Abstract The inter-annual variability of global ocean air-sea CO 2 fluxes are non-negligible, modulates the global warming signal, and yet it is poorly represented in Earth System Models (ESMs). ESMs are highly sophisticated and computationally demanding, making it challenging to perform dedicated experiments to investigate the key drivers of the CO 2 flux variability across spatial and temporal scales. Machine learning methods can objectively and systematically explore large datasets, ensuring physically meaningful results. Here, we show that a kernel ridge regression can reconstruct the present and future CO 2 flux variability in five ESMs. Surface concentration of dissolved inorganic carbon (DIC) and alkalinity emerge as the critical drivers, but the former is projected to play a lesser role in the future due to decreasing vertical gradient. Our results demonstrate a new approach to efficiently interpret the massive datasets produced by ESMs, and offer guidance into future model development to better constrain the CO 2 flux.