Search for: All records

In order to fully understand current and future climate impacts from rising carbon emissions, it is crucial to accurately quantify the air-sea CO2 flux and the ocean carbon sink in space and time. Air-sea flux estimates from observation-based data products used in the Global Carbon Budget show a large spread, and suggest a stronger carbon sink than global ocean biogeochemistry models (GOBMs) in the last decade. Output from GOBMs and Earth system models (ESMs) can be used as ‘testbeds’ to better understand current estimates of ocean carbon uptake in time and space through sub-sampling experiments. Recent testbed studies show improvement in reconstruction skill with increasing observational coverage, but the direction (over- vs. underestimation) and magnitude of bias for ocean carbon uptake vary significantly. Here, we use a collection of CMIP6 ESMs as a testbed to better understand the causes of the spread of sink estimates in observation-based products. Specifically, we assess how the choice of hyperparameters for the machine learning algorithm and the testbed structure impact reconstruction skill of surface ocean pCO2 (spCO2) using the pCO2-Residual method. We find that, when negative mean squared error (nMSE) is used as error metric during hyperparameter optimization, the reconstruction significantly underestimates spCO2 over 2017-2022, irrespective of which CMIP6 ESM is used as a testbed; this results in an overestimation of the global ocean sink, assessed through comparison to the ‘testbed truth’. If hyperparameters are selected based on bias as the error metric, this trend of increasingly negative bias is eliminated. When applied to real-world SOCAT data, this leads to a significantly weaker global ocean carbon sink in 2021-2022 (up to ~ 0.5 Pg C/yr), and less divergence from GOBM estimates. This suggests that the increasingly stronger sink showed by the pCO2-Residual method in recent years might not represent a real trend, but may be due to algorithmic design choices in the context of sparse and biased observational coverage.