Abstract
<jats:p>Abstract. The ocean acts as a critical carbon sink, but its efficiency in absorbing anthropogenic CO2 varies significantly over multiyear to decadal timescales. Accurately predicting this variability is essential for anticipating atmospheric CO2 growth and establishing the unperturbed baselines necessary for verifying climate mitigation efforts, such as marine Carbon Dioxide Removal (mCDR) activities. This study introduces a fully coupled physical-biogeochemical prediction framework, which integrates the NOAA GFDL Seamless System for Prediction and EArth System Research (SPEAR) with the COBALTv3 ocean biogeochemical model. We conducted ensembles of uninitialized historical simulations, data-assimilative reconstructions, and retrospective initialized predictions of global air-sea CO2 flux, evaluating their skill against observation-constrained products for a recent 40-year period (1984–2023). The uninitialized ensemble skillfully predicts the amplitude of the observed historical increases, but struggles to resolve multiyear and decadal variability. We show that initialization significantly improves prediction skill. Globally, skill is enhanced for lead times up to two years, extending up to five years in specific higher-latitude regions. Through Average Predictability Time (APT) analysis, we isolated distinct physical drivers of the dominant predictability. We find that skillful predictions up to two years are primarily governed by the El Niño–Southern Oscillation (ENSO) and its impact on tropical upwelling. Moreover, we identified a multidecadal, potentially predictable signal linked to long-term changes in Eastern Boundary Current upwelling, as well as Southern Ocean mixed layer depth and sea surface temperature. However, verifying this long-term potential predictability remains fundamentally constrained by the sparsity of multidecadal observational records in these remote or nearshore regions. This underscores the critical need for sustained, optimized ocean observations and an improved understanding of the uncertainties associated with existing observational data.</jats:p>