A new study in Biogeosciences shows that neural networks can produce better ocean carbon model calibrations than traditional data assimilation when the physical forcing data is noisy. The ocean absorbs about 2.8 gigatonnes of carbon per year, and the biological carbon pump depends on plankton dynamics that are hard to simulate. The researchers built a 1D testbed with known answers and found that neural schemes kept bloom timing errors within roughly 10 days, while the variational method could be off by 28 days. Parameter estimates also became more consistent under realistic forcing uncertainty. The strongest result came from a hybrid approach. Feeding the neural network's corrected states and parameters into the variational scheme halved reconstruction errors for phytoplankton and zooplankton, and the ensemble uncertainty ranges became far more reliable. The tradeoff is that the network must be retrained for each observing configuration, and scaling to 3D global models remains computationally expensive. Still, the study gives concrete evidence that machine learning can improve ocean carbon accounting, especially where sampling is sparse and physical reanalyses are imperfect.
