Berkemeier et al. present two surrogate model approaches, polynomial chaos expansion and neural networks, used to predict discrete reaction times of kinetic multi-layer models and used for sensitivity analysis and inverse modelling. The paper is well-organized, concise, and appropriate for GMD.  I recommend its publication after addressing the comments below.

One general comment: It is not clear in a first read what the surrogate models are predicting (in other words, what the targets are).  On a second read, it seems like the targets are the order of magnitudes of 3 different reaction times: the 90% reaction time, half-life, and 10% reaction time.  Using these targets instead of concentration targets makes sense if the application is machine learning aided sampling rather than full model replacement.  However, many readers (including myself) might initially assume that surrogate models replicate the output of their reference model, so this non-standard target should be addressed earlier on.

Specific comments

1. The methodology would be easier to follow if the output and purpose of the surrogate models was described early on. This could be done on line 5 of the abstract, for example “we show how machine learning methods can generate inexpensive surrogate models \add{to predict reaction times for}\remove{based on} the kinetic multi-layer model of…”, or something to an equivalent effect.

2. Not required, but a suggestion that might improve clarity: could the authors perhaps include a table for the 3 outputs, similar to Table 1 for the input parameters? This might help new readers understand the targets more easily.

3. There is existing recent work towards network surrogate models in aerosol applications for atmospheric chemistry and physics. Though I have no objection to the mention of applications outside the field, e.g. chemical engineering and materials science, the authors could include several relevant references to round out the discussion around line 60, which would help embed this work in recent literature on ML surrogate modelling for aerosol applications.

a. Surrogate model of MOSAIC, which includes heterogenous chemistry and mass transfer: https://doi.org/10.1029/2020JD032759

b. Mass-conserving surrogate model of mass transfer for bulk/surface partitioning: https://doi.org/10.5194/gmd-15-3417-2022

c. Physics-informed surrogate model of aerosol microphysics, specifically the M7 module: https://doi.org/10.1017/eds.2022.22.

Berkemeier et al. train surrogate models on a more granular kinetic multilayer model, for use in sampling rather than a solver / operator replacement (thus predicting a different output instead of concentration distributions).  Inclusion of these more relevant references will allow the authors to make clear what the contribution of this paper is in the context of the field.

4. Line 65: “In the PCE approach, the full model is represented as an infinite series…” I would get rid of the word infinite here, or perhaps say “In the PCE approach, the full model can be approximated by a series…”.  The PCE approach assumes the model can be represented as an infinite series, but is not applied in practice.  This isn’t clearly stated until mention of truncation around line 136.

5. Figure 2: Are these correlation coefficients for just the log10 of half-life or for all reaction times, as line 205 suggests? Also, maybe “halflife” could be hyphenated to match the text? The inset for the laboratory-experiment relevant timescale is a nice addition.

Review of Berkemeier et al.

Berkemeier et al. present work using neural networks and polynomial chaos expansion to emulate complex models of multiphase kinetics for atmospheric aerosols. They find that both techniques are suitable for emulation, and assess the benefits and drawbacks of each approach. In general, the work is well-presented, addresses an important topic, and provides a valuable contribution to the scientific literature. I have a few comments that I believe should be addressed before this paper is accepted for publication.

Motivation

The motivation of this work makes it challenging to understand exactly what the emulators are being used to emulate. This becomes clear on a more careful read, but I suggest the authors clarify for readability.

Emulator Performance

The authors demonstrate very good bulk performance of their emulators in predicting the target quantity. However, the error of the methods as a function of the input parameter space is important to assess (e.g., error as a function of initial gas phase concentration). If the error is largely independent of the input parameter space, a sentence clarifying this would suffice.

L54-56

Much work on NN in the atmospheric science has been done since 1998. I suggest improving the citations here.

L93

Were other normalization strategies explored (e.g., log-normal outputs)?

L190

These packages need citations.

Table A2

Some of these hyperparameters selected are at the upper boundary of the sampled parameter space. This suggests that additional exploration would benefit model performance. Particularly for the shallow NN number of neurons parameter.