This manuscript describes the design of the Copernicus Atmosphere Monitoring Service with respect to the European regional analysis, reanalysis and forecasts,  based on an ensemble of 11 regional air quality models.

As the authors write, the manuscript is merely a ‘description’ than an ‘evaluation’ paper. For that reason the emphasis for this review is on a clear description of all the features, and less on questioning the actual choices made throughout, both in individual models contributing to the system, and in the overall design of the Ensemble product; Those may only be addressed through follow-up activities addressing topical discussions on individual aspects. Some questions an comments concern some aspects of overall design of the service, although also here it is merely a matter of clarifying, and discussing the choices that have been made, rather than expecting changes in the model setup.

One of the questions I had when reading the manuscript concerned the overarching ideas on how to decide on the best use and exploitation of knowledge that is encompassed by the large consortium that is contributing to this service. Providing an ensemble median product based on results from any of the individual contributing members appears very robust as a baseline solution, but I could imagine that, especially for products of more  experimental nature, other choices could be made, by making use of distributed expertise and knowledge across teams. 

Table 1 and general description: It would be useful to add a row in the table to compare the assimilation aspects to specify if this includes  only the optimization of IC or also optimization of other processes, such as emissions and deposition, as it appears that some of the models also optimize the surface fluxes. Also, this emission optimization procedure appears to be inconsistent with one of the ‘strict requirements’ set to the ENSEMBLE members, i.e. to use specified emissions. Can the authors comment on this potential discrepancy?

Throughout the individual, contributing model descriptions, please try to adhere better to the same formulation, esp. when referring to the boundary conditions for chemical and aerosol compounds. Either refer to it as the IFS-COMPO forecasts or the CAMS-Global forecasts. 

line 157 ff: What defines this selection of trace gases / aerosol types that is requested from individual models? For instance, I wonder if there is an interest in nitric acid, and sulfate as fraction of PM. On the other hand, I am a bit surprised to see a use case for glyoxal as an official product. There are different quality assurance limits

line 1228: “the evaluation is performed on about one third of the stations, deliberately left out of the assimilation workflow” : is this selection of stations fixed over time, or does it vary? 

Also, how does the distribution of uncertainty in the CAMS regional system look like, e.g. in terms of spatial, temporal variability, and for different compounds..  if not going into specific details, could you indicate to what extent you assess this.

line 1261: “For this example, for the year 2021, the ENSEMBLE median has the best success ratio, but some individual models still outperform in terms of probability of detection.”. Given the importance of a good performance of the CAMS analysis for regulatory purposes, and considering that the ENSEMBLE is not the best product with respect to this metric, do the authors have recommendations on either selecting an alternative, individual ensemble member if one wants to get the best product? related, do you have an understanding what is the cause for this, with the aim to improve the ENSEMBLE product?

To what extent is performance improvement seen for the final reanalysis product compared to NRT analysis and interim analysis?  

For further (mostly technical) comments I refer to the annotated manuscript attached here.

This is a nice overview of the design, implementation, and products of the CAMS regional air quality service. While the manuscript intentionally presents little analysis providing insights into what physical or chemical processes drive forecast skill and ensemble variability and could therefore be targeted for specific model improvements, the comprehensive documentation of the different modeling systems contributing to the service and the standardized way in which this documentation is provided nevertheless are an important contribution to the community. I don’t have any major comments regarding the manuscript, it is organized very clearly and the information it aims to convey is presented in a straightforward manner. The authors may want to consider the minor comments listed below when revising the manuscript.

Line 41: add “and” before “SILAM”

Line 49: remove “of” before “in the description”

Line 99: add “and” before “SILAM”

Line 109: introduce the “ADS” acronym referenced later

Line 148: change “disseminating” to “disseminated”

Figure 1: Are the individual analysis and reanalyses performed by the different modeling groups or by Meteo France (NRT/AN)  and INERIS (IRA and VRA)? I think it’s the different modeling groups. In that case, maybe the white text boxes “Near Real Time NRT” and “Reanalyses” within the “Production” box could be expanded to “Centralisation of Near Real Time NRT” and “Centralisation of Reanalyses” for further clarification.

Line 155: does the 48 forecast horizon start at midnight UTC or at 08:00 UTC when the forecast is released?

Line 159: add “and” before “total Peroxy-Acetyl Nitrates (PANS)”

Line 166: if the NRT/AN product is created every day, why is it currently only available for 2021 on the ADS? Make sure to define ADS earlier as noted above.

Lines 164 – 170: It might be good to mention that the methods to create the NRT/AN product will be described in Section 3.

Lines 171 – 177: Are the IRA and VRA products generated by the individual modeling groups using the same methods used to generate their NRT/AN product (just with updated observational data), or is this performed at INERIS with potentially different methods than those used to generate the NRT/AN product?

Lines 191 – 192: Please provide details on how this split between distributed vs. withheld data is performed. Is it based on a random selection of stations? Does the selection change every day / month / year or is it fixed in time? Is the split the same between the observational datasets distributed for NRT/AN vs. IRA and VRA?

Line 219: add a comma before “and the atmosphere data store (ADS) of Copernicus”. Also see my note above on defining ADS earlier in the manuscript.

Line 240: Please provide details (or a reference providing details) on the methods used to adjust the reported emissions that are several years old to current conditions.

Line 242: No comma is needed before “(NMVOCs)”. In addition, NMVOC is already defined on line 159 does not need to be redefined here.

Line 246: Should “ventilation” be “speciation”? Which chemical and aerosol mechanism(s) is the default speciation for NMVOC and PM species provided for? Is there any guidance or harmonization on the vertical allocation of emissions from power plants and industrial sources?

Line 288: Section 2.5 stated that the IFS meteorology is available on a roughly 9 km grid, and Section 2.4 stated that all CTM operate on a domain with about 0.1 degree resolution. Why is a 0.2x0.2 degree grid mentioned here for the forcing meteorology?

Lin 294: add “and” before “snow depth”

Line 364: Same question as for the CHIMERE input meteorology - why is a 0.2x0.2 degree grid mentioned here for the forcing meteorology when the IFS forecast is available at 9 km resolution?

Line 462: In which cases are IFS chemical boundary conditions not available? How often does this happen? Similar information for how this situation is handled in EMEP should be added for all other models, too.

Line 495: Is this indeed 1300 species, or should it be 130? It seems the Bergström reference shows 158 reactions.

Lines 531 – 534: This seems to contradict the statement in Section 2.5 that all models use IFS for their forcing meteorology. In this case, IFS is only used for initial and boundary conditions. Please provide more details on the differences in physics options, land/surface model, land use characterization, etc. between IFS and WRF. Such differences mean that using WRF instead of IFS is not just an improved temporal and spatial interpolation to the EURAD geometry, but actually a distinctly different way to represent meteorology compared to most of the other models used in the ensemble.

Lines 608 – 613: Similar to the comment above regarding EURAD-IM, this also seems to contradict the statement in Section 2.5 that all models use IFS for their forcing meteorology. In contrast to EURAD-IM, it seems that the relaxation of the GEM internal fields towards the IFS target fields with a 3 hour time scale will maintain greater consistency between GEM and IFS than with the use of just initial and boundary conditions in EURAD-IM, but it would still be good to document differences in GEM physics options and IFS physics options that may influence comparisons to other models.

Lines 627 – 628: Please provide a reference for these profiles.

Lines 629 – 630: Please discuss why it is desirable to avoid the effects of short-term variability in biogenic VOC on simulated air quality. Couldn’t such effects be important in some situations?

Line 694: Why 15 km instead of the available 9 km noted in Section 2.5

Line 695: “augmented by” instead of “substantiated by”?

Line 702: Please see my earlier comment on line 462 for the EMEP model.

Lines 742 – 743: Has this been implemented now?

Lines 756 – 759: If “pairs” of IFS layers are combined to generate the MATCH layers, why are the lowest 76 IFS layers lumped into only 26 rather than 38 MATCH layers?

Line 762: which model processes use information from CLC/SEI (e.g. deposition?), and how is this information related to the “model geometry” section?

Line 765: Why is coarser resolution meteorology used for analysis compared to forecasting? If the native IFS resolution is about 9 km, wouldn’t interpolating these fields to a coarser 0.2 x 0.2 degree resolution introduce dynamic inconsistencies in the resulting fields?

Line 737: Please see my comments above about discussing how often (and why) this situation happens and then also provide information on how this situation is handed for all other models.

Lines 786 – 788: This seems to be a discussion of chemistry rather than emissions and should therefore be moved to Section 3.7.8

Lines 859 – 862: Are “diffuse emissions” gridded area emissions? Maybe “combined with” would be clearer than “summed up to” to describe the fact that point source emissions were combined with these gridded emissions prior to horizontal interpolation, and then those combined emissions were vertically allocated using GNFR-sector specific vertical allocation profiles provided by TNO in the absence of more detailed plume rise information for point sources.

Line 917: please specify the chemical species represented in the AERO3 module.

Line 947: Does MOCAGE not require any other meteorological parameters like PBL height, precipitation, soil moisture and temperature, roughness length, albedo, etc?

Lines 980 – 983: How does the deposition module obtain the necessary meteorological, surface, and soil fields given the short list of forcing meteorology listed in Section 3.9.3?

Line 1037: Why is the IFS data not used at its native resolution of 9 km (Section 2.5)?

Line 1204: change “model results deliver” to “models deliver”

Line 1209: do any models miss any of the required species? If so, how is this handled in the generation of the ensemble?

Line 1210: change “offer” to “offers”

Line 1215: Change “Figure 1” to “Figure 2”. Also discuss the reasons why the percentage of forecasts delivered on time is higher than the percentage of analyses delivered on time, given that the timeline for the forecasts is actually stricter than the timeline for the analyses.

Line 1227: Insert “and” before “PM2.5”

Line 1229: Please see my earlier comment on Section 2.3 to please more details on how the split between shared vs. withheld observational data is handled and whether it varies across time and/or NRT/AN vs. reanalysis.

Line 1236: suggest changing “a couple” to “several”

Line 1238: Will such analysis be performed and shared with the community in future publications?

Line 1247: Please provide a reference for this Key Performance Indicator and also define its acronym.

Figure 4: please define “frequency bias”

Line 1300: change “use” to “uses”

Line 1309: change “European” to “Europe”