Review of the article titled “Deep-Learning-derived Planetary Boundary Layer Height from Conventional Meteorological Measurements” by Tianning Su and Yunyan Zhang for publication in atmospheric chemistry and physics.

The authors have used 27 years of data collected by variety of instruments at the ARM SGP site to determine PBL height using machine learning. The method uses the PBL height derived by radiosondes, ceilometer, doppler lidar etc. at variety of temporal resolution to derive PBL height as hourly resolution. The results compare well with the evaluation data. The method is then applied to data collected during two field campaigns, CACTI and GoAmazon showcasing reasonable results. The authors argue that this demonstrates the utility of the deep learning models in predicting PBL height (Line 39). The article is well-written, and a lot of work has gone into it. However, I find some flaws with it and encourage the authors to revise it as it will make it better.

Major Comments

It is unclear to me whether the article is about highlighting the uniqueness of deep learning model or it is about implementing the model to derive PBL height for atmospheric science. From the abstract and discussion, it seems that it is an article demonstrating the uniqueness of machine learning model, which is fine but might make it unsuitable for ACP. If it is for doing science from the derived high resolution PBL height values, then maybe some more analysis should be included in the paper.

Table 3: The table lists feature importance of the input variables. Thereby it should highlight the variables that are most important for predicting the PBL height. The values are very small, and it is unclear why they don’t add up to one. I highly encourage the authors to normalize the values before presenting them in the table. Please see the paper below for more information. Something like their Figure 7 would be great.

Gagne II, D. J., S. E. Haupt, D. W. Nychka, and G. Thompson, 2019: Interpretable Deep Learning for Spatial Analysis of Severe Hailstorms. Mon. Wea. Rev., 147, 2827–2845.

The second author Dr. Zhang has done a lot of work on the SGP site, especially on shallow cumulus clouds and their controls pertaining to land-atmosphere interactions. It will be great if the authors can use either the shallow cumulus case library made by Dr. Zhang, or the shallow cumulus cases simulated by LASSO activity to probe how the new DNN derived PBL heights compare with cloud boundaries.  As of now it is hard to tell whether the DNN derived PBL heights are physically consistent.

Line 240: you have used the lidar derived PBL height when the radiosonde data are not available, with a caveat that they agree within 200m. Can you please tell us how many of them did not agree within the 200m threshold and what was done for those periods? Thanks.

Minor Comments

Line 117: add height above mean sea level.

Line 119-120: Add references to the field campaigns.

Please see attached for the detailed review.