In this paper, the authors used machine learning technique to produce a best-estimate PBLHT by integrating four PBLHT estimates derived from remote sensing measurements at the DOE ARM Southern Great Plains observatory. The paper is generally well written, and I have some minor comments for the authors to consider.

• Line 101: There is not ‘absolutely true PBLHT’ because there is no unique definition of boundary layer height. Different definitions or looking at different aspects of atmospheric boundary layer could lead to different boundary layer heights.

• Table 1: ‘PBLHT-sone’ should be ‘PBLHT-Sonde’ to be consistent with other places.

• Line 134: ‘the Heffter method (PBLHT-Heffter) determines PBLHT as the height of the base of the lowest inversion layer’ seems is not consistent with PBLHT-Heffter in Figure 1, which shows that PBLHT-Heffter is obvious higher than the base of the inversion layer.

• Line 139: define delta s.

• Line 184: delete ‘the’.

• Figure 1c: PBLHT-Heffter seems missing.

• Line 206: delete ‘a’

• Line 217: need to clarify what is ‘local gradient minima’. Does it mean the smallest gradient which is close to 0?

• Line 225: Is each quality index different with others? What if there are more than one PBLHT candidate that have the highest quality index?

• Line 284: Why did the PBLHT-THERMO only use the Heffter method? Did you try the Liu-Liang method and Bulk Richardson number method? In addition, it is the exactly the same as PBLHT-SONDE because PBLHT-SONDE uses three methods.

• Line 289: This might also because PBLHT-THERMO uses the Heffter method, which overestimate PBLHT at large PBLHT as shown in Figure 2a.

• Figure 4d, again the overestimation from PBLHT-THERMO might be caused by using the Heffter method.

• Line 418: delete ‘However’.

• Line 440: Change “Gain.” to “Gain”.

• Line 457: Please be specific what does ‘PBLHT-THERMO’ dominates?

• Line 461: I believe other environmental variables presented in the previous section should also be included as inputs to the ML models.

• Line 506: change it to “and use it to evaluate the predicted PBLHT-BE-ML”.

• Line 569: change the comma to ‘and’.

• Line 574: water vapor profiles are not used in the study.

This is a well-written paper describing the use of remote sensing measurements to derive estimates of PBL height.  The authors describe the use of various measurements and algorithms to derive the PBL height and how the remote sensing measurements can be used in machine learning algorithms to improve retrievals of PBL height.  Recommend publications after the authors address the minor comments below (pay particular attention to item 21).

1. Lines 51-55. The references provided in lines 51-55 refer to ground-based remote sensing measurements. The authors should also include references to airborne remote sensing measurements.
2. Line 60. This should also mention aircraft as well as ground stations and satellites.
3. Line 66. More reliable than what? Radiosonde profiles also generally have higher vertical resolution than many remote sensing techniques.
4. Line 81. This statement says that “…combining PBLHT estimates from multiple remote sensing techniques can lead to more accurate estimates under various conditions.”  If multiple remote sensing techniques each provide a different estimate of the PBLHT, how does one combine these, especially if each technique may have some advantages and disadvantages? Are these combined using different weights for the different techniques?
5. Line 101. What is the true PBL height if (as mentioned in line 40) there are several well-accepted definitions of the PBL?  The authors should indicate the PBL height varies depending on the method used to define the PBL so there is no “true” PBL height.
6. Line 115. Should read “…most extensive ground-based climate research facility”.
7. Line 132. The Heffter technique requires a critical potential temperature lapse rate as well as a potential temperature difference between the top and bottom of a layer.  What was the critical lapse rate used in this study and was this the 0.001 K/m recommended by DelleMonache et al. (2004) for the SGP site instead of the more typical 0.005 K/m?
8. Figure 1d. The green triangle seems to correspond to a critical Richardson number of 0.15 rather than 0.25.
9. Figure 2 shows PBLHT estimates compared with their median. The shaded regions correspond to the kernel distribution estimates. It would be helpful if the paper provided a short description of the kernel distribution  estimate (KDE), how this is computed, and how to interpret the values.
10. Line 215. There is likely also some attenuation by water vapor at 910 nm, which may also cause the ceilometer to have a lower S/N than the MPL.
11. Line 224. Since there are often Cumulus clouds at/near the top of the daytime BL, why would a higher quality index be assigned when there were no clouds detected near the boundary layer? Is this because clouds interfere with the remote sensing measurements? See also item 24.
12. Line 224. How do the various techniques (MPL, Ceil, radiosonde, thermos, DL, etc.) deal with the presence of clouds?
13. Line 244. Should read “σ_w² remains low at night and in the early morning.”
14. Line 264. What is (are) the uncertainties in the RL temperature retrievals? How do these uncertainties compare with the temperature uncertainty in the radiosonde temperature measurement? How do the RL temperature uncertainties vary with altitude? With daytime vs. nighttime operations?
15. Line 284. Does PBLHT-THERMO use only the Heffter method? If so, why not also compute PBLHT-THERMO1 (using Heffter) and PBLHT-THERMO2 (using Liuliang) and use the mean of these? This could increase the agreement between PBLHT-THERMO (mean) and PBLHT-SONDE.
16. Line 372. In the case of missing values, why not also just leave out those cases where one or more pieces of data were missing, rather than trying to  fill in these missing data values, especially for those cases that were used to train the ML model?
17. Line 400. Hopefully the results shown in Figure 7 correspond to data that were NOT included in the training set. Is this true?
18. Line 432. Rather than using local time, could you have used instead solar zenith angle to avoid issues with other locations?
19. Figure 8. Based on these results, is it safe to say that satisfactory results can be obtained solely using the PBLHT-THERMO, PBLHT-DL, and PBLHT-MPL results without the use of the other measurements?
20. Line 446. Isn’t it likely that PBLHY-THERMO emerges as the most significant feature because two of the three items used to compute PBLHT-SONDE use potential temperature profiles as the metric rather than aerosol or water vapor profiles?
21. Line 454 and Table 1. This sentence implies that the higher S/N of the MPL led to a better performance when measuring aerosol gradients, which in turn, led to a better performance in determining the PBL height.  If this is the case, then why were Raman lidar measurements of aerosol profiles (e.g. aerosol attenuated backscatter, aerosol unattenuated backscatter, extinction, and/or depolarization) not used in any of these analyses?  The Raman lidar S/N should be much greater than the MPL S/N and hence be more likely to see weaker aerosol gradients than may correspond to PBL HT, especially at night.  (If for some reason this was not true the paper should explain what was wrong with the RL.) These measurements have inherently high vertical and temporal resolution that could be very useful for such analyses. Furthermore, the RL provides (or is at least supposed to provide) high temporal and vertical resolution water vapor profiles during both daytime and nighttime operations. It would be interesting to see if the gradients in water vapor would be more or less useful than aerosol gradients to determine PBL HT, especially at night. This reviewer was surprised (and disappointed) that such RL measurements were not included in these analyses.   This omission is even more surprising when considering that RL measurements of temperature were included. Consequently, the paper should address why such RL measurements of aerosols and water vapor were not included when RL measurements of temperature were included.    
22. Line 504. Tis line seems to indicate that the RL temperature retrievals have large uncertainties (at least for the altitudes associated with the strong convective periods). The paper should indicate what these uncertainties actually are and perhaps show a comparison of potential temperatures from THERMO as compared to those derived from the radiosondes. How small should these temperature uncertainties be to derive accurate PBLHT from the THERMO method?
23. Line 519. This line suggests that the performance of the ML model may decrease for periods different from those included in the training sets. Perhaps more useful would be to compare the values of the various variables with those used in the actual cases. For example, comparisons of the BL heights, potential temperatures, potential temperature gradients, aerosol graidients, etc. for the data included in the training sets vs. the data that are in the data to be analyzed.  This would one to see where the ML model must interpolate vs. extrapolate results.  One would expect poorer performance for extrapolation vs. interpolation.
24. Lines 532 and 542. If PBLHT-THERMO is derived solely from the AERI data using the TROPoe algorithm, how well can this be expected to work with ubiquitous stratocumulus clouds in coastal California, and is this a reason why the PBLHT-THERMO overestimated PBLHT there?
25. Line 574. I don’t recall the use of water vapor profiles in PBLHT-THERMO.
26. OK, so given these results, is there a plan to use this PBLHT-BE-ML algorithm operationally at the SGP site, or at other sites like ENA and/or BNF? If not, why not?