This paper applies a machine learning approach to try to interpret the POA and SOA data that is observed at the Mace Head atmospheric observatory. These have been reported in previous publications, but this uses a different approach to try to analyse and interpret a long-term dataset.

There are advantages to this type of approach over other methods like PMF, in particular computational cost, interpreting dependencies on independent variables and the ability to gap-fill where necessary. However by using a supervised approach, this can be said to be less objective. This is a particular concern because quite a lot of a priori information is brought into the analysis, in particular how POA is defined, where a number of assumptions are made based on previous experience analysing data. So that being the case, it is perhaps not surprising that the model performs so well. Although it should be noted a certain amount of objectivity is brought in through the use of clustering to generate the representative POA. Furthermore, given that this specific site has proved fairly unique in producing observations such as these (owing to its location), it would remain to be seen whether this technique was applicable to other sites.



All that said however, I still found this an interesting and informative paper. The relationships between the POA and SOA and the other variables in particular, but it was also interesting comparing these data products to the outputs of the HTDMA data, which could have implications for marine CCN populations. I do have some concerns, perhaps the biggest being that I didn't find enough technical information accompanied how the models had been set up (see below) but besides that my comments are pretty minor and I recommend publication after these are addressed.

Major comments:

The article, as it stands, is severely lacking in the technical details of how the FCM and SVR methods were applied to the data. While it is not necessary to post the code that was used, certain technical details were missing from the article and the supplement. Specific things relating to the FCM I found missing were the distance metric (while Euclidian distance is the default it cannot be inferred), whether there was any pre-treatment to weight or linearise variables, and how the optimum number of factors were determined. Likewise with the SVR, more explanation should be given regarding the specifics of the input data and parameters. More details such as these should be included, in the supplement if necessary.

Minor comments:

L185: While the Monte Carlo bootstrapping is a powerful method of assessing random uncertainties in the data, it does not address the issue of systematic uncertainties, which given that these AMS OA types have been reported from few locations, is potentially substantial. This should be noted.

L224: A note of caution should be added regarding treating nss-SO4 as a secondary marker because while it is indeed formed through secondary timescales, the precursors and formation timescales are different to SOA and the relationship between the two is inconsistent when looking at terrestrial environments. There is a case to be made for them being correlated in the marine boundary layer if they are both assumed to originate from biological activity in the sea surface, but explicit correlation should not necessarily be expected.

Figure S3: I actually found this one of the more interesting aspects of this work, and I would consider moving it to the main article.

Figure S7: The figure caption doesn't currently make sense. Reword.

Please see my review, attached.