Machine Learning with Gradient-Based Optimization of Nuclear Waste Vitrification with Uncertainties and Constraints

Round 1

Reviewer 1 Report

Thank you for submitting this work. I found the potential for the work and this type of paper focused more on application and accessibility something that is worth publishing. However, I have some reservations about the suitability and rigour of the demonstrations, in addition to the clarity of the statement of novelty.

In particular, since GEKKO is focused on dynamic optimisation, why is the problem shown not a dynamic one? Perhaps I am mistaken, but the problem appears to be one of optimisation over a single function without constraints (beyond bounds). To demonstrate the functionality, I feel as if the paper is incomplete with only a single, limited case study.

Additionally, there have been several efforts recently to include surrogate modelling within other optimisation platforms and these have not been explicitly mentioned or cited. I felt the statement of novelty (and I do believe that there is some novelty) was not very explicit and did not seem to make note of the significant efforts in the engineering and optimisation communities to perform optimisation incorporating surrogates.

Beyond these more significant concerns, I also found the following issues arise while reading the paper:

Abstract, line 4: replace ‘effort’ with ‘study’ or ‘paper’

abstract, line 8. Remove ‘from this endeavour’.

Well-written abstract that gets to the point. I advise adding a bit more on results (with quantification of improvements) and which techniques performed best.

P1. Line 23 – typo. ‘optimizations’ should be ‘optimization’.

P.2, line 38 – replace ‘don’t’ with ‘do not’. Remove word ‘actual’

Line42: replace ‘effort’ with ‘study’ or ‘paper’

You use ‘have not been thoroughly explored’ a few times. So what exploration has been done? What makes previous work not thorough. References and discussion around why this is a novel contribution should be included.

There is little to no mention of other method approaches for optimisation considering surrogates, apart from Gekko. There is a vast literature on this topic and it seems to be totally overlooked. A more convincing literature analysis and novelty statement is needed.

In Algorithm 1, there is inconsistency in k vs K.

Line 201 – misplaced word, ‘with’

Line 386 – ‘immediately and quickly.’ Can it be both? They are different.

Throughout the manuscript there is an inconsistent use of capitalisation. Sometimes neural network is capitalised, other times not, the same with many other terms.

389 - ‘However, due to tree-based regressors not being compatible with the framework of Gekko, this option and others were not thoroughly explored’. This surely means they were not explored at all. Why not say ‘were not implemented’?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper proposes the integration of several methods and integration of Python machine learning tools and Gekko. The authors failed to persuade the reviewer that the work is innovative enough to warrant publication in a journal. I suggest that the authors reduce the part of the paper in which they describe known methods such as Gaussian processes and ANN and focus more on physics-based learning and its challenges. The authors need to do a better review of the scientific machine learning field (not only Python related) and to identify novelties after performing a new literature review.

 

Minor comments:

The primary focus has been training and optimization of these algorithms, but little has been done to use trained ML algorithms to solve complex physics-based numerical optimizations problems.

->The amount of work is limited in comparison with training traditional ML algorithms.

A lot has been done both in Python and in Julia communities – please check scientific machine learning with Julia https://sciml.ai/. The solutions also provide uncertainties – however there is no discussion about them in the paper.

 

The purpose of this effort is to integrate these two tools: well-known and widely used machine learning algorithms and the Python-based optimization suite, Gekko.

->The integration itself does not warrant a journal paper.

 

A 95 % significance level indicates that 95 % of the true property values lies within the uncertainty boundaries.

             This is true for Bayesian approach while in traditional statistics it indicates that out of 100 experiments, 95 will have results within uncertainty bounds.

 

Uncertainty quantification is not properly defined and explained from the very beginning.

Conformity scores are not defined.

 

 

Why do you decide to use these uncertainty quantification methods?  

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Dear Authors

 

Initially, I congratulate you on the proposal of using ML to solve real problems within the industry. The proposed model adds to the myriad of models offered annually. I want the authors to highlight the advantages and disadvantages of similar models. I see great potential in this work. In addition to the above highlight, I suggest that the authors heed the following suggestions:

1.           The authors have presented their suggestions for future work in the conclusion, I suggest mentioning these future works in the abstract.

2.           I suggest the authors mention which are the solver alternatives to Gekko. And, what are the advantages and disadvantages of Gekko, concerning the others, that led the authors to choose this solver?

3.           Add a short paragraph at the end of the introduction section and give a short description of what the reader will find in the other sections.

4.           I suggest including a section where the authors present a more detailed overview of the application of ML in the studied waste area, its results, and what makes it different from the proposal of this work.

I wish the authors a good review.

 

Reviewer.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Many thanks to addressing my concerns. I think the paper is now suitable for publication.

Author Response

Again, thank you for your time. Your comments helped us prepare the manuscript for publication and made it a better article overall.

Reviewer 3 Report

Dear Authors

After reviewing the revised paper, I found that the authors have implemented the suggestions indicated by the reviewers. Thus, I do not see any other improvements in this version.

I congratulate you on the proposed model.

Best Regards

Author Response

Again, thank you for your time. Your comments helped us prepare the manuscript for publication and made it a better article overall.