README.txt

@@ -29,11 +29,7 @@ The corresponding mask is also provided in the data frame.
 We also apply the robust scaler that sklearn provides to scale our data set.
 
 The used data set is available for download at 10.5281/zenodo.7568673 . This step is necessary for all the further steps.
-------------------------------------------------------------------------------------------------------------------------
-Acknowledgments
 
-This work was supported by the Deutsches Zentrum für Luft- und Raumfahrt (DLR) as SOHO/CELIAS 50 OC 2104.  
-We further thank the science teams of  ACE/SWEPAM, ACE/MAG as well as ACE/SWICS for providing the respective level 2 and level 1 data products.
 
 ------------------------------------------------------------------------------------------------------------------------
 
@@ -58,7 +54,7 @@ For the resulting clusterings the obtained prediction based on the possibly down
 The instead of the 100 trials the default number of trials for the downsampled version is 5.
 If you want change it please change the number of trials in kmeans.py in line 2410.
 When using the downsampled version please note that when the fraction is below 10% the results can differ from the results in the paper.
-The interpretation provided in the section discussing a possible interpretation of the different solar wind types is most likely not the same for different versions of sklearn or different fractions.
+This holds espacially for the interpretation provided in the section discussing a possible interpretation of the different solar wind types.
 When training on the reduced data set, please make sure to use the same fraction in all further steps. 
 
 
@@ -97,31 +93,31 @@ When training on the reduced data set, please make sure to use the same fraction
    For the seven cluster case please call option 70 to 73 and 75 to 80 : it splits 100 Monte Carlo runs into 10 pieces
    If you only want to do a reduced version with 3 Monte Carlo runs
    please call upon:
-   the option 103 for three clusters (using this option you can skip step 8)
-   and option 107 for seven clusters (using this option you can skip step 8)
+   the option 103 for three clusters
+   and option 107 for seven clusters
    With this option, there is no need to do an extra evaluation step because it is included in the respective option
    Note: when using the downsampled version: the scores are NOT calculated on the whole data set in this case.
          please make sure that option 7/3 and 43/42 are run before (if downsampled version is used please make sure to run those options also in the downsampled version).
 
 8. For the evaluation of the 100 Monte Carlo runs please use option
-   34 for three clusters (requires option 3, 42, 30, 31, 32, 33)
-   74 for seven clusters (requires option 7, 43, 70, 71, 72, 73, 75, 76, 77, 78 ,79, 80)
+   34 for three clusters
+   74 for seven clusters
    These steps are not necessary for if you used the reduced option with 3 Monte Carlo run.
  
 
 8. To recreate the elbow plot call on option from 2 to 13 to do the bigExperiment for the respective number of clusters.
    Here, the number directly defines the number k for k-means.
 
-9. To evaluate the the results call on option 90 (requires option 2, 4, 5, 6, 8, 9, 10, 11, 12, 13)
+9. To evaluate the the results call on option 90
    here the experiments except for 3 and 7 clusters are evaluated and big plots are created
 
-10. call with the option 45 to create the elbow plot as seen in the paper (requires option 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 42, 43, 90)
+10. call with the option 45 to create the elbow plot as seen in the paper
     Note: when using the downsampled version: the Calinski-Harabasz and Davies-Bouldin scores are NOT calculated on the whole data set but on the downsampled version.
 
-11. For the scores in the text call option 101 (requires option 3, 7, 42, 43)
+11. For the scores in the text call option 101
     If the clusterings are calculated on the downsampled version please make sure to use here also the downsampled version. 
 
-12. For the scores in the table call option 102 (requires option 3, 7, 42, 43)
+12. For the scores in the table call option 102
     If the clusterings are calculated on the downsampled version please make sure to use here also the downsampled version. 
 
 Reminder how to call an option. Chose a terminal of your liking and type: python kmeans.py <number of option> you want to do for

paper/aa.bib

@@ -2704,21 +2704,3 @@ booktitle = {American Institute of Physics Conference Series},
   year={2002},
   organization={IEEE}
 }
-
-@MISC{Teichmann2023-bu,
-  title     = "Feature selection with k-means on solar wind data",
-  author    = "Teichmann, Sophie and Heidrich-Meisner, Verena and Berger, Lars
-               and Wimmer-Schweingruber, Robert",
-  publisher = "Zenodo",
-  year      =  2023
-}
-
-@MISC{Berger2023-gd,
-  title     = "Solar Wind properties measured with instruments on the Advanced
-               Composition Explorer ({ACE})",
-  author    = "Berger, Lars and Heidrich-Meisner, Verena and Teichmann, Sophie
-               and Wimmer-Schweingruber, Robert F",
-
-  publisher = "Zenodo",
-  year      =  2023
-}

paper/coverletter.tex

@@ -1,417 +0,0 @@
-\documentclass{article}
-\usepackage{graphicx, natbib}
-\usepackage{psfrag,pstricks,epsfig,tabularx}
-\usepackage{epstopdf}
-\usepackage{hyperref}
-\newcommand{\markme}[1]{#1}
-\newcommand{\ac}{a_{\textrm{col,p-p}}}
-\newcommand{\Tp}{T_{\textrm{p}}}
-\newcommand{\verena}[1]{\textcolor{blue}{#1}}
-\newcommand{\reply}[1]{\textbf{Reply:} #1 \newline
-}
-\newcommand{\commentA}[1]{{\bf Comment: }
-
-#1 \newline
- }
-\newcommand{\commentB}[2]{{\bf Comment: \textit{#1}}
-
-#2 \newline
- }
-\begin{document}
-
-
-\section*{Cover letter for the revision of manuscript number AA/2019/37378}
-
-Dear Michael Balikhin, Dear Referees,
-
-Together with this cover letter, we resubmit the revised manuscript
-entitled ``Influence of solar wind parameters on unsupervised solar wind classification with k-means''.
-We would like to thank you and the referees for the constructive, helpful and
-detailed comments.  Based on these comments, we have made changes to
-the manuscript, which are detailed below. All changes are highlighted
-with boldface in the newly submitted manuscript.
-
-The most important changes are:
-\begin{itemize}
-\item
-\end{itemize}
-
-We are very grateful for the thorough and critical comments from the referees. A detailed reply to both referee reports  can be found below.
-
-With best regards,
-
-% hier kommt dann dein Name hin :)
-
-
-\section*{Detailed changes and reply to the comments of Referee 1}
-
-\commentA{This paper discusses the application of an unsupervised machine learning method using k-means applied to solar wind observations from the ACE spacecraft to identify different types of solar wind. Essentially, different clusters of solar wind with similar properties are identified in the analysis using various combinations of up to seven solar wind parameters. In the paper, results for three and seven clusters are presented. The "similarity" (quantified using various scores) between the clustering obtained with different parameter combinations and that obtained when all seven parameters are included is used as a criterion for assessing how successfully the combinations of parameters classify the solar wind types. The technique is interesting and appears to produce a classification that is physically meaningful. For example, the three cluster method identifies slow and coronal hole solar wind and the compressed plasma inside stream interaction regions, though I note that such a classification is hardly original and may be inferred from examining the original data (e.g. in a 1971 study noted below).}
-
-\reply{todo}
-
-\commentA{
-I do have some reservations:
-The authors should be aware that the reader may not be familiar with the ML method described and with the specific Python program used. There are places where details are not sufficiently explained and others where program settings are given (maybe details of the python code could be in an appendix?)}
-
-The figures and table are generally not described adequately in the text, though the captions are fairly informative. In particular, often specific features in the figures are mentioned, but there is no guidance provided to the reader on how to locate and identify these features. This is a particular problem since many of the figures are complex and of a format that is likely to be unfamiliar to most readers. The figures/tables are also placed far from the related text for no clear reason.
-
-The results are described in considerable detail, e.g. for different combinations of solar wind parameters, and can become repetitive. It requires some motivation to read through all the details, and the more important results are not always clear. Similarly, the "Discussion" section goes through the influence of the different solar wind parameters on the results. A persistent theme here is the need for more accurate solar wind measurements, and there are critical comments on current instrumentation. Again, this became rather repetitive, and I question whether it is particularly relevant as a discussion point for this paper. I would suggest cutting some of this discussion and focusing more on the results obtained in this paper.
-
-I was disappointed and surprised that, unlike most solar wind classification schemes, the authors have chosen to remove ICMEs and therefore focus just on analysis of the background solar wind. They do not really discuss why they do this, but this artificially removes an important class of solar wind structures that have properties different from those of the ambient solar wind and are major drivers of space weather at Earth. Identifying clusters associated with ICMEs would presumably be of value in identifying these structures in real-time data. Perhaps the authors will consider including ICMEs in future work? A major conclusion is that certain parameters are more important in this classification (in particular the density) but would this also hold in a more realistic scenario where ICMEs are not excluded by some prior analysis? Also, the authors use an ICME catalog to remove these intervals, but it is possible that some events might have been missed in the catalog, Also, shocks and post shock sheaths that are not followed by an ICME will still be included in the analyzed periods, though maybe they are sufficiently rare that they do not influence the clustering.
-
-The emphasis in the paper is on comparing the results for different combinations of solar wind parameters rather than on understanding the implications of the results for solar wind physics (as the authors comment in line 451). Personally, I would have been more interested in knowing more about whether this method provides any new insight or just uncovers features that are known from previous studies.
-
-In summary, the paper demonstrates an interesting method of classifying solar wind using ML, but I suggest that some of the discussion of the results could be trimmed and be more focused on the results of this study rather than regretting that we don't have ideal measurements to work with.
-
-I also have a number of specific comments that should be addressed in a revision.
-
-\commentB{Line 12}{Say what specific measurements are required. But I suggest de-emphasizing this aspect of the paper.}
-
-\reply{We added:
-  \begin{quote}\it \markme{... , in particular for the proton temperature and the charge state composition.}
-  \end{quote}
-
-
-  \verena{we need to think about this. Maybe: keep this here, but repeat it less often in the paper?.}
-
-The noise resulting from the measurement uncertainty inhibits the accuracy of any solar wind categorization. This is not intended as a critic on the available instruments, but rather to continue the trend of improvement towards higher wquality measurement already implemented by the respective modern instruments in Parker Solar Probe and Solar Orbiter.  
- }
-
-\commentB{Line 16}{Suggest something like: based on properties that are, however, affected by processes during transport from the Sun.}
-
-\reply{Thank you for the comment! We changed this to:
-
-  \begin{quote}\it
-    Most available solar wind classifications focus on the solar origin, based, in part, on properties that are, however, affected by processes during transport from the Sun.
-  \end{quote}
-
-  Since the charge state composition which is also frequently used for solar wind classification is not affected during the transport from the Sun. However, all proton properties and the magnetic field strength are. }
-
-\commentB{Line 31}{Is there a limit on the length of the summary? This seems longer than most I've seen.}
-
-\reply{There is indeed a limit on the length of the plain language summary which was however stated during the submission process. We have shortened the plain language summary further which is now well below the limit of ...}
-
-Line 35: These properties
-
-Line 39: "not influenced by..." This isn't true - these ions are subject to effects such as expansion, etc. However, this comment presumably refers to their charge states, which should be made clear.
-
-Line 42: "darker" in what sense. Also, high speed streams were associated with coronal holes long before charge state observations were made, for example by field line mapping and comparing the magnetic field direction in coronal holes and the associated streams.
-
-Line 43: "The corona..." The solar wind is the expansion of the corona, so maybe this comment could be incorporated into the first sentence of the summary.
-
-Lines 45-7: There has been a lot of previous work on classifying SW structures from such properties, so clearly they are known to be of value. The new feature of this work is the application of ML.
-
-Line 48: What bias?
-
-Lines 69-71: Maybe this sentence referring to charge states could be moved to/merged with line 83?
-
-Line 70: charge states of the ions observed
-
-Line 72: "the majority" - they are the majority ion species, but don't forget electrons.
-
-Lines 73-74: Suggest: and proton temperature, as well as the magnetic field strength... [The magnetic field isn't a proton property].
-
-Lines 84-91: Suggest breaking up this long sentence with multiple parentheses.
-
-Line 91: "In the absence of ..." - Further context could be provided for this comment. I suspect this refers to the partial failure of the ACE/SWICS instrument in 2011. Also, if I recall, Xu and Borovsky didn't use charge states in their classification because of the lack of SWICS data at the time they made their study but so that they could use the more extended set of standard solar wind parameters for their analysis as charge states are only available for a limited period of near-earth solar wind observations.
-
-Lines 108-9: References to support these statements?
-
-Line 114: I'm not sure that a field peak is a typical signature of the interface. Usually the magnetic field profile in SIRs is quite variable. Also, the density tends to decrease at the interface while the temperature and speed increase; there are also compositional signatures of course.
-
-Line 115: "an acceleration region" -what is being accelerated? I presume that this refers to energetic particles, but in the classic picture discussed by several of these references, particle acceleration occurs at the shocks bounding the SIR, not at the interface. However, some authors have suggested that particles may be accelerated at the speed gradient in the SIR. But this topic and SIR shocks aren't relevant to this study.
-
-Line 117: I somehow doubt that Rankine and Hugoniot, well before the discovery of the solar wind, discussed shocks formed at SIRs. In fact none of these references is particularly relevant to the generation of shocks at SIRs.
-
-Line 119; I found this paragraph somewhat confusing, complex (many details are noted but not explained) and disorganized. It doesn't convey to the reader the motivation for this study particularly well.
-
-Line 138: Why is a ML approach appropriate for this problem? What advantages does it have over previous methods?
-
-Line 146-7: "and in accordance with ..." Explain what this means. The reader may not be familiar with the papers referenced.
-
-Line 151: This seems inconsistent with previous statements explaining why the ML approach is used here.
-
-Line 177: When I first read the introduction, I was surprised that it doesn't mention ICMEs. I would suggest adding a couple of comments noting that just the background solar wind, excluding ICMEs, is considered in the analysis and explaining why.
-
-Line 185: This paragraph could be moved later. It seems unnecessary here - just continue with the description of the data sets. Also, I suspect that most of these points are made again later.
-
-Line 197: Why is this period chosen? Presumably the end corresponds to the problem with SWICS, but why start in 2001?
-
-Line 201: Unfortunately, SWEPAM does have instrumental issues and has degraded with time. Is this also a reason for ending in 2011?
-
-Line 204: The study mentions multiple times that certain parameters are subject to transport effects, but on the other hand, all the observations are made at 1 AU, so the sampled plasma has suffered similar effects (such as expansion) such that signatures of the solar source may still be evident.
-
-What are the averaging times of the data sets used? This is mentioned later, but it would be natural to include this information in the data description.
-
-Line 217: Brackets around references.
-
-Line 218: Delete "with the".
-
-Equation 1: Define the parameters in the equation. It's strange to have units in the denominator. There's a gap in the divisor line.
-
-Line 230: I don't really understand from this brief comment why ICMEs are excluded. In particular, why don't the same arguments (no temporal information and variable properties) apply say to high speed streams, which are also variable from event to event, and, as Figure 3 shows, even with the lack of temporal information, continuous high speed stream intervals can be identified.
-
-Line 234: Are ICME sheath regions also removed? i.e., does the preceding 6 hour buffer extend before the shock (not really necessary as there's no signal of the ICME ahead of the shock, except maybe some waves) or the ICME leading edge (in which case, prior to 6 hours could still include the sheath)?
-
-Line 239: "To estimate their impact on our results..." - the details of the analysis haven't yet been presented, and some related knowledge is needed to understand many comments in this paragraph. I suggest moving this detailed discussion to later.
-
-Line 244: Why is a 12 minute averaging period chosen for this study? Wouldn't the SW parameters on this time scale be significantly influenced by small scale structures/waves/turbulence that are unrelated to the large scale structures of interest in this paper? Would the choice of 12 minutes vs. say 1 hour averaging impact the results of this type of analysis?
-
-Line 264: The slow and dense streamer wind is unrelated to SIRs. However, it may be swept into SIRs.
-
-Line 265: My understanding is that the "streamer belt" includes the heliospheric currrent sheet and hence a sector boundary, so it's unclear why it is suggested to be equivalent to a pseudo-streamer (without a polarity change) here.
-
-Figure 1: Y-axis label: "Median"; 'k-means"; missing dashes in the score names. Suggest using initial upper-case letters.
-
-Figure 2: It's not clear what the left column of colored stripes indicates. Are these different combinations of solar wind data for seven different runs, or are the different data sets merged into one run? Does it represent a specific case in Figure 6?
-
-Line 270: This was the main message of Neugebauer et al. (2016), that the classification schemes they examined produced inconsistent results.
-
-Line 274: Is the number of clusters k decided "manually" at the start of each run or does it emerge naturally during the run? Explain how the clustering addresses the problem of SW classification. E.g., are the different clusters that result expected to represent different types of solar wind? (Also line 284.)
-
-Line 297: This explanation is unclear if you are unfamiliar with the code. e.g., what is a hyper parameter? Can it also be explained in plain language for the reader? Or maybe discuss the operational details in an appendix?
-
-Line 320: Folkwes or Fowlkes? Both are used throughout the text and figure labels. It should be Fowlkes. Please check the text/figures carefully.
-
-Line 335: exceptions discussed in Section
-
-Line 346: Should this be (see e.g., Figure 6) up to seven input parameters?
-
-Line 352: What does the "internal choice of starting points" mean?
-
-Line 354: "Different colors" where in Figure 2?
-
-Line 355: How and why is the data "shuffled"?
-
-Line 356: Explain Figure 3.
-
-Line 386: Would choosing samples at different solar activity levels (e.g., solar maximum vs. solar minimum) also influence the clustering? (Maybe less so since ICMEs are removed?)
-
-Line 400: It was, more generally, the limited charge state observations compared to other parameters rather than just the problem with SWICS that caused Xu and Borovsky to use just the three parameters not including charge states.
-
-Line 412: The other properties are also "source-dependent", but the relevant point is that the charge states are preserved during transport to 1 AU.
-
-Figure 5: Why is there little evidence of a peak for cluster 3 in all of the left-hand panels? Is this a problem with the color scale? Or do the values have a wide spread?
-
-Line 419: What are the two types here? Does this mean that, if the compressed solar wind in SIRs is identified as a separate type of solar wind, then the two types of fast and slow wind are insufficient.
-
-Line 423: Already mentioned in line 414.
-
-Line 432: median of each score is computed
-
-Line 433: The rest of this paragraph is unclear, e.g., where is an elbow "visible" (line 434)? "we identify the comprising elbow between 7 and 9 clusters (lines 436-7) " - Are these features the reader can find somewhere, perhaps in figure 1? What exactly is an "elbow" in this context?
-
-Line 446: The reader might note a mismatch between k=3 and the four types of solar wind in Xu and Borovsky, but presumably this is because ICMEs are ignored? - this might be noted.
-
-Line 448: The elbow in Figure 1 could be pointed out to the reader.
-
-Line 452: How do the results of these other papers relate to the current study, e.g., do they use the same analysis method?
-
-Line 460 and following. A lot of numbers (scores) are given in the text but it's not clear what the reader is supposed to deduce from them. Are they necessary?
-
-Line 460: ])
-
-Line 467: Explain what these figures show, which are also called out of sequence. Or maybe this comment should be postponed until these figures are discussed?
-
-Line 471: There could be more explanation of Figure 5 in the text. In particular, explain to the reader how to read the distributions for a particular parameter. It took some examination of the figure to figure out what it shows.
-
-Would a time series figure similar to Figure 3 also be useful to demonstrate how the clusters relate to these three types of solar wind?
-
-Line 485: What specific properties indicate that cluster 1 is polar coronal hole wind, and similarly for the other clusters (add some supporting references)? Are these inferred from Figure 5 or examination of Figure 3, or both? Although the authors don't consider this interpretation of the data a priority, from the point of view of a skeptical researcher reading the paper, indicating that this analysis does identify different types of solar wind similar to those found from previous studies is a strong selling point for this type of technique. So I would suggest expanding this discussion.
-
-Line 488: Cluster 4 apparently corresponds to "regions close to stream interfaces" - but the solar wind properties change abruptly across the interface so I would have expected the plasma on either side to fall into different clusters.
-
-Line 494 and following: The rest of the paragraph revisits some of these clusters in more detail. Maybe these details could be merged with and explain the initial description of the clusters earlier in the paragraph. Perhaps there could be a paragraph for each cluster e.g.,
-
-Cluster 1 is characterized by low O7/O6 charge state ratio, high proton speeds, high proton temperature and low proton density. These properties are consistent with polar coronal hole wind. (Add references to support this?)
-
-Cluster 2 contains all the very high proton density observations suggesting that it corresponds to the highly compressed slow solar wind. It includes only 1.71% of all the observations, making it the rarest of the seven solar wind types determined by K-means clustering.
-
-Cluster 3 has a higher O7/O6 ratio than the other coronal hole-associated clusters (1 and 7) and fits the properties of solar wind from equatorial coronal holes including Alfvenic slow solar wind. Etc
-
-Line 495-7 (and elsewhere): O7/O6 charge ratio
-
-Line 499: Some of the slow high density regions might be associated with the heliospheric plasma sheet and not necessarily due to compression. And compression of the HPS could lead to even higher densities.
-
-Line 509: Showing the solar cycle variation would also be a nice demonstration of the power of the technique to the skeptical researcher. I would encourage the authors to consider expanding this comment.
-
-Line 513: "Both for ..." Does the following comment only apply to the seven cluster case?
-
-Line 515: The meaning of this comment is unclear without any knowledge of the observations of Heidrich-Meisner et al., or a discussion of why the high Fe charge state should indicate the influence of "variations of the upper part of the solar corona" - the reason for this is not obvious.
-
-Line 521: "importance" in what sense?
-
-Line 545: Explain further why a comparison of the similarity to the reference results using all parameters is a valid way to assess the results using fewer parameters.
-
-Line 551: Please avoid statements like "it is obvious" when presenting results which are likely to be unfamiliar to the reader and without any explanation in the text of what the figure shows. Explain the results sufficiently so that the reader can identify this "obvious" result. I'm guessing that the reader is supposed to be comparing the distributions of all the various scores for different combinations of parameters, but some guidance would be useful.
-
-Line 553: "overlap" in what sense?
-
-Line 563: Explain to the reader where the one parameter results are in Figure 6. Presumably, they are in the region numbered 1 but if so, it's not obvious why the collisional age scores highest. Some more explanation/guidance would be useful.
-
-Line 567: If I understand the figure correctly, doesn't B also have a very low similarity?
-
-Line 568: It is unclear what this comment relates to. Is this also discussing some feature in Figure 6?
-
-Line 599: charge state parameters
-
-There is a lot of detailed information for the reader to digest in the sections discussing the different parameter combinations, but it's not clear what the overall message is, and the important points for the reader to note.
-
-Line 659: trials
-
-Line 672: show
-
-Line 673: This last sentence is unclear.
-
-Line 676: clusters
-
-Line 678: including both Tsw and nsw (?)
-
-Line 688: Not sure this makes sense: two parameter combinations containing "all but one" parameter.
-
-Line 689: Where can the reader find evidence for the "influence of the starting points" in Figures 6 and 7?
-
-Line 696 and following: Again, there's a lot of details, but it's not clear what the reader should take away from these results.
-
-
-Line 786: except for the combinations without $v_{sw}$ ...(?)
-
-Line 820: "systematic differences..." But this is also the case for other parameters, most obviously $v_{sw}$. Does the dynamic range of a parameter play a role in the "importance" of a parameter? For example, the speed may range over a factor of around two between slow and fast solar wind (say ~300-600 km/s), but the density may vary by a larger factor.
-
-Line 821: Is it true that density is a tracer for slow/coronal hole wind in Xu and Borovsky as they use combinations of a number of parameters and Tp (which is correlated with Vsw)?
-
-Line 823: There is only compression in SIRs, but there is rarefaction in the declining speed regions of coronal hole streams.
-
-Line 828: I don't follow why the results underline the importance of accurate density measurements since the successful analysis described in the paper is made with observations that potentially have the issues discussed here. This paragraph sounds like a general rant about measuring proton densities that isn't especially relevant to this paper. In particular, is the correction expected to be small since it's largely been ignored? Also, do these comments apply to the ACE data used here? Furthermore, note that Wind provides accurate density measurements from the radio instrument which can be used to check the calibration of the Wind plasma instrument.
-
-Line 847 and following: Again, is this discussion of the physics influencing the proton temperature relevant to this study?
-
-Line 856: It's been known for decades that the proton temperature is correlated with the solar wind speed, and that the temperature increases in compression regions. Figure 13 of Belcher and Davis, 1971, https://doi.org/10.1029/JA076i016p03534 identifies the regions mentioned in lines 586-8. Also, there have been more recent studies e.g., by Heather Elliot, assessing the influence of compression on the proton temperature.
-
-Line 860: Again, it's not clear why improving proton temperature measurements, is particularly relevant to this study which successfully demonstrates a classification using existing data.
-
-Line 867: Is it really true that the speed has been frequently used alone for solar wind classification. And this study unfortunately ignores ICMEs which can be a source of fast, transient, non-coronal hole solar wind to be distinguished from coronal hole flows.
-
-Line 879: Where has the "perpendicular variability of the magnetic field" been used as a candidate for solar wind classification?
-
-Line 893: My understanding is that the differences in slow and coronal hole solar wind originate in their different sources (e.g., note their different compositions and charge states), not from collisions as implied here.
-
-Line 894: What does a compact velocity distribution mean? Is this just another way of saying "slow/low temperature solar wind".
-
-Line 896-7: Although a single parameter, the collisional age is just another combination of observed parameters. In Figure 3, it is evident that the collisional age is approximately anticorrelated with speed, which is expected from equation 1 since speed tends to be correlated with Tp, and anticorrelated with np. I don't see that it's really providing any new insight into solar wind structures.
-
-Line 915: Does the "stream boundary" mean the stream interface? if so, I don't believe it's accurate to say that the field strength peaks at this location. The behavior of |B| seems much more variable from event to event.
-
-Line 924: Explain what table 1 shows, and guide the reader to the results discussed here - there's a lot of information in the table!
-
-Line 944: To be more accurate, it's the ratio of oxygen charge 7/6 states that is used here.
-
-Lines 945-6: (Struck through O =?) the O7/O6 charge state ratio is more important than
-
-Line 947: Struck through O again.
-
-Lines 965-6: This conclusion is perhaps inconsistent with the first sentence of this paragraph.
-
-Line 980: Which are the "entropy based scores" in Table 1?
-
-Line 985 and following: Presumably this is related to the well-known observation that SW charge states typically change abruptly in the vicinity of stream interfaces at the transition between slow and coronal hole wind in SIRs.
-
-Line 1007: This variation in $T_{SW}$ is also dependent on instrumentation and data analysis, e.g., there are differences in the temperatures derived from ACE and say Wind plasma data. But these pleas for more accurate measurements are not especially relevant to this study.
-
-Line 1016: I don't get the point of this paragraph.
-
-Line 1022: Note also that one parameter is actually a ratio of charge states.
-
-Line 1025: We also know from many other studies that there are variations in the charge states and composition of different types of solar wind.
-
-Line 1029: What results presented indicate that the temperature is an "indirect proxy" for wave activity? Could this also be said for the solar wind speed which is closely correlated with the temperature?
-
-Line 1067: unsupervised machine learning method
-
-Line 1071: Maybe this comment on radial dependence could be moved until after the method and "our results" have been summarized?
-
-Line 1113: approximation to what?
-
-Line 1123: Give the sources (websites/URLs, etc) of the data (e.g., the ACE Science Center?)
-
-Line 1133: It's not clear what "respective" refers back to.
-
-Line 1135: Several references have "others" in the author list rather than et al. This reference also appears to be missing information e.g., Volume 415 of what publication? SOHO should be in upper case. (Note that not having " " around a title in BibTeX can remove the upper case characters. I suspect this is a problem for very many references, e.g., line 1172 as just one example.)
-
-Line 1146: Is more information than just "Zenodo" needed to find this data?
-
-Line 1163: Something strange here.
-
-\section*{Detailed changes and reply to the comments of Referee 2}
-The manuscript 'Influence of solar wind parameters on unsupervised solar wind classification with k-means' aim at identifying the most relevant solar wind parameters when a unsupervised clustering with k-means is performed. The authors make the important point that some solar wind parameters are related to the solar origin of the wind, while others are subject to transport phenomena during solar wind acceleration.
-
-The study is potentially interesting but, in my opinion, the methodology, results and conclusions are not strong enough to warrant publication in JGR.
-
-\commentA{The paper is extremely repetitive and not a pleasure to read. The explanation of results (page 17- 23) is very pedantic and excessively lengthy (everything could be summarized in a few tables). And, more importantly, the final discussion is not at all conclusive.}
-
-\reply{Thank you for your critical comments. We have extensively revised the manuscript to improve the readability and to sharpen the discussion.}
-
-\commentA{I have major issues with the methodology that seems to follow a brute force strategy. Although the authors claim to not use any ground-truth for solar wind classification, they eventually use their 7-parameters classification as ground truth (or a statistical average of an ensemble of k-means runs). Even assuming that that is a sensible choice, they end up running thousands of k-means runs (with and without noise, etc.) In my opinion the whole problem could have been more elegantly solved by using sensitivity analysis tool (the sensitivity of each individual parameter is, at the end of the day, what the authors are looking for). Moreover, once the centroids of the "ground-truth" have been established, it would be relatively simple to look (analytically) at the relative importance of each parameters, since k-means uses a simple Eucledian distance.}
-
-\reply{
-
-  Yes, indeed we a treating the 7-parameter classification as a stand-in for the ground truth. We emphasized this more strongly in the revised manuscript.
-
-  We are interested in a situation where the results of the solar wind classification depends on several solar wind parameters that are not independent from each other. Thus, approaches that assume independent variables (as simple sensitivity analysis, feature selection or feature importance approaches do) are not appropriate for our case. The approach that we follow here, i.e. testing all sub-sets of parameters individually, is also frequently applied for feature selection of feature importance \verena{(add cites :P)}. Other approaches that are applicable to dependent variables tend to be more complicated and time consuming (for example, evolutionary algorithm) and are intended for cases where a brute-search approach is not feasible since instead of a few (in our cases seven), hundreds or thousands of features are considered.
-
-  Sensitivity analysis deals with analyzing the effect of uncertainty of each feature on the overall model. Most approaches are interested only in local sensitivity, that is the effect of the uncertainty of a single feature. Even most global sensitivity analysis methods assume independence of the different features from each other. 
-
-  
-  We added a section discussing alternative approaches to the introduction.
-
-  Disregarding the Monte Carlo computations for the moment, we require 127 runs of k-means, which takes on a \verena{add computer and running time}. This would increase considerably, if we added more than seven features.
-
-  The additional (thousands) runs are solely for the purpose to derive errors that indicate the effect of the measurement uncertainty. For this step a sensitivity analysis would be appropriate. However, we would still need to deal with dependencies between the individual features and would require a similarly involved/ comprehensive approach like, for example \verena{select something ..} It is also feasible to compute the error propagation through the k-means updates analytically. However, already with seven parameters that can all depend on each other this is a quite lengthy operation. 
-
-
-  
-
-}
-
-Instead, the authors choose different comparison criteria, which are not very well introduced or explained, and end up giving contradictory results.
-
-Minor comments:
-The English needs to be polished and some sentences/concepts are repeated too many times. E.g. line 35: cancel "previously mentioned"; line 42: hole --> holes; line 71: remain --> remains; line 72: is --> are; line 108: need --> needs, etc.
-
-Line 51 and elsewhere: use consistently element symbol or word.
-Line 164: Bloch et al. did not use k-means but Bayesian Gaussian Mixture;
-
-Eq.1 : write units in parenthesis, after formula
-
-Line 239 and ff: are the different measurement uncertainties considered independent? This should be explained
-
-Figure 1: I do not see any error bar
-
-Line 297 and elsewhere: I suggest to separate the description of hyperparameters with the specific python package used for this work (which could become obsolete in the future). In fact, I suggest to dedicated a separate section (Open Research) to mention all packages and precise values of functions used.
-
-Line 317 and ff: all of these score need to be better defined.
-
-Line 354: isn't k-means insensistive to the order of the data set? Why shuffle it? (Unless you work in mini-batches?)
-
-Line 377: not clear where the standard deviation comes from? Is there an assumption of Gaussianity?
-
-Line 378-380: Not clear.
-
-Line 387-392: Once again, this goes into a dedicated section.
-
-Line 394-407: This discussion is oddly placed here. It goes before in the introduction.
-Line 408 and ff: This is a repetition of concepts already discussed.
-
-Figure 5: The hotmap does not work very well here. These are all 1D probability density functions that can be plotted as lines.
-
-Line 429: Please discuss the shorcomings associated to the elbow method (see https://arxiv.org/abs/2212.12189)
-
-Line 437: ... and in fact I do not clearly see any elbow here!
-
-Line 456-467: much better to put all of this info in a Table.
-
-Figure 6: there is too much information on this figure to be understandable. As a minimum I suggest to sort the column within a given number of parameters with respect to one of the criteria (but then again I also suggest to not use these criteria)
-
-Line 531-538: this paragraph can be removed.
-
-\end{document}

paper/discussion_agu.tex

@@ -129,6 +129,7 @@
     \label{tab:scores}
     \end{table*}
     
+ 
   \subsection{Implications for the proton speed $\vsw$ }
     Compared to the proton density and the proton
     temperature, the proton velocity has proven to be a useful

paper/methods_agu.tex

@@ -121,4 +121,4 @@ contains with the $\nOsix$ the most frequent solar wind ion (heavier than He).
    
    We tested the stability of the resulting clustering on randomly selected 10\% of the complete data set and found that the resulting similarity to the full data set is on average with $\sim 90-95\%$ very high. The emerging deviations from the results based on the complete data set depend on the parameter combination. If we consider more clusters, we see a higher deviation from the results based on the complete data set. The resulting $1\sigma$ corresponding percentiles denote a bigger distribution as the results based on the measurement uncertainties. But since we only considered 10\% of the full data set we observed a remarkably high stability. If the reduced data set is biased (e.g. leaving out data point with low counting statistics of Oxygen ions) we see more systematic changes in the clusterings.
    
-  We provide the code we used to calculate the clusterings and generated the plots online \cite{Teichmann2023-bu} . The data set including the ICME masks and the respective errors for $\mcsFe $ and $\nO$ is published online \cite{Berger2023-gd}.  All figures in this study are created with matplotlib version 3.3.4 and the respective median and percentiles over all trials are shown wherein the $14.9$th- and $85.1$th-percentile are chosen as corresponding to a $1\sigma$-interval.
+  \markme{We provide the code we used to calculate the clusterings and generated the plots at ... . The data set including the ICME masks and the respective errors for $\mcsFe $ and $\nO$ is published at ... } All figures in this study are created with matplotlib version 3.3.4 and the respective median and percentiles over all trials are shown wherein the $14.9$th- and $85.1$th-percentile are chosen as corresponding to a $1\sigma$-interval.

paper/test.tex

@@ -1,11 +0,0 @@
-\documentclass{article}
-\usepackage[utf8]{inputenc}
-
-\title{Bericht Dynamics}
-\author{Sophie Teichmann}
-\date{January 2023}
-\usepackage[usenames,dvipsnames]{color}
-\begin{document}
-hello \color{green} world
-
-\end{document}

reduced_kmeans.py

@@ -35,8 +35,8 @@ pgf_with_latex = {
   "pgf.texsystem": "pdflatex",
   "pgf.preamble": "\n".join([
    r'\usepackage{color}',
-   r'\usepackage[notextcomp]{stix}', r'\usepackage{siunitx}',r'\usepackage{amsmath}', r'\DeclareUnicodeCharacter{2212}{\ensuremath{-}}',
-   r'\definecolor{orange}{rgb}{0.93, 0.53, 0.18}',r'\definecolor{brown}{rgb}{0.59, 0.29, 0.0}',r'\definecolor{purple}{rgb}{0.5, 0.0, 0.5}',
+      r'\usepackage[notextcomp]{stix}', r'\usepackage{siunitx}',r'\usepackage{amsmath}', r'\DeclareUnicodeCharacter{2212}{\ensuremath{-}}' #,r'\usepackage[utf8]{inputenc}'#\DeclareUnicodeCharacter{2212}{-}'
+  ,r'\definecolor{orange}{rgb}{0.93, 0.53, 0.18}',r'\definecolor{brown}{rgb}{0.59, 0.29, 0.0}',r'\definecolor{purple}{rgb}{0.5, 0.0, 0.5}',
    r'\definecolor{cyan}{rgb}{0.0,1.0,1.0}',r'\definecolor{ngreen}{rgb}{0.133,0.545,0.133}',r'\definecolor{pink}{rgb}{0.99,0.412,0.706}',r'\definecolor{yellow}{rgb}{0.9,0.9,0.00}'])
 }
 rcParams.update(pgf_with_latex)
@@ -2107,6 +2107,7 @@ def pre_plot_big(parameter_combi = ("tsw",),nclusters = 3, clabel = '_3', size =
     fig, axs = plt.subplots(1,3, figsize=(16,8), gridspec_kw= {'width_ratios': [42, 1, 7]}, sharey = True)
 
     for i in range(len(sclist[0])):
+        print(sclist[i])
         axs[0].scatter(xlist,sclist[:,i],marker=markerlist[i])
         error = percentilelist[i]
         upper_error = error[1]-medianlist[i]