Application of Machine Learning Methods to Neutron Transmission Spectroscopic Imaging for Solid–Liquid Phase Fraction Analysis

Round 1

Reviewer 1 Report

We read the paper on the application of machine learning to neutron transmission with some interest. I am a specialist in neutron methods with no machine learning background.  For the general reader the paper is well thought out. The results of the solidification interface seem consistent with the machine learning derived parameters. Is it possible to derive suitable uncertainties in this method ?  For the general reader section 2 is a little longwinded and unclear for a non specialist I do not follow the reduced dimensionality arguments and this should be made clearer how this is done.

I believe that the english could be improved .

Examples (I do not include all) ‘it becomes that the machine learning can be applied’ -Line 18

‘made prediction directly’ -Line 165

In figure 10 what are the colours representing? Is there a significance in the three data sets plotted ?

What is the resolution of the instrument RADEN ? Both wavelength and spatial.

Author Response

Thank you for your review and valuable suggestions.

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

In this study, Kamiyama et al. proposed a machine learning framework for solid-liquid phase fraction analysis from neutron transmission spectroscopic imaging. Both supervised and unsupervised learnings have been assessed to produce a promising performance. However, there are some major issues that need to be addressed in this manuscript:

1. The authors mentioned "we adopted the KNN algorithm, which needs less cost of computation", but in my opinion, ETR and SVM are both not too costly. The authors should have a comparison to see differences in time-consuming among different methods.

2. More literature reviews should be added to show related works on this problem.

3. The authors assess the model performance using RMSE, it is necessary to provide more regression metrics i.e., MAE, MSE, or R2.

4. The authors should have validation data.

5. Why did the authors not compare their predictive performance with previously published works on the same problem/data?

6. There must have some baseline comparisons among different methods to show the significance of their selected method.

7. Besides 'k', the authors should also tune the distance learning function, which is important in kNN algorithm.

8. More discussions and limitations should be added.

9. Machine learning has been used in previous studies i.e., PMID: 33036150, PMID: 33717604, PMID: 33735760, PMID: 33664984, and PMID: 31518859. Thus, the authors are suggested to refer to more works in this description.

10. Source codes should be provided for replicating the methods.

Author Response

Thank you for your review and valuable suggestions.

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

I have no objections to the main content. The methods and results are clearly explained and presented. However, the quality of language is the major issue. On the whole, it is far from unreadable, but there are lots of words, phrases and word forms which make the paper more difficult to follow (as if the words were selected based on google-translate instead of their subtle but appropriate meaning). I give here only a few examples. In any case, the paper must be read over by someone with firmer grasp of English, not only of the English vocabulary, but of English phraseology.

- Line 18 in the abstract: “it becomes that the” (this is only one example of the un-English phrase; there are more mistakes even throughout the abstract).

- Just an example of subtle nuances in meaning: at line 102 “plural spectra” should be “multiple spectra”. These seem like small differences, but they scream to the reader.

- I would prefer that the word “learn” (when used in relation to neural networks) be avoided and that “train” be used instead (instead of “neural networks learn” I prefer “the neural networks are trained”).

- A general comment: “data” is plural. So “data are”, instead of “data is”.

- When describing the functionality of neural networks (in multiple places), the phrases are set in quotation marks, such as at the bottom of page 3 (and elsewhere). This sounds either as if it is meant in some metaphorical sense (which it should not be, since this is an exact description) or as if quoting some “universal instruction manual for neural networks”. The second case leaves the impression that authors have no understanding of what they are doing (as if they don’t actually “believe” in the neural networks), and are only blindly quoting some dogmatic source. Please rewrite this (at the very least, remove quotation marks).

- There are some phrases or even whole sentences that are completely incoherent. Some examples are:

[line 165] …data is learned and made prediction directly… (this is also an example of “data” used in singular)

[line 302] …model is not failed about he overlearning

[line 245] Note that each principal component is computationally found, then the unit vector cannot be defined. (I cannot even extrapolate what this is supposed to mean.)

 

I repeat, these are only a few illustrative examples and the whole paper must be systematically read over and corrected.  

Author Response

Thank you for your review and valuable suggestions.

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

My previous comments have been addressed.