This manuscript presents a statistical analysis of near surface ozone observations over a 23 year period at Villum Research Station in the high Arctic, utilizing local meteorological observations, backward air mass trajectory modeling, and statistical analysis to elucidate mechanisms controlling observed ODEs. The dataset and analysis are interesting, and the majority of the discussion section is really well done. However the data analysis suffers from some major issues that need to be rectified. 

The way this paper is written suggests a fundamental misunderstanding of the role of sea ice regions in halogen activation and ozone depletion chemistry. Sea ice has snow on it! I'm sure the authors are aware of this fact, but the analysis and discussion give the impression that they believe snow only exists on land. The physical surface of the sea ice itself does not have a pH conducive to halogen activation chemistry (Abbatt et al 2012, Wren et al 2013, Pratt et al 2013). It is the snow in sea ice regions that drives the halogen chemistry. Your analysis and discussion of snow vs sea ice needs to be completely reworked to reflect the complexity of sea ice regions. An analysis of the surface temperature along the back trajectory would potentially help with determining the potential for halogen activation along the back trajectory. 

The selection of the time period for further analysis seems arbitrary, as ODEs don't necessarily follow a clear Mar-May pattern as seen in Fig 2, particularly at these high latitudes. The paper would be strengthened if the time period analyzed were empirically defined utilizing the first to last ODE day. You could choose the earliest and latest over the whole study period to have a consistent time frame across years. It might end up being March to May still but at least you would have a better justification for the choice. 

Ozone seems to be persistently below background through the summer months, this is an interesting finding that merits more discussion/analysis. In my view this is a big missed opportunity by the authors particularly given the low number of ozone observations at this latitude and the discussion of the potential role of iodine motivated by the MOSAIC papers (e.g. Benavent et al 2023). 

The description and utility of a SHAP value needs to be in the main text as the whole ML discussion relies on the reader having an understanding of those values and being able to interpret them. Additionally, Section 3.4 needs to be revised for clarity, I've read it a few times and I'm not entirely sure what I am supposed to be taking away from this section, especially figure 10. Maybe folks with a background in machine learning will find value here, but the broader community I think is going to be lost. 

Minor points: 

Line 284: Given that high wind speed enhances vertical mixing it is not suprising that ozone would not be depleted during those conditions. 

References:

Benavent, N., Mahajan, A. S., Li, Q., Cuevas, C. A., Schmale, J., Angot, H., et al. (2022). Substantial contribution of iodine to Arctic ozone destruction. Nature Geoscience, 1–4. https://doi.org/10.1038/s41561-022-01018-w

Abbatt, J. P. D., Thomas, J. L., Abrahamsson, K., Boxe, C., Granfors, A., Jones, A. E., et al. (2012). Halogen activation via interactions with environmental ice and snow in the polar lower troposphere and other regions. Atmospheric Chemistry and Physics, 12(14), 6237–6271. https://doi.org/10.5194/acp-12-6237-2012

Wren, S. N., Donaldson, D. J., & Abbatt, J. P. D. (2013). Photochemical chlorine and bromine activation from artificial saline snow. Atmospheric Chemistry and Physics, 13(19), 9789–9800. https://doi.org/10.5194/acp-13-9789-2013

General comments

This paper describes a study on the parameters impacting ODEs in the high Artic. It uses a comprehensive set of atmospheric parameters and sea ice conditions together with back trajectory analyses to investigate the sources of ODEs. Apart from a statistical analysis, a machine-learning algorithm is used to identify the most important parameters affecting ozone.

The paper is very well written and addresses important processes in a part of the Earth's atmosphere that is most vulnerable to climate change. It is therefore of high scientific relevance and fits well into the scope of ACP. The way the data analysis is performed and the results are discussed are appropriate given the complexity and multiphase nature of halogen release and ozone depletion events in Polar Regions. There are, however, a few points that should be addressed before final publication:

The authors do not pay appropriate credit to former studies on tropospheric ozone depletion which apply very similar methods as the present study. In particular, Frieß et al. [2023] presents statistical analysis of multi-decadal ozone (and BrO) observations based on back-trajectories and sea ice data, just as this present study, but for the Antarctic. The paper would benefit a lot from a discussion on the possible similarities and differences between the drivers of ODEs in both hemispheres via comparison of the results from both studies.

I am not an expert in machine learning and I have to admit that I was quite lost while reading Section 3.4. I could imagine that other experts in atmospheric physics and chemistry, but not in machine learning, would experience the same. I therefore feel that Sections 2.6 (ML methods) and 3.4 (ML results) require substantial revision as discussed in more detail in the specific comments below. 

The abstract is quite short. It would be important to provide some more specific information on this study (e.g., measurement site, observation period, etc.). 

Specific Comments

L17: It is not clear what you mean with "increasing". Is this a seasonal tendency or an increase over the years?

L64: Please explain what you mean with "relative rate principle"

Section 2.4: It should be pointed out that the back-trajectory analysis applied here is very similar to the methods by Frieß et al. [2023]. 

Sections 2.6: The description of the machine learning algorithm is far too short. I think the reader should be able to get at least a basic understanding of the model without going through the detailed description in the supplemental material. See also my comments to Section 3.4 below.

L267: It is not clear to me what you mean with "monthly hours within the same bin".

Section 3.4: This section is hard to understand for readers inexperienced in machine learning. I do not have any clue what to learn from Table 1, except that high numbers are good. What is a "cross validation score"? What is "Aera Under Curve Receiver Operating Characteristics"? What does "Recall" mean? The following discussion is mainly based on SHAP values. The explanation of this parameter should therefore be moved from the Supplemental to Section 2.6. 

L615ff: You state that in situ radiation measurements would not be available for the entire measurement period, and would also not be indicative for the radiation along the trajectory. Is there any reason why you do not use radiation along the trajectory, which is part of the Hysplit model output?

L622ff: It would be worth mentioning here that an important process that promotes bromine release at lower temperatures is carbonate precipitation from the sea ice, which reduces its buffer capacity and facilitates acidification [Sander et al., 2006]

L649: It is not true that a relationship between RH and ODEs has not been reported before - see Frieß et al. [2023].

L825: This is not a new finding. Replace "Our results show..." with "Our results confirm...".

Technical Comments

L40: "long range" -> "long-range"

L120: What is the meaning of "i.d."?

L148: Maybe the term "accept" would be more appropriate than "require" here.

L152: Start a new sentence after "horizontal plane".

L179: Either state "below mixed layer HEIGHT" or "within the mixed layer".

L192: "A trend analysis of trends...": please rewrite.

L295: Remove "For temperatures" at the beginning of the sentence.

Figure 10: It seems that the y-axis scale refers to the lines (SHAP values), but the histograms have different units. So probably a second y-axis on the right needs to be added for the histograms.

L572: What do you mean with "SS"? Define acronym/abbreviation.

References

Bognar, K., Zhao, X., Strong, K., Chang, R. Y.-W., Frieß, U., Hayes, P. L., McClure-Begley, A., Morris, S., Tremblay, S., and Vicente-Luis, A.: Measurements of Tropospheric Bromine Monoxide Over Four Halogen Activation Seasons in the Canadian High Arctic, Journal of Geophysical Research: Atmospheres, 125, e2020JD033015, https://doi.org/https://doi.org/10.1029/2020JD033015, 2020.

Frieß, U., Kreher, K., Querel, R., Schmithüsen, H., Smale, D., Weller, R., and Platt, U.: Source mechanisms andtransport patterns of tropospheric bromine monoxide: findings from long-term multi-axis differential optical absorption spectroscopy measurements at two Antarctic stations, Atmospheric Chemistry and Physics, 23, 3207–3232, https://doi.org/10.5194/acp-23-3207-2023, 2023.

Sander, R., Burrows, J., and Kaleschke, L.: Carbonate precipitation in brine - a potential trigger for tropospheric ozone depletion events, Atmos. Chem. Phys., 6, 4653–4658, https://doi.org/10.5194/acp-6-4653-2006, 2006.

In this study, Pernov et al. investigate ozone and ozone depletion events (ODE) over a long time period (1996-2019) at the Villum Research Station located northeast of Greenland. A statistical analysis and machine learning (ML) approach is used to analyze the relation of ozone and meteorological variables as well as back-trajectories to study air mass history and surface properties. ODE frequency and duration were found to be highest in May, declining in April and March. Sunny and calm conditions connected with northerly winds seem to favor ODEs in Villum. The ML model revealed that radiation, time over sea ice, and temperature seem to be the most important variables for modeled ODEs during spring time.

To my knowledge, there has been no study which applied an ML approach to investigate ODEs at a specific location. This approach adds some further information regarding the interaction between variables and indicates threshold values for some variables that contribute to ODEs. However, since ML is still a fairly new method and probably relatively unknown to some in this community, I suggest, to include large parts of the ML description from the supplements into the main text, especially the explanation of SHAP values (see below in ‘Specific comments). Further, I am uncertain about why one week was chosen for back-trajectories, as the ODEs at Villum are usually limited to a few hours, with only a few exceptional cases extending to several days. In addition, it was found that ODEs mainly occur under calm and stable meteorological conditions, which would suggest only minor transport of air masses during ODEs. This long time span could bias the analysis, particularly when examining the time above the mixing layer, which occurs towards the later part of the trajectory (see below in ‘General comments’ and ‘Specific comments’).

But overall, this paper is a pleasure to read, particularly the results and discussion parts are very well-executed. Therefore, I recommend publication in ACP with minor revision.

General comments:

Why were 7-days backward trajectorys chosen? On page 27, lines 809-815, you already list the problems of these long backward trajectories (uncertainties, distortion due to the predominant time over the mixed layer further back along the trajectory). Furthermore, it is relatively unrealistic that air masses from 7 days before have a direct influence on ODEs in Villum, especially when they seem to occur mainly during calm and stable conditions and are therefore less affected by transport of air masses. Accordingly, it is also quite unlikely that the Chukchi Sea and the Beaufort Sea are relevant source regions for the "average" ODE in Villum. This could be the case in situations with a lot of transport (e.g. cyclone), but this seems to be the exception here.

I would suggest to shorten the "Summary and Outlook" section, especially the last four paragraphs. In general, all of the topics mentioned in these four paragraphs are relevant and related to tropospheric ozone in the Arctic, but in some cases they are not directly related to what you did in your study (e.g. radiative forcing, AMDEs, cloud cover, etc.) so they come a bit out of the blue and lack context.

Specific comments:

Page 4 Line 132: It would be more coherent to use consistent units for the uncertainties, either % or ppbv.

Page 6 Chapter 2.6: Include parts of the Supplement in here: missing data imputation, machine learning, model, ML explain ability approach

Page 9 Lines 265-268: This sentence is very long and hard to read. I would suggest to split it up in several sentences.

Page 10 Line 325 and following: Maybe include a ‘snow on land’ to every ‘snow’ in the text, to make clear that no snow on sea ice is analyzed.

Page 11 Figure 4: I suggest to only have 2 images per row, to make the individual plots bigger. Even when zooming in, the numbers on the bars are very hard to read.

Page 13 Line 366: Is it really a ODE source region and not just a origin of the air masses? (see above General comments)

Page 13, line 367: Perhaps it should be emphasized that Greenland plays a more important role for the Non-ODE source region (due to the higher trajectory frequencies) compared to the Arctic Ocean.

Page 15, Line 415: Maybe list different surface types here and mention already that land without snow does not play a role. This came as a bit of a surprise further down in the text.

Page 16 Lines 428/429: Shouldn’t it be ‘… start to descend earlier ...’?

Page 17, Line 455: What is meant by ‘model performance’ here? I only see an increase in the ‘Recall’ variable from March to May.

Page 17, Line 458: The difference in the train and test data set and how the model is trained should be explained more detailed in Chapter 2.6.

Page 17 Line 466: Are the mean SHAP values meant by ‘The mean ...’? Should be specified.

Page 18 Line 485: The relationship between SHAP and ambient values and the information its results provide for this study should be explained more detailed (maybe with an example).

Page 18 Line 497: Does ‘negative effect on model prediction of ODEs’ mean the model predicts ODEs wrong when RH is below average?

Page 19 Line 505: Maybe ‘after this bin’ should be replaced with ‘towards lower temperatures’

Page 20 Figure 10: I suggest to include a legend as it was done in Figure 4. An explanation of what the lines represent in the images should be included in the figure description.

Page 22 Lines 598-600: Have there been any investigations into halogen release during ODEs in Villum (generally or specifically for this study)?

Page 22 Lines 606-609: I suggest to make 2 sentences out of this long one.

Page 22 Line 616/617: Might it be possible to use the ERA 5 solar radiation to investigate solar radiation along the trajectory path?

Page 23 Line 662: Maybe include a rough location of the buoys, so one can assume where northerly/easterly/westerly is located.

Page 24 Line 673: Which relationship is meant here?

Page 25 Line 734: What is meant by ‘higher values’ here?

Page 25, Lines 748/749: Are ODEs meant here and not air masses? If the air masses have low ozone levels, these can only be observed in Villum.

Page 26, Lines 772-777: Split this into two sentences.

Page 26, Lines 774/775 & 782: The acidity as an additional factor for ODEs comes a bit out of the blue here. I would suggest to include some sentences about acidity and its impact on halogens/ODEs in the Introduction or exclude the acidity part from the text.

Page 27 Line 812: Have you tried what happens if you take shorter (e.g. 3 days) back-trajectories?

Page 28 Line 835: ‘high time’ → long time?

Page 28 Lines 840-842: see Page 13 Line 366

Supplement:

S1 Machine learning modeling methodology

Second paragraph:

‘We imputed missing data using the median value for the hour of the day for that day

of the year.’ This sentence is very hard to follow, maybe describe it with an example.

Fourth paragraph:

‘Tuning was performed for 1000 trials and the best parameters were selected.’ Is parameters or hyperparameters meant here? If parameters, maybe shortly explain the difference.

Figure S3 description Lines 3/4: Which blue bars?

Technical corrections:

Page 4 Line 120: please define i.d.

Page 6 Line 203: please define RH (first mention)

Page 7 Line 234: please define CA (first mention)

Page 8 Line 246: please define CL (first mention)

Page 8 Line 248: CL instead of Cl

Page 8 Line 258: (d) bold

Page 9 Line 291: ODE instead of ODEs

Page 23 Line 660: please define AK (first mention)

Supplement:

Figure S19 → Figure S9