The submitted article develops a method to estimate PM2.5 values over China using a linear combination of three machine learning model. The innovative of this approach is the method to have an ensemble PM2.5 data from multiple machine learning model outputs. The research method is solid, and the results are convincing.  However, there are several caveats in the paper:

1. The background of the research does not cover all of the most recent machine learning produced PM2.5 products over China and provide convincing reason of why this approach is superior to the rest products. The big advantage of using AHI is the high temporal data (sub-hourly), however, the results section does not reflect this advantage.
2. The most contribution of this study is the linear hybrid ML model. However, the paper does not explain details of this procedure. For example, why using linear combination, and how are the coefficients are determined? Instead of a simple regression, complexed error evaluations of individual ML PM2.5 data may provide insights on a better way of combining these model outputs.
3. The parameter impotency is listed but no further explanation of parameter selection is mentioned.
4. Bias analysis as functions of other influence factors is needed to better understand the uncertainties in PM2.5 product.

You'll find my comments in the attached file.