Improved Quantification of Ocean Carbon Uptake by Using Machine Learning to Merge Global Models and pCO2 Data

Abstract

The ocean plays a critical role in modulating climate change by sequestering CO2 from the atmosphere. Quantifying the CO2 flux across the air-sea interface requires time-dependent maps of surface ocean partial pressure of CO2 (pCO2), which can be estimated using global ocean biogeochemical models (GOBMs) and observational-based data products. GOBMs are internally consistent, mechanistic representations of the ocean circulation and carbon cycle, and have long been the standard for making spatio-temporally resolved estimates of air-sea CO2 fluxes. However, there are concerns about the fidelity of GOBM flux estimates. Observation-based products have the strength of being data-based, but the underlying data are sparse and require significant extrapolation to create global full-coverage flux estimates. The Lamont Doherty Earth Observatory-Hybrid Physics Data (LDEO-HPD) pCO2 product is a new approach to estimating the temporal evolution of surface ocean pCO2 and air-sea CO2 exchange. LDEO-HPD uses machine learning to merge high-quality observations with state-of-the-art GOBMs. We train an eXtreme Gradient Boosting (XGB) algorithm to learn a non-linear relationship between model-data mismatch and observed predictors. GOBM fields are then corrected with the predicted model-data misfit to estimate real-world pCO2 for 1982–2018. The resulting reconstruction by LDEO-HPD is in better agreement with independent pCO2 observations than other currently available observation-based products. Within uncertainties, LDEO-HPD global ocean uptake of CO2 agrees with other products and the Global Carbon Budget 2020.