Predicting Properties of Imidazolium-Based Ionic Liquids via Atomistica Online: Machine Learning Models and Web Tools

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This manuscript by Armaković et. al.  presents a timely and valuable contribution to the field of ionic liquid (IL) research. The work presents an integrated computational workflow for predicting the density and viscosity of imidazolium based ILs and most significantly deploys these models and associated data as open access tool on the online platform. The manuscript's primary strength lies in its clear commitment to open source science and the development of a really useful tool for the research community. The density models (IonIL-IM-D1 and IonIL-IM-D2) appear robust and well validated, demonstrating strong predictive performance.

However, the most significant flaw is poor generalizability of the viscosity model (IonIL-IM-V) and the misleading claim of its performance. The discrepancy between the reported test set performance and the cross-validation results is severe and indicates significant overfitting, a fact that is not adequately addressed or reflected in the manuscript's conclusions. Therefore I recommend a resubmission by addressing following major comments, 

1. The manuscript's central and most severe flaw lies in the presentation and interpretation of the IonIL-IM-V model's performance. The claims made about this model are not supported by the data presented and are therefore scientifically misleading. The authors report a test set R² of 0.907 for the IonIL-IM-V model which at first glance suggests high accuracy. However they also report that 10-repeat 5-fold cross validation yielded a mean R² ~ 0.5497. This massive discrepancy is a definitive sign of severe overfitting. A high R² on a single, static test split can be an artifact of a randomly "lucky" split, where the test set happens to be particularly easy to predict based on the specific training set. Cross validation (CV) is the gold standard for assessing a model's true generalizability because it averages performance over multiple, different partitions of the data, providing a much more robust estimate of how the model will perform on new, unseen data. A drop in R² from ~0.91 to ~0.55 is not a sign of "moderate robustness" as the authors claim; it is a clear indication that the model has learned noise and specific artifacts from the training data and fails to generalize. While the authors' aim to create a highly accurate model is understandable, its performance is not acceptable.
   
   
2. The manuscript does not specify units or transformation (e.g. log scale) for viscosity. This must be clarified.
3. 
   Authors must note that the viscosity dataset is imbalanced, with a heavy concentration of low viscosity ILs and far fewer samples at higher viscosities. However, they fail to connect this observation directly to the model's poor performance. This imbalance is a likely contributor to the severe overfitting observed. A model trained on such data will naturally learn to fit the dense, low viscosity region very well, while the few high viscosity points may be treated as noise. The discussion must be expanded to explicitly link the dataset imbalance to the observed overfitting and the large discrepancy between test and CV scores. 
4. Given that the IonIL-IM-V model is poorly generalizable and overfit, the interpretations derived from its SHAP analysis are also questionable. SHAP is a powerful tool that explains what a model has learned. If the model has learned non physical artifacts and noise to fit the training data, then SHAP will explain these learned artifacts, not necessarily the true underlying physics governing viscosity. The authors must explicitly state that “these interpretations are derived from an overfit model and therefore may not represent true physical structure-property relationships but rather reflect the patterns the model learned to fit the specific training data”.
5. The manuscript also fails to describe how the data was split into training and test sets, which is crucial for any ML- paper. In cheminformatics a simple random split can lead to "information leakage" where structurally analogous compounds (e.g. members of a homologous series with slightly different alkyl chain lengths) are present in both the training and test sets. This makes the prediction task artificially easy and can leads to  overly optimistic performance metrics. The authors must add a detailed subsection to the Methods section describing their data splitting procedure. 
6. The choice of the GFN-FF force field for geometry optimization of the ionic liquids is stated but not justified. GFN-FF is a generic force field designed for broad applicability and speed, which makes it attractive for high-throughput studies. However, ionic liquids are a particularly challenging class of molecules due to their strong and complex electrostatic interactions, significant charge delocalization and polarization effects. The accuracy of a generic, non-polarizable or partially-polarizable force field for such systems cannot be assumed. The authors should add a paragraph to the Methods section justifying their choice of GFN-FF. This should include a discussion of its advantages and its potential limitations for highly charged, polarizable ionic systems. 
7. To ensure full reproducibility specific hyperparameters for the SVR models (e.g. kernel type, C, gamma, epsilon values) should be provided
8. Font size for labels and tick marks in the plots (Figures 2-10) is quite small and should be enlarged for clarity.

Comments on the Quality of English Language

• The manuscript contains several informal languages like "a smart and practical strategy was adopted", "it is logical and can be easily explained" and the repeated use of "excellent predictive performance". These should be revised
• There is a typo in Figure 1: "Gemmation of L structures" should  be "Generation of IL structures". The model name is spelled inconsistently as "lonIL-IM-V" and "IonIL-IM-V" throughout the manuscript.

Author Response

 

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

Please see the attached file. 

Comments for author File: Comments.pdf

Author Response

 

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

The study addresses an important and relevant problem in computational chemistry and cheminformatics—predicting the physicochemical properties of imidazolium-based ionic liquids using machine learning and making these tools accessible through a web platform. The topic is relevant to computational chemistry, chemical informatics, and materials science. However, while integrating ML models with a web interface is practically useful, the methodological novelty is somewhat limited. Similar ML-based IL property prediction studies already exist, many of which also discuss accessibility through online or open-source tools. To strengthen the contribution, the authors could benchmark their system against leading IL prediction models in the literature. The integration into atomistica.online is a positive step toward democratizing access to predictive modeling. However, the methodological description requires more detail, such as dataset preparation, feature selection, validation strategy. For the conclusions presented, they are generally supported by the reported results, but the evidence could be improved and enhanced by clearer discussion of limitations. Regarding the references, some citations here are outdated. Please consider citing  some recent works on graph neural networks and deep learning for IL property prediction. For the tables and figures, they need to be improved by adding clear column headings and using higher resolution to avoid pixelation. With these suggestions, it would enhance the manuscript’s impact and credibility.

Comments on the Quality of English Language

The manuscript is generally understandable, but the English requires moderate revision to improve clarity and flow. Especially in the Introduction and Discussion section, some long and complex sentences should be simplified to increase the readability. I would recommend a thorough language edit by a native or professional scientific editor.

Author Response

Dear Reviewer 3,

We are using this opportunity to sincerely thank you for your time and expertise to evaluate our manuscript and to provide constructive feedback. We have revised our manuscript according to all of your comments and suggestions. All changes in the main manuscript file are marked with yellow color for your convenience. Bellow, we are providing detailed replies to all of your comments. 

Thank you once again.

Report for Reviewer

Comment 1: To strengthen the contribution, the authors could benchmark their system against leading IL prediction models in the literature

Reply: Thank you very much for this valuable suggestion. We fully agree that benchmarking against other models is a powerful way to validate performance and highlight strengths. However, in this case, a direct comparison is not feasible because our models are specifically curated for ionic liquids based on the imidazolium cation, whereas most existing models are designed to cover a much broader range of ionic liquids. Given the prevalence, availability, and importance of imidazolium-based systems, we deliberately focused our dataset on this class, with the expectation that such specialization would enable higher accuracy even when using relatively simple features. Nevertheless, to provide context, we have now included in the revised manuscript a discussion of the reported accuracies of leading models from the literature, which allows the reader to see how our results compare to existing approaches. Additions relevant to this comment are in lines 95 - 115.

Comment 2: However, the methodological description requires more detail, such as dataset preparation, feature selection, validation strategy

Reply: Thank you for pointing this out. In the revised version of our manuscript, we have added more detailed explanations regarding dataset preparation, feature selection, and the validation strategy. Namely, the chapter “Workflow and computational details” is now almost doubled with important information allowing the reproduction of the results. Additions relevant to your comment are in lines 217 – 222.

Comment 3: For the conclusions presented, they are generally supported by the reported results, but the evidence could be improved and enhanced by clearer discussion of limitations.

Reply: Thank you for your comment regarding the conclusions. In the revised version of the manuscript, we have strengthened the conclusion section by providing a clearer discussion of the study’s limitations. In particular, the relevant addition is in lines 593 – 598.

Comment 4: Regarding the references, some citations here are outdated. Please consider citing  some recent works on graph neural networks and deep learning for IL property prediction.

Reply: Thank you for this helpful suggestion. In the revised version of our manuscript, we have added more recent references, including studies where graph neural networks and deep learning techniques have been applied to predict the properties of ionic liquids. In these regards, we cited additional eight references.

Comment 5: For the tables and figures, they need to be improved by adding clear column headings and using higher resolution to avoid pixilation

Reply: Thank you very much for this comment. Our manuscript does not contain any tables. As for the figures, we prepared all of them at a minimum resolution of 300 dpi, and the slight pixelation observed is likely due to PDF conversion by the publisher’s tools. Of course, if the publisher requires figures in even higher resolution, we will be more than happy to provide them.

Comments on the Quality of English Language

The manuscript is generally understandable, but the English requires moderate revision to improve clarity and flow. Especially in the Introduction and Discussion section, some long and complex sentences should be simplified to increase the readability. I would recommend a thorough language edit by a native or professional scientific editor.

Reply: Thank you very much for this comment. The revised work contains in total the addition of around 110 lines of text. In the revision, we gave our best to modify sentences, to make them clearer and less complex. In the same time, we paid attention for the added text to be clear and grammatically correct, and to describe the methodology and results in spirit of English language.

Thank you once again.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have been exceptionally thorough and responsive in addressing the concerns raised in the initial review. They have successfully rectified all major scientific and methodological issues, significantly improving the quality, transparency and overall merit of the manuscript.  Based on my review of the revised manuscript and the authors responses, my recommendation is to “Accept” the paper for publication. 

Specifically, they have:

• correctly reframed the performance of the viscosity model to reflect its limited generalizability, basing their conclusions on the more reliable cross validation results.  
• added all requested methodological details, including the data splitting procedure and a justification for the force field choice, which enhances reproducibility.  
• included appropriate explanation regarding the interpretation of the overfit viscosity model.  

The revised manuscript is now a scientifically valuable contribution to the field.

Author Response

Dear Reviewer,

We are using this opportunity to sincerely thank you for your time and valuable feedback. Your comments and suggestions greatly improved the quality of the manuscript.

Wishing you all the very best.

Sanja Armaković

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript is in good shape and ready for publication.

Author Response

Dear Reviewer,

We are using this opportunity to sincerely thank you for your time and valuable feedback. Your comments and suggestions greatly improved the quality of the manuscript.

Wishing you all the very best.

Sanja Armaković