Hybrid Graph Learning Reconstructs Global Ocean Oxygen Spatiotemporal Changes

Climate change and anthropogenic activities have exacerbated hypoxic conditions in the global ocean, threatening marine ecosystems and earth sustainability. Investigating the history of ocean deoxygenation is crucial for developing coping strategies. However, due to the severe sparse in-situ observations and spatiotemporal complexity of deoxygenation mechanism, precise modelling of the oxygen trends, especially in large scale, remains a long-standing problem. Here we propose Jingwei, a hybrid graph learning framework that integrates observation data and ocean climate model in a complementary manner, accurately reconstructing global oxygen levels over the past six decades. A knowledge retrieval module is designed to learn the hidden laws that dominate the complex mechanism of oxygen changes from model-based simulations. Constrained by those laws, a knowledge propagation module spatiotemporally connects the sparse data, capturing dynamic and spatial heterogeneous oxygen loss. With a joint consideration of invariant laws and variant factors, Jingwei significantly reduces reconstruction error, achieving equal accuracy with the dense cruise surveys. It can quantitatively trace the expansion process of oxygen minimum zones (OMZ), revealing potential drivers, effectively supporting on-target strategies. We foresee that the implications of such a graph-based hybrid framework extend beyond ocean deoxygenation and pave way for next-generation earth system modelling.