Review of "ML-TOMCAT: Machine-Learning-Based Satellite-Corrected Global Stratospheric Ozone Profile Dataset from a Chemical Transport Model" by Dhomse et al., Earth System Science Data

GENERAL COMMENTS

First, I think that what this team has done here is excellent, very much needed, and that the data set arising from this work will be very useful. It is something I have always wanted to do myself but, lacking infinite time, have been unable to. What is presented in this paper is what I see as an ideal combination of our knowledge of stratospheric chemistry with all available stratospheric ozone profile measurements to create a definitive 42-year history of the vertical distribution of ozone in the stratosphere. The only superior approach that I could think of would be a version of TOMCAT that has a 4Dvar assimilation of all available ozone profile measurements. You may want to say something about how your approach differs from what would be achieved from a data assimilating version of TOMCAT and why nobody has gone down the route of creating a data assimilating version of TOMCAT for this purpose.

I found that the description of exactly how the ML-TOMCAT database was created was too terse and I would like to see much more detail on that, i.e. sufficient for a reader to reliably replicate what was done.

SPECIFIC COMMENTS

Line 58: Rumour has it, but who knows whether it is true, that the true expansion for BDBP is "Birgit's Database of Biblical Proportions". 

Lines 95-96: Does this mean that the resultant ML-TOMCAT ozone profile database is not longitudinally resolved? That's a little disappointing. I thought that with TOMCAT being 3D you would be able to create a 3D ozone profile database.

Section 3: I suspect that most readers of your paper will be like me, i.e. experts in stratospheric ozone with some (or maybe even little) expertise or knowledge of machine-learning methods. When I think of a 'random forest' I think of someone who has planted trees somewhere arbitrary to claim carbon credits. All I am saying here is that you may need to be more pedagogical in your approach to Section 3. You may need to describe in greater detail, and at a higher level, the machine-learning methods that you are using. Don't miss the opportunity to educate your reader to the level that they require to understand your paper, being cognizant of the level of knowledge your typical reader is likely to have. If you do it well, they will be eternally grateful. Test the material on non-ML-experts to see whether it has been pitched at the appropriate level. Just as one small example, on line 113 you refer to 'the model'. I suspect here that you are no longer referring to TOMCAT but now to some model underlying the random forest? Line 113 also refers to 'learners' for the first time without describing what these are. You will need to rewrite this section being extra careful to maintain clarity.

Line 117: Isn't the full correct name for the package scikit-learn?

Line 132: In the context of regression modelling, would a better term for 'learner' be 'predictor' or 'basis function' or 'explanatory variable'? I would suggest that, to the extent possible, you use terms that will be more familiar to the majority of your audience (atmospheric scientists) rather than terms used more by the data science community.

Equation (1): shouldn't there be coefficients at the front of each of these terms as in a standard regression model, these being the regression model coefficients? You refer to these coefficients on line 191. Otherwise this equation doesn't make sense to me. It doesn't even make sense in terms of units, e.g., dO3 would be measured in ppm and dTCO in DU.  OK, maybe you are let off the hook by the 'and so all the predictor time series are detrended and normalised between 0 to 1' clarification at the end.

Line 140: I assume that TCO fields are not available from ERA5?

Line 163: And this is no small caveat. I think that it would be worth pointing readers to the SPARC S-RIP activity here.

Figure 1: Why start in December and finish in November? Here, our year runs from January to December.

Line 169: You say '(2006–2020)' but the figure caption says 2001-2020.

I think that throughout the paper things would be clearer if you replaced 'the model' with 'TOMCAT' wherever that makes sense since there is also the regression model and you don't want the readers to constantly be having to figure out which one you are referring to when you say 'the model'.

Line 175: I find this confusing. On line 170 you state that positive biases occur in the upper stratosphere and negative biases occur in the lower stratosphere. But then on line 175 you explain the source of negative ozone biases in the upper stratosphere (even though this is not what is observed) and, similarly, on line 180 you explain the source of positive biases in the lower stratosphere whereas what is observed is negative biases. Am I missing something here?

Line 195: And presumably these RF-derived biases extend to regions of the globe for which observed ozone profiles are not available?

Line 202: Here you refer to TOMCAT-SWOOSH differences whereas on line 170 you said 'SWOOSH minus TOMCAT'. I find this confusing.

Figure 2: Presumably the sum of all 5 fields other than the R^2 field should sum to the R^2 field? Has that been confirmed?

Line 209: I think that it would be worth saying more about why the HCl basis function has such high explanatory power in the tropical lower stratosphere.

Line 238: This suggests that in Figure 3 there should be stippling to show where the differences are statistically significantly different from zero at the 2 sigma level.

Line 247: I assume you mean significant improvements compared to TOMCAT? You should then say so.

Line 252: Why the need to use SAGE II and HALOE when UARS MLS always covers the lower latitudes (34S to 34N)?

Line 262: You say with respect to GOZCARDS but in the title of the figure it says "w.r.t. SAGE-CCI-OMPS"?

Line 266: It wasn't clear to me what you meant by 'uneven variations'.

Line 275: You need to cite a paper to support this assertion that there is a bias in the ERA5 reanalyses in 2020.

Lines 276-278: As it is currently written, this sounds like conjecture. You either need to do the analysis to investigate whether or not this is the case or cite something that gives some credence to the assertion.

Line 279: I would prefer to see this referred to as 'Comparison with ozone concentrations reported on altitude levels'. In fact I would prefer to see the word 'altitude' rather than 'height' used throughout the paper, unless you are specifically referring to geopotential height rather than geometric altitude in which case I think that you should always be explicit and say 'geopotential height'. After all, Figure 6 refers to altitude rather than height.

Line 288: Yes, but if I remember correctly, the TOMCAT profiles are not used directly in the BSVert ozone data set but only as a transfer standard to calculate relative biases. So there is no reason why BSVert ozone cannot be biased against TOMCAT. I would suggest that you read Hassler et al., 2018 very carefully.

Line 289: This is not true. The negative biases extend to 25N.

Line 310: Should this be 'ozone number density profile'?

GRAMMAR AND TYPOGRAPHICAL ERRORS

I have suggested some, but not all, grammar and typographical corrections. I am hoping/expecting that further corrections will be made by the author(s) and the ESSD editorial staff.

Line 3: 'Satellite instruments obtain stratospheric ozone profile measurements' sounds rather convoluted. Why not just say 'Satellite-based instruments measure stratospheric ozone profiles'.

Line 37: Replace 'which could variability in ozone trends' with 'which could induce variability in ozone trends'.

Line 40: Replace 'As there is no' with 'As there are no'.

Line 54: Replace 'data was' with 'data were'.

Line 65: replace 'merged data' with 'merged data set'. Likewise on line 72 replace '(SWOOSH) data' with '(SWOOSH) data set'.

Line 75: Replace 'One of major' with 'One of the major'.

Line 81: Replace 'firstly' with 'first'.

Line 111: Would it not be clearer to replace 'algorithm by splitting observations' with 'algorithm to split observations'?

Line 119: Replace 'on to' with 'onto'.

Line 120: Replace 'data is' with 'data are'.

Line 152: Isn't the IUPAC convention that it is 'sulfate' rather than 'sulphate' irrespective of which side of the Atlantic you're one?

Line 198: Replace 'that RF regression model' with 'that the RF regression model'.

Line 216: Replace 'This means although' with 'This means that although'.

Line 226: Replace 'data was processed' with 'data were processed'.

Line 227: Replace 'geopotential' with 'geopotential height'.

Line 320: Replace 'is able to polar' with 'is able to model polar'.

Lines 323-325: This sentence lacks a verb.

Line 350: I believe this should be 'Hassler et al., 2018'? And please fix the reference accordingly - you appear to have used first names rather than surnames.

The manuscript “ML-TOMCAT: Machine-Learning-Based Satellite-Corrected Global Stratospheric Ozone Profile Dataset from a Chemical Transport Model“ by S.S. Dhomse describes a new stratospheric ozone profile dataset and the machine-learning method used for its creation. The basis for the new dataset are profiles from a simulation provided by the chemical transport model TOMCAT. Biases between these profiles and profiles from the well-established ozone profile dataset SWOOSH are then calculated and used as input for a Random Forest regression method with five learners (or basis functions). The obtained coefficients for these five learners are then used to bias correct the entire TOMCAT simulation. The resulting ML-TOMCAT dataset is then compared to several different available observational datasets and differences are discussed. The manuscript is very well written, well structured, and the topic lays within the scope of the ESSD journal. However, there are a few things that I think would help to improve the manuscript, and that I would suggest the authors to consider while revising the manuscript. Comments highlighting these issues are outlined below.

General comments:

• As the authors state in the manuscript, merging techniques for ozone datasets can result in significant uncertainties within the created dataset, and can be a considerable source of differences between merged datasets. The applied machine-learning method to correct the biases between the TOMCAT simulation and SWOOSH is therefore one of the main parts of the manuscript. However, at the moment it feels too short and not detailed enough considering its importance for the creation of the dataset. Are the five learners applied in the same form to all latitudes and altitudes/pressures? How big are the residuals after the Random Forest regression has been applied? Why was the tropospheric part of the profile also put through the Random Forest regression although there is no data from SWOOSH and the ML-TOMCAT dataset is recommended only for stratospheric values?
• It is not clear enough where the focus of the ML-TOMCAT dataset is placed. The title of the manuscript suggests that it would the stratosphere, and in the summary section the recommendation is given to only use the dataset between the tropopause and 0.1hPa, but the creation of the dataset is described for all pressure levels between 1000hPa and 0.1hPa. This is confusing. Please add more explanations why either the tropospheric levels are necessary for the creation of the dataset or why they were created but should not be used.

More specific comments:

• Line 18: Should be “…within uncertainties of the…” rather than “…within uncertainties in the…”?
• Line 29: Should be “…similar or larger magnitude…” rather than “…similar magnitude or larger…”?
• Line 37: Word missing? “…which could variability in ozone trends.”
• Line 66: Remove “s” from “profiles”
• Line 102: Add “the” between “…throughout troposphere.”
• Line 117: “Random Forest” has been introduced already and does not need to be written out here anymore.
• Line 121: The selection of the years to train the machine-learning model (1991-1998 and 2005-2016) seems very random here and is not explained. It is explained later in the manuscript, but I think it would be good to add an explanation here as well (not just that it is a 20-yr period, but WHY exactly these 20-yrs were chosen).
• Line 129: Please explain here what you mean with “passive ozone”
• Line 169: Here it states that the climatology was calculated for the period 2006-2020, but the figure caption (Figure 1) states that the climatology was calculated for the period 2001-2020. Which is correct?
• Line 201: the sentence part about positive and negative ozone biases is confusing here since it refers to Figure 1 and the paragraphs explains details of Figure 2. Maybe add a reference to Figure 1 at the end of the sentence?
• Line 209: What could be the reason for the significant HCl coefficients in the tropical lower stratosphere? You mention that they are present there but don’t offer an explanation.
• Line 225: Remove the comma after “Since, …”
• Line 246: Reference to Figure 1 is not correct, I guess, since Figure 1 does not show ML-TOMCAT data.
• Line 321: “;” between the two references should probably be an “and”
• Line 322: I guess you mean “North Polar region” instead of “North Pole”? It might be good to mention here as well which latitudes are covered by the North Polar regions.
• Line 323: I think “South Polar region” is more appropriate than “South Pole”.
• Figure 2: The differences in the blue color range a very hard to distinguish. It might be helpful to change the color scale somewhat to help the reader understand the discussion about these differences from Section 4.2.
• Figure 5: The title says “w.r.t. SAGE-CCI-OMPS”, but it should be “w.r.t. GOZCARDS”
• Figure 9: Last line of the figure caption: the “)” is missing after “70°S”.