Award: OCE-1818501

To date, the ocean has absorbed about 40% of all fossil CO2 emissions, and this sink is critical to limiting the future accumulation of carbon in the atmosphere. In the portion of the work completed at Columbia University and the Lamont Doherty Earth Observatory, we have used simulations and observations together to reduce uncertainties in ocean uptake of anthropogenic carbon. Both of these approaches to understanding the ocean carbon sink have strengths and weaknesses, and our goal is to take advantage of the complementary strengths to the greatest extent possible. We have used Earth System Models to reduce mechanistic uncertainty in projections of the ocean carbon sink by carefully studying how the North Atlantic can sustain its intense carbon uptake. We find that waters flowing north at about 400m below the Gulf Stream have a very low anthropogenic carbon content. When these waters are mixed to the surface in the subpolar gyre to the south of Greenland, they can absorb carbon. In the future, as the Arctic melts, the exposure of these waters will be reduced and uptake will decline. Observations are required to understand how the ocean carbon sink is operating at present, and high-quality observation-based products are critical for ground-truthing models. For the ocean carbon sink, surface ocean partial pressure of CO2 (pCO2) observations are required, but these data are very sparse. The largest international database has pCO2 observations that cover fewer than 2% of the global ocean since the 1980s. Thanks to the advent of data science, these data can be processed through a neural network to fill in the gaps in pCO2 so that the remaining 98% of the ocean can be covered. This is a major advance, but the work is incomplete without understanding the uncertainty in the resulting observation-based product. In this project, we take advantage of the realistic representation of ocean carbon processes in the Earth System Models to evaluate uncertainties in a neural network gap-filling approach. We assess how well the neural network performs by subsampling modeled pCO2 in the same pattern as the real-world observations, re-running the neural network, and then statistically evaluating the result against the full-field original model. With this approach, we find that the neural network has high fidelity for the mean and for the seasonal cycle, but that data sparsity creates significant uncertainty for the amplitude of decadal variations. Finally, we combine a suite of ocean models and observation-based products with theoretical considerations to assess what drove the ocean carbon sink to slow down in the 1990s and then to recover after 2000. Agreement is strong between all approaches (see Figure). The fact that the ocean sink in the 1990s was barely larger than in the 1980s can be explained only with the slowdown in the atmospheric pCO2 growth rate in the early 1990s. The eruption of Mt Pinatubo in 1991 modified the temporal evolution of the sink in the 1990s, but did not impact its total magnitude. This finding implies that as fossil fuel emissions are lessened and this reduces the atmospheric pCO2 growth rate, the ocean carbon sink will immediately slow down. Last Modified: 04/15/2020 Submitted by: Galen A Mckinley