Reducing Model Biases with Machine Learning Corrections Derived from Ocean Data Assimilation Increments

Large ocean Data Assimilation (DA) increments in the mixed layer reflect systematic biases in vertical mixing. These increments can be leveraged to derive bias corrections. We develop a machine learning (ML) model trained on DA outputs from the GFDL SPEAR system to predict temperature corrections addressing vertical mixing biases. The trained ML model is implemented online in MOM6 to produce DA simulations and free runs. Compared with the existing Ocean Tendency Adjustment (OTA) scheme, which applies climatological corrections, ML more effectively reduces analysis increments in DA runs. Combining ML with OTA further improves DA performance by addressing residual biases. In 10-year free runs, ML outperforms OTA, reducing overly deep mixed layer depth biases by capturing restratification. When combined, ML and OTA further improve the long-term mean sea surface temperature. This workflow is generalizable to any climate model with an ocean DA system, offering a promising pathway to reduce model biases.