Against the backdrop of the rapid development of marine big data and artificial intelligence， constructing global high-resolution predictive models for ocean physical fields has become feasible， helping to address the issues of slow computation and insufficient prediction accuracy of numerical models. To this end， this paper systematically reviews the fundamental theories and key technologies of large ocean models， including basic ocean theories and data characteristics， feature representation and data algorithms， as well as applications of large ocean models in ocean phenomenon forecasting， extreme event warning， and downstream industrial sectors. The results indicate that although the field of deep learning-based large ocean models is still in its early stages， it has already significantly improved the computational efficiency and prediction accuracy of ocean models， achieving progress in areas such as mesoscale ocean phenomena， early warning of ocean storms and tropical cyclones， fishery forecasting， and shipping route planning. However， challenges remain in task-specific modeling of ocean fields， short-range， medium and long-range ocean forecasting， as well as physical consistency and interpretability. The study suggests that future efforts in constructing and applying deep learning-based large ocean models should focus on using heterogeneous or complex data structures to fully represent ocean fields， incorporating physical constraints or mass-energy conservation equations to ensure physical consistency during evolution， and predicting interannual to decadal long-period ocean phenomena， thereby promoting the continued development of AI-based oceanography in fundamental marine research.