As someone who is inexperienced in all but the simplest of machine learning, I thoroughly enjoyed reading this manuscript. It is very well organized, takes the reader step by step through the various stages of algorithm development, testing and application, and addresses most of the questions that I might have had about the parameters chosen, implementation and application. The use of the algorithm to measure the heterogenity of mixed phase clouds I found especially intriguing. I commend the authors on a very nicely done study, well presented.

There is, of course, a "however".  I have the following questions bear considering:

1) A number of previous studies were mentioned that used various methodologies for pattern recognition of OAP images. I am assuming that the authors were made aware of many of these as a result of our workshop summary in BAMS back in 1985. Can the authors explain why they chose Holroyd as opposed to Hunter, Rahman or Duroure?  These latter three approach seemed to hold a lot of promise but were limited by computing power at that time. I certainly don't expect the authors to try and test their algorithm against these others, but a word of explanation would be helpful.

2) I could not find an explanation of why the particular 15 parameters were selected for the training, in particular not only is the use of interarrival time (IAT) puzzling, it is also somewhat troubling since there seems to be only two possible physical links between this parameter and the phase of a particle: 1) ice crystal shattering or water droplet splatter or 2) inhomogeneity in clouds. It was briefly mentioned that images with IATs shorter than a constant threshold were rejected as artifacts but Korolev et al have shown that a constant threshold should not be used but that a dynamic threshold should be set according to the number concentration. His studies show that the distribution of IATs from shattered particles often overlap those of real particles, leading to some uncertainty in artifact rejection. Is it possible that the IAT is a significant factor because it is identifying/reject liquid or ice solely because they are artifacts?  Alternatively, the IAT has been used to identify inhomogenieties in cloud by looking if they are Poisson distributed.  In another publication Korolev has postulated that due to the WBF effect, there might be clusters of droplets between single ice particles. Perhaps this is why the IAT is a factor.  Could the authors comment on why they chose the IAT in the first place and why it should have an impact?

3. Pervious efforts to develop algorithms used simulated images where their shapes, orientation, roughness, size could be precisely defined. Did the authors give any consideration to this as a complementary approach?

A final recommendation is I think it would be very helpful to show a real image with all the parameters defined. Some are obvious, others not so much.

Thanks again for a delightful read.

The paper „The University of Washington Ice-Liquid Discriminator (UWILD) improves single particle phase classifications of hydrometeors within Southern Ocean clouds using machine learning” by Atlas et al. introduces a random forest algorithm referred to as UWILD for the discrimination of ice and water particles recorded with a 2D-S. It is shown that UWILD outperforms two other existing classification algorithms and the strengths but also weaknesses of the algorithm are discussed in detail. Accurate phase discrimination is important to draw conclusions from measurements on research questions. I support publishing this paper in AMT and only have a few comments, which should be addressed before the final publication

1) Independency of the test set

For training and testing the algorithm, the authors used liquid-only and ice-dominated periods because hand-labeling the data, which introduces a human bias, can be avoided like this. I generally think that this is a good approach. However, the authors should be aware that the training and test sets are not completely independent like this. The particles were all measured during similar conditions (temperature, RH, one-phase). This would be only a minor issue for liquid droplets as they should always look the same and only the size distribution can change, which was partly considered by dividing the test set into different size ranges. However, ice crystal habit strongly depends on the temperature and RH but also on the liquid water content as riming as an effect as well. The conditions during the ice-dominated period are therefore not completely comparable to other periods. Therefore, I suggest making an additional test by hand-labeling a random sample from all other periods (not used for training) and compare the three algorithms, being aware that a human bias is introduced like this. However, it would be a more fair comparison between the different algorithms.

2) Page 5, first line (something went wrong here with the line numbers, which is why I can not refer to a specific line number)

Lawson et al. (2006) only mention a 10µm pixel resolution and not an optical resolution. Please verify if “optical resolution” is the correct term.

3) Page 6, line 138

How can the measured voltage from RICE be used to identify flight periods where the hydrometeors are most likely of the same phase? The authors later say that ice-dominated periods are defined as having no RICE response but do not explain why that is the case.

The word “from” is repeated two times after each other in the same line.

4) Page 8, line 213 “This variable would be more difficult to incorporate in an image-based deep learning model.”

To my knowledge, it is not difficult and uncommon to add variables to image-based deep learning models.

5) Page 11, line 307: “Holroyd and UWILD classify too many medium-sized (0.1 mm < D_eq 3 mm) and large (D_eq >0.3 mm) particles as ice at warm temperatures…”

Holroyd and UWILD classify too many medium-sized particles compared to what?

6) Page 14, lines 399-405 and Figure 10:

In the text and the description of the figure, the solid lines are referred to as the bootstrapped median and the dashed lines as deterministic. However, if I understand the figure correctly, it should be the other way round. Furthermore, the text refers to a,b,c… but the different plots in the figure are not labeled with letters.

7) Page 17, Line 520-522 “Large droplets are necessary for Hallett-Mossop rime splintering and droplet freezing and estimating their concentrations from liquid classifications of 2D-S images can aid in understanding the ice-production process.”

I am not sure if one can refer to “the ice-production process” as there are several ice-production processes and the processes described in this sentence are secondary-ice-production processes.