This manuscript applies CNN-LSTM to replace outliers in PM10 mass concentrations from 12 monitoring sites in China, evaluates spatiotemporal patterns of PM10 mass and OP, and performs source apportionment for four of the sites. It provides a valuable dataset that covers urban, suburban, rural and remote sites. The topic is relevant to ACP, however, several methodological and interpretational issues currently limit the manuscript’s robustness and impact. Substantial revision is required before the manuscript can be considered for publication in ACP.

Firstly, although the title and Methods emphasize the CNN-LSTM model, it is used solely for outlier replacement. The manuscript does not yet demonstrate what additional insights the deep learning approach offers beyond conventional gap-filling techniques such as linear regression or random forest. The authors are encouraged to supply a comparison table that shows the CNN-LSTM’s performance relative to other simpler methods. It might also be helpful to conduct an independent cross-validation, for example, leave-one-site-out to confirm that the network reproduces physically meaningful variability rather than site-specific bias.

Regarding the source apportionment, PMF analysis is only conducted for four of the twelve sites. The manuscript should explain the basis for this selection. Authors should also include more comprehensive error estimation for the PMF analysis. Authors should expand the diagnostics in Table S3 to report whether > 80% of factor elements are mapped in BS runs, and summarize BS-DISP error estimates.

Furthermore, the discussion on OP lacks depth. OPv reflects a combination of PM mass concentration and particle intrinsic toxicity. The current discussion on OP focuses almost exclusively on emissions. Authors are encouraged to discuss how emission sources influence OPv differently from their share of PM10 mass. For example, integrating Fig. 11 and Fig. 12 will help to reveal the intrinsic toxicity associated with different emission sources.

Specific comments:

#33, “due to its small particle size”, this statement does not seem valid for PM10.

#62, photochemical aging can either decrease or increase OP, for example, this paper reports a decrease after O₃ aging: Ma, S., Cheng, D., Tang, Y., Fan, Y., Li, Q., He, C., Zhao, Z. and Xu, T., 2025. Investigation of oxidative potential of fresh and O3-aging PM2. 5 from various emission sources across urban and rural regions. Journal of Environmental Sciences, 151, pp.608-615.

#67-69, this statement is inaccurate. Furthermore, CNN-LSTM is used solely for dealing with missing data, and thus referring to traditional source attribution methods in this context is misleading.

#190-191, the criterion for flagging outliers appears arbitrary. The summed species exceeding the measured PM10 mass does not necessarily indicate outliers considering measurement uncertainties.

#331, which six monitoring stations are being referred to?

#353-355, any reference that supports the temperature-dependent partitioning of ammonium sulfate?

#365-366, the high OPv in Gucheng is related to its high PM10 mass concentration.

3.4, 3.4.1, 3.4.2 source appointment should be source apportionment.

#471, Liu et al. 2023 is cited in the text but missing from the reference list.

Figure 10, The agricultural activities factor shows very different OC/EC loadings at ZZ vs. GC. Please provide supporting literature or discuss why the profiles are different.

#510-512, please clarify what pathways.

#526-529, these statements are inconsistent with the results in Fig. 12. In GC, the contribution of coal combustion to OPv ranks second among the four sites, and the secondary aerosols contribution at GC is smaller than LFS. Please re-examine the conclusions.

The present manuscript aims to characterize the chemical composition, sources and oxidative potential of PM10 at 12 representative sites in China. First, anomalous data were predicted and replaced using a deep learning model that combines CNN and LSTM approaches. Then, the chemical compositions and OP of PM10 were described at the different sites. Finally, a PMF model was applied for the PM10 source apportionment and to evaluate the contributions of the identified sources to OP at 4 selected sites. The collected datasets and the results obtained are of interest for the spatial characterization of PM10 across China. However, the paper faces several issues related to the methodologies employed and the discussion of the results. These critical concerns must be addressed before I can recommend publication in ACP.

Major comments:

• The title and introduction suggest that CNN-LSTM analysis supports PM10 source apportionment, whereas it is actually used only for replacing anomalous data. It should be clearly emphasized what is the added value of such pre-processing step for the subsequent source apportionment.
• The used PMF approach lacks a proper statistical description. Key aspects such as number of factors selection, residuals and error estimation should be presented, at least in the supplement section. Moreover, PMF analysis was conducted for only 4 sites, despite data being available for 12. Performing PMF across all sites could yield more robust results, highlighting the contrasts between urban, suburban, rural and remote areas. Otherwise the authors should clearly demonstrate that the selected sites are fully representative of the different typologies.
• The discussion about OP might be improved. For example, in section 3.3, the link between the chemical composition and OP is not addressed. This could help to understand the elevated OP level observed in Chengdu during summer or in Jinsha during spring. Conducting PMF at these sites might provide further insight. Additionally, the sources contributions to OP were obtained from the PMF analyses, but the reported percentages should be clarified. Do they account for the explained variability of the model ? What are the contributions to the unexplained part ? Moreover, given that OP concentrations are often known to be associated with biomass burning  (Daellenbach et al., 2020), the absence of any such contribution for LFS and ZZ is surprising and should be discussed.     
• The authors are encouraged to carefully revise the text as there is no mention of supplementary materials (Tables S1, S2, S3) in the main text.

Specific comments:

All acronyms must be clearly defined before being used for the first time, even in the abstract, and not re-defined after it (ex: lines 52, 211, 238, 337, 361). Please revise it carefully.

Line 21: Please indicate these “highest OP values”.

Line 48: spaces are missing several times in the text: lines 48, 51, 54, 74, 86, 165, 239, 324, 371.

Lines 60-61: These health effects are already stated earlier (line 46).

Lines 66-81: It is mentioned here that to face challenges in source apportionment methods, deep learning technics can be used to deal with anomalous data. However, the CNN/LTSM model was only applied for correcting the total PM10 data, and not the data inputs used for PMF. This should be clarified.  

Line 95: This is not correct, no MLR model was used in this study for evaluating the PM10 sources contributions to OP.

Lines 190-191: I’m curious how relevant is this criterion. Metals such as Al, Fe, Si, K are not measured and can significantly affect the total PM10 mass concentrations during some specific periods, e.g. during desert dust events. I would recommend some more justification.

Line 216: n, m and p are not described.

Line 231: Is it “PM10 concentration measurements” or “reconstructed PM10 concentrations”? If PM10 were measured at all sites then how ?

Line 232: Please justify the use of OM =1.4*OC.

Lines 232-235: This is already stated earlier, and needs to be better explained.

Line 236: What percentage of the total data do the outliers represent ?

Line 246: Please remove “W”.

Line 247: Be consistent with significative digits of Table 2.

Line 260: Results of Figure 5 would need more discussion. Does the slight deviation occurs for the same sites ?

Line 270: Change with “Due to its location”.

Line 305: Figure 6 is not referenced in the main text, and doesn’t provide any additional information beyond what is presented in Table 3. I would suggest to remove it. How was the “unknown components” part determined ?  

Line 307: Providing the standard deviations to these values would help assess the significance of the observed differences.

Figure 7:  the axis labels are difficult to read.

Line 321: Change “Seasonal” by “Monthly”.

Line 340: What does the “Chinese atmospheric particulates” term refer to ?

Figure 8: Y-Axis units are unclear and not displayed in the figure.

Line 375: Change “northern Chinese sites” by “northern China”.

Lines 375-380: Unclear, are the factors presented to explain the high PM10 and OP levels in winter and autumn applicable only to the northern Chinese sites ?

Line 396: How are the northern/southern regions defined ?

Line 412: What does “more complex” mean ?

Line 415: Figure 11 reference is mentioned before the Figure 10.

Lines 459-460: No contributions from EC and OC were found in the Biomass Burning factor at ZZ site, which is quite unexpected. It should be discussed.

Line 464: Please add a reference supporting the coal combustion origin.

Lines 469-470: What about NO3- ? Is there a reason why NH4+ is only present in the agricultural activities factor at LFS?  

Lines 475-476: Please add a reference supporting this.

Line 495: Transition metals are not measured for this study.

Line 511-512: Several studies highlighted that ammonium nitrate is not redox-active, but can be present with species inducing OP. It should  be clarified.

Daellenbach, K.R., Uzu, G., Jiang, J., Cassagnes, L.-E., Leni, Z., Vlachou, A., Stefenelli, G., Canonaco, F., Weber, S., Segers, A., Kuenen, J.J.P., Schaap, M., Favez, O., Albinet, A., Aksoyoglu, S., Dommen, J., Baltensperger, U., Geiser, M., El Haddad, I., Jaffrezo, J.-L., Prévôt, A.S.H., 2020. Sources of particulate-matter air pollution and its oxidative potential in Europe. Nature 587, 414–419. https://doi.org/10.1038/s41586-020-2902-8