The authors present a thorough investigation of the variability in surface ozone of two related CAMS products compared to a comprehensive set of ozone observations distributed over Iran. To account for the fact that the global model simulations are not optimized for these conditions the authors have developed a downscaling approach based on a so-called LSTM neural network method, with, apart from modeled ozone, also assimilated meteorological quantities, as well as lagged O3 observations. They show the benefit of the LSTM method compared to using the raw CAMS products for providing O3. Also particularly the importance of the lagged O3 observations was quantified.

I consider this manuscript well suited for publication in GMD, considering the comprehensive analyses presented here, including the development of the LSTM method, and the analysis of the different model versions to represent the ozone variability at various temporal scales and different stations and regions over Iran, which is also of wider interest. My only main hesitation concerns the difficulty to follow exactly what approach the authors have taken in their methodology. Any further revisions that help to clarify (and improve readability) their methods is still welcome. Also some further discussion on the implications of your study, e.g., for the use of coarse-scale global model data such as CAMS for policy applications (?) or possibly the scope of the methods developed here for wider application (?) should be better highlighted, to better place this work in a wider context. Part of the scope of this work is indeed mentioned on line 56-60 of the introduction, but there is no explicit answer to these interesting resarch questions in the abstract or the conclusions - only indirectly.

More specific comments:

l. 16: “correspondence precision” - not clear what this is - suggest to use another wording here.

end of abstract (and end of conclusions): I expect a sentence that briefly describes the implications of your study. Same comment holds for the (end of) the conclusions.

l. 41: suggest to change this sentence to: “In recent years, the Copernicus Atmosphere Monitoring Service (CAMS) has been mainly developed to assimilate observations of chemical composition to provide analyses of tropospheric ozone and aerosol concentrations,… ”

l. 43L “and a control run (no assimilation)” -> “and a control run (without assimilation of atmospheric composition)”

l. 75: “…using a four-dimensional variational (4D-Var) scheme as…

L. 80 “MERAA”-> ”MERRA”

L84 “It is noteworthy that newer versions of data have been frequently adopted in CAMS.” : it is unclear what the authors want to convey in this sentence. Is it that different satellite data have been used in the Reanalysis product, or that different CAMS reanalysis products exist, with CAMSRA the latest and most comprehensive, to date? 

l. 91: “Compared to CAMSRA, in CAMSFC only the initial conditions of each forecast are obtained from reanalysis datasets, i.e.,..”

consider change to 

“Compared to CAMSRA, in CAMSFC the initial conditions of each forecast are obtained from analyses of atmospheric composition in near-real time, i.e.,…”

l98: “Biomass burning injects from GFAS” -> “biomass burning emissions are based on GFAS”.

l102: “from 9 July 2019 onwards,…”

l149-150: “KZ(35,5)”  - I understand that 35 here refers to ‘m’, the window size. But can the authors please explain why they choose the value of 35 here? (and a value of 5 in the definition of S in eqn. 2) Does this correspond to a filtering time scale of 35 x 3hr = approx. 105 hr, i.e. 4 days?

l189-191: As a modeler on initial reading I find this split in definition between ‘explained’ and ‘unexplained’ error a bit artificial. Different to what is suggested, I would also not have a direct understanding of the cause of ‘explained error’. After reading the manuscript, I think I better understand the arguments of calling errors either ‘explained’ or ‘unexplained’, but it might help to allude to that.

l.235: “for most of the stations…”

l. 246 and 248 The authors refer here to ‘opochs’. please provide an explanation what an ‘epoch’ is exactly, in this context. I missed that.

l. 250-251: “That might reflect that the more predictors, the better the model would not be.” : as also reflected in the conclusions, I find this an important finding indeed. I’d suggest to stress this a bit better, also by re-formulating this sentence a little - now it reads a little clumsy. This finding may be worth a bit more statistical analysis, i.e. do the authors have any quantitative metric arising from the method which provides insight as to how much each of the individual parameters contributes to the quality of the end product? It might be a useful exercise to exclude some of the (physically) less obvious parameters from the list of fitting parameters, such as U, V, W, MSLP (?). Here an analysis of station 22 (Yazd), which performs relatively poor, while it uses an excessive list of input data, suggests indeed the limitations of this work. Can the authors comment?

l. 282 “lagged O3” please specify here (again) that this refers to lagged O3 from actual observations, to help the reader understand.

l 301: typo ‘products’

L328: suggest to drop the sentence “These values…” - no need?

l364:”peroxides”->”proxies”

l377: Lines 377-380: I’d expect here a comment on the implication of these findings, e.g. the importance of observed (lagged) O3 as predictor (?) and/or the potential use cases of the methods as the authors have developed.

Table 1: “single level” -> “surface level”

Table A3:  typo in units for UV

This study uses observational and CAMS reanalysis and forecast data to investigate O3 variability and errors in CAMS global systems for several stations over Iran. To this end, the observed and CAMS O3 time series are decomposed in three spectral components short, medium, and large, with the latter one not examined. Subsequently, an LSTM neural-network is applied to downscale the components and then investigate the CAMS performance and error sources.

Overall, I find this to be an interesting study, yet, with several points that need clarification and improvement, to suggest publication in GMD.

Main comments

1. Title of the paper: The authors use surface-ozone data from both observations and CAMS (RA and FC). Is this correct? If yes, I don’t understand the rationale behind the use of term “tropospheric ozone” in the title and in the abstract. Tropospheric ozone variability is governed by different processes and spatiotemporal scales compared to surface ozone, thus, I believe that the title is somehow misleading.

2. I feel that more interpretation of the results is needed in the manuscript. For example, uncertainties in CAMS emissions inventories and deposition, and the fact that the two products use different emissions inventories should be discussed, especially for the S term.

3. The role of stratospheric ozone contribution to the surface ozone is only discussed once in the Discussion. Stratospheric ozone can affect surface ozone levels indirectly through vertical downward transport of ozone from the lower stratosphere and/or the upper troposphere in larger time scales (Zanis et al., 2014) or directly through intense stratospheric intrusions (rarer) (Akritidis et al., 2010, Chen et al., 2022). The broader Iran region is a well-known hot spot of stratosphere-to-troposphere transport that might affect day-to-day O3 variability in some cases. The CAMS reanalysis product includes a tracer for stratospheric ozone that might be useful in explaining part of the unexplained ozone variance. This is a suggestion that the authors might consider for the present or future studies. At least, the authors should consider for the discussion that by not including such a source (stratospheric ozone) of surface ozone is a potential source of error.

4. I agree with the main comment raised by Reviewer 1. The manuscript should be more oriented to presenting the main objectives of the study and then the main implications. For sure, this will make it more reader friendly.

Comments

- P1, L21: A recent review of air pollution impacts on health is worth cited here by Pozzer et al. (2023).

- P5, L135-136: “whereas long-term and seasonal variation is mainly related to solar radiation.”

What about long-range transport and stratosphere-to-troposphere transport (Monks et al., 2000).

- P6, L166-167: I appreciate the fact that the authors performed a hyperparameter tunning instead of using default values.

- P6, L175-176: Regarding the training of the LSTM model:

1. Are the data shuffled prior to the training process?
2. Did you apply "early stopping" during training to help avoid overfitting?

- P9, L250-251: “That might reflect that the more predictors, the better the model would not be.”

Or that the M term is of less complexity and easier to be modeled?

- P26, Table 2: How should someone interpret the fact that for station 23 only Q (specific humidity? Not included in Table A3) is important. Moreover, there are stations that the O₃^RA (or O₃^FC in Table A4) is not important; how should someone also interpret this?

- P29, Table A3: The sea surface temperature is listed here as a meteorological variable for CAMSFC. As the SST fields are only over sea, for which coordinates are the SST data extracted for each station?

- A small paragraph on what might drive the differences between O₃^FC and O₃^RA in the discussion is needed.

P12, L364-365: “The most relevant peroxides were found by screening several meteorological variables and chemical species.” I don’t understand this sentence. Please explain or rephrase.

Minor comments

L9: Atmospheric -> Atmosphere. This is by CAMS definition.

L11: datasets -> time series

L20: please delete “, or tropospheric ozone at ground level,”

L30: level is -> levels are

L36: provide: provides

L37: satellite-> satellites, and also remove “computer”

L38: technique-> techniques

L90-91: 50 chemical species and seven different aerosols. It provides outputs for several meteorological variables as well. -> 50 chemical and seven aerosol species, providing also several meteorological parameters.

L101: . They-> and

L127: Both reanalysis and forecast datasets were..

L196: Add equation number.

L224: was extracted-> were extracted

L247 explained variability-> explained variance . Please apply this were applicable.

L260: tiny->small

L260-261: so both models show similar performance. -> with both models exhibiting similar performance.

L307: SDS?

References

Akritidis, D., Zanis, P., Pytharoulis, I., Mavrakis, A., and Karacostas, Th.: A deep stratospheric intrusion event down to the earth’s surface of the megacity of Athens, Meteorol. Atmos. Phys., 109, 9–18, doi:10.1007/s00703-010-0096-6, 2010

Chen, Z., Liu, J., Qie, X., Cheng, X., Shen, Y., Yang, M., Jiang, R., and Liu, X.: Transport of substantial stratospheric ozone to the surface by a dying typhoon and shallow convection, Atmos. Chem. Phys., 22, 8221–8240, https://doi.org/10.5194/acp-22-8221-2022, 2022

Monks, P. S.: A review of the observations and origins of the spring ozone maximum, Atmos. Environ., 34, 3545–3561, 2000

Pozzer, A., Anenberg, S. C., Dey, S., Haines, A., Lelieveld, J., & Chowdhury, S. (2023). Mortality attributable to ambient air pollution: A review of global estimates. GeoHealth, 7, e2022GH000711. https://doi.org/10.1029/2022GH000711

Zanis, P., Hadjinicolaou, P., Pozzer, A., Tyrlis, E., Dafka, S., Mihalopoulos, N., and Lelieveld, J.: Summertime free-tropospheric ozone pool over the eastern Mediterranean/Middle East, Atmos. Chem. Phys., 14, 115–132, https://doi.org/10.5194/acp-14-115-2014, 2014