A Deep Neural Network for Accurate and Robust Prediction of the Glass Transition Temperature of Polyhydroxyalkanoate Homo- and Copolymers

Round 1

Reviewer 1 Report

This paper discusses the prediction of Tg for PHA homo- and co-polymers by a DNN model. The authors' conclusions that the chemical structure and copolymer composition are important and that molecular weight and PDI do not contribute so much to the Tg values, are very well known. However, the use of a DNN model to show them seems to be scientifically a new result. Although I am not an expert in this field, I found it interesting as a new area of materials science and judged this is worthy of publication in Materials. The followings are some minor comments. Please consider them before publication.

• It would be helpful to show the chemical structure of PHA in the general form.
• PHA has 150 type monomers and all of them might be converted to binary digit vectors (128 length) as molecular fingerprints. However, it is difficult to see how molecular fingerprints are related to Tg. I think it would be better to first show the validity of the authors’ method on homopolymers and then discuss including copolymers. For example, it would be helpful to distinguish between homopolymer and copolymer data in Figure 3.
• It is generally considered that Tg is determined by the stiffness of polymer chains and the interaction between polymer molecules. In the Authors' method, these factors may be included in the chemical structure and copolymer compositions. However, in the case of copolymers, if a strong intermolecular interaction exists between A and B monomers (such as hydrogen bonds), Tg may increase in a specific manner. I wonder if this kind of interaction between hetero species could be incorporated or not by the authors’ method. It would be appreciated if the authors could explain it.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Dear Authors,

The article presents work on developing a Deep Neural Network for prediction a glass transition temperature of Polyhydroxyalkanoate. The article describes the process of creating a data-driven machine learning model in an extremely comprehensive manner. The authors, in a layman-friendly way, describe the various stages of network creation and training and provide a simplified interpretation of the results for the learning parameters. This is especially important for the use of data by chemists, biologists, materials engineers, etc.

The article is of high scientific value, but I have some comments on the text:

Major errors:

• The entire chapter of the introduction should be supported by more citations. The authors present many theses and statements that are not supported by appropriate citations or there are only single citations. There is a need to better position research in the current scientific context
•  Authors should clearly refer to their previous research presented in citation [14], and state the novelty of the research presented. After reading the quoted article, I state that the reviewed work has elements of novelty and is suitable for publication, but the element of novelty itself should be more emphasized in the work, so as not to leave any illusions.

Minor errors:

• Citation [16] should contain a broader description 
• The signs "a), b) ..." in the graphs diverge and have different formatting than the rest of the text 
• Table 6 is broken into 2 pages, it should be on one page
• Acknowledgment is not completed, this text is the same as  in a sample text in MDPI article formatting example
  

To summarize, this paper has the potential to be interesting but more work is necessary. Therefore, I recommend a major revision.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors have used a deep neural network model to predict glass transition temperatures of polyhydroxyalkanoate-based homo- and co-polymers. They encoded information of monomers by digital fingerprints and subsequently inputs to model together with other factors affecting the glass transition temperature. The hyperparameters of model were optimized using Bayesian optimization schema with different strategies evaluated to cope with missing data including removal of the input node, estimation with the mean, minimum, or maximum values of the dataset. The results indicated that the model is robust in achieving high accuracy even with missing data and that the predicted glass transition temperatures by the model are in good agreement with the experimental results.

The paper is well-written, has data needed to appreciate the effectiveness of the model developed and can be recommended to publication.

There is however an important comment that authors are invited to address before publication.

Authors are invited to elaborate on the main parameter investigated by machine learning – the glass transition temperature. From the manuscript is not evident what is the glass transition temperature used for calculations. Very often the glass transition is defined as the temperature where the viscosity is 10E12 Pa*s – see e.g. Angell C.A. Relaxation in Liquids, Polymers and Plastic Crystals - Strong/fragile Patterns and Problems. J. Non-Cryst. Solids, 131−133, 13−31 (1991). This was shown however as misleading and even wrong – see e.g. D.S. Sanditov D.S. et al. Glass transition criterion and plastic deformation of glass. Physica B, 582, 411914 (2020). “The IUPAC Compendium on Chemical Terminology”, Cambridge: Royal Society of Chemistry (1997) treats the glass transition as it manifests itself, namely as a second-order phase transformation in the sense of Ehrenfest classification. Depending on the kind of measurement performed, the glass transition is thus revealed either as a continuous change of first order thermodynamic properties such as volume, enthalpy, entropy, or as a discontinuous variation of second-order thermodynamic properties such as heat capacity or thermal expansion coefficient across the glass transition range. Although the paper specifies (line 78) that the glass transition temperature is “where polymers turn from a rigid state to a rubbery state” it is not specified the technique used and the meaning of glass transition temperature as a parameter analysed. If authors relate to glass transition temperature based on viscosity data, then some of the results could be even wrong in case of complex molecules see the above Physica B reference. Therefore, it is important to have the glass transition temperature clearly defined and the method of data obtaining specified.  

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors have made all the corrections I suggest. I believe that the article is ready for publication in its current form.