This arXiv preprint (not yet peer-reviewed) demonstrates that replacing conventional eddy parameterization with a data-driven machine learning approach can substantially alter the equilibrium climate state in an idealized, fully coupled CESM simulation. The study implemented the Zanna and Bolton (2020) scheme, which uses offline-trained models on high-resolution data to represent unresolved mesoscale eddy effects on momentum, inside an eddy-permitting MOM6 ocean component (roughly 1/4° resolution) coupled to a simplified atmosphere and sea ice.

Methodology involved running long-term control and parameterized simulations until statistical equilibrium was reached in the idealized geometry featuring a zonally reentrant Southern Ocean channel. No large ensemble was used; results come from single long integrations. The authors performed a regional heat budget analysis to separate the roles of ocean heat transport versus surface fluxes. Key limitations include the highly idealized domain (no realistic continents or bathymetry), potential overfitting of the data-driven scheme to its training data, and lack of assessment in fully realistic global configurations or varying climate states.

The ZB20 parameterization produces upgradient (backscatter) momentum fluxes that energize rather than dissipate eddies. This strengthens mesoscale activity, boosts poleward ocean heat transport, and intensifies the meridional overturning circulation near 60°S. The ocean response creates a clear dipolar surface temperature pattern: cooling at mid-latitudes and warming at high latitudes, driven mainly by anomalous meridional heat transport. The atmosphere compensates by reducing its own poleward heat transport and shifting the mid-latitude jet equatorward in response to altered meridional temperature gradients.

Original coverage of this work (and the abstract itself) underplays the broader pattern this fits into. Traditional Gent-McWilliams-style parameterizations usually diffuse potential vorticity downgradient and often produce overly sluggish Southern Ocean circulation and heat uptake biases seen across CMIP models. By contrast, the data-driven scheme captures inverse energy cascades and backscatter that high-resolution models exhibit, as first quantified in Zanna and Bolton (2020, Journal of Advances in Modeling Earth Systems).

Synthesizing this with related work, a 2022 study by Bachman et al. in Ocean Modelling showed that adding stochastic backscatter to MOM6 improved ACC transport and reduced temperature biases in realistic domains. Similarly, a 2023 review by Schneider et al. in Nature Reviews Earth Environment on machine learning for Earth system modeling highlights how data-driven closures for subgrid processes (from convection to eddies) are emerging as a systematic way to reduce long-standing structural uncertainty. What the current preprint misses is explicit discussion of how these circulation changes would affect carbon uptake, sea-level projections, and transient climate response in a warming scenario.

The deeper significance is that AI is no longer just accelerating simulations but actively changing the simulated climate itself. By learning directly from high-resolution truth, these methods can break decades-old parameterization compromises that have hampered confidence in regional projections. However, care is needed: if the ML model is trained only on present-day statistics, it may fail under strong climate change. This work therefore marks an important early benchmark for next-generation hybrid physics-ML climate models, showing that better eddy representation alone can flip hemispheric asymmetry and jet positions, effects large enough to matter for policy-relevant questions.