General comment

Authors implemented their previously developed numerical model to simulate the dynamics of charge fraction during NPF events.  They also implemented machine learning algorithms to expediate their simulation.  Forward simulation outputs the charge fraction as a function of particle diameter, and the reverse simulation outputs the fraction of ion-induced nucleation (IIN) from given set of charge fraction vs size.  Authors systematically analyzed how each input parameters of NPF (e.g, GR, N_ion, J_IIN, Coag_S) affects the size dependence of charge fraction over 3-20 nm particle diameter range.  It was also surprising to see that charge fraction goes below steady state value when the recombination among charged particles was included, and I hope that authors will have opportunities in their future to prove the observed tend with those measured during atmospheric aerosol sampling.  Authors have shown the conditions in which the reverse simulation can estimate the value of F_INN, and indicated that the charge fraction at smallest measurable size is needed to predict F_INN.  Authors also showed that the sensitivity of the reserve calculation is influenced by the criteria used to select training data.  My comment is positive, and I recommend the publication of this study in ACP after some minor corrections are made.

Specific comments

The left hand side of Equation 1 should be the rate expression. 

Line 276 to Line 277

I think that how far from the initial charge state to the steady state value also affect the value of tao.

In the paragraph between Line 268 to Line 277, the statements refers to Equation S12 multiple times.  It helps readers to be able to see Equation S12 in the main article.   

In the paragraph named “Section 3.3 ResNet Assisted inference of F_IIN”  It is recommended that author provide additional explanation about why and how the difference in the value of the Greek letter “Chi” affects the sensitivity S_x. 

Minor comments

 Line 164, “additional” => “addition”

Line 225, “Neutral” => “Neural”

Line 372 “Resnet” => “ResNet”

The authors present a numerical model to simulate the charging dynamics of newly formed aerosol particles. They use machine learning to infer the fraction of ion-induced nucleation. The model is valuable and the paper is interesting and likely useful. The language is generally clear and the paper is well-written and well-referenced with useful supplementary material. The explanation of the machine learning approach and its results is good. Despite this  I found some details of the model description missing and some aspects difficult to understand, and recommend the following revisions to help put the model and its results in context.

Major comments

How is water handled in the model? In dry versus humid environments, how will the microphysical process rates differ?

What were the conditions for the simulations shown in Figure 2? Does this figure represent a large number of calculations performed with monodisperse size distributions?

Figure 3 and Figure 4: I found this ratio r_c hard to interpret. It would be helpful to add a right-hand y axis to all of these figures to show how the absolute value of the steady-state singly charged fraction varies with the variable plotted on the x axis. Generally through the paper, I found it hard to get a feeling for this pervasive r_c ratio without knowing what the steady-state charge distribution actually is.

It would be beneficial to put the results into the context of the dynamics of the atmospheric ions. During a nucleation event, ion concentrations are likely to decrease. This is not simulated, but it would be helpful if the authors could describe the scale of the changes expected during a realistic NPF event and relate them to the changes they are studying here. Also, a lot of the timescales are very long compared to other expected changes.

The authors should make their model code public or explain why they cannot do so and provide the code to the reviewers as described here: https://www.atmospheric-chemistry-and-physics.net/policies/data_policy.html. Also, the authors should comply with “The data needed to replicate figures in a paper should in any case be publicly available”.

Minor comments

L53 Recent papers by Mahfouz and Donahue are relevant here.

L56 :INN typo for “IIN”

L104 Eqn 1 does not appear to be dimensionally consistent. It is an ion balance of sources and sinks. Presumably there is some multiplication by timestep or derivative missing. The similar equations in the supplement make more sense.

L110 please explain where in the cited López-Yglesias and Flagan paper equation 2 comes from, or otherwise explain how it was determined. Also specify dimensions (units) of beta values.

Please use and define consistent expressions for Boltzmann’s constant between eq 4 and eq 8

L121 specify equation number in Gopalakrishnan and Hogan

L134 specify what is meant by a Lagrangian approach (or at least specify that the sectional scheme is double-moment, if that is correct).

It would be helpful to summarize, perhaps in a table, what is held constant in the model – vapor concentrations, ion concentrations, perhaps? What are the implications of this for how realistic the results are?

Figure 2: it would be helpful to state how these characteristic times compare to the appropriate recombination lifetime and ion-aerosol attachment lifetimes. Do these mean some of the parameter space shown in the figure is not relevant?

L331 It is not clear to me that this statement would be true at the very high ion concentrations produced by a typical ion source in an SMPS, eg. Polonium. What ion concentrations are realistic for an SMPS?

Conclusion: perhaps let other researchers determine whether or not the study is “pioneering”.