The Machine learning (ML) revolution is becoming established in oceanographic research, but its applications to emulate marine biogeochemical models are still rare. We pioneer a novel application of machine learning to emulate a highly complex physical-biogeochemical model to predict marine oxygen in the shelf-sea environment. The emulators are developed with intention of supporting future digital twins for two key stakeholder applications: (i) prediction of hypoxia for aquaculture and fisheries, (ii) extrapolation of oxygen from marine observations. We identify the key drivers behind oxygen concentrations and determine the constrains on observational data for a skilled prediction of marine oxygen across the whole water column. Through this we demonstrate that ML models can be very useful in informing observation measurement arrays. We compare the performance of multiple different ML models, discuss the benefits of the used approaches and identify outstanding issues, such as limitations imposed by the spatio-temporal resolution of the training/validation data.