Abstract:

The meridional overturning circulation (MOC) plays a crucial role in the global distribution ofheat, carbon, and other climate‐relevant tracers. Monitoring the evolution of MOC is essential for understandingclimate variability, yet direct MOC observations are sparse and geographically limited. Although satellitemeasurements have shown potential for short‐term monitoring of the MOC, it remains unclear whether MOCvariability on decadal and longer timescales can be detected remotely. In this study, we leverage machinelearning to reconstruct long‐term MOC variability from satellite‐measurable quantities, using climatesimulations under pre‐industrial conditions. We demonstrate that our proposed non‐local dual‐branch neuralnetwork (DBNN) effectively reconstructs both the strength and vertical structure of the Atlantic MOC (AMOC)and the Southern Ocean MOCs across sub‐annual to multi‐decadal timescales. Using a neural networkinterpretation technique, we identify ocean bottom pressure near the western boundary and along dense‐waterexport pathways as the dominant input features for MOC reconstruction. This indicates that DBNN's predictionscan be interpreted as an approximation of geostrophic balance. The DBNN also effectively reconstructs theAMOC in the equatorial region, where geostrophy breaks down. This success is attributed to the capability ofDBNN in utilizing latitudinally non‐local ocean bottom pressure information and the meridional coherence ofAMOC variability. Additionally, the DBNN accurately reconstructs Southern Ocean MOCs using only seasurface height and zonal wind stress as inputs, thereby avoiding reliance on ocean bottom pressure, which issubject to considerable measurement uncertainty in practice. This work demonstrates the possibility ofcontinuous, long‐term MOC monitoring using satellite measurements.

Publisher's Version