The oceans mitigate climate change by absorbing approximately 25% of anthropogenic carbon emissions. Decadal variability in the ocean carbon sink, such as a weakening in the 1990s and a strengthening in the 2000s, has been suggested by pCO2-based reconstructions, but its causes remain poorly understood. This variability is also not well represented in climate models, raising concerns about our ability to accurately project future changes. To address potential biases from sparse observational data, machine learning methods have been applied to surface pCO2 and interior dissolved inorganic carbon (DIC), but global reconstructions of full-depth DIC remain lacking. We aim to determine whether ocean carbon sink variability is real and to understand the role of interior DIC inventory changes in the carbon budget. Using neural networks trained on GLODAPv2.2023 observations and predictors like atmospheric CO2, location, temperature, and salinity from EN4 analysis, we reconstruct full-depth global DIC distributions from the 1990s to the 2010s using a residual neural network (ResNet). Validation through prediction of independent datasets show an improvement over previous products. Validation with the ECCO-Darwin dataset results in an average RMSE of 15.1 µmol/kg and bias of -0.3 µmol/kg. The global average uncertainty is 16.85 µmol/kg. The global change in the DIC inventory exhibits pronounced peaks in decadal variability, especially in the early 2000s driven primarily by intermediate waters at depths of 300-1200 m, particularly in the Atlantic, Indian, and Southern Oceans, and to a lesser extent in the Pacific. The accumulation rate of DIC increases steadily from the mid-2000s.

The Antarctic margin is a critically under-observed region despite its importance to the global climate. Here, in-situ ocean observations are difficult to obtain and clustered in easier-to-access regions. In addition, autonomous salinity measurements have to be corrected for drift or bias after collection. In this work, we introduce a new method that uses neural networks to identify and correct errors in ocean observations. Salinity estimates from a neural network trained on ship-based data are evaluated against Argo and seal measurements around Antarctica. We find that Argo salinity observations lie within the bounds of the ship-based data uncertainty, validating existing quality control processes for Argo. However, salinity data from seal-mounted sensors has a salty bias of up to 0.13 g/kg below 250m, which peaks at 5 months since sensor deployment. Our results showcase a new, flexible and computationally efficient way to assess in-situ ocean data, with potential for global implementation.