simulating extremes, notably an overestimation of atmospheric blocking frequency and duration with the underestimation of peak intensity. Our findings underscore that coupled DL climate models require in-depth and physically-grounded validation, analogous to traditional numerical models, to build confidence in their use for physical climate analysis.

Figures (7)

• Figure 1: (a–b) Climatological sea surface temperature (SST) differences of the DLESyM free-running simulation relative to (a) the ERSSTv5 observational dataset and (b) the CESM2 Large Ensemble (LENS2). The black contour indicates the pacemaker masking region. (c) Normalized spectral variance of the Niño 3.4 index following the methodology of Cresswell-Clay et al. 2025 CresswellClay2024ADL. The correlation coefficient ($r$) denotes the lag-1 autocorrelation of each Niño 3.4 time series derived from the models and ERSSTv5. The dashed line represents the theoretical red-noise spectrum corresponding to the observed lag-1 autocorrelation ($r$). The gray shaded region indicates the 2–7 year ENSO band. For DLESyM, $n$ denotes the number of ensemble members, and the shaded envelope represents one standard deviation across the ensemble. (d) DJF ENSO (El Niño $-$ La Niña) composite of SST anomalies from ERSSTv5.
• Figure 2: ENSO composites (El Niño $-$ La Niña) of DJF Z500 (gpm) for (a) the ensemble mean of the 10 CESM2 Pacific Pacemaker (PACE) simulations, (c) DLESyM-PACE with CESM2 SST climatology, (d) ERA-20C, (f) DLESyM-PACE with DLESyM SST climatology, and (h) the DLESyM free-running experiment. Panels (b), (e), (g), and (i) show their respective differences relative to ERA-20C. For panels (a), (c), (d), (f), and (h), values not significant at the 5% confidence level based on a two-sided $t$-test are shaded in gray. For panels (b), (e), (g), and (i), areas without shading indicate regions where the observed composite lies below the 5th percentile or above the 95th percentile of the bootstrapped ENSO composites. Red contours denote regions where the ERA-20C composite lies outside the range of all individual simulations.
• Figure 3: ENSO composites (El Niño $-$ La Niña) of DJF Z500 (gpm) for prescribed global SST (GOGA) experiments using (a) CESM2 and (b) DLESyM. Difference maps showing the contribution from freely evolving SSTs outside the pacemaker region, calculated as PACE minus GOGA, for (c) CESM2 and (d) DLESyM (CESM2 Climo). In panels (a) and (b), gray shading indicates regions where the composite is not significant at the 5% confidence level based on a two-sided t-test. In panels (c) and (d), gray shading indicates where the difference between the PACE and GOGA composites is not statistically significant at the 5% confidence level.
• Figure 4: The 5%–95% confidence intervals (CIs; gpm) of the Z500 ENSO composites for (a) the CESM2 Pacific Pacemaker (PACE) simulations, (c) DLESyM-PACE with CESM2 SST climatology, (d) ERA-20C, (f) DLESyM-PACE with DLESyM SST climatology, (g) the DLESyM atmospheric control experiment, and (i) the DLESyM free-running experiment. Panels (b), (e), (h), and (j) show their respective differences relative to ERA-20C. Gray shading indicates regions where the observed composite lies within the spread of values from the individual simulations. CIs are derived from 2000 bootstrapped samples for both the model simulations and ERA-20C.
• Figure 5: DLESyM Pacific Pacemaker experiments comparing two different SST climatologies: one using the CESM2 climatology and the other using the DLESyM climatology within the Pacific region outlined by the black contour. (a) Ensemble-mean SST difference during El Niño (EN) between the DLESyM and CESM2 climatologies; (b) same as (a) but for La Niña (LN); (c) ENSO composite (EN--LN) of SST using the DLESyM climatology; (d) ENSO composite (EN--LN) of SST using the CESM2 climatology; (e) difference of ENSO composites between (c) and (d); (f) difference in the DLESyM Z500 ENSO response between the DLESyM and CESM2 climatologies (corresponding to Fig. \ref{['fig:mean response']}c -- \ref{['fig:mean response']}f). Values not significant at the 5% confidence level based on a two-sided $t$-test are shaded in gray.