The Single-particle soot photometer (SP2) is a widely recognized instrument for quantifying the mixing state of black carbon (BC). However, deriving BC mixing state from SP2 measurements remains challenging. This study introduces a user-friendly SP2 inversion method based on machine learning. Notably, the machine learning approach does not merely replicate the results of physical inversion methods but also utilizes previously unexploited signals. It overcomes the low signal-to-noise ratio issue in input signal prevalent in conventional methods. This advancement will benefit the development of BC mixing state observations and radiative effect assessments. Overall, the manuscript is well-organized, and I recommend its publication after minor revisions.

1. There appears to be a correlation between the deviation of predicted values from the true values and particle size, as observed in Figure 3c. It would be beneficial to further characterize the relationship between prediction accuracy and particle diameter (Dp). This analysis could provide valuable insights into the model's performance across different particle size ranges and potentially identify any systematic biases or limitations in the prediction methodology.
2. Does the deviation between the predicted values and the true values refer to the test set, or does it also occur in the training set? What could be the underlying reasons for this? Please clarify.
3. Does the inversion of Dc and Dp in BC-containing particles utilize multiple outputs from the same trained model, or from different models? Additionally, does Dc influence the inversion of Dp?
4. Could you elaborate on the rationale behind selecting LightGBM over alternative models?