Through adopting a rolling  strategy, an artificial intelligence model is established by using a U-Net architecture, and global SWH prediction is produced by the model after trained using ERA5 data.

Specific comments: 

1. I just wonder if  the RMSEs in Figure 2c and Figure 4d are smaller than that of climatology.

2. For the sake of comparison, the colorbar scales in  Figures 6-7 should be the same as those used in Figures 4-5.

3. The units of variables on the vertical axis are missing in (b) and (c) in Figures 2-3.

Major comments:

1. It is not clear what goal this AI SWH model would like to achieve. Is it to make good 0-10 days SWH predictions similar to traditional numerical wave models? If yes, we should use predicted wind fields from GFS (or IFS) as the forcing instead of the ERA5 reanalysis wind fields. In the current manuscript, one year of SWH series was generated by the AI SWH model. But it essentially functions like post-processing the ERA5 wind analysis to 0-6h short tem SWH forecasts. This is not the regular 0-10-day SWH forecasts we normally expect.

2. Section 2.4 and other relevant parts: (1) data assimilation will improve initial conditions, but it is not related to the establishment of this AI SWH model and it is suggested to remove this part from this manuscript; (2) the data assimilation method here is too simple. Nowadays, we would expect at least a 2DVAR method that considers the uncertainties of the background and the observations.

Comments: 

Lines 48-49: "However, …" The statement is not accurate. The SWH prediction is both a forcing and an initial value problem. When focusing on the short-term (a few hours to a day) forecasts, the initial value will have a large impact. Zhou et al., 2021 and Ouyang et al., 2023 targetted 24 h and 3 days forecasts respectively. 

Line 124-125: "Particularly, if a long series…" Could you clarify which part of Song and Jiang, 2023 drew this conclusion?

Lines 133-134: "We believe…": Evidence, instead of a subjective "belief", is expected here to demonstrate why U-Net is suited for this work.

Lines 135-136: "The processes of both…": In AI for NWP practices, lots of literature has demonstrated that AI struggles to resolve different scales at the same time (for example, smoothing to get better medium-range forecasts and hence not able to resolve smaller scales features). So how can it be assured that "The processes of both local wave generation by wind and wave propagation in space can be captured by convolutional kernels at different scales“? More discussions on this are needed.

Lines 139-142: It looks like there is a problem with the "-190 degree to 190 degree" trick: How do waves at 180 degrees propagate to -179 degrees in this method?

 Line 161-168: (1) How does the epoch ensemble method work? Need more theoretical discussions here. Generally, we would like to use all available data to train the best AI model (instead of splitting them into different epochs). (2) We can add different perturbations to generate ensembles or use ensemble wind forcings (such as from GEFS)  (3) The reduction of RMSE from the 4 ensembles shown in the manuscript may come from the smoothing effect.

Fig 2: It looks like the "AI model with data assimilation" results are NOT free forecasts but analyses performed every 6 hours along with short-term (0-6h) forecasts. 

Line 220: "This suggests that the simple AI model can function independently, at least, in certain scenarios." What does this mean and what scenarios it refers to? Need clarifications.

Fig. 3, 4, 5: Are the results in these figures from the AI model with or without data assimilation?

Line 378-379: "why the AI model can slightly outperform the NWM in these areas." If we want to draw this conclusion, we will need to run both the NWM and AI models with the same settings and compare them directly.

LInes 386-390: The initial value problem has a predictability limit (around 10 days for the wave forecasts). So one would not expect the impact of initial SWH will last beyond ~10 days.

Line 423: "An important advantage of the AI SWH model proposed here is its low computational cost compared to traditional NWMs.": It will be good if we can give concrete examples of the computational resources needed by the AI SWH model and the traditional NWMs.

Lines 445-446: "While training such a model would be challenging, it is not an impossible task, and the rapid advancements in AI may make this goal more achievable in the future": This sentence is too subjective, Consider revising.

Edit: 

Line 304: "clearly evident"  ->  "evident"