Ocean models often struggle to represent how water mixes vertically, leading to persistent temperature and circulation biases. In this preprint , Danni Du and colleagues use machine learning (ML) to correct those biases by learning directly from data assimilation outputs in NOAA’s GFDL SPEAR system. When integrated into the ocean model, the ML corrections improved temperature and mixing accuracy, outperforming existing correction methods. Combining ML with traditional approaches produced even better results, leading to more realistic sea surface temperatures and ocean structure. This approach can be applied to other climate models, offering a powerful new way to make ocean simulations more accurate.