The Southern Ocean is central to the global circulation of the ocean: it is a location where waters that have been confined thousands of meters deep, sometimes for centuries to milennia, finally return to the sea surface.  Following interactions with the atmosphere and cryosphere, these waters then return to both intermediate (~1000m) and abyssal (~4000m) depths, and ultimately fill over half of the global subsurface ocean. This "meridional overturning circulation" (MOC), illustrated in the figure below, carries globally significant quantities of heat and sequesters around 40% of humanity's carbon emissions. Despite its global importance, no direct measurements of the Southern Ocean' MOC are currently available. This stands in contrast with the MOC in the North Atlantic, which now is monitored by arrays at two different latitudes. Instead, estimates of the MOC have been derived from "inverse models", which construct a best approximation to the MOC using spatially and temporally sparse measurements of the ocean's properties. The overarching aim of this project was to estimate past and ongoing changes in the Southern Ocean's MOC by quantifying its "signature" in measurements that can be made using NASA satellites: variations in the elevation of the sea surface surface, the temperature of the water at the sea surface, and the pressure at the sea floor (or equivalently the "weight" of the ocean).  

The global ocean overturning circulation, as diagnosed from NASA's ECCO state estimate version 4, release 3, averaged over the period 1992-2015. Colors indicate the strength of the streamfunction in units of Sverdrups (1 Sverdrup = 1,000,000 cubic meters per second), and each color gradiation corresponds to a transport of 1 Sverdrup. The flow follows the contours of the streamfunction in the direction indicated by the black arrows. Dark gray lines correspond to average depths of selected potential density surfaces, and approximately indicate the upper/lower bounds and core of each overturning "cell". Light gray areas correspond to the sea floor/land.

One of the project's key findings was that export of dense water across the Southern Ocean , a key component of the MOC, is largely controlled by zonal surface wind stress. Specifically, we analyzed NASA's ECCOV4r4 state estimate (a simulation of the ocean that is constrained to match available measurements as closely as possible). found that north/south transport  anomalies at the sea floor anomalies are (somewhat counter-intuitively) driven by fluctuations in surface winds. These transport anomalies are comparable to the multi-decadal mean transport and primarily occur at frequencies shorter than 2 years. The reason for this connection between winds and deep ocean transports can be understood in terms of the need to balance the east-west forces exerted on the ocean by the winds versus pressure at the sea floor. A key implication of our findings is that Southern Ocean MOC variability can be almost entirely reconstructed from surface winds alone. This represents a significant step toward our central project goal of indirectly inferring AABW export from satellite-measureable quantities. 

Building on this effort, we sought to identify more general "fingerprints" of MOC variability in  satellite-measurable properties of the ocean like sea surface elevation and ocean bottom pressure, but found that these "fingerprints" were difficult to separate from other forms of ocean variability. We therefore approached the problem from a different angle, applying Machine Learning techniques to indirectly infer MOC variability from satellite-measureable quantities, again using NASA's ECCO state estimate as a test bed. This approach turned out to be very successful: we showed that training a simple "Neural Network" (NN) on just ~10 years of historical satellite measurements and MOC variability, the NN could predict more than 90% of future MOC variability from future satellite measurements. Furthermore, this approach works not just in the Southern Ocean, but throughout all of the major ocean basins. The success of our machine learning technique has motivated us to pursue further development of this approach via another NASA ROSES project.

This work was supported by the National Aeronautics and Space Administration (NASA) ROSES Physical Oceanography program under grant number 80NSSC19K1192.