Global ocean deoxygenation trends are largely undetectable in machine learning reconstructions due to climate predictor sensitivity

Machine learning (ML) models have become a widely used tool for reconstructing global ocean 15 dissolved oxygen (O2) concentrations. Recent studies, however, suggest a divergence in deoxygenation trends between process models and reconstructed O2 from observations, which is still unresolved. The use of a non-physical measure of temporal changes, such as year of observation or number of months, as a ML model predictor and its effect on reconstructed deoxygenation trends has not been systematically assessed. Here, we develop eight feed-forward neural networks corresponding to four different climate 20 predictor configurations at 0.25° and 1.0° monthly resolution and obtain gridded products of O2 concentrations from the surface to 1500 m depth between 1993 and 2022. We show that while all experiments perform comparably well against a held-out test set, climate predictor choice is the dominant source of methodological uncertainty in reconstructed deoxygenation trends. Including year as a predictor inflates deoxygenation trends by introducing a spurious temporal association with O2 concentration 25 independent of physical drivers, as evidenced by systematic drift in model bias at two ocean time-series stations. Using inter-experiment consistency as a measure of signal detectability, we show that robust detection of global ocean deoxygenation over 1993–2022 is currently limited to the surface ocean, while in the ocean interior methodological uncertainty renders robust detection of deoxygenation trends largely unachievable. Lastly, we recommend the use of a physical variable, such as temperature anomaly at 30 different water depths, instead of non-physical measures of temporal changes, in future machine learning