Improving the near-term estimate of air-sea CO2 exchange by merging models and observations with machine learning

The ocean plays an important role sequestering CO₂ from the atmosphere, which reduces the effects of climate change now and in the future. Quantifying the CO₂ flux across the air-sea interface requires time-dependent maps of surface ocean partial pressure of CO₂ (pCO₂). To assess the ocean uptake, output from general circulation models and observational-based data products provide estimates of the temporal evolution of global pCO₂ and the air-sea CO₂ exchange. However, model-based estimates of pCO₂ are biased high compared to in situ ship-based observations and there is disagreement in the variability between these two approaches on decadal time scales, particularly in the Southern Ocean. Here, we merge model output and observations using supervised machine learning techniques to provide the best estimates of temporal evolution of surface ocean pCO₂ and the air-sea CO₂ exchange. We train a neural-network to learn a relationship between the model-data mismatch using satellite observed ocean/atmosphere variables as predictors. Once trained, evaluated, and tested, we create monthly maps of the model-data mismatch at all spatial locations. We rectify the original model output by adding the learned mismatch to the original output. The final product, which is a combination of observations and models, is in better agreement with independent in situ observations than the original models. When compared to pCO₂ data withheld from training, the rectified model has a bias of only 1 µatm while the original model has bias of 11 µatm. Our final flux estimate takes advantage of recent innovations in data science and decades of high-quality observations, and additionally benefits from the strengths of modern ocean models.