The manuscript is great in its approach and compares with previous models. The application of the work is also in a major area of concern.  Adding a bit more technical details to the manuscript will be good for the readership of the article. 

This manuscript presents a new deep-learning storm surge forecasting (HIDRA2) for the Northern Adriatic which is an “updated” version of a previous deep learning architecture HIDRA1. The authors show that HIDRA2 outperforms its previous version HIDRA1 and it also performs better that the Copernicus Mediterranean Reanalysis. The paper is well written, and I think it must be published after some minor changes.

Major comment:

The authors compare their new HIDRA2 with CMEMS Reanalysis, that is a 3D model in a regular horizontal grid. I recommend to the authors to compare the performance of HIDRA2 against numerical simulations specifically design to determine the sea surface elevation. As examples, the authors may use the global simulation by Muis et al. (2020) which has, in the Mediterranean, a 1.25 km of coastal resolution; or they can use the newest available simulation performed by Toomey et al. (2022) specifically designed for the Mediterranean Sea with a coastal resolution of 200 m. The latest hindcast includes the wave setup component that maybe relevant when computing the total elevation in the studied area.

Minor comments:

- The authors use the term storm surge to refer to the combined effect of atmospheric pressure, winds, and astronomic tides. I strongly recommend the authors to follow the terminology detailed in Gregory et al. (2019). In that paper, they write: “N9 Storm surge: The elevation or depression of the sea surface with respect to the predicted tide during a storm.” And they also write: “Sea-surface height (SSH) can be greatly elevated during a storm by a storm surge, and the consequent extreme sea level is sometimes called a storm tide”. So, the combination of storm surge + astronomical tide should be referred as storm tide, if we decide to follow the definitions in Gregory et al. (2019).

- Line 17: this line needs, at least, two references where I have included (Ref. XXXX): “Global mean sea level rise, related to anthropogenic climate change (Ref. XXXX), is causing a worldwide increase in coastal flooding frequency and is leading to a myriad of negative consequences for coastal communities, civil safety and economies (Ref. XXXX).”

- Fig. 1: Please change the title of the figure to “Adriatic Sea topo-bathymetry”.

- Line 35-36: put the citation to Medvedev between parentheses.

- Line 48: “HIDRA1 ensemble (Žust et al., 2021) is a million times faster than the operational numerical ocean model ensemble based on NEMO engine”. Although it maybe true, it is not fair to compare the computational time of a 3D model that computes water dynamics (currents, temperatures, salinities, …) in several vertical layers with HIDRA1 or HIDRA2 that gives SSH information only and in a single point. The authors should compare the computational time with, for example Muis et al. (2020) or Toomey et al. (2022), and multiply their computational time by the number of nodes that the other two studies are computing.

-Fig. 10: it is difficult to appreciate the differences between the different datasets. I would recommend the authors to apply a filter to the spectrum. There may be a newer reference but I usually do it following the Chapter 5- Time-series Analysis Methods (https://www.sciencedirect.com/science/article/pii/B978044450756350006X) from Data Analysis Methods in Physical Oceanography by Emery and Thomson.

Some thoughts:

1) Could this system be scaled up and applied to estimate SSH values at a Mediterranean scale?

2) How does the predictability change if the system is fed with the first predicted day?

References:

Gregory, J.M., Griffies, S.M., Hughes, C.W. et al. Concepts and Terminology for Sea Level: Mean, Variability and Change, Both Local and Global. Surv Geophys 40, 1251–1289 (2019). https://doi.org/10.1007/s10712-019-09525-z

Muis S., Apecechea M. I., Dullaart J., de Lima Rego J., Madsen K. S., Su J., et al. (2020). A high-resolution global dataset of extreme sea levels, tides, and storm surges, including future projections. Front. Mar. Sci. 7. doi: 10.3389/fmars.2020.00263

Toomey, T., Amores, A., Marcos, M., & Orfila, A. Coastal sea levels and wind-waves in the Mediterranean Sea since 1950 from a high-resolution ocean reanalysis. Frontiers in Marine Science, 1873. doi: 10.3389/fmars.2022.991504

This manuscript shows an implementation of a neural network (HYDRA2) to predict sea level at the Koper tidal station. This neural network is an improvement over HIDRA1 and it is compared to its predecessor and to a numerical ocean model NEMO. Particularly notable in this manuscript is the detailed validation of the performance of the model. I can recommend publishing this manuscript after minor changes.

To facilitate the reading of the manuscript and the interpretation of the figures and table I would recommend the captions clarify if the authors show an independent validation (data not used during training and not used for the optimization of hyperparameters, if this is the case) or validation with dependent data. Likewise I think it would be useful to mention this also with the skill scores mentioned in the abstract (starting at line 7) whether these error reductions are obtained from the independent test data or not.

I don’t not have any doubts about the scientific soundness of the results, but adding this information would help readers understand the results of the manuscript more quickly.

Minor comments:

Line 5: “single member of ECMWF atmospheric ensemble”: is  this the central forecast or any single member (chosen at random)?

Line 47: “HIDRA1 ensemble (Žust et al., 2021) is a million times faster than the operational numerical ocean model ensemble based on NEMO engine (Madec, 2016) at Slovenian Environment Agency”: There is not a lot of context to understand this comparison. NEMO will provide you with a sea level estimate over the whole domain. Is this also the case for HIDRA1 or would it provide the sea-level for a single location?

Line 99: “HIDRA2 does not require explicit annotation of whether a location point belongs to land or sea, thus land masks are not generated.”

I am wondering if the land-sea mask would still be a useful feature to provide to the neural network as a wind over land would not generate seiches. I guess that the neural network compensates for this by learning the land-sea mask internally.

Line 103: “three-days prediction lead time” I think that your ML model will give in one application the full 3-day time series. Can you confirm? Or do you rather need to apply the ML model iteratively to obtain the 3-day time series? Can you also clarify this in the manuscript?

104: “full ECMWF three-day forecast” -> “full” refers to the full ensemble (i.e. all ensemble members)? 

143: “prototype matching layer” Can you provide more information and a reference ?

section 4.1.1: this is an interesting and surprising result. Can the authors speculate why this is the case? (predicting full SSH leads to better results for extreme events). Could it be that the neural network internally limits its output range when working on anomalies? Do you expect this outcome to remain should one have more training data?

Section 4.1.2 “Ablation study”: can you clarify that you retrained the network for the different test cases (without tides encoder, without atmospheric encoder,..)  or you do rather zero-out the output of the corresponding encoder without re-training.

Typo:

Line 205: 1°/24 -> 1/24 °