Inverse Design of Aluminium Alloys Using Genetic Algorithm: A Class-Based Workflow

Round 1

Reviewer 1 Report

I think that this work is kind of useful to inform that making the inference model accurate is very important to have more practically optimal pareto front. Therefore, I think the authors did not convince me that the corresponding content meet up with the title “A Genetic Algorithm Approach for Inverse Design of Aluminium Alloys” where I think class-based should emphasized too.

Some comments are listed as following:

(1) Emphasis the novelty of the present work in the introduction.

(2) It seems that Multi-targeted random forest regressor is not suitable to work out a prediction capability towards too many features. Therefore, some proof should present that when class-based strategy is not used, simple machine learning model is not capable of fully capturing details of training dataset. Important bonus points are to be rewarded if the authors contribute more efforts on this point, e.g., train different models (SVR, GPR, NN) on the present dataset, and prove that simple models are hard to achieve good prediction accuracy when considering too many features.

(3) Please reveal the corresponding convergence sequence of the hypervolume so as to prove the convergence of a multi-objective optimisation algorithm.

(4) Explain why the pareto front of different classes in Figure 5 is discontinuous, i.e., for Class 1.

(5) Quality of figures should be improved, e.g., Figs. 6, 7 and 8. Could the notions (a), (b), … be placed at the upper left conner of the figure for the sake of clarity?

(6) Please reveal the implementation of the optimization processes, e.g., exact corresponding libraries being used, the machine learning framework and etc.

(7) Why the corresponding dataset leads to 8 classes classification result.

No Comment.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

The reviewed manuscript entitled “A Genetic Algorithm Approach for Inverse Design of Aluminium Alloys” presents the usage of various machine learning models to predict alloys with higher tensile strength and elongation. Although the method presented cannot substitute for experimental procedures, it can provide some framework for guiding future Al-alloys design. It would be very interesting for a potential reader if the authors could define and provide ideas for future plans to optimize properties for critical quantities such as temperature or aging time using the methods presented. In my opinion, the manuscript can be published in the Journal after minor amendments. The research is detailed and systematic, but it will benefit from resolving the following issues:

1.      Page 2, lines  63-64: What does it mean: “(...) which resulted in the design of alloy with 0.74% Mg, 0.78% Si and 0.37% Cu.”?

2.      Page 3: I recommend the authors give a short description concerning the differences between the classes and/or similarities of features inside the class. The citation to Ref. [27] (self-citation) is O.K., but not enough.

3.      General remark: The manuscript lacks references to figures and tables. Instead, reviewers find: “Error! Reference source not found.”, see: page. 3, line 100; page 5, lines 182, 185, 184, 190; page 6, line 195; page 9, line 229; page 10, lines 257, 266, 267; page 11, line 279; page 13, line 299. I hope that a potential reader will find citations to Fig. 1 (p.3, l.100), Table 1 (p.5, l. 182), Fig. 3 (p.3, l. 185), Fig. 4 (p. 3, l. 188), Fig. 5 (p. 9, l. 229), Fig. 6 (p. 10, l. 257), Fig. 7 (p. 11, l. 279), and Fig. 8 (p. 13, l. 299), etc.

4.      Why does class 7 possess the lowest MSE value for 1 feature?

5.      Why does class 4 show a different behavior than other classes? The MSE is (more-or-less) an increasing function of the number of features selected (starting from 2).

6.      Can the authors calculate (or estimate) values of the errors or uncertainties (for Figs. 5-8)? They can discuss it briefly.

7.      Minor errors/misprints found: the manuscript uses different fonts (in places), see page 6, line 200, page 11, lines 287-293.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

In the present article, the authors bring in discussion an interesting way to design Al alloys with targeted properties, in this case an optimal balance between mechanical resistance and ductility.

Nowadays, there are a huge quantity of data about the mechanical properties of metallic alloys, so that to find and to use a method by which a metallic alloy with desired mechanical behavior can be designed is a worthy to note research work.

When such issue is approached, first, it is important to establish what is meant by the ”design of an alloy”. Usually, it is understood the chemical composition of the alloy, but, at least to the same extent, the morphology and volume fraction of the microstructural phases must be considered. An alternative way to take in account the microstructure of the alloy, which is also used in the present article, is to consider the processing routes of the alloys which lead to microstructural changes. Briefly, the mechanical behavior of an alloy depends by the chemical composition and the microstructural features (morphology, crystallographic structure, volume fractions of phases etc.). In other words, ”to design on alloy” means to design both the chemical composition and microstructural features which occurring for this chemical composition, and which together provide the required mechanical behavior.

The comments are as follow.

a) the authors should include in the manuscript a brief discussion on the relationship between chemical composition-processing route-microstructure and mechanical behavior; the authors should point out whether the microstructural features were taken in account for data set.

b) in section ”Predicted compositions”, the authors should include for each class an example of chemical composition and processing route for which a ”optimized” strength/ductility ratio would be obtained.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Authors’ effort on providing more details are appreciated. In the revision, additional explanation is given for illustrating the “pronounced discontinuity is apparent in the Pareto front”, though the conclusion is likely not confirmed. I suggest that further classification should be applied within class 1, i.e., by clustering the corresponding dataset into two categories, and then computing the corresponding Pareto front for the newly classified categories. Therefore, the discovery will satisfy the objective of the manuscript, that accurate prediction model will likely result in more superior Parent front. I recommend acceptance of the manuscript if the such a simple but essential comment is reasonably addressed.

Author Response

Response: Thank you for your constructive comments and suggestions. We acknowledge your observation regarding the possibility of distinct clusters within Class 1, which could explain the pronounced discontinuity in the Pareto front. Previous research applying an iterative label spreading method for clustering did not reveal any distinct subclusters within this class (Bhat, et al., 2023). To substantiate these findings, we employed the k-means clustering algorithm and the Elbow Method to find the optimal number of clusters. The Elbow plot demonstrated a gradual decrease in distortion with increasing clusters, suggesting a lack of distinct clusters within the Class 1 data set.

Revision: We have added the Elbow plot to our supplementary information. We have expanded our discussion in the results section to address the comment raised.