In this paper, the authors introduce an air pollutant (ozone and PM2.5) forecasting model system which based on the deep-learning method. With the implementation of ground observations and the outputs of 3D chemical transport model, this model can forecast more accurate concentrations of ozone and PM2.5. Further, this model system can extend the prediction of air pollutants from individual station to a regional forecasting by considering the temporal characteristics of the time series and spatial relationships among different stations. By comparing the result with the observations, the results of this model show a more accurate and reasonable distributions of ozone and PM2.5, which indicates that this model system can work as a feasible and efficient option to improve current forecast performance. I think the authors did an interesting work. And this study is within the scope of GMD journal. Some problems need to be solved before it can be published.

Comments:

1. “The ground monitoring stations with at least 90% valid records” (Line 92~93) and “the ground monitoring stations with at least 95% valid records ……were selected as the source stations” (Line 95~96). How are these threshold values determined? Please describe detailed information about these threshold values determination methods.
2. the data of source stations is also used for model training. What is the difference between the data of source stations and which of training stations?
3. Line 114. I am confused that why the authors use the WRF-CMAQ data with the next two days to training the model?
4. In this model system, large quantities of observations are used for training model. Please discuss the impact of the number of training stations and source stations on the model performance?

This manuscript (gmd-2022-164) proposed a new approach to refine air pollutant forecasting using deep-learning techniques. Based on the LSTM technique, a novel broadcasting layer was introduced to provide results with spatial coverage. This study not only extends the LSTM-3D-VAR method (Lu et al., 2021), but also demonstrated an alternative to achieve spatial coverage. The ability to extend single site results to multiple locations  is useful for real-world applications and of great interest to the atmospheric community. The performance of the new integrated modal was carefully evaluated, and the results are promising. The reviewer thinks the manuscript can be published on GMD after addressing the following issues.

1. Line 144. Please provide the full name of “The SC method”.
2. The authors mentioned that the LSTM-3D-VAR model (Lu et al., 2021) required substantial computation power. How about the computational efficiency of LSTM-Broadcasting compared with LSTM-3D-VAR? Does LSTM-Broadcasting consume less computation power than the LSTM-3D-VAR model? If possible, please include the direct comparison of the two models developed by the authors.
3. Line 304. “As in Section 3.3, with GPU acceleration…”. Please check, section 3.3 cannot be found in the manuscript.