ANN-Based Inverse Goal-Oriented Design Method for Targeted Final Properties of Materials

Round 1

Reviewer 1 Report

This paper presents ANN-based Inverse Goal-oriented Design Method for Targeted Final Properties of Materials. Overall this paper explores the design space exploration framework to predict performance, properties, microstructure and processing. Details for one example are presented. For potential reader benefits:

1. Additional figures are suggested with ANN schematic for each stage. At this moment, this information is embedded in Figure 4.
2. More details on ANN structure.
3. Predefined acceptance criteria should be considered instead of “comparison of the predicted output with the actual values present in the dataset shows that our proposed ANN model has acceptable generalization capability”.
4. There is not enough discussion on end material property estimation. Presently only hardens and tensile strength are included as the input parameters.

Author Response

Please see the attachment

 

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper (ANN – Based Inverse Goal – Oriented Design Method for Targeted Final Properties of Materials), implements Machine Learning (ML) methodology, specifically an artificial neural network with a sigmoid activation function, in order to correlate the mechanical properties of a Boron steel with the heat treatment that its subjected to and thus its microstructure. In my opinion, the manuscript can be considered for potential publication after the following comments are addressed:

Article – Review

1. The “prediction number” in the provided figures isn’t consistent across the graphs making it impossible to compare/ view as a whole.
2. The used “statistical techniques” aren’t named nor described. The general mention of “some statistical techniques” is inadequate. Please provide more information in the text.
3. The acceptable margin of errors should be mentioned and be substantiate.
4. The most recent references are from 2018-2019 (3 papers). Isn’t there available any work in literature after that?
5. The accuracy of the model isn’t directly compared with the physical models so the argument about its improved general accuracy. Please provide some comparative diagram/table to further support your findings.
6. There are zero (0) tables in the manuscript despite naming 14 of them in the text. Where are the tables?
7. The labels/ legends of the graphs are unreadable. Please provide graphs and figures of higher quality and bigger fond size.

Specific comments

1. In the text there are many instances that the “traditional” modelling techniques are referred to (almost exclusively) as “incorrect, incomplete and infidel” that claims needs to be supported with the proper analysis and references in order to be accepted as is.
2. In the end of the second paragraph (regarding the exclusion of the MPU from the study) it should be explained how this is expected to affect the results.
3. In the second section, when referring to forming the term “macrostructural transformation” is used. Are you referring to the deformation of the part?
4. In the second paragraph of the section 2,
   1. The term “distribution of the microstructure” usually refers to grain – size distribution. However, this seems to not have been taken into account. Is the term here referring to something different?
   2. The sentence “Boron steel (22MnB5) in the austenite phase has low flow stress at elevated temperatures.” Implies that other steels may behave differently which is incorrect. Please rephrase.
5. In the section 3.B.Step 1: Please explain the term physical simulations in the text
6. In the section 3.B.Step 2: The first sentence is generic and doesn’t provide any information about the methodology.
7. In the section 4.1: If the needed data cover all the scenarios and are un-biased, why are the ANN needed for decision making?
8. In the section 4.1.2: There is an inconsistency with the names of the data sets.
9. In the section 4.1.3: “the rule of thumb method, which says “the number of neurons in hidden layers should be between the size of the input layer and the size of output layer”.” Please provide reference or state that it’s an assumption.
10. In the section 4.1.4: The number of combinations (c) is claimed to be nxn. Why isn’t the statistical equation used?
11. In the section 4.1.4: Please provide more information on how the number of iterations is chosen and clarify if the “stopping criteria” refers to convergence of the model or as an safety measure in case of non-convergence.
12. In the equation (6) why isn’t the equation for feature scaling (provided below) used?
13. In the section 4.6: the phrase “For each individual phase fraction, we multiplied its hardness by its percentage” is an assumption since the effect can be non – linear.
14. Page 15/19: If the target is to achieve maximum martensite why isn’t the case that the martensite fraction is less than 1 (given that we can heat the material high enough so that the hot pressing will be finished above the A₃). Please explain in the text

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper presents an interesting and new topic which concerne the topic. It is well developed, interesting, with a good methodology that support conclusion. My advice is to include a research question which is currently missing. This will also give further robusteness to the final conclusion.Good job and good luck.

 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 4 Report

Tables 1 a 13 are not attached.

There aren't Figures 12, 13 and 14. Figures 6-7 and 10-11 have incorrect order.

Figures 6 onwards present are of poor quality. The text is not well distinguished.

Figures 4 and 5 present the same information. Condense them into one

What is the prediction number indicated on the X-axis in Figures 6 onwards?

It is claimed that the use of ANN generates better results than methods based on empirical experience. But then it is proposed to make the training set data for ANN dependent on running simulations with their validity and accuracy unproven (Item 3.B.Step 1). This is inconsistent.

Item 3.B.Step 2: it discusses the use of statistical techniques to collect data for the ANN training set. You have to specify which one you are referring to

item 4.1.3: the way of dimensioning the hidden layers of the ANN could be improved. The Bayesian regulation method has been well known for 20 years.

Claims about the results are not substantiated, perhaps because they lack the tables cited.

To propose the addition of the results predicted by this methodology without experimental validation to extend the ANN training dataset is a very serious methodological flaw.

 

 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Dear authors,  

Thank you for in-depth response to all the raised points. In general, your answer were convincing. However, there are three points that we’d like to discuss further.

Article – Review

On point #1. Figure 7 and 9 don’t, respectively, add up to 100%

On Point #3. If there is no discussion about the accuracy of your model, how can it be proposed to be implemented from other research teams? It is within the scope of this work to quantify the error and discuss it.

Specific Comments

On point 1. The claim that the traditional, non – AI, computational methods are “incorrect, incomplete and infidel” is very bold. Either support this claim better with references in the text or omit it whatsoever from everywhere in the text. In case that you decide to support it, please, prefer to use references that are less than 10 years old.

Author Response

Respected Reviewer, I hope you are doing well. We really appreciate your time and comments on our paper. here are point by point responses in the attachment. Please see the Attachment

Author Response File: Author Response.pdf

Reviewer 4 Report

The legend of figure 8 should have the same format as the other figures.

Jeff Heaton's comment on the size of hidden layers is from a book published in 2005. Bayesian regularization answers that question (see "Bayesian regularization of neural networks", by Frank Burden and Dave Winkler, 2008). While Heaton proposes a pragmatic way of sizing hidden layers in his book, it is clearly not currently optimal. This should be made clear in the text. This comment does not discredit the results obtained in the article but the claim that it is the optimal procedure to work with.

One cannot add ANN results to the ANN training set for future training without experimentally validating those results (they are predictions, not experiments). To do so, as proposed in the article, is a very serious methodological error that discredits the proponent.

 

Author Response

Respected Reviewer,

I hope you are doing well. We really appreciate your time and comments on our paper. here are point by point responses. Please see the attachment

Author Response File: Author Response.pdf