This is an excellent study, which is very well-written and clear and with very useful implications for atmospheric inverse modeling, particularly in urban areas.  A few small questions and concerns for clarification should be addressed before final publication:

• To use the VSI approach, does one also need an inventory containing the vertical height profiles of all point source emissions? This would be great to have, but in practice, this may currently exist in Europe only.  (For example, I don’t believe that the Vulcan product for the USA contains height of emissions sources now, nor other products like FFDAS or ODIAC.)
• What are the additional computational requirements of the VSI relative to the SSI approach? Also, how would one go about creating a footprint from a single tower with a mix of the VSI approach for nearby point sources and the SSI for farther-away emissions sources?  How would one do that practically with the WRF-STILT framework?
• I was left wondering what are the relative impacts of mixing assumptions versus PBL height errors when using night-time measurements. Could you include a small theoretical example to demonstrate the impact of realistic mixing height errors with the VSI approach and nighttime observations?

Other small comments:

• Abstract, line 28: “to fall below 0.1 ppm” à during day or nighttime or both?
• Page 3, line 61: “nighttime situations showed a relative bias of more than 50%” -> in which direction is this bias?
• Is 100 particles enough for this study? I assume you would get the same results using 500 particles or more, but it might be worth a small check for sensitivity here.
• Figure 1: This is a nice map, although it’s a bit hard to see the country outlines and the actual distance from point sources to measurement locations. Consider additionally including a histogram or barplot of distance to nearest point source(s) for each measurement location?  To what extent do existing measurement locations follow the ICOS recommendations to stay at least 40 km away from strong anthropogenic sources?  (And how did ICOS derive this recommendation in the first place?)
• Page 3, line 61: “a relative standard deviation of about 40%” in mixing height, or errors in mixing height?  Also, please clarify for following sentence.
• Page 4, lines 62-64: if the uncertainty in daytime mixing height translates into uncertainties of ~3 ppm and 30% of the simulated biogenic signal during summer, what does this tell you about nighttime uncertainties? Just complete the thought here.  Also, in reference to the previous comment, this article develops a better approach to dealing with mixing assumptions in STILT but doesn’t address or improve mixing height errors.  So, what is the relative impact of these two types of errors on both daytime and nighttime measurements?
• Page 5, lines 88-95: this is a great explanation for why the ability to use nighttime observations in inversions would be very useful and is a prime rationale for your study.  I suggest adding a statement to this effect in the abstract about why this work would be very helpful for other researchers for the reasons laid out here.
• Page 7, line 141: please spell out what TNO stands for, for those not familiar.  In general, it might be nice to describe this inventory in a bit more detail for non-European audiences, especially because you are relying on the height profiles in this inventory to implement your VSI approach.  Also, for the differing spatial resolutions between Germany and the rest of Europe, is this how it’s produced in Super et al, 2020, or do you aggregate emissions yourself for the purposes of this study?
• Page 9, lines 189-191: How would time-varying emissions affect these TNO height profiles (e.g. with some emission sources starting and stopping again)? Also, do the TNO height profiles shown in Figure 3b represent sector-specific averages?  Or are heights included for individual point source locations as well?
• Figure 7f: it is nice to have a consistent y-scale with the subplot above (7c), but it’s a bit confusing with the arrows and negative values.  Consider changing the y-scale to include negative values for both.
• Page 11, lines 247-250: is there a physical reason why these errors would be lower in summer than in winter?  I think this could be interesting for the reader.
• Page 16, lines 349-357: it’s a bit hard to follow the argument here.  For example, the statement “However, the power plant within a 5 km radius yields lower ffCO2 contributions during stable PBLH < 500 m conditions than during PBLH > 500 m situations” à is this referring to the VSI approach?   And the opposite is true for the SSI approach?  It sounds like it from the statement in the next paragraph that “the SSI approach simulates on average almost 5 ppm larger ffCO2 contributions than the VSI approach for the closest power plant during stable conditions.”  This is just for the 30-m tower, correct, and not the 200-m tower?  Also, the possible explanation mentioned for the VSI behavior, is this in the model, in reality or both?
• Page 16, line 362: what are the typical inlet heights for ICOS tower stations?
• Page 20, line 461: “inaccurate representation”
• References: please use better indentation to distinguish each reference.

General

The manuscript "Effects of point source emission heights in WRF–STILT: a step towards exploiting nocturnal observations in models" by Fabian Maier and co-workers describes the use of vertical emission profiles for point sources in the time-reversed application of the Lagrangian particle dispersion model WRF-STILT. The authors convincingly show that ignoring vertical emission profiles and assuming surface emissions only, as done in many applications of LPDMs, may lead to serious biases for sites influenced by elevated point sources. The study is an important contribution for regional-scale inverse modelling of greenhouse gas emissions as it directly address shortcomings that can easily be remedied without major modifications on the transport description in LPDMs. The manuscript is well organized and written, methods and results are presented in an appropriate manner. Some minor concerns and considerations remain that I would like the authors to consider in a revised version of the manuscript.

Major comment

The way the introduction (L84-95) and section 4.2 states the problem of using nighttime observations tends to suggest that including vertical emission profiles ('volume source approach') alone may enable modelers to use such observations in inverse modelling studies. However, an important prerequisite, and this is only mentioned rather weakly and hidden (e.g., citation of Geels et al, 2007), is the models ability to realistically reproduce nighttime stable boundary layers and the erosion of these stable layers in the morning hours. Analysis of simulated diurnal cycles and, where possible, vertical gradients against observations are inevitable before assimilating nighttime observations. This fact should be highlighted with more emphasize (introduction, section 4.2, and conclusion). The vertical emission profiles will not solve anything if, for example, the nighttime stable layers are only formed to weakly.

Minor comments

L55 & : This is specific for STILT. Other LPDMs (for example NAME, FLEXPART) use fixed sampling heights that do not vary with the boundary layer height.

L66f: There is another issue with point sources in time-reversed LPDM simulations. The source sensitivities (footprints) are usually stored on a horizontal grid with limited resolution. This adds to model uncertainties as well, since the limited resolution of the footprints may lead to false attribution of point source emissions in cases where a higher resolution footprint may actually have missed the point source. Since STILT is using an adaptive output grid that becomes coarser with distance to the receptor location, this problem may be more important for distant sources, but also for near sources and an inappropriate output resolution false attribution may happen. I think this issue deserves mentioning at this point.

L89-93: Another important point is that the average daytime footprint will differ significantly from the average nighttime footprint. Especially for tall towers the nighttime footprint is usually larger, sampling more distant sources, whereas the daytime (convective) footprint is often dominated by more local sources. Similar to point 2 this may lead to sampling of different source mixtures. The use of nighttime data would certainly extent the 'field of view' of tall tower sites in any inverse modelling study. One requisite is however that the diurnal cycle of boundary layer heights and mixing are captured correctly in the LPDM (point source representation or not; see main comment above).

L119-121: Could these point sources be highlighted in the map? Maybe panel b should be zoomed even further, in order to clearly see the location of these four sources relative to the site.

Figure 2: I find the depiction of model/emission domains a bit confusing. There seem to be two different resolutions and domains for WRF and TNO. However, the figure somehow can be read as if there are 3 WRF domains. Maybe just indicate the higher resolution nests on the left (yellow and black rectangles, as is, but label them only with TNO 1km and WRF 2 km, respectively). Then produce a high resolution zoom that is smaller than the WRF high resolution domain in order to show the nearby point sources (see last comment).

L138: One hundred released air parcels per hour seems to be very small. How can one statistically resolve any vertical gradients with these? The VSI approach requires five different layers as applied here, the lowest two with a thickness of only 100 m. How can you be sure that one hundred air parcels can robustly represent any vertical gradient in such thin layers? The previous h_pbl/2 method may have allowed for such small air parcel numbers because no vertical gradient below h_pbl/2 had to be represented. The improved results with the VSI approach seem to justify the small number of air parcels, but they may merely result from improved separation between stable PBL and lower free troposphere at night. The lack of improvements during daytime (from SSI to VSI) may indicate that residence time gradients during the day are not well represented by the limited number of air parcels.

L139: Considering the outer WRF domain, this backward integration time seems to be rather short. What is the reason for the selected 3 days? How frequently do particles remain within the domain after 72 hours?

L139: There is no information here about the output resolution of STILT. Was this identical to the input resolution of WRF (2 km) or to that of the TNO emissions (1 km). As far as I know STILT output resolution varies with distance from the release location. What was the typical output resolution at distances covered by the synthetic source experiment?

L157: As mentioned before: other LPDMs use a fixed sampling heights in the order of 50 m to 100 m. A smaller sampling height actually assures that the assumption of instant vertical mixing is met. However, it also may require the use of larger particle ensembles in order to sufficiently represent particle distributions in more shallow layers.

L173: This may need further explanation. Wouldn't h' depend on may factors like wind speed, stability, etc. Is this 'effective mixing depth' used in any LPDM? Maybe this should be discussed in the previous section when STILT's sampling height is introduced. It seems to be unrelated to the 'volume source' approach.

Section 2.2.2: This being GMD, I have a technical question: How are the height profiles implemented in STILT? Are these fixed levels according to the TNO suggestions or is the user able to specify them.

L214f: This baseline assumption is most likely not very accurate for the eastern domain border and easterly advection. A note of caution should be added here.

L244 ("The standard deviation …"): What does this aim at? Put observed variability and unexplained variability into perspective? Shouldn't this be done by simply giving the coefficient of determination of linear regression or by an analysis of variance?

L243ff: The use of RMSD may be a bit misleading as it contains the bias as well. A bias-corrected (centered) RMSD (CRMSD) would allow analysing if the representation of variability beyond the bias was improved or not (CRMSD = sqrt(RMSD^2 – BIAS^2)). A quick check of the values in Fig4 suggests that the VSI approach mostly improves the bias but not the representation of variability.

Figure 4, caption: 'standard error of the mean': Not clear which mean this refers to. That of the observations? Why give the standard error? The standard deviation could be more easily compared to RMSD.

Figure 6: Are these surface footprints or obtained from the VSI approach for a specific height?

L315: Which times does this refer to? Hours of release at the receptor or hours of the backward calculation? One particle back-trajectory seems to be a rather poor sample size for a LPDM. How can this be justified.

L318: Are these the PBLH regimes at the time of arrival at the receptor? They may not be representative for the whole transport duration and the considered domain. Travelling times from the furthest power plant were probably larger than 8 hours. So arriving at nighttime in Heidelberg could mean that the power plant plume was still well-mixed over a large boundary layer during the previous day. Could be one reason why differences between the two PBL regimes seem to become smaller again for the largest power plant distance.

Figure 7: Wouldn't it make sense to show the differences as relative differences (e.g. 2(A-B)/(A+B))? The different logarithmic axis (compared to the individual SSI and VSI plots) make it difficult to judge if the differences are important. It would also allow for a more detailed discussion in the text.

L334 and elsewhere: Since sub-panels of Figure 7 are labeled, please refer to them as a,b,c in the text as well.

L363f ('The SSI approach ...'): I don't see this. In Fig 7f (SSI-VSI) most differences are positive, exceptions being power plants at 15 and 20 km, but those are not the closest. Discussion in relative terms would be helpful. See comment on Figure 7.

L386 ('quite low'): Please quantify and explain why these would be smaller. Uncertainty type 3 seems clear since the location is known. But especially uncertainty type 4 could be an important factor when it comes to the current study.

L394: Connected to the previous comment. Is this true? What is the diurnal profile that is applied to large point sources in the present simulations? This may well be different for different point sources. Can you be sure that the nearby point sources mentioned in section 2.1 can be represented well by the applied diurnal profile? A day/night difference of 50 % seems likely for a combined heat and power plant that can react relatively quickly to energy demands.

L461ff: I am not sure if this is the right location for this paragraph. Could also go to the introduction. It certainly does not fit the section title. Maybe also in the context of the Brunner et al 2019 publication L87, which also highlights the importance of representing point sources correctly in Eulerian models.

Code availability: It is not clear to me if the modifications of the 'volume source approach' have been made available in the latest STILT version and how they can be activated. A few more technical details on how to use the approach and from which version these are available would be appreciated.