The spatiotemporal evolution of the global interior ocean’s anthropogenic carbon sink: reconstructed through machine learning

The oceans mitigate climate change by absorbing roughly 25% of the anthropogenic carbon that is released. Past reconstructions of air-sea CO₂ flux based on surface pCO₂ observations have indicated that this carbon sink exhibits decadal variability, appearing to weaken during the 1990s and strengthen in the 2000s. However, the causes of this variability are unclear, and it is poorly represented in climate models and the future climate projections they generate. It also remains uncertain whether the estimated variability is a product of bias due to the limited availability of biogeochemical observations. To address the challenge posed by sparse data, machine learning techniques have been applied to surface pCO₂ as well as interior dissolved inorganic carbon (DIC). However, reconstructions of DIC and anthropogenic carbon for the full depth of the global ocean have not yet been achieved.

Our objective is to determine whether the variability in the ocean carbon sink is real and to understand changes in the interior carbon inventory as part of the carbon budget. To this end, we use neural networks to predict the spatiotemporal distributions of full-depth DIC and C* from the 1990s to the 2010s. C* is a quasi-conservative tracer that corrects DIC for biological activity by applying Redfield stoichiometric ratios. ΔC*, the difference in C* between two time points, has been used as a proxy for added anthropogenic carbon.

The neural network is trained on observations from the GLODAPv2.2023 database. We make predictions of DIC and additional C* components - total alkalinity, oxygen, and nitrate - based on the location, depth, temperature, and salinity from the EN4 reanalysis product and atmospheric CO₂. Here, we present findings on the spatiotemporal evolution of full-depth interior carbon in the global ocean, providing a quantification of the anthropogenic carbon sink and its variability over time. The interior carbon inventory changes are then compared with current air-sea CO₂ flux products. In further work, the results are being combined with a water mass based inverse method to investigate the drivers of variability.