Identifying Key Predictors of Ocean Oxygen Variability Using Artificial Intelligence

Global warming is expected to alter the ocean's dissolved oxygen (O₂) distribution, but quantifying these changes remains challenging due to the sparse and irregular nature of in situ measurements. Machine learning approaches offer new opportunities to reconstruct O₂ profiles from routinely measured ocean parameters and to identify the dominant physical and biogeochemical drivers. In this study, we evaluate Artificial Neural Networks (ANNs) for reconstructing vertical O₂ distributions using a 300-year simulation from the Earth System Model UVic2.8. The networks were trained on 18 predictors spanning physical, chemical, and biological variables. To interpret predictor influence, we applied Hierarchical Agglomerative Clustering (HAC) to the ANN weight matrix, enabling identification of functional hierarchies and correlations among predictors. This framework was then used to reduce the predictor set and quantify the uncertainties associated with O₂ reconstructions. Results show that ANNs reproduce vertical O₂ profiles with high fidelity and consistent performance across multiple realizations. The ANN-HAC approach robustly identifies the most influential predictors, while also capturing regional differences driven by local physical circulation and biogeochemical processes. The reduced-predictor configurations retain strong reconstruction skill, underscoring the potential for efficient application to observational datasets. Our findings demonstrate the feasibility of combining ANN and HAC techniques for reconstructing dissolved O₂ distributions and for extracting mechanistic insight from high-dimensional predictor spaces. This methodology provides a promising pathway to bridge observational gaps, evaluate Earth System Models, and improve projections of oxygen variability and deoxygenation under climate change.