The rising atmospheric CO2 concentrations, primarily due to fossil fuel burning and land-use changes, necessitate understanding the ocean's role as a carbon sink. The Southern Ocean, known for its significant carbon uptake, shows decadal variability in this uptake. While surface ocean CO2 measurements indicate a strengthening of the carbon sink in the 2000s, the changes in carbon storage below the surface remain poorly understood due to sparse in-situ measurements and the complexity of modeling biogeochemical cycles. This research addresses this challenge by developing a deep-learning model to estimate DIC concentrations in the Southern Ocean's interior using readily available surface data.

Concerns have been raised about a declining trend in Southern Ocean carbon uptake from the 1980s to the early 2000s. However, recent studies using surface ocean CO2 measurements have shown a reversal of this trend, suggesting an increase in uptake during the 2000s, linked to changes in ocean circulation driven by shifts in wind forcing. The export of carbon from the surface to the ocean interior is crucial for long-term carbon storage and impacts marine chemistry, leading to ocean acidification. Existing models struggle to accurately capture these biogeochemical processes due to the sparse spatial and temporal distribution of oceanic measurements.

To overcome the limitations of sparse observations, a deep-learning model was developed to predict DIC concentrations in the upper 4 km of the Southern Ocean. The model employs readily available surface and near-surface variables: sea surface temperature, surface flow velocity, sea surface height, near-surface wind velocity, and surface CO2 partial pressure (pCO2). The model's training was conducted in two phases. Phase 1 used data from the Biogeochemical Southern Ocean State Estimate (B-SOSE), a high-resolution data-assimilating ocean circulation model, to provide a large volume of data, especially for deeper layers. Phase 2 incorporated DIC measurements from Global Ocean Data Analysis Project version 2 (GLODAPv2) shipboard measurements and Southern Ocean Carbon and Climate Observations and Modeling (SOCCOM) biogeochemical Argo floats to correct for biases in the B-SOSE model. The model architecture was adapted from a U-net model, incorporating convolutional and recurrent neural networks to capture spatial and temporal dynamics. A total of 22 U-net models were trained to cover the 48 vertical levels from the surface to 4 km depth. The model was trained to capture the relationship between surface predictors and DIC fields at different depths, and combined both shipboard and float measurements for improved accuracy. The model was then used to compute 5-day-averaged DIC concentrations over the period 1993-2019 south of 30°S.

The deep-learning model revealed decadal trends in DIC concentrations in the Southern Ocean's three major basins (Atlantic, Pacific, and Indian) between 1993 and 2019. Near-surface DIC concentrations increased poleward, largely following neutral density surfaces in the interior. The Pacific and Indian basins showed higher DIC concentrations than the Atlantic basin due to their older, bottom-sourced waters. From 1993 to 2009, DIC concentrations decreased in the ocean interior, especially in the Pacific sector. This decrease in surface DIC, lowering pCO2, is consistent with the previously observed strengthening of the Southern Ocean carbon sink in the 2000s. However, the trends were not zonally uniform, indicating distinct mechanisms in different basins. Since the 2010s, DIC trends reversed, showing increases, particularly near the surface. During the 1990s, DIC mostly increased in the upper 1 km over the Pacific within the Antarctic Circumpolar region (50–60°S), likely due to stronger Westerlies intensifying upwelling of DIC-rich waters. Zonal differences were observed, potentially due to variations in atmospheric forcing. In the 2000s and 2010s, stronger Westerlies increased northward Ekman transport, bringing colder, fresher water, and impacting water-mass transformations. Increased melting of advected ice caused buoyancy gain in SAMW, while brine rejection led to buoyancy loss in CDW, with significant zonal differences. These transformations help explain DIC trends, with weakening CDW upwelling decreasing DIC delivery to the surface in the 2000s. The 2010s saw a reversal in near-surface DIC trends, turning negative in the Pacific, while DIC built up below 1 km. The Pacific sector exhibited increased near-surface DIC in the 2010s, potentially due to increased contribution from cooling of subtropical low-DIC waters, rather than freshening of high-DIC CDW waters. The Atlantic and Indian sectors showed warming and heat storage, stabilizing the water column and weakening wind-driven upwelling. In the Atlantic, decreasing DIC trends along upwelling density isosurfaces were observed in the 1990s and 2000s, and continued in the 2010s subsurface. The Indian sector displayed similar negative trends south of 50°S but positive trends near the surface to the north. The weakening of the Atlantic Meridional Overturning Circulation (AMOC) since the 1990s is also consistent with our findings of progressively decreasing trends along upwelling density isosurfaces.

The study's findings show long-term changes in Southern Ocean DIC concentrations and carbon uptake. Weakened upwelling and connectivity between deep and surface waters, potentially hindering carbon export, were observed across different sectors. The varying underlying mechanisms highlight non-uniform responses to future circulation changes. While the model doesn't separate anthropogenic carbon uptake from natural variability, the near-surface DIC decrease suggests enhanced atmospheric carbon uptake, followed by increased near-surface DIC and potentially weakened export to the interior. The results align with previous studies showing decreased CO2 uptake in the 1990s and increasing uptake in the 2000s, indicating a continuation of increasing uptake potential into the 2010s. The model's high spatiotemporal resolution reveals spatial patterns in temporal trends, offering advantages over previous aggregate decadal averages.

The study demonstrates long-term changes in Southern Ocean DIC and carbon uptake, highlighting the impact of weakened upwelling and connectivity between ocean layers. The model provides a valuable tool for future monitoring of the ocean carbon sink, enabling estimation of DIC concentrations in the ocean interior using new satellite data. Continued monitoring is needed to assess the long-term impacts of DIC accumulation on anthropogenic CO2 storage. The zonal differences in underlying mechanisms suggest that future responses to circulation changes may also be non-uniform. Further research should integrate methods from prior studies to separate the effects of anthropogenic carbon uptake and natural variability.

The model's accuracy depends on the quality and availability of input data. Seasonal variations in biological activity, not fully captured by the model, may affect the accuracy of DIC estimates, particularly in areas with limited wintertime data. The model does not explicitly account for all biogeochemical processes that influence DIC concentrations (e.g., sinking rates, organic matter remineralization, total alkalinity, calcification). Additionally, separating changes in DIC due to anthropogenic carbon uptake and natural circulation variability requires further investigation.