The authors propose to use artificial neural networks to correct air-sea heat fluxes in the NEMO ocean model. This is a novel approach compared to traditional bias correction techniques or climatology-based methods. It involves developing a machine learning-based state-dependent correction approach. Learning the corrections from nudging SST increments allows for online inference without relying on direct observational data in operational forecast settings or in historical periods with scarce data.

The use of artificial neural networks (ANNs) to model nonlinear relationships between atmospheric and oceanic state predictors supported by stationary predictors and heat flux errors appears remarkably effective. It would be of interest to include in the manuscript information on the performance of the neural network in predicting the nudging increments using standard metrics like R2 or normalised mean square error or error maps. It appears that the network is able to learn a substantial part of the corrections. Figure 4 provides a hint.

The method has been demonstrated to be effective in the ocean only model. Non-linear feedbacks in the coupled models could affect the effectiveness of the ANN-based corrections. Do you foresee the method to work well in the coupled models? Perhaps applicability in the context of coupled modelling/coupled reanalysis could be discussed.

Have you tested other ANN architectures like CNN? While CNN may be more difficult to apply online, it would be interesting to assess if it would be superior to the simple column model?

Please see attached pdf