This study presents the development of a new reduced isoprene oxidation scheme for application in a large-scale atmospheric model, using a benchmark state-of-the-science full description of the isoprene chemistry as a starting point, utilising a novel graph-theory based approach. The mechanism is then optimised and evaluated against its benchmark and other reduced schemes specifically designed for use in US regulatory models as well as limited chamber data, in box models as well as being incorporated into the US EPA Community Multiscale Air Quality modelling system (CMAQ v5.3.3) and evaluated against NE US air quality data.

This interesting study highlights the process of transparently developing a hierarchy of chemical schemes traceable to a benchmark mechanism that reflects the state-of-the-science in chemical understanding (Kaduela et al., 2015, doi:10.1016/j. atmosenv.2015.10.031). 

The authors demonstrate a directed graph path-based automated model reduction approach, going through the necessary steps needed to reduce complex atmospheric chemical mechanisms, such as that for isoprene degradation, including optimisation and evaluation.  This approach is certainly one of the main ways forward that atmospheric chemists should be using for dynamically constructing chemical mechanisms for a range of applications. 

The motivation and application of this study are well founded and reasonably well executed.  However, I am not clear on what the main focus of this work is.  Is it to demonstrate the first steps in using graph theory in the development of reduced chemical mechanisms from benchmark descriptions of (atmospheric) chemistry, or is it the development of a new reduced isoprene scheme, optimised and evaluated for specific conditions for use in US regulatory models?  It could be both, but the paper should then be split into two clear sections, focused on these two motivations. 

For example, the graph theory method sections are really interesting and initially set the paper off as a description of a new method for reducing complex chemical mechanisms. This is great and it's clear that a lot of effort has gone into the approach. However, this is largely undone by the requirement for the manual steps outlined in sections 2.3.5 and 2.4. The changes made here are quite substantial (adding a handful of reactions to a mechanism with 22 reactions soon represents a high proportion of the total mechanism) and so this paper does not demonstrate a method for automated mechanism reduction but rather is aiming to show off a new isoprene mechanism, optimised for a certain specific range of conditions.

With this in mind, it feels like the validation/evaluation of the mechanism in section 3 focuses quite heavily on comparison to other mechanisms rather than comparison to measured data (whether from limited chamber experiments or ambient measurements). Again, if this were a paper displaying a mechanism reduction technique then it would be reasonable to make comparisons only to the mechanism you have reduced, but since there have been extensive manual changes to form the mechanism, this seems like more of an exercise to produce a good isoprene mechanism, in which case comparison to real-world data is required.

I also do not think that it has been demonstrated how well this mechanism performs under different atmospheric conditions for applications in other regional models.  The input parameters used in the pathway importance algorithm and the model scenarios outlined in Table 2 and Table 3 do not demonstrate an ability to work in high-NOx environments. The High-NOx case included is 5 ppbv. How would this mechanism work, for example, in urban conditions in China (where ozone titration of NO effects the NO/NO2 ratio as the day progresses) or even modelling atmospheric chemistry over the Bornean rain forest?

Having said this, this work is certainly of great interest to the atmospheric science community and does indeed show a potentially interesting way forward for the future development of reduced chemical mechanisms.  The work is therefore ideally suited for publication in GMD.  I would recommended publication after the above and following comments have been considered by the authors.

Specific Comments

• Table 1. The authors have chosen to compare the new scheme against those included in mechanisms primarily designed and optimised for use in US Regulatory models (except for the benchmark MCM and CalTech mechanisms).  Why have you not also included comparisons to the Common Representative Intermediate (CRI) mechanism (CRIv2.2; https://doi.org/10.1016/j.atmosenv.2019.05.055)?  This is a reduced scheme developed from the MCM (using a different lumping approach) for use in the EMEP MSC-W chemistry-transport model.
• L40: CRI should also be mentioned (and referenced) as a further example of a “lumped” approach.
• L85 “Full Mechanism Input”: There are no files included in the supplementary giving the “full” mechanism, as described. It would be useful to give some examples of the extended/”missing” chemistry needed in the supplementary material.  “We have done some preliminary testing of this mechanism in comparison to experimental data, but it was difficult to attain meaningful results. Thus, further assessment of this updated mechanism is required”  This is a worrying statement. What are the issues in obtaining “meaningful results?  The integrity of the reduced mechanism depends on the benchmark “full” description of the chemistry, and so testing and evaluation of this scheme is a key step in the process and needs to be evaluated/shown.
• L109: “NO_x” is not an oxidant. Do you mean “NO3”?  However, NOx (i.e. NO and NO2) are key species in the radical propagation cycles (NO to NO2 conversion efficiencies are an important metric) and so should probably be included.
• L125: Include important representative references of the “many prior works”
• Figure 2 (and within the text): In what respect is NO an “oxidant” here?  NO is often stated as an oxidant in the text, which I don’t think it is (RO2 is the oxidising species).  If you want/need to describe NO as an oxidant you need to define why.
• L156: Which “oxidants” are held constant?
• L189: A visualization of the yield estimation algorithm is not shown in Figure S.4.
• L196: Again, not comfortable with NO and NO2 being described as “oxidants”.
• Table 2 shows that the elevated NO and NO2 conditions are very low. While I appreciate that the role of these elevated conditions is not to represent ambient conditions, setting these elevated values too low could result in pathways being deemed "not important" as the yields of species produced under really high NOx conditions (10s of ppbv) would not be given a chance to change from the baseline conditions. How is “solar intensity” defined?
• L220: “Care was taken in the selection of inputs to balance breadth of input conditions with relevance to the atmosphere” Do the conditions outlined in Table 2 really cover the optimal range of atmospheric conditions?
• Table 3 shows that all of the models start with 0 initial O3 except for the High O3 model. Given the effect of O3 on NOx partitioning, it seems that this could have quite an impact on NO concentrations in all of the models. It could be good to run at least a model with (properly) High NOx and high O3. Also in Table 3, I don't think it is specified what photolysis conditions the models are run under (or what low hv actually means, is it 0 photolysis?).
• L371 and onwards describe the addition of IHN as a priority species, but outlines that this addition was done manually. Why was IHN not included as a priority species from the beginning? It seems like this could lead to some changes in the mechanism since IHN has formation routes from OH and NO3 but the current mechanism only includes formation from OH.
• L382: “Table 4”
• L392-3 states that "the {NO3, HO2, hv} pathway was determined to be unnecessary...". If this is the case, why wasn't it excluded by the pathway importance algorithm? What was the rationale behind removing this pathway and replacing it with those listed?
• Figure 3: “pbb” to “ppb”. Give reaction numbers which are adjusted. How do the comparisons highlighted in Figure 3 look for other conditions (NO)?
• Section 2.4.4 outlines the Total Error Metric. This metric seems quite arbitrary, particularly with regards to the weightings applied to each group (Table 5). To avoid issues with the decisions made around weightings for each species in the total error calculation, I would suggest that the error for each species should be presented in Table 6 (particularly since the specific error values are listed below for many of the species anyway). The total error could still be kept in Table 6, though I don't know how much use it is since different use cases would prioritise better predictions of different species/groups.
• L446: Why not 5% for GLY?
• L472: Why only show comparisons for HCHO and HO2? What about for other important AQ species? (O3, HOx, NOx, GLY/MGLY?)
• L474 says "Figure 3.1" where it should say "Figure 4"
• L507 says "shwon" instead of "shown"
• Figures 4 and 5: Why was the box model not “spun-up” first to radical steady state conditions (then show comparisons for non-radical important AQ species)?  Comparisons seem to show new mechanism doing a better job than RACM but not CB6…

• L515 should reference the table in the SI with a number. However, I can't actually find the full error tables for each mechanism in the supplementary.
• Section 3.2 (box model comparisons) is very short considering this is the first comparison of the mechanism to real data. The authors state that the "Caltech Full mechanism matched the concentrations of all measured species from the chamber study" but provide no evidence of this. It would be good to see some plots of major species of interest with the modelled and measured data included. This is especially necessary since their "Caltech full mechanism" is an adaptation of the published Caltech mechanism, so has the potential to show important differences. Including this data would also allow for a comparison of the new mechanism to measured data for species other than IEPOX (e.g. how does it do for the species groups listed in Table 5?).  This is a very limited comparison to one specific chamber experiment under specific conditions.  Why have the authors not compared to other isoprene chamber data?  There must be significant amounts of isoprene chamber experiments available in the ICARUS (https://icarus.ucdavis.edu) and ACTRIS/EUROCHAMP (https://data.eurochamp.org) chamber databases.

• Similarly, in Section 3.3, the authors only show comparisons for formaldehyde and O3. Although they state that there were no changes for NOy, OC, and HNO3, it would be good to see plots of this (probably in the SI).

• One final comment… The authors demonstrate the use of graph theory for the (significant) reduction of complex chemical mechanisms.What would be the outcome if you started from using the reduced from of the CalTech mechanism?

General comment

The authors present a work describing the construction of a very reduced isoprene oxidation mechanism by use of an original semi-automatic algorithm also developed by the authors. The manuscript is well written, and the reduction algorithm is explained well. I only have concerns (detailed below) model performance at low NO conditions.  

Major comments

• The mechanism presented has not been developed and tested for real low NO conditions (< 200 ppt) that are not that seldom in high-isoprene environments. In general, unimolecular decompositions are not considered among the “oxidation pathways”. However, Bates and Jacob (2019) report ISOPO2 isomerization to be 22% of the total loss on average. Reaction 2 in Table S.2 look like to be such a unimolecular reaction, but it is not found in the final mechanism. This is likely to be the main reason the authors need to add 0.6 HO2 to the ISOP + HO2 reaction in order to improve the comparisons with the Caltech mechanism as shown in Fig. 3. However, the comparison for ambient conditions is presented for HO2 and HCHO only. The deviations of AMORE are not small but may be accepted given the extreme computational cost saving. What is worrying is the absence of a comparison for predicted OH, especially at low NO. It is by now well known that photolysis of HPALDs and other H-shifts (unimolecular decompositions) are key in determining OH-recycling and OH levels under relevant atmospheric conditions. The authors should show a comparison for OH. I mean, model bias reduction for ozone and formaldehyde simulated by CMAQ are significant (as shown in Sect. 3.3) but I wonder that could be partly for the wrong reasons. For instance, at low O3 the model bias worsens. Why is that? Anyway, errors in predicting OH cause errors in magnitude and timing of VOC oxidation and SOA production. A hint could be provided if plots like the ones in Fig. 3 and 4 were done for OH and ISOP. In Fig. 6 the error for “HOx” should be split in the error for OH and HO2.
• The pathway importance algorithm is said to be very sensitive to initial conditions (p. 9, l.219). However, the elevated values for mixing ratios of OH and NO2 does not seem to be give a balanced breadth of atmospheric conditions. Measurements of OH and NO2 often exceed the respected elevated values in Table 2 (2.5E6 molec/cm3 and 10 ppt, respectively). I would like to see how the results change by using values like 1E7 molec/cm3 and 1ppb for OH and NO2, respectively.

Minor comments

• The authors seem neither to explore nor to be aware that by neglecting most intermediate species (which deposit on surfaces, react in aqueous media and are transported away) any 3D-model would commit an error in the predictions of the priority species. I think this aspect should be discussed or at least mentioned.