fCO2-Residual
Beginning with v2026, we now release both a fCO2-Residual and a newly developed method, ΔfCO2-Residual.
Reconstruction products begin in 1982 at monthly, 1x1 degree, near-global coverage. End year is updated to most recent SOCAT release each summer (v2026 extends through 2025).
Description
The common approach for estimating full-coverage fCO2 is to train a machine learning algorithm on sparse in situ fCO2 data and associated physical and biogeochemical observations. While these associated variables have understood relationships to fCO2, it is often unclear how they mechanistically drive fCO2 around the world. The LDEO fCO2-Residual method takes the basic approach and adds advances that enhance connections between physical understanding and reconstructed fCO2. The novel approach used here includes applying pre-processing to the fCO2 data to remove the direct effect of temperature, a relationship well-documented in literature and lab experiments (Takahashi et al. 2002). This enhances the biogeochemical/physical component of fCO2 in the target variable (now fCO2-Residual) and reduces the complexity that the machine learning must disentangle. The resulting algorithm has physically understandable connections between input data and the output biogeochemical/physical component of fCO2 (Bennington et al. 2022).
New in v2026, the ΔfCO2-Residual method! The LDEO carbon group is constantly working to improve our reconstruction methods. Heimdal et al. 2026 evaluates the impact of various algorithmic design choices on reconstruction skill of surface ocean fCO2 using the fCO2-Residual method. Based on these results, in v2026, we have updated the method to include optimizing the hyperparameter selection strategy to specifically reduce reconstruction bias. Additionally, we now present an alternative product, called the ΔfCO2-Residual approach, which selects ΔfCO2 as the target variable (fCO2ocn - fCO2atm). We find general agreement between the original method and the two v2026 updates for the majority of the analysis period, but there are significant differences towards the beginning and end of the reconstruction time period. We encourage users to refer to Heimdal et al. 2026 for further details.
Beginning in v2026, we present both options for users however we recommend utilizing the ΔfCO2-Residual approach.
References
Bennington, V., Galjanic, T., & McKinley, G. A. (2022). Explicit physical knowledge in machine learning for ocean carbon flux reconstruction: The pCO2-Residual method. Journal of Advances in Modeling Earth Systems, 14, e2021MS002960. https://doi. org/10.1029/2021MS002960
Heimdal, T.H., Fay, A.R., Shaum, A.P., Acquaviva, V., Bennington, V., Sharples, A.E., Joenson, N.M. and McKinley, G.A., 2026. An update of the LDEO fCO2-residual data product: algorithmic choices improve ocean carbon sink estimates. Machine Learning: Earth, 2(2), p.025005. doi.org/10.1088/3049-4753/ae86f8
Takahashi, T., Sutherland, S. C., Sweeney, C., Poisson, A., Metzl, N., Tilbrook, B., et al. (2002). Global sea–air CO2 flux based on climatological surface ocean pCO2, and seasonal biological and temperature effects. Deep Sea Research Part II: Topical Studies in Oceanography, 49(9), 1601–1622. https://doi.org/10.1016/S0967-0645(02)00003-6