Abstract
We evaluate the impact of various algorithmic design choices on reconstruction skill and estimated air–sea CO2 flux using the Lamont–Doherty Earth Observatory fCO2-Residual machine learning data product. We reconstruct fCO2 globally over the period 1982–2023 by optimizing the hyperparameter selection strategy to specifically reduce reconstruction bias (ResidualOPT) and/or using ΔfCO2-Residual (subtracting fCO2atm from fCO2ocn) as a target variable (ΔResidualOPT). We compare these approaches to the original fCO2-Residual methodology, which uses fCO2-Residual as a target variable and selects hyperparameters based on reducing mean-squared error (ResidualORG). We find general agreement between the three approaches for the majority of the analysis period, but there are significant differences towards the beginning and end. Compared to the original methodology (ResidualORG), ΔResidualOPT and ResidualOPT show stronger ocean carbon uptake in the 1980s, with a global mean difference over 1982–1987 of 0.6 and 0.4 PgC yr−1, respectively. In 2023, the global mean air–sea CO2 flux for both these approaches is reduced by 0.4 PgC yr−1. The test error metrics show that the highest reconstruction skill is achieved for ΔResidualOPT. We present an additional analysis of reconstruction fidelity for the three approaches by using a testbed of CMIP6 Earth system models to reconstruct fCO2. The testbed analysis is in agreement with the observation-based test error metrics; both updated approaches improve reconstruction skill compared to the original methodology (ResidualORG), but ΔResidualOPT leads to the lowest reconstruction errors for the full analysis period (1982–2023). We conclude that, for the fCO2-Residual method, the impacts of algorithmic design choices (i.e., target variable or hyperparameter selection strategy) on reconstruction skill and estimated air–sea CO2 fluxes are significant. This indicates that it is possible to improve surface ocean fCO2 reconstruction fidelity even without the availability of additional observations.
Funding Information
  • National Oceanic and Atmospheric Administration (NA20OAR4310340)
  • European Space Agency
  • Simons Foundation (12871)
  • Schmidt Sciences
  • NSF (#2019625)

This publication has 38 references indexed in Scilit: