Global Gridded Air-sea Oxygen Flux Inferred from a Machine Learning-based Dissolved Oxygen Product

Abstract. We present estimates of the monthly open ocean air-sea O₂ flux from 2004 to 2024 on a 1°×1° grid spanning 64.5°S to 79.5°N. The flux is computed using 4-dimensional ocean dissolved oxygen (DO) fields derived from Argo float and shipboard observations via machine learning algorithms (GOBAI-O₂, Sharp et al., 2023). Flux uncertainties are quantified by generating a large ensemble of flux estimates, first propagating errors from the dissolved oxygen product, then computing fluxes across all combinations of three gas exchange parameterizations that account for bubble-mediated flux and three wind products. We apply DO adjustments to align our resolved global annual mean fluxes with those derived from scaling global ocean heat uptake and regional annual mean fluxes with those derived from ocean inversions. Our results show larger seasonal flux variations at high latitudes than at low latitudes, with clear differences between major ocean basins. Both adjusted and unadjusted annual mean flux estimates exhibit strong ocean O₂ sinks at high latitudes, weak sources in the low-to-mid latitude subtropics, and weak sinks near the Equator. The annual mean adjustment significantly enhances ocean O₂ uptake in the northern high latitudes and tropical regions while reducing outgassing in the northern subtropics. We evaluate our flux seasonal cycles and annual mean values by comparing forward transport simulations with atmospheric O₂ observations from global airborne surveys and surface sampling stations. The simulations reproduce the observed mean seasonal cycles well, but some differences remain in the annual mean latitudinal gradients. We analyze fractional variance contributions from DO, wind, and gas exchange scheme uncertainties and their interactions at regional and hemispheric scales for both climatological monthly and annual mean fluxes. This dataset marks a major improvement over existing air-sea O₂ flux products, as it includes the resolution of interannual variability in the flux seasonal cycle, the use of advanced machine learning-based DO fields that better represent complex spatial and temporal patterns, and robust uncertainty quantification across various scales.