This paper addresses the role of the driving model in an urban-area LES model that

covers an 8 km square. In this study multiple configurations of the driving model

with different biases were used to provide initial and boundary conditions to PALM.

the main finding was that the biases do propagate into the PALM results, but PALM

can also alleviate some of them. The main verification was wind and temperature for

soundings and the surface, but the results were also compared to the spread of the

driving model. Many results were presented in tables and supplementary plots.

Overall the paper is worth publishing because this type of downscaling is becoming

a common tool for urban modeling that addresses some important problems affected by

local details in built environments. It is useful to emphasize to what extent the

driving model matters, and this is quantified by the methods adopted here.

I will only have some minor comments to address, but I think some may be useful

additional thoughts for consideration.

Minor Comments

1. Table 1. It would have been useful to put the month with the episode number here.

2. Line 135. Does the urban model include anthropogenic heating and a building

energy model?

3. Line 225. Define MRT, PET and UTCI.

4. Figure 2. Is it one point for each time and profile-average and each simulation, 

or each day? It is not clear how many points to expect to see here?

5. Supplementary Figures. It is noticed that the night time stable layer development

is missed by both models, nor does PALM seem to show any improvement when in calm

conditions it should be able to develop cooling if it was able. I am not sure that

the driving model could be solely blamed for missing these and more likely both 

models are missing something. This should be discussed more even if it is just

seen in Supplementary plots.

6. Line 318. "fringe-like pattern". Is this because the flow is from the west and the

variation is along the west boundary?

7. Figure 3. We see a strong local signal which is presumably an entirely urban area,

but it is hard to know because no maps show the WRF urban area. Could an outline be

added for WRF urban grid cells?

8. Figure 5b and others. It is interesting to note that the more urban cells have

less difference due to driving data. I think this implies that the buildings are

deterministic in some way in the local climate. This could be an interesting aspect

of the results to discuss. What characteristics of urban density do these areas have?

9. Figure 6. Some of these rows are redundant being just the reverse of other rows.

It can be reduced to 6 rows without loss of information.

10. Line 405. As above, I think it is not completely clear that the local errors in

both models can be blamed on the mesoscale model in cases where both miss the night-time

stability. There is no indication that PALM has tried to correct this bias even in

conditions of light wind where local effects should dominate.

11. Line 412. Also related to this as stated earlier, dense urban areas appear less 

affected by the driving data, so in these areas the parameters may matter more.

12. Line 444. Yes, WRF has recently added the local climate zone map which should

improve urban morphology. LCZ could be mentioned here.

13. General comment. I think the large number of tables is not critical to the bulk of

the paper while some soundings in the Supplementary data are. I think a better balance

could be achieved. Or maybe some way could be found to condense the table information

more.

Overview

• The role of initial and boundary condition (IBC) uncertainty on high-resolution large-eddy simulations (LES) is a topic of high relevance for the LES modelling community, with many relevant open questions.
• The manuscript investigates the effect of IBC uncertainty arising from the boundary layer physics parametrisations of the parent model on LES simulations with the state-of-the-art modelling system PALM.
• In general, the model experiment the manuscript describes is systematic, is methodologically sound, uses relevant metrics and suits well in the scope of GMD. Studies like of that presented in the manuscript, investigating different aspects of IBC sensitivity in high-resolution LES, are much welcomed.
• However, there are some issues stated below which should be addressed prior to publication. In addition, a careful review of the language and proofreading is required. Given that the issues are addressed by the authors, the manuscript can be recommended for publication in GMD.

Major comments

• The manuscript addresses a certain subset of the open questions on IBC-related uncertainties. More specifically, it analyses those arising from incomplete description of boundary layer physics by the WRF model schemes. There are still a lot more open questions related to IBCs in LES though, such as sensitivity to errors in synoptic scale forcing, improper representation of highly dynamic features such as fronts by temporal interpolation of discrete output time steps of the parent model and sensitivity of the LES-resolved dynamics on the domain size, to mention some. The limitations of the study regarding remaining open questions should be discussed further and more clearly. Also in this context, rephrasing the aims and objectives presented in the introduction and in the conclusions to better reflect the research questions that can be addressed using the described model experiment would benefit the manuscript. Similarly, a more specific title for the manuscript, reflecting that only a certain subset of the open questions related to optimal IBCs for LES was studied, could be considered.
• Given that the specific focus of the manuscript is IBC uncertainty, a detailed description of how exactly the offline IBCs were processed from the WRF model outputs should be added to the manuscript. As the IBCs are the central theme for the manuscript, this should be included in the manuscript itself, especially because PALM’s own documentation is unfortunately not that detailed on the technicalities. The description should include, but not limit to, description of e.g. the spatial and temporal interpolation methods, the temporal resolution of WRF output data, topography matching, applied mass-flux correction etc. Also, the limitations of the used synthetic turbulence generator should be discussed. This would allow for more in-depth critical discussion of the processing pipeline in the review and also in the case that the topic is discussed further in later studies. It would also make the process more clear to the readers, without need of browsing through prior literature and incomplete documentation.
• I don’t see that LES domain size sensitivity is discussed in the manuscript. The domain size dictates how much freedom the LES-based dynamics and PALM model physics have to diverge from the IBCs, and is thus a crucial component for IBC sensitivity. Has a domain size sensitivity test been performed? I regard domain size sensitivity a higher concern than the sensitivity to grid resolution, as it has more direct influence on the analysed metrics, study objectives and research questions in this particular case. Furthermore, the influence of initial conditions on the LES results decreases over simulation time, and I don’t see this being discussed in the manuscript.

Specific comments

L32-40: I don’t think there exists “the” “large eddy simulation for urban environments”. The current phrasing lifts uDALES above other urban-scale LES models, incl. PALM, for an unknown reason. I think the authors could consider how the whole paragraph could be revised for better communication of the relevant information.

L56: The concept of “PALM-4U” is not introduced prior to this. I recommend using just “PALM” or “PALM model system” throughout the manuscript, as PALM-4U components are part of PALM. In any case, the terminology should be consistent and clear to the readers.

Table 1: The way to report the wind speeds here is a bit ambiguous. Consider reporting the mean geostrophic wind speed for 3000 m (from the pressure gradient, assuming it doesn’t change a lot) and e.g. an interpercentile range of 25-75 for the 10 m wind speed.

L126-137: It should be clearly separated which of the facts are specific to the model itself and which are specific to the used model configuration. E.g. PALM implements three different models for subgrid-scale turbulent diffusion, and one is selected for use by the authors.

L135: The PALM documentation refers to the “building surface model” as an “urban surface model”, abbreviated USM, not BSM. Is there a reason why the authors opted to use a different name and abbreviation than the official documentation? Admittedly, the authors’ choice actually describes the model better, but this might confuse those familiar with PALM.

L140-142: It could be added how this spinup is forced. The clear-sky assumption can be a source of bias for the surface fluxes due to excess heat storage if the real-life conditions have been cloudy.

Figure 1: I would suggest using LCZs instead of the Urban Atlas codes for better interpretability. Also, it would be beneficial to show the boundaries of the nested WRF domains and the PALM domain in the figure. Usage of LCZs complemented with the domain boundaries would assist the readers in interpreting the spatial patterns in the results.

L176: With STG, the generated turbulence is not only “started” artificially, like when imposing disturbances until a certain disturbance energy level is reached, but rather continuously imposed on the non-turbulent BC. The physical basis/principle of the Xie and Castro STG method could be described with a couple of phrases here as well.

L181: The information on domain extents/sizes is missing.

L194: Abbreviation NOAH LSM should be defined and cited even though it is just an example here.

L225-227: I think it would be beneficial for the readers if the authors would define MRT, PET and UTCI. This would give the readers a better understanding of what factors are considered in the given metric (thus a better understanding of the potential error sources). It would be also good to reason why these particular metrics were selected for analysis and comparison.

L246-247: Are instantaneous values here used for the models as well or did the authors perform any temporal averaging for these?

FIgure 2: It would be interesting to see Taylor diagrams of the results. In general, I find Taylor diagrams highly useful in model error comparisons. Especially in this case, as a turbulence-resolving model PALM can be assumed to be more “active” than WRF.

Figure 3, Tables 4-5: A suggestion: the July episode could be added to Figure 3 as a second column. Also the mean values could be then included in the figure. Then Tables 4 and 5 could be then moved to the appendix.

Figure 3b: The wave-like pattern is interesting here. Was Rayleigh damping applied above the BL? If not, did the authors ensure that there are no (artificial) gravity waves in the simulations?

Figure 5: I would use LCZs here as well instead of the Urban Atlas codes, as they are more well known in the urban climate community and include fewer categories.

Figure 6: This is a bit hard to read. Please consider increasing the font size so that it would be closer to that of the other figures. The spacing between the x-axis ticks could be increased, e.g. every 3 or 6 hours. The upper panel could also have its own x-axis, as now at least I had to rely on the aid of a ruler.

L400-408: I don’t think this discussion is tied well enough to the results. This could be moved either to the introduction with appropriate references or it should be better tied to the results.

L410-412: I would argue that the magnitude of sensitivity really depends on the metric you study. The diagnostic surface outputs are naturally very sensitive to surface parameters. However, these can’t be alone used to assess the overall sensitivity of the PALM.

L421-423: While in principle I do agree that the offline nesting performance across the mesoscale models could be studied and it is beneficial to have interfaces to as many models as needed by the users, I think it should be clearly mentioned here what kind of contribution to our pre-existing knowledge could be expected from this.

Section 5: A bit of a repetition from the general comments, but the sensitivity of the results on (PALM) domain size is not discussed here. The larger the PALM domain, the more freedom the LES-based dynamics have to deviate from the IBCs. Furthermore, as initial conditions are included as a theme in the manuscript, the sensitivity on how long the LES model is run before the data is obtained from it is not discussed. If it is the case that these have not been studied in this study as it seems, they are a significant limitation for the study and should therefore be included in this list.

L451-454: I would not recommend making statements that are not either well-known, shown in the manuscript, included in the supplementary data nor published prior.

L455-460: It is great to see that a leakage of the validation data into the model inputs is recognized here, even though the consequent bias in this case can be safely assumed to be minimal.

Technical corrections

L2: Should be “the PALM model system” instead of “PALM”. Also, PALM is not an abbreviation anymore, but an independent name.

L58: Underscore (_) is unnecessary.

Table 1: “scattered” has an extra “s”.

L633: The Zenodo link is incorrect and leads to the 100 m LCZ dataset.