The authors are presenting here the first version of the Air Control Toobox that allows flexible emission scenario analysis based on a sparse ensemble of CHIMERE emission scenario simulations. The approach is interesting and certainly useful for analysing pollution episodes, although it is true that having the emission reductions applied uniformly over the entire Europe remains a limitation for more practical uses. The paper also nicely illustrates how the ACT tool can be used for analysing the source allocation and chemical regimes. I think it should be accepted for publication after addressing the few minor comments below.

Main comments :

Abstract and introduction : Overall, the paper is well written but the authors should introduce more clearly the objective of their study and the general form of the tool they aimed to build. While reading the paper, it took me some time to understand precisely which type of tool they were developing. One reason may be that in the abstract and introduction, the authors are often talking about a “surrogate model” to tackle the complexity of geophysical chemistry-transport models (CTM). I was a bit confused at the beginning since this tends to suggest that they want to build a CTM surrogate model (able to predict concentrations based on emissions) while actually they want to develop an emission scenario surrogate model (able to predict the concentration changes in response to emission changes). Although some sentences indeed refer to the later, some others refer to the former, which maintains the confusion (e.g., P2L6 : “an emulator, or surrogate model, of a comprehensive air quality model”). While reading the paper, we progressively understand what the authors were referring to before, but the first part of the paper could be clarified so that things are clear earlier in the text. I would personally understand better and easier the goal of their study by introducing it in such a way :

The overall objective of the authors is to build a tool for conducting flexible emission scenario analysis on an every-day basis, for exploring the impact of different mitigation strategies on the current forecast. More specifically, they want the tool to estimate the impact of a given set of emission reductions (in this case, looking at the options available for the user on the ACT web interface, 21 emission scenarios of emission reduction bins of 5%, namely 0%, -5%, -10%..., -95%, -100%) applied to a given set of emission sectors (in this case, 4 emission sectors). Simulating explicitly all possible combinations would require a total of 21^4=194,481 emission scenario simulations (or 101^4=104,060,401 simulations with emission reduction bins of 1%), which is obviously not feasible for computational reasons. This motivates the need for developing, in each cell of the domain and for each day of forecast, (1) a regression model able to estimate the pollutant concentration responses to different emission reductions based on (2) a reasonable number of emission scenario simulations that could be run in parallel to the reference CTM simulation on an every-day basis.

Then, this naturally justifies the type of regression model chosen. Indeed, although more complex regression models can eventually achieve better accuracy, they require larger training datasets (in this case, more emission scenario simulations). As a compromise between these two aspects, this motivates the choice of multivariate second-order polynomial models with interaction terms able to capture some of the non-linearities existing between concentration changes and emission change (at least in some emission sectors) while remaining simple enough, in the sense that only a few coefficients need to be fitted, which allows relying on a limited number of emission scenario simulations.

Section3 : I agree with the first reviewer that section 3 could be shorter. Also, it is not very clear to which extent the step-by-step approach - (1) univariate first/second/third order model, (2) bivariate second order model with interaction terms, (3) quadrivariate first/second order model with interaction terms - is followed only for pedagogical purposes or if (at least part of) the methodology is really sequential (it seems so). Please clarify. If I understand correctly, pairs of emission sectors with most important interactions are chosen in section 3.2 based on different two-variable polynomial regression models, but how this ensures that it remains then the most appropriate choice in the four-variable polynomial models developed in section 3.3? The authors acknowledge that “by doing [such a step-by-step approach], [they] assume that the optimal structure and training scenario remains valid when including interactions whereas there is a possibility that the addition of an interaction term could change the selection of univariate terms.” However, in the type of model proposed in section 3.3, the authors keep only the interaction terms selected before. Also, to test this final model, they retain only the subset of emission scenario selected before. These different methodological aspects should be clarified.

I wonder to which extent not considering any emission scenario with reductions applied to 3 or 4 sectors simultaneously (besides the scenario with all 4 sectors reduced by 100%), for the training or at least the testing is a strong limitation in this study. This should be discussed (and ideally tested).

Additional comments :

P1L13 : “flexibility” : please be more specific on which type of flexibility you are talking about (here, flexibility on the emission forcing)

P4L11 : “Note however that those periods do not constitute a specific training period for the surrogate model which is intended to be re-fitted to new CTM simulations every forecasted day in an automated machine learning approach.” Do the authors mean that all regression models in this study are trained only based on the current forecast day? Please clarify. In total for this study, the authors have performed 46 CHIMERE simulations for each of the 4 months mentioned in Section 2.1, is it correct?

P4L25 : “operational analyses of the IFS” : not IFS forecasts?

P5L20 : “we conclude that the surrogate model will be at most a third order polynomial, less if interaction terms are accounted for.” Please clarify that this is motivated by the number of coefficients to be fitted.

P6L4 : “emissions are applied uniformly over Europe” should be “emission reductions…”

P7L16 : “both dates” : which dates are you referring to?

P9L1 : “testing and validation” should be “training and testing”

FIG3 : Please describe in the legend the meaning of the numbers into brackets (I guess they correspond to the minimum and maximum error over the domain?)

P10L1 : “perform” should be “performs”; “largest emission reduction” should be “largest emission reduction”

P11L6 : “yield” should be “yields”

Sect. 3.1 : Results are discussed for PM10 and briefly mentioned for O3max but not for the other pollutants

Table 2 : a comma is missing between IND30 and IND60

P12L10 : It seems to me the authors are actually talking about lines 3 and 4 where no interactions are included in the training set, please correct if needed.

All formulas : In order to make formulas easier to read (especially the longest ones), I suggest to introduce another variable such as DELTA^sector=EPSILON^sector-EPSILON^ref.

P14L13 : “This final model is tested the” should be “This final model is tested against the”

P17L6 : “ammonia emissions December is” should be “ammonia emissions in December are”

P18L4 : “the allocation we display here” : where?

P18L14 : “assess” should be “assessed”

ACT web interface : Why only O3 is proposed and not O3max or O3avg? Also, the source allocation model is useful for analysing pollution episodes and I would suggest to include it on the ACT web interface in the future.

This paper addresses an important topic, is interesting and well written. The authors develop a surrogate model for a full CTM to investigate ambient pollution changes in response to sectoral emission reductions. Currently the application is limited to episodes and uniform emission changes over Europe which, as also remarked by the other anonymous referee, limits the practical applicability for policy analysis, but the authors also emphasize that they are planning to extend the tool in this direction. It will be interesting to follow this development.

I have a few comments which should be addressed, but these are mostly clarifications and should not require major changes before publication.

General comments

1. I confess I am not an expert for machine learning but I was wondering whether the title is really appropriate. The toolbox as described in this paper uses a parametric regression (first and second order with interaction terms) to approximate the behaviour of a full CTM, and while I think the methodology is sound, I’m not sure if this can be really termed machine learning.

2. I struggled a bit to understand how the regression is set up. When looking at Eq 1, one could get the impression that everything is known, and the equation can simply be solved with a unique solution in each grid cell for the parameters alpha, beta, gamma (if C_i,j and ε are scalar values averaged for the time period of interest) and no regression is needed if the right number of training scenarios is used (2 for a linear fit, 3 for quadratic etc). Is this what is done, or is there a regression involved in the sense of solving an overdetermined equation system and estimating optimal alpha, beta, gammas? Also, should Eq 1 not have a sectoral dimension index in alpha, beta, gamma and epsilons?

3. Besides the limitation that emission reductions which can be explored with this tool are uniform across Europe, also they are specified as reduction of emissions by sector since the coefficients are expressed by sector and not by pollutant, meaning that all pollutants emitted from this sector are reduced proportionally. In practice, emission control policies may reduce different pollutants from the same SNAP sector to a different extent, which cannot be simulated with the current setup. It would be good to mention this somewhere.

Detailed comments

P 5, L 6: The number of training simulations should be constrained by more fundamental principles than the availability of computing time?

P 7 – I would emphasise a bit more the complete reverse response for TRA reductions wrt O3 avg vs peaks. In this context it would be good to mention what is assumed regarding NO/NOx ratio in industrial vs traffic emissions, this may play a role here.

P10 L 22: I am confused what is actually calculated and shown here. Maybe I’m too long after my last statistics class but I don’t exactly understand why a confidence interval is calculated here (for what estimated quantity?) and where the formula comes from.  Wouldn’t it be more useful to show just the 95th or 97.5th percentile of this error distribution?

Section 3.4: Is the final model setup that the authors arrive at in Section 3.4 stable or is this re-evaluated for new time periods in the operational tool?

Section 4.2.2: The conclusion that the December episode is mainly driven by agricultural emissions is rather unexpected, I would have expected a higher contribution from residential heating. How well does the CHIMERE model in this configuration perform in reproducing the measured concentrations of these episodes? The authors note that the emission inventory used may underestimate residential sector emissions, which is an important caveat to highlight when drawing conclusions for wintertime episodes.

Technical corrections

• P 6, L 5. “Emissions”: emission changes / reductions
• P9 L21 performance (singuar)
• P10 L12: in order to reduce the number of performances in this sentence, I suggest to replace the middle one with ‘done’
• Fig 3: The legend is messy and the line colors are not distinguishable. I suggest to change the figure layout. For example, unify to one table with scenario numbers like in Fig 4.
• Fig 4: color scales are exhausted for some of the figures
• P13 L19 – “We are left already” -> We are still left
• Figure 7 caption: raw -> row
• P14 L13 – is tested with / against
• Figure 11 caption: 2015903 ?
• Figure 12 top right - caption unclear
• P16 L 10: NOx from industry and heating
• P16 L18: in Western Europe