In this article, the author proposes a high-precision grid prediction by fusing regional-feature selection machine learning prediction (RFSML v1.0) and CTM prediction. The set CTM prediction driven by the disturbance emission inventory is used to represent its spatial covariance statistics to solve the error, and the covariance expansion algorithm is designed to amplify the integrated disturbance and reduce the error, and the model evaluation is carried out on the basis of independent measurement. The prediction fusion based on EnKF is significantly improved than pure CTM. The manuscript has a good innovation, the content of the manuscript is detailed, the discussion is explicit, and the conclusion is clear. It is suggested to add some content for minor revision and publish it in the journal Geoscientific Model Development.

1. What are the input variables of machine learning model (RFSML)?
2. It is suggested to add time-based verification results, that is, to use the data of two months for modeling and the data of the third month for verification.
3. It is suggested to use a schematic diagram to show the variable input and output relationship between the machine learning model and the chemical transfer model, so as to facilitate readers' understanding.
4. In what aspects is the feature selection of machine learning region reflected?
5. What is the maximum forecast time for the prediction effect of the model to ensure certain reliability? Can it be applied to short, medium and long term forecast?
6. What are the main reasons for the different prediction errors in different regions in Figure 9? For example, why does the prediction effect of PRD region have greater error and smaller R than that of other regions?
7. What is the role of the integrated Kalman filter (EnKF) based on Bayesian theory?
8. Does Figure 2 show that the prediction effect of fusion is worse than that of RFMSL according to actual observation.
9. It is suggested to add the advantages, disadvantages and prospects of the article.
10. What is the effect if the model is applied to other seasons?
11. You’d better analyze the prediction effect in different regions. Which regions are better predicted?
12. In the introduction and other section, it is suggested to add some references on the application of machine learning and prediction models (such as: https://doi.org/10.1016/j.scitotenv.2022.160928).

General Comments

This manuscript develops an EnKF-fusion method combining RFSML prediction and chemical transport model outputs to improve air quality forecasts, which has some good applications. However, the manuscript needs some major revision for publication. First, English languages need to be improved overall. Some of the sentences and words are causing confusion. Also, the organization of the manuscript can be improved to be more concise and clear. Please see the specific comments below.

Specific comments:

1. The title is kind of confusing, it doesnot show the purpose of the work, yes, it is a method, but what it is used for?
2. The abstract needs to be revised. It is not clear what the method used is and what the improvement is like from this method. Also, there is no need to discuss a lot of the technical details in the abstract. It should also focus on the significance of this work.
3. P2, Line 1-2: the atmospheric environment has been improved, you mean air quality has been improved?
4. P2, Line 3: how do you know if it is primary PM₅ or secondary PM_2.5?
5. P2, Line 13: what do you mean by typical cities?
6. P2, Line 15-25: These few sentences are quite confusing, I am not sure what the purpose is. Are you trying to say the RFSML is good or not?
7. Paragraph 3 and 4 can be combined to discuss why model forecasts and data assimilation are needed. For example, line 30-35 on P2 mentioned a lot of previous modeling work but did not mention how well the models did, which is the point.
8. P3, line 5: what are the same challenges here?
9. P3, line 17-26: this paragraph is not clear. I cannot tell what the innovation is and what exactly the method is.
10. P4, line 1-5: this paragraph is repetitive from the last paragraph of the introduction. Instead, you can briefly mention the study domain and period here to give an overview.
11. P4, line 10: six categories, you mean 6 regions?
12. P4, line 15: the winter of 2019, can you give the specific study period instead?
13. This work has tested the predication skill of the method for 6, 12, 18, 24h, would it work for longer time periods such 48h or more?
14. I am not sure if Figure 2 can really explain much of the fusion method. For section 2.3, it is not clear how the RFSML prediction works, for example, what are the inputs used? Also, how do you first acquire the RFSML prediction and then do the fusion?
15. Figure 4, time series of surface PM₅, right? This should be added in the caption. Also, where are the ground observations on the plot?
16. For Figure 6, can you also use scatter plots or Taylor diagrams to show the overall performance compared with ground observations?
17. Section 2.5: the radius tested is 3°, 5° and 10°. The distance is quite far, even using 3°, the distance between a ground station and grid would be about 300 km, the correlation between these two places would be low, if I understand the method correctly. If you reduce the radius, I assume you can see some impacts in the city area where ground observations are clustered. There is still value using the interpolation method. Also, what would be the computational cost for the interpolation method vs the fusion method? It is not quite convincing here just to make this comparison and I think it depends on the application.
18. What is the computational cost of this fusion method? What is the application of this method? Some discussions are needed for the implication of this work.