From the journal Digital Discovery Peer review history

Dear Dr Yeo:

Manuscript ID: DD-REV-04-2023-000078
TITLE: Computational and data-driven modeling of solid polymer electrolytes

Thank you for your submission to Digital Discovery, published by the Royal Society of Chemistry. I sent your manuscript to reviewers and I have now received their reports which are copied below.

I have carefully evaluated your manuscript and the reviewers’ reports, and the reports indicate that major revisions are necessary.

Please submit a revised manuscript which addresses all of the reviewers’ comments. Further peer review of your revised manuscript may be needed. When you submit your revised manuscript please include a point by point response to the reviewers’ comments and highlight the changes you have made. Full details of the files you need to submit are listed at the end of this email.

Digital Discovery strongly encourages authors of research articles to include an ‘Author contributions’ section in their manuscript, for publication in the final article. This should appear immediately above the ‘Conflict of interest’ and ‘Acknowledgement’ sections. I strongly recommend you use CRediT (the Contributor Roles Taxonomy, https://credit.niso.org/) for standardised contribution descriptions. All authors should have agreed to their individual contributions ahead of submission and these should accurately reflect contributions to the work. Please refer to our general author guidelines https://www.rsc.org/journals-books-databases/author-and-reviewer-hub/authors-information/responsibilities/ for more information.

Please submit your revised manuscript as soon as possible using this link:

*** PLEASE NOTE: This is a two-step process. After clicking on the link, you will be directed to a webpage to confirm. ***

https://mc.manuscriptcentral.com/dd?link_removed

(This link goes straight to your account, without the need to log on to the system. For your account security you should not share this link with others.)

Alternatively, you can login to your account (https://mc.manuscriptcentral.com/dd) where you will need your case-sensitive USER ID and password.

You should submit your revised manuscript as soon as possible; please note you will receive a series of automatic reminders. If your revisions will take a significant length of time, please contact me. If I do not hear from you, I may withdraw your manuscript from consideration and you will have to resubmit. Any resubmission will receive a new submission date.

The Royal Society of Chemistry requires all submitting authors to provide their ORCID iD when they submit a revised manuscript. This is quick and easy to do as part of the revised manuscript submission process.   We will publish this information with the article, and you may choose to have your ORCID record updated automatically with details of the publication.

Please also encourage your co-authors to sign up for their own ORCID account and associate it with their account on our manuscript submission system. For further information see: https://www.rsc.org/journals-books-databases/journal-authors-reviewers/processes-policies/#attribution-id

I look forward to receiving your revised manuscript.

Yours sincerely,
César de la Fuente, Ph.D.
Associate Editor, Digital Discovery

************

Reviewer 1

In this review, the authors present an in-depth exploration of machine learning concepts and their applications in material modeling, specifically focusing on solid polymer electrolytes and related materials. The review encompasses frequently used ML algorithms and their relevance to screening and prediction in material design. Furthermore, optimization algorithms commonly utilized in materials design are discussed, followed by an extensive examination of the integration of data-driven methods into computational simulation tools like density functional theory, molecular dynamics, and coarse graining . The manuscript effectively reviews cutting-edge computational methodologies with ML algorithms, demonstrating the authors' expertise in the field. With this in mind, I believe that the points listed in the attached file (Comments_to_the_Author.pdf) should be addressed prior to publication to enhance the impact of the manuscript.

Reviewer 2

This review focuses on recent progress in using machine learning methods in a specific research area (polymer electrolytes, PE). Such a topic should be of great interest to the polymer electrolyte community and even the broader battery electrolyte community. Especially considering that the development of ML has experienced fast progress in the past decade, the majority of researchers including experimentalists and computational chemists in this specific area do not have any background, such a review is important and welcome. As the scope of this journal include both experimental and theoretical research across different disciplines, it is expected that this review should be introductory to audience with and without relevant background. The review should discuss the most commonly used ML approaches in polymer electrolyte research with selected examples that cover the most important challenges in this area. I feel that this review focuses more on presenting the methodology rather than the research questions and research progress. For example, it introduces how can ML was used with DFT, MD, CG methods, whereas the detailed research on polymer electrolyte architecture design, performance improvement and interfacial issues are not discussed much here. This point should be clarified in the Introduction.

In the introduction part, the authors first discussed the importance of polymer(solid) electrolyte research, which is not very necessary and can be shortened. I feel that the emphasis here should be on why machine learning is becoming increasingly important in this research area, what are the limitations of current research approaches in addressing the great challenges of polymer electrolytes that can potentially solved by ML, i.e. how ML approaches can fill the gap. Then it can summarize the main problems that are currently solved by developing/using ML approaches and what progresses have been made before discussing more details in methods and applications in the next sections.

In Section 2, the authors discuss a number of machine learning methods. For readers without relevant background, this part of the content is difficult to understand and follow, especially it involves detailed description of different concepts, methods. I'm not sure if all of these methods have been adopted in PE research, as some of them don't include any examples. Maybe this part should focus on the most relevant/common methods and give corresponding examples. Or you can categorize different methods according to PE research questions, which will allow readers more targeted when reading. A diagram to explain different methods in different research questions will be more useful than figure 3.

The reviewer is more enjoyed reading sections 3 and section 4. Please double-check the latest references to make sure you include the latest outcomes. For examples, Nature Communication 2022; 13: 3415, Nature communications, 2023, 14, 2789, ACS Cent. Sci. 2023, 9, 2, 206–216. More specific comments for these two sections:

Section 3:
1. In DFT section, can you expand the discussion regarding the specific examples you mentioned in reference 27, 156, 158, 159, and explain what DFT datasets and how were they used to train ML models to study those PE properties (wide electrochemical stability window, high ionic conductivity, and good thermal and mechanical stability).
2. How to use PES from QM calculation to enable fast and accurate MD simulation? Can you explain more clearly here?
3. How do free energy surfaces of a molecule relate to PE research, can you provide some examples?
Section 4:
consider to add some suggestions for experimentalists and computational chemists, for example, what we can do to help in developing this area.

Kindly refer to the attached response to reviewers.

This text has been copied from the Microsoft Word response to reviewers and does not include any figures, images or special characters:

Response to Reviewers
We appreciate the reviewers’ thoughtful comments and suggestions. They are an essential part of improving the quality of this paper. We have revised the paper based on these comments. Reviewers’ comments are given in black italicized font, and specific concerns are numbered. Point-by-point responses are provided below, and changes/additions to the manuscript are also provided as highlighted text in the revised manuscript. We have also added some comment-relevant information in the main manuscript.

Referee 1:
Comment 1
In Section 3.3, the authors introduce top-down CG simulations such as the Kremer-Grest model and dissipative particles dynamics for modeling SPEs. However, it may be beneficial to discuss the advantages of bottom-up CG simulations in capturing local interactions, such as the polarized effect in SPEs, and preserving chemical specificity. To provide further insights, the authors could consider referencing recent review/perspective articles that discuss integrating bottom-up CG models with machine learning (ML) algorithms. Some relevant articles include:
[1] S. Dhamankar and M. Webb, “Chemically specific coarse-graining of polymers: Methods and prospects” J. Polym. Sci., 2021, 59, 2613.
[2] J. Jin et al., “Bottom-up coarse-graining: principles and perspectives” J. Chem. Theory Comput., 2022, 18, 5759.
[3] C.-I Wang and N. Jackson, “Bringing quantum mechanics to coarse-grained soft materials modeling” Chem. Mater., 2023, 35, 1470.
[4] W. G. Noid, “Perspective: advances, challenges, and insight for predictive coarse-grained models” J. Phys. Chem. B, 2023, 127, 4174.

Response to Comment 1
We have added the additional references and mentioned the key differences between top-down CG methods and bottom-up CG methods in the main manuscript:

“It should be noted that both K-G model and DPD model are considered top-down CG approaches, i.e., the explicitly proposed simple potentials are tuned to match macroscopic thermodynamic properties. (Ref 244: W. G. Noid, 2023) In contrast, bottom-up CG approaches employ more complex potentials that are parameterized with information from atomically detailed simulations. Therefore, bottom-up CG can be better at capturing local interactions, such as polarized effect in SPEs, and preserving chemical specificity. (Ref 245: S. Dhamankar & M. Webb, 2021) There are some recent reviews that have discussed bottom-up CG approaches in more detail. (Ref 246: C. -I Wang & N. Jackson, 2023. Ref 247: J. Jin et al. , 2022)”

Comment 2
On page 13, the authors briefly discussed common exchange-correlation functionals used in DFT calculations, with the exception of the ωB97 series. However, it is worth noting that the ωB97 functionals have gained significant recognition for their improved efficiency and accuracy over the past decade. Additionally, a highly informative review article by Bursch et al. titled "Best-Practice DFT Protocols for Basic Molecular Computational Chemistry" (https://doi.org/10.1002/ange.202205735) offers valuable guidelines for selecting appropriate DFT functionals and basis sets. Including a brief mention of this article would greatly assist readers in establishing reliable settings for their DFT calculations.

Response to Comment 2
To highlight the importance of the ωB97 series, we have added two more references and modified the text as follows:

“Commonly used functionals include VWN, PW91, M06-class, ωB97-series, (Ref 150: J. Da-Chai & M. Head-Gordon, 2008) B3LYP, etc., belonging to different families—linear-density approximation (LDA), generalized gradient approximations (GGA), meta-GGA, and hyper-GGA, etc., with their strengths and weakness. Many DFT textbooks and articles have provided valuable insights on selecting functionals and basis sets with examples of applications. (Ref 155: M. Bursch et al., 2022)”

Comment 3
The authors are advised to remove redundant sentences, such as “The main text of the article should appear here with headings as appropriate,” “The main paragraph text follows directly on here,” and “The main text of the article should appear here with headings as appropriate” found at the beginning of each section on pages 1, 8, 9, and 16.”

Response to Comment 3
We have made corrections as suggested by the reviewer.















Referee 2:
Comment 1
Section 1: “I feel that this review focuses more on presenting the methodology rather than research questions and research progress specific to SPEs. The authors should make the point clear in the introduction part.”

Reply to Comment 1
To emphasize that this review predominantly centers on computational and data-driven approaches. We have modified the last part of our introduction as follow:

“Herein, we are hoping to provide insights for both experimentalists and theorists in this area and foster more collaboration between them to facilitate the development of advanced SPEs. As such, our review primarily emphasizes the methodologies of computational and data-driven techniques, with examples on how they are employed in the SPE system. First, we review some basic concepts about machine learning (ML), including frequently used algorithms and how they are applied to materials modelling, with an emphasis on screening and prediction. Subsequently, we review optimization algorithms that are commonly used for materials design. We provide specific examples on how certain algorithms are tailored to SPE research. Next, we discuss how data-driven methods are incorporated into computational simulation tools, such as density functional theory (DFT), molecular dynamics (MD), and coarse graining (CG). Lastly, we provide a summary and outlooks on using computational and data-driven approach for modelling of SPEs.”

Comment 2
Section 2: “The authors discuss a number of ML methods. For readers without relevat background, this part of the content is difficult to understand and follow, especially it involves detailed description of different concepts, methods. I’m not sure if all of these methods have been adopted in SPE research, as some of them don’t include any examples. Maybe this part should focus on the most relevant/common methods and give corresponding examples. Or you can categorize different methods according to PE research questions, which will allow readers more targeted when reading.”

Reply to Comment 2
To include more detailed examples in section 2, we introduced a new Section 2.4 to demonstrate the application of ML algorithms to SPE systems using experimental data:

“In the preceding sections, we have mentioned a series of ML algorithms and optimization methods and how they can be applied to broad topics for materials design and polymer informatics. These topics demonstrate the effectiveness of data-driven approach. In this section, we delve into specific case studies, focusing on the practical application of these algorithms to SPE systems with experimental data.

Back in 2011, feed forward NN was applied to fit the ionic conductivity data obtained via experiments. Ibrahim et al. measured the conductivity of PEO, LiPF6, ethylene carbonate and carbon nanotubes mixtures under different temperatures. During training, the chemical compositions and temperatures were used as inputs and ionic conductivities as outputs. The simple NN was able to predict the ionic conductivity of such a system well, as the predicted value can be further validated with new experiments.

Hatakeyama-Sato et al. employed ML methods to explore superionic glass-type SPEs with aromatic structures. They constructed a database including 104 entries about ionic conductivity. First, GNN was utilized to truncate molecular descriptors and extract useful features. The NN is pretrained on a database of randomly generated de novo polymers and monomeric compounds. The goal of this pretraining is to predict 2000 molecular descriptors from these compounds using only 32-dimensional vectors. This vector was then used to represent the feature of each compound for further ML processes. Subsequently, GP was used for establishing the relationship between chemical features and ionic conductivity. GP was able to output conductivity values along with confidence intervals. Combining GNN and GP, the authors successfully yielded glass-type polymer complexes with high conductivity that was later validated via experiments.

Bradford et al. built a chemistry-informed ML model that could predict SPE ionic conductivity based on the electrolyte and composition. They gathered data set of SPE ionic conductivity values from 217 experimental publications. They adopted a message passing NN, which is a special type of GNN, to learn optimal representations of the molecular components. The input of the NN took vectorized SPE features including polymer structure, salt structure, polymer molecular weight, salt concentration and temperature. The authors encoded the Arrhenius equation, which describes temperature dependence of ionic conductivity, into the readout layer of the NN and found that this chemically informed layer would increase prediction accuracy of the NN. After training the NN, they used the model to screen over 20,000 potential SPEs composed of commonly used lithium salts with synthetically accessible polymers and identified promising candidates. The predicted ionic conductivity exhibited good agreement with two types of in-house synthesized polymers. Furthermore, they extended their predictions to encompass various anions within PEO and poly(trimethylene carbonate), showcasing the model's effectiveness in identifying descriptors for solid polymer electrolyte (SPE) ionic conductivity.”

Comment 3
Please double-check the latest references to make sure you include the latest outcomes. For example, Nature Communications 2022, 13, 3415; Nature Communications 2023, 14, 2789; ACS Central Science 2023, 9, 2, 206-216.

Reply to Comment 3
We added the above-mentioned publications into our review and updated new studies wherever appropriate.

Ref 145: G. Bradford, J. Lopez, J. Ruza, M. A. Stolberg, R. Osterude, J. A. Johnson, R. Gomez-Bombarelli and Y. Shao-Horn, ACS Cent. Sci., 2023, 9(2), 206-216.

Ref 164: K. Li, J. Wang, Y. Song and Y. Wang, Nat. Commun., 2023, 14(1), 2789.

Ref 232: T. Xie, A. France-Lanord, Y. Wang, J. Lopez, M. A. Stolberg, M. Hill, G. M. Leverick, R. Gomez-Bombarelli, J. A. Johnson, Y. Shao-Horn and J. C. Grossman, Nat. Commun., 2022, 13(1), 3415.

Comment 4
Section 3: “In DFT section, can you expand the discussion regarding the specific examples you mentioned in reference 27, 156, 158, 159, and explain what DFT datasets and how were they used to train ML models to study those PE properties (wide electrochemical stability window, high ionic conductivity, and good thermal and mechanical stability).”

Reply to Comment 4
We provided a more detailed discussion regarding the specific examples:

“Specifically, fully exploiting the development of the extensive DFT datasets for training ML models shows great potential to enable the design and discovery of novel electrolyte systems containing polymer and lithium or other alkali metal compounds with wide electrochemical stability window, high ionic conductivity, and good thermal and mechanical stability in a fraction of the time.27,156,158,159 For example, Li et al. developed a ML workflow embedded with DFT and GNN to discover promising ionic liquids as additives for SPEs. DFT was employed to calculate the training data of electrochemical stability window based on HOMO/LUMO theory. The authors further verified a subset of selected candidates and measured the performance using experiments.164”

Comment 5
Section 3: “How to use PES from QM calculation to enable fast and accurate MD simulation? Can you explain more clearly here?”

Reply to Comment 5
We provided additional examples with more details in the manuscript:

“MLPs interpolate ab initio calculations by training the ab initio or DFT dataset and thus extend the system size and time scale in MD simulations. For example, Musaelian et al. recently introduced a deep NN interatomic potential architecture to achieve simultaneously accurate and computationally efficient parameterization of PES. In one of their testing cases, the authors simulated the Li-ion migration in a Li3PO4 electrolyte. Compared to AIMD, a mean absolute error in energies of 1.7 meV/atom was obtained for the proposed MLP. The authors further demonstrated the superior scaling ability of this method by running a system containing 421,824 atoms on multiple GPUs.172 Fu et al. benchmarked a collection of state-of-the-art MLPs under different practical scenarios. Apart from force and energy prediction errors, the authors suggested other metrics to evaluate MLPs such as radial distribution function (RDF) and diffusivity coefficient for LiPS dataset.173 A more comprehensive review of recent advances in MLPs was given elsewhere.174-176”

Comment 6
Section 3: “How do free energy surfaces of a molecule relate to PE research; can you provide some examples?”

Reply to Comment 6
Indeed, the difference between potential energy surface (PES) and free energy surface (FES) is very subtle yet critical. In section 3.1, we have talked about using ML algorithms to estimate potential energy surfaces (PES) from DFT data. In section 3.2, we underline that ML can also be employed to construct free energy surfaces (FES). We have modified the first two paragraphs of section 3.2 to clarify these terminologies:

“As the compositions and structures of SPEs become increasingly sophisticated, the investigation of ion transport kinetics within SPEs and across SPE-electrode interface are growing in importance.210-212 While PES describes the potential energy landscape of a system and can be primarily used for structural optimizations of molecules (e.g., the rearrangements between isomers), the free energy surface (FES) includes information of both potential energy and entropy contributions and can be used for assessing kinetics and thermodynamics of bulk molecular systems (e.g., protein folding) at a given temperature.213, 214 Currently, there’s ample research opportunities for constructing FES of SPE systems.

An accurate description of the free energy is key to understanding complex systems that have many intrinsic degrees of freedom.204,205 The relevant configurations of such systems and the transition between them can be captured by reducing the high-dimensional PES to a low-dimensional FES. Concretely, for a large system that contains N atoms, it requires roughly 3N degrees of freedom to describe the PES. Yet, we aim to employ collective coordinates with significantly fewer dimensions than 3N to encode information.208 This is particularly helpful for description of the chemical processes and the validation of computational models.215,217”

Comment 7
Section 4: “Consider adding some suggestions for experimentalists and computational chemists, for example, what we can do to help developing this area.”

Reply to Comment 7
In section 4, we mentioned several challenges in this field and provided corresponding suggestions for each challenge. We further strengthened insights on these challenges:

“With such repositories, experimentalists are expected to document accurate experimental data about SPEs such as ionic conductivity, mechanical properties, and morphological information. Meanwhile, computational chemists are expected to modify existing descriptors or design new descriptors that are compatible for more complicated polymer systems that can be beneficial for establishing a more comprehensive database.”

“Moreover, it is essential for computational chemists to open access to their code for public use, thereby reducing the barriers to implementing ML models.”

“Once more, a robust partnership between experimentalists and computational chemists remains essential in crafting such workflows. Computational chemists can assist experimentalists in designing experiments and ensure that experiments are efficient, cover a wide parameter space, and provide meaningful data. Concurrently, experimentalists can help computational chemists gain a deep understanding of SPEs to adjust and improve their models.”

Dear Dr Yeo:

Manuscript ID: DD-REV-04-2023-000078.R1
TITLE: Computational and data-driven modeling of solid polymer electrolytes

Thank you for submitting your revised manuscript to Digital Discovery. I am pleased to accept your manuscript for publication in its current form. I have copied any final comments from the reviewer(s) below.

You will shortly receive a separate email from us requesting you to submit a licence to publish for your article, so that we can proceed with the preparation and publication of your manuscript.

You can highlight your article and the work of your group on the back cover of Digital Discovery. If you are interested in this opportunity please contact the editorial office for more information.

Promote your research, accelerate its impact – find out more about our article promotion services here: https://rsc.li/promoteyourresearch.

If you would like us to promote your article on our Twitter account @digital_rsc please fill out this form: https://form.jotform.com/213544038469056.

We are offering all corresponding authors on publications in gold open access RSC journals who are not already members of the Royal Society of Chemistry one year’s Affiliate membership. If you would like to find out more please email membership@rsc.org, including the promo code OA100 in your message. Learn all about our member benefits at https://www.rsc.org/membership-and-community/join/#benefit

By publishing your article in Digital Discovery, you are supporting the Royal Society of Chemistry to help the chemical science community make the world a better place.

With best wishes,

César de la Fuente, Ph.D.
Associate Editor, Digital Discovery

Reviewer 1

Comments:
The author has addressed all the reviewers’ comments and has clarified the ambiguity. This work presents a comprehensive review of modeling solid polymer electrolytes through the use of machine learning and data-driven approaches. Therefore, I recommend the publication of the manuscript.