This work presents a new method for predicting organic molecule properties (carbon number, mean oxidation state, oxygen-to-carbon ratio, vapor pressure) for compounds measured by gas chromatography and mass spectrometry but not listed in mass spectral databases. The novel idea is to use a statistical model trained on compounds listed in such databases together with parameters extracted from measurement, which is available regardless of the compound being listed in such a database. The authors include the caveat that this approach requires two-dimensional gas chromatography measurements that capture both volatility and polarity and on dimension is insufficient, which does require more complicated instrumentation than commonly deployed. Nonetheless, the general idea can be useful for the atmospheric science community, and is recommended for publication in Atmospheric Measurement Techniques. I note a few comments regarding the generalizability below.

The authors refer to their "extrapolation set" in they are not included in the training set, but in reality it appears that new samples span a subset of the feature domain spanned by the training set. The consideration of whether extrapolation in this sense is happening or likely to happen in new data sets is relevant because random forest is not capable of such extrapolations - and this would limit the model's utility substantially. Can the authors clarify this point? 

I assume the results are solely applicable to samples run on the same instrument with the same protocol, as the retention time is dependent on the operating procedure. For any new protocol a new model would have to be trained. Can this model be used to generate predictions using measurements on similar instruments using the same protocol, or does a new model have to be trained on each instrument? For publication, the authors should include a statement regarding what is required for adaptation by other users.

This study developed new machine learning techniques to characterize unidentifiable organic compounds using GC-MS and GCxGC-MS techniques. The authors provided a detailed discussion and demonstration of this model and its potential to improve the current understanding of undefined organic species in the atmosphere. This new method is able to improve the quantification accuracy compare with manual proxy modeling, which will lead to a better understanding of atmospheric organic aerosols formation and chemical properties. I'm supportive of this paper and recommend for publication in Atmospheric Measurement Techniques. 

Here are a few minor comments:

Authors have mentioned that the vapor pressures were calculated for model training and evaluation, a few external standards test set and extrapolation set species were incompatible with vapor pressure prediction, can the author provide more explain more about how it is incompatible and the evaluation process for vapor pressure? Based on Figure 4, it seems like the predicted vapor pressure has more variability than other perimeters, and the more accurate vapor pressure can improve the model accuracy. 

Author mentioned that the model underestimated the high carbon oxidation state region and the high carbon number region, but there is no predicted data shown in the plot? it also seems like the model is a little bit overpredicted for the carbon number region between 20 and 30, can author comment on that? 

Does this model capable of any GC-MS system or is there any specific requirement for the instrument? Can author add some discussion of the limitation of this model as well?