Decadal variability in the Southern Ocean carbon sink: Reconciling the mismatch between hindcast models and observationally-based products

The ocean plays an important role in sequestering CO₂ from the atmosphere, therefore mitigating the effects of climate change now and in the future. Because the ocean carbon sink cannot be remotely sensed, indirect methods for estimating the sink’s strength must be used, these include output from ocean models and observationally-based products. These methods provide estimates of the temporal evolution of the global air-sea CO₂ exchange. However, the observationally-based products tend to suggest greater decadal variability than do the models, particularly in the Southern Ocean. Here, we present a testbed using the large ensembles of four coupled Earth System models to statistically evaluate how well a leading neural-network based data product can reconstruct CO₂ fluxes across timescales and climate states. Our results suggest the leading neural-network based method performs well in large regions of the global ocean, with high correlations (>~0.9) across timescales in the subtropics. However, this approach overestimates decadal variability in the Southern Ocean by more than 40%. We provide evidence that additional sampling in the Southern Ocean would result in a reduced decadal variability by the neural-network, which would be a more accurate estimate. If it is a common feature of the observationally-based products that they overestimate Southern Ocean decadal variability due to spatiotemporal gaps in observations, this would help to reconcile the existing mismatch between these products and ocean models.