Machine learning-guided discovery of ionic polymer electrolytes for lithium metal batteries

As essential components of ionic polymer electrolytes (IPEs), ionic liquids (ILs) with high ionic conductivity and wide electrochemical window are promising candidates to enable safe and high-energy-density lithium metal batteries (LMBs). Here, we describe a machine learning workflow embedded with quantum calculation and graph convolutional neural network to discover potential ILs for IPEs. By selecting subsets of the recommended ILs, combining with a rigid-rod polyelectrolyte and a lithium salt, we develop a series of thin (~50 μm) and robust (>200 MPa) IPE membranes. The Li|IPEs|Li cells exhibit ultrahigh critical-current-density (6 mA cm−2) at 80 °C. The Li|IPEs|LiFePO4 (10.3 mg cm−2) cells deliver outstanding capacity retention in 350 cycles (>96% at 0.5C; >80% at 2C), fast charge/discharge capability (146 mAh g−1 at 3C) and excellent efficiency (>99.92%). This performance is rarely reported by other single-layer polymer electrolytes without any flammable organics for LMBs.

The strong advantage of this work is the combination of computational and experimental studies. It is very common in the Materials Chemistry field for the researchers to focus either on strictly theoretical or experimental studies, and a work showcasing the complementarity of the two is an excellent idea. The development of ionic polymer electrolytes incorporating ionic liquids as enhancers of their mechanical or physicochemical processes is currently a hotspot of scientific interests, with a plethora of papers being published every day. However, the authors of this work have managed to create an algorithm which can predict and classify the ionic liquids with the desired set of properties (something which is directly expandable to other research fields), but also actually create a set of fuel cells with very promising behaviour.
-Will the work be of significance to the field and related fields? How does it compare to the established literature? If the work is not original, please provide relevant references.
While the idea behind the simultaneous classification/prediction machine learning algorithm is brilliant, I am not convinced that this work will be of great significance for the field. The main reason for this is the absence of adequate validation of their predictions, which makes then their classification ambiguous. The authors are using a machine learning algorithm to predict the properties of interest in their system (here conductivity and electrochemical window), which however is at no point compared to literature data for the same compounds. I understand that since the studied dataset comes as a result of ion combinations, some of those ionic liquids are not commercially available, or maybe they have never been reported at all. But a comparison on a selected subset of ionic liquids is crucial.
Regarding the experimental data, from a pool of 40 recommended ionic liquids the authors select 4 to use in their fuel cells. However, I believe that the selection of ionic liquids is not in agreement with the scope of the work. All 4 selected ionic liquids are commercially available and I believe that they were chosen so that the authors avoid synthesising their own compounds. The whole point of this work is to discover an optimised selection of ions for this application, but then the authors decide to work with 1-ethyl-3-methylimidazolium triflate and tetrafluoroborate, which are 'old' ionic liquids and they have been used since the 1990's for electrochemical applications. I believe that as a proof of concept they should have tried at least one 'unconventional' combination of ions from their list. The results they obtain on their fuel cells are indeed very promising, but for an expert on the ionic liquid electrochemistry field it is not surprising or unexpected.
-Does the work support the conclusions and claims, or is additional evidence needed?
The experimental investigation is thorough and their results well-explained. The prediction and classification algorithm part though needs further work and validation in order to prove that the algorithm works properly. As I mentioned in the previous section, there is no cross-reference for the prediction accuracy of the calculated properties. This should definitely be amended before publishing this article.
-Are there any flaws in the data analysis, interpretation and conclusions? -Do these prohibit publication or require revision?
During the 'supervised learning' section the authors introduce an additional filter to validate their data, by calculating the interaction energy between the anions and the cations. Figure 3c and Supplementary Table 1 show the results of these Energy calculations. To my opinion these calculations are problematic and they need revisiting. According to their formula, the energy values should be negative because the energy of the ion pair (E[+] [-]) should be more negative than the individual energies of the ions. However, all the energies that are shown in Supplementary Table 1 are positive. The values of E[+] [-], E+ and E-should also be shown in the ESI, for comparison reasons. Moreover, from my experience, the scale of those energies is unexpected, as I would expect the interaction energies in the range of minus a few hundreds kJ/mol. Depending on the geometry optimisation process the interaction energy can differ, so the graph shown in Figure 3c is expected to have an error bar of ±50 kJ/mol, which makes it impossible to extract certain results from it, as the two box plots mostly overlap. Before the article is considered for publication, the authors need to revisit their calculations, explain properly their calculation methodology and answer questions such as what convergence criteria they use, specify grid accuracy, did they check for absence of imaginary frequencies etc?
Finally, something that is not so much of a flaw in the data analysis but more a part that needs further clarification is the part about the ion exchange with the ionic liquid electrolyte (rows 232-235). It is not clear to me why the incorporation of Li to the IPE is done with LiFSI and not with Li salt with the corresponding anion as the IPE (LiOTf and LiBF4 accordingly) and also why the ion exchange happens in [C3mpyr] [FSI]? I assume that the authors have tried different solvents and have a reason for performing the ion exchange in another IL and not in a conventional volatile molecular solvent, but this should be clearly explained in the text.
-Is the methodology sound? Does the work meet the expected standards in your field?
Generally, the methodology of the work follows a reasonable pipeline, although some gaps exist (they are extensively discussed in the previous sections). These gaps need revising before considering this work for publication. I am not convinced where the novelty and the importance of this work lies. There are many published articles on both ML algorithms for property prediction and conductive ionogels. I think that the authors need to add a paragraph (probably in the conclusions) highlighting the novelty of their work, what are the advantages of their algorithm compared to others and why their methodology/IPEs is superior to others reported in the literature.
-Is there enough detail provided in the methods for the work to be reproduced? Yes, apart from the quantum chemistry calculations discussed above, the rest of the methodology is adequately explained and can be reproduced.
Regarding the ML algorithm, and while not being an expert on the field, I believe that the algorithm is sound and reasonable. I am not convinced that one could reproduce the algorithm just by reading the article's methodology section, since there are several gaps on the actual coding, but since the code is available on GitHub one could use the pre-made code.
-General Comments 1. There are several grammatical/syntax errors throughout the text, language needs to be corrected.
2. The company selling the ionic liquids is IoLiTec (in the text it is written as IoLiTech).
3. The authors refer to their setup as full cell in several points in the text. I believe it is supposed to be fuel cell. 4. The authors need to amend their abbreviations for their ionic liquids. For example 1-ethyl-3methylimidazolium triflate, here it is abbreviated as C2mimTFO, the established abbreviation for that is either [Emim][OTf] or [C2C1im] [OTf]. The authors need to correct their ILs abbreviations according to the literature. 5. Figure 4. The embedded photos are too small for anyone to see them properly.
6. Row 52. 'Large population of ion pairs' is wrong. It should be a large population of IL candidates 7. Row 52. Here they use the abbreviation ML but they define it later in text.
8. Row 181. Geometric not gyometric. 9. Row 232. If the authors claim that their films are mechanically strong, they need to perform a tensile stress analysis.
10. Row 350. The Machine Learning methodology needs to be more detailed in order to be more reproducible.
11. Row 360. The quantum chemistry calculations methodology needs more details to be reproducible.
12. Rows 365-373. The purity/water content of the purchased ionic liquids needs to be stated. Is Canrd in row 373 correct?
13. Rows 389. The thickness/diameters of the electrodes need to be stated.
14. Row 534. The github link is not working.
Reviewer #2 (Remarks to the Author): In this manuscript, machine learning approaches are applied to discover ionic liquids with high ionic conductivity and electrochemical window. Some of the top candidates are blended with a polymer and a Li salt. Experiments are conducted to assess the performance of the ionic polymer electrolyte. Given the tremendous interest in Li-ion batteries and favorable attributes of ionic liquids as electrolytes, the topic of the manuscript has broad appeal. Developing electrolytes based on ionic liquids requires careful tuning of properties such as ionic conductivity, electrochemical window, viscosity etc., which are probed in this manuscript. Furthermore, machine learning approaches are becoming mainstream in the field of ionic liquids, so the manuscript holds special interest.
Although the topic of the manuscript is of great importance, there are several aspects of the manuscripts which deserve attention and substantial revisions.
-Very few details on the development of machine learning models are provided. I also find the development of the three machine models rather problematic as the data set is not split into a training and test data set making it difficult to assess the predictive capability of the model. This is especially concerning as it has been pointed out in the manuscript that the models tend to overfit due to over-representation of one or more cation/anion types.
-RDKit generates a large number of molecular descriptors for a given ionic liquid. Please include a description of how the total number of features was reduced to 60. Also specify if the cation and anion in an ionic liquid pair were modeled with different set of descriptors.
-Provide a reasoning for the basis set selection. Please include the level of theory for electronic structure calculations, -Why was a particular threshold of ionic conductivity and ECW selected?
-It is confusing that output properties from models such as ionic conductivity and electrochemical windows are listed as features.
-It is not clear that spherical anions will yield liquids/solid from Figure 3a. In fact, there does not seem to be a correlation between the sphericity of the anions and phase.
-How were top 10 important features determined?
-The results on 20 ILs is not enough to establish the correlation. For 1000 ILs, it is not difficult to conduct quantum calculations. I would recommend at least 20% of the ILs on which these calculations are carried out ensuring that cation types and anions are well represented.
-Why is it necessary to choose hydrophillic ionic liquids for Li-ion batteries? In fact, this property will actually be detrimental to the performance of batteries if water is absorbed.
-Provide the rationale for using PBDT as a liquid crystalline polyelectrolye.
-The Github link is not provided.
-Provide a caption for Supplementary Table 2. Include units for ECW and ionic conductivity. Specify the temperature at which ionic conductivity is predicted. For many of the ionic liquids, measured ionic conductivities are available from the NIST ILThermo Database. A comparison must be made between predictions and the measurements.
Statement: This manuscript offers a novel and combinatorial approach using data science and machine learning for prediction and classification of ionic liquids with optimised properties, which are then used as property enhancers in solid state electrolytes for lithium-metal batteries. Although the concept behind this work is promising and widely applicable, I personally believe that this research needs further work before reaching the publication stage, even if this is a Communication and not a full Research Article. If this Communication is to be published, I would recommend that, at a minimum, the comments below be addressed.

Response:
We thank the reviewer for recognizing the potential importance of this work.
The comments proposed by the reviewer are significant and insightful. Thus, we hope that the substantial revision to this manuscript based on the comments will be satisfactory.
Comment #1: What are the noteworthy results? The strong advantage of this work is the combination of computational and experimental studies. It is very common in the Materials Chemistry field for the researchers to focus either on strictly theoretical or experimental studies, and a work showcasing the complementarity of the two is an excellent idea. The development of ionic polymer electrolytes incorporating ionic liquids as enhancers of their mechanical or physicochemical processes is currently a hotspot of scientific interests, with a plethora of papers being published every day.
However, the authors of this work have managed to create an algorithm which can predict and classify the ionic liquids with the desired set of properties (something which is directly expandable to other research fields), but also actually create a set of fuel cells with very promising behaviour.
Response: Thanks again to the reviewer for recognizing the uniqueness of our work. In terms of the noteworthy results of this work, the developed machine learning protocol not only relieves the issue of data scarcity, but also confirms the importance of machine learning in materials design and optimization. In addition, this work also emphasizes the development of materials from experimental perspective; meanwhile, proposing great potential of the developed IPEs in functional devices. For example, the assembled LMBs using IPEs coupled with commercial LiFePO4 cathode (with high loading 10.3mg cm -2 ) and bare Li metal anode deliver outstanding capacity retention for > 350 cycles (> 96% with 0.5 C at RT; > 80% with 2 C at 50 ° C), fast charge/discharge capability (146 mAh g -1 with 3 C at 80 ° C) and ultrahigh coulombic efficiency (> 99.92%). This performance is rarely reported by any single-layer polymer electrolytes without any organic plasticizers/oligomers for LMBs. (Please refer to Comment #2 for   more details about the performance compared to the state-of-the-art literature)

Response:
We appreciate the reviewer for raising these concerns. We will illustrate the significance and compare this work to the literatures from two perspectives, including the machine learning model and the experimental performance, correspondingly.
In terms of the machine learning model, this work is original and complementary. This is essential to improve the efficiency to target promising ILs for practical applications. Compared to previous literatures, instead of focusing on individual properties and predicting the absolute physical properties of the IL pairs purely, for example, melting point 3 , viscosity 4 and ionic conductivity 2 , we first combine the factor of ionic conductivity with the electrochemical window, which are critical properties for battery electrolytes. This novel conceptual design is also insightful and can be easily applied for related research areas.
In terms of the material development and performance evaluation, the promising experimental results reported in this work represent the performance of the state-ofthe-art polymer electrolytes for Li metal batteries. We conclude a comparison to recent literatures [1][2][3][4][5][6][7] as shown in Table R1. Overall, the IPEs reported in this work outperform from a comprehensive perspective, including the current density, the cell cycling life and especially the high cathode loading required for practical applications.
In summary, this work is original in terms of the machine learning model, the quantum chemistry computational approach, the design of the materials and the excellent performance achieved in LMBs.

Modifications to the manuscript and supplementary information:
(1) We added a paragraph (Page 25, Line 409 -431) in the end of the manuscript to emphasis the novelty and importance of this work from the two perspectives.
(2) We inserted Table R1 in the main manuscript as Table 2.  ILs between the test dataset and the NIST ILThermo Database at 25 °C.
The electrochemical window (ECW) is another popular topic as discussed in the literatures. The ECW value in this manuscript is calculated based on the HOMO/LUMO theory rather than predicted values from the ML model. We note that the ECW is not provided by the ILthermal Database. Thus, we directly compare the calculated ECW with the results scrapped from IoLiTec. In Figure R2, the mean absolute error (MAE) between the calculated results and the experimental results are < 1.1 V. As we know, it is still challenging to estimate ECW for ILs accurately in the field. 12 Besides, the measured ECW values are highly dependent on experimental conditions, thus we believe that the accuracy of the calculated ECW is overall satisfactory and of significant reference to the field. . Figure R2. Comparison of ECW based on IoLiTec to ECW based on HOMO/LUMO theory for the cation(blue) and anion(red) types, correspondingly.
Modifications to the manuscript and supplementary information: (1) We added Figure R1 and Figure R2 in the manuscript as Figure 3d and Figure 2 c,d.
(2) We added more discussion about the validation of predicted ionic conductivity The results they obtain on their fuel cells are indeed very promising, but for an expert on the ionic liquid electrochemistry field it is not surprising or unexpected.

Response:
We thank the reviewer for mentioning these concerns. H2O/DMF mixture, we found that this "unconventional" IL is also very promising. We added this IL in the Figure 4 of the manuscript as shown below in Figure R3. The detailed experimental performance is summarized in Figure R4. We are still conducting in depth investigation on this IL for future work. There are many interesting ILs in the final recommendation list. However, we cannot cover all the combinations experimentally. Above all, we believe this screening method is significant and inspiring for related research areas. Figure R3. We added the new ILdata in Figure 4 of the main manuscript. Modifications to the manuscript and supplementary information: (1) We updated the Figure 4 in the manuscript with Figure R3 by inserting the new results based on IL Dems TFSI.
(2) We added Figure R4 in the Supplementary Fig. 6 to support the efficiency of this IL screening workflow.

Comment #5: Does the work support the conclusions and claims, or is additional evidence needed?
The experimental investigation is thorough and their results wellexplained. The prediction and classification algorithm part though needs further work and validation in order to prove that the algorithm works properly. As I mentioned in the previous section, there is no cross-reference for the prediction accuracy of the calculated properties. This should definitely be amended before publishing this article.
Response: Thanks again for these important suggestions. We believe we have included the answer for these comments in our responses to previous Comments. We have included the cross-reference in our manuscript. Please also refer to our responses to Comment #3.
Comment #6: Are there any flaws in the data analysis, interpretation and conclusions? -Do these prohibit publication or require revision? During the 'supervised learning' section the authors introduce an additional filter to validate their data, by calculating the interaction energy between the anions and the cations. Figure  3c and Supplementary , E+ and E-should also be shown in the ESI, for comparison reasons. Moreover, from my experience, the scale of those energies is unexpected, as I would expect the interaction energies in the range of minus a few hundreds kJ/mol. Depending on the geometry optimization process the interaction energy can differ, so the graph shown in Figure 3c is expected to have an error bar of ±50 kJ/mol, which makes it impossible to extract certain results from it, as the two box plots mostly overlap. Before the article is considered for publication, the authors need to revisit their calculations, explain properly their calculation methodology and answer questions such as what convergence criteria they use, specify grid accuracy, did they check for absence of imaginary frequencies etc?

Response:
We greatly appreciate the reviewer for indicating this error in the manuscript.
Yes, we totally agree that the interaction/binding energy between the cation and anion should be negative. We revisit the calculation and append the corrected results and calculation details below. We initially optimize the structure of the cations and anions  Figure R5 and Table R2 and Table R3. The lower binding energy (~ -400 kJ mol -1 ) of the solid further confirm our conclusion that the binding energy of the solid ion pair usually shows lower binding energy, which means strong interactions between cations and anions.
In terms of the prediction results for the remaining 72 ILs, as shown in Figure R3, we divide the predicted results into categories, including liquid and solid-x, where x (x = 1, 2, 3) is the number of ML models with prediction results being in solid phase, thus the higher the number, the larger possibility for the IL being in solid phase. We observe that the predicted liquid cluster showing the highest average binding energy. As x increases, we observe lower average binding energies that further confirms our demonstration.
Modifications to the manuscript and supplementary information: (1) Figure R5 was added to the main manuscript. More discussion was added to (Page 11, Line 200 -216) in the manuscript.
(2) Table R2 and Table R3 were added to the Supplementary Table 1 and Table 2.    Table 4, we observe that the absolute value of the binding energy of LiFSI is significantly smaller than that of LiTfO and LiBF4, this explains why the solubility of LiFSI in ionic liquids is significantly greater than that of LiTfO. Note: The used theory and basis set is M062X/6-311(+)G(2d,p).
We avoid using any conventional volatile solvents in the ion exchange procedure, (1) We added Table R4 and more discussion about the solubility of Li salts in the Supplementary Note 1.
(2) We added the explanation about the choice of LiFSI and C3mpyrFSI in the manuscript (Page 4-5, Line 80-89) .
Comment #8: Is the methodology sound? Does the work meet the expected standards in your field? Generally, the methodology of the work follows a reasonable pipeline, although some gaps exist (they are extensively discussed in the previous sections). These gaps need revising before considering this work for publication.
Response: Thanks again for these comments. After the careful validation and detailed explanation of the models included in our responses to previous Comments. We believe that this methodology is suitable and reliable.
Comment #9: I am not convinced where the novelty and the importance of this work lies. There are many published articles on both ML algorithms for property prediction and conductive ionogels. I think that the authors need to add a paragraph (probably in the conclusions) highlighting the novelty of their work, what are the advantages of their algorithm compared to others and why their methodology/IPEs is superior to others reported in the literature.
Response: Thanks for this suggestion. We have added the following paragraph to conclusion of the manuscript.
Modifications to the manuscript and supplementary information: Comment #10: Is there enough detail provided in the methods for the work to be reproduced? Yes, apart from the quantum chemistry calculations discussed above, the rest of the methodology is adequately explained and can be reproduced. Regarding the ML algorithm, and while not being an expert on the field, I believe that the algorithm is sound and reasonable. I am not convinced that one could reproduce the algorithm just by reading the article's methodology section, since there are several gaps on the actual coding, but since the code is available on GitHub one could use the pre-made code.

Response:
We thank the review for raising this concern. We added more details for the ML models in the experimental section of the manuscript. We also update the GitHub link for the projects. The class object called ILP can be reused and reproduced for future research.
Modifications to the manuscript and supplementary information: (1) We added more details about the ML algorithm (Page 26, Line 400 -454) in the manuscript.
(2) The updated GitHub link is https://github.com/wangyingxie/ILP Comment #11: There are several grammatical/syntax errors throughout the text, language needs to be corrected.

Response:
Thanks for raising these concerns. We have scanned the manuscript again and corrected corresponding the grammatical/syntax errors in the manuscript and the supplementary information.

Comment #12:
The company selling the ionic liquids is IoLiTec (in the text it is written as IoLiTech)

Response:
We thank the reviewer for pointing out the errors of this work. We have corrected this mistake correspondingly.

Modifications to the manuscript and supplementary information:
(1) We corrected 'IoLiTech' to 'IoLiTec' and other related errors in the manuscript.

Comment #13:
The authors refer to their setup as full cell in several points in the text. I believe it is supposed to be fuel cell.

Response:
We would like to first appreciate the reviewer for raising this point. Please Response: We greatly appreciate the reviewer's comment that improves our work a lot.
Modifications to the manuscript and supplementary information: (1) We corrected 'a large population of ion pairs' to 'a large population of IL candidates' in updated row 52.
Comment #17: Row 52. Here they use the abbreviation ML but they define it later in text.
Response: Thanks for this comment. We have modified the corresponding error in the manuscript.
Modifications to the manuscript and supplementary information: (1) We defined ML in row 51 and we also removed the definition of ML in row 53.  The corresponding DMA curve for IPE with 10% PBDT with C2mim TfO from −50 to 300 °C.

Modifications to the manuscript and supplementary information:
(1) We added Figure R6 to the Supplementary Fig. 3.
Comment #20: Row 350. The Machine Learning methodology needs to be more detailed in order to be more reproducible.

Response: Thanks for this comment.
Modifications to the manuscript and supplementary information: ( Modifications to the manuscript and supplementary information: (1) We replaced the "Canrd" mentioned in (Page 28, Line 478) with the company's full name.
(2) We added the purity of all the ionic liquids purchased from IoLiTec in Line 472-

477.
Comment #23: Rows 389. The thickness/diameters of the electrodes need to be stated. provided. I also find the development of the three machine models rather problematic as the data set is not split into a training and test data set making it difficult to assess the predictive capability of the model. This is especially concerning as it has been pointed out in the manuscript that the models tend to overfit due to overrepresentation of one or more cation/anion types.

Response:
We thank the reviewer for these comments. We added more details about the machine learning models in the experimental section. For the data splitting question, we employed the 5-fold cross validation for the three machine learning models to prevent the potential overfitting. Cross validation is a resampling method that uses different portions of the data to test and train the model. The 5-fold cross validation will split the data into 5 parts, every time 4 parts of the data are used to train the model, the remaining 1 (20% percentage of the data) is used for the validation dataset. Thus, the performance shown in Table 1   Modifications to the manuscript and supplementary information: Comment #5: It is not clear that spherical anions will yield liquids/solid from Figure   3a. In fact, there does not seem to be a correlation between the sphericity of the anions and phase.

Response:
We thank the reviewer for proposing this comment. We agree that Figure 3a is not clear to show the dependence, so we reevaluate the correlation between the sphericity of the anions and the phase of the ILs in Figure R7. We divided the predicted ILs into the four categories, including liquid and solid-x, where x (x = 1, 2, 3) is the number of ML models with prediction results being in solid phase for the ILs, thus the larger the number, the higher possibility for the IL being in solid phase. To discover the key features, we start with the features indicated by the feature importance score of the model one by one, finally we observe that the sphericity index of cation and anions will increase with increasing possibility for the cation-anion pair to be solid. The ELUMO of the anion seems like another important feature to determine the phase of the ILs. Indeed, the explanation of the model is quite challenging since the molecular descriptor are highly corelated in the ML models. However, we can still get some basic idea about the important features according to the calculated feature importance from Random Forest and XGBoosting models.
Modifications to the manuscript and supplementary information: (1) We added Figure R7 in the manuscript as Figure 3b. XGBoosting models directly, the calculated importance score indicates the contribution of each feature in the model. After careful examination of the model, we found that the multicollinearity issue is very severe when we try to explain the importance of the features, though it will not influence the accuracy of the model. The ranking of the importance is fluctuated, but the critical features are usually stable, which give us clue to find the relationship as shown in Figure R7. Thus, to prevent confusion, we decide to remove the Figure 3b from the main manuscript.
Modifications to the manuscript and supplementary information: (1) We removed original Figure 3b in the manuscript.

Comment #7:
The results on 20 ILs is not enough to establish the correlation. For 1000 ILs, it is not difficult to conduct quantum calculations. I would recommend at least 20% of the ILs on which these calculations are carried out ensuring that cation types and anions are well represented.
Response: Thanks for this significant suggestion. We have expanded the pool for the calculation. To ensure the representative of the types, we selected 91 ion pairs, including 7 cations and 13 anions from the main cation and anion types in the dataset to validate our prediction results. The results shown in Figure R5, Table R2 and Table   R3 strongly confirms the established correlations. Please refer to Comment#6 of Reviewer #1 for more details.
(1) Figure R5 was added to the main manuscript. More discussion was added to (Page 11, Line 200 -216) in the manuscript.
(2) Table R2 and Table R3 were added to the Supplementary Table 1 and Table 2. developed IPEs were placed in a vacuum oven at 80℃ for more than 24h to adequately remove water before assembled in the batteries. The ion exchange process was finished in an Ar-filled glove box (< 0.01 ppm H2O). Here we mainly need to measure the H2O in the dried membrane (10% PBDT C2mim TfO). Here, we used 1H NMR and DSC to carefully measure the water content in the membrane. As shown in the 1H NMR spectra in Figure R8a, we cannot observe a distinct signal that belongs to H2O, which usually appears around 4.9 ppm, for the membrane. As shown in the DSC curve in Figure R8b, no significant heat absorption peak for H2O was observed. The excellent battery cycling performance in the manuscript also confirms that the effect of H2O can be neglected in the IPEs. we observe no apparent heat absorption peaks above 100 °C, which indicates that H2O molecules were successfully removed after the vacuum drying step.
Modifications to the manuscript and supplementary information: (1) Figure R8 was added to Supplementary Fig. 4, which is mentioned at (Page16, Line 291 -293) in the manuscript.
Comment #9: Provide the rationale for using PBDT as a liquid crystalline polyelectrolye.

Response:
We thank the reviewer for this important question. Our previous work has illustrated the importance of PBDT in IPEs (Nature Materials 20.9 (2021): 1255-1263). Briefly speaking, the local parallel packing of charged PBDT rods can serves as the assembly templates not only offering mechanical integrity, but also endowing nanoscale structuring in the composite, ensuing the fast Li + transpotation 15 . In our response to Comment # 2 of Reviewer #1, we conclude a comparison to recent literatures as shown in Table R1. Overall, compared to other polymer matrix systems, the IPEs based on PBDT reported in this work outperform from a comprehensive perspective, including the current density, the cell cycling life and especially the high cathode loading required for practical applications.
Furthermore, in our latest study, we found that PBDT at Li metal surface significantly reduces the interfacial resistance and improves the charge-transfer kinetics.
Modifications to the manuscript and supplementary information: (1) We added a paragraph (Page 25, Line 409 -431) in the end of the manuscript to emphasis the novelty and importance of this work from the two perspectives.
(2) We inserted the Table R1 in the main manuscript as Table 2. the ionic conductivity is predicted at 25 °C, which has been added to the manuscript.
(Line 119) The comparison with the NIST ILThermo databases is also included in the Figure R1 in our responses to Comment#3 of Reviewer #1.
I would like to thank the authors for addressing my comments and providing detailed answers to all of them. Their results are very promising and all the additions have complimented the paper nicely. However, there are still some issues regarding the statistical evaluation of the importance of their classification/prediction algorithm.
In the previous review round (Comment 3) I asked to see a comparison between the predicted properties and experimental measurements available in the literature. The authors kindly did that both for conductivity and ECW. Conductivity is an accurate physical property and their comparison with the experimental data is not great. I understand that there were problems finding experimental data for their ionic liquids (from their 992 predicted ionic liquids, only 17 were on the ILThermo Database), which is a very small percentage to be used as an accurate validation. However, even for those ionic liquids it seems that the accuracy of their prediction is low. The R2 of the linear predicted-to-experimental conductivity is 0.76, but especially for the low conductivities this leads to huge deviations. It is in the discretion of the Editor to judge whether the 0.76 R2 factor is satisfactory for publication.
Very accurately the authors noted that the ECW cannot be predicted accurately, as it is significantly influenced by the experimental process -and for that I completely agree. However, the fact that the prediction of ECW is very accurate for some ionic liquids and not for others could be an indication that the model is undertrained or overfitted to specific structures. It would be very interesting for the authors to prove statistically the nature of this deviation. Also, in Figure R2 they should show the number of structures for each ion family studied; for example we see very accurate predictions for pyrrolidinium ILs, while not so much for imidazolium. How many pyrrolidinium ILs and how many imidazolium ILs have been studied?
Similarly, in Figure R5 and Table R2 where the authors show the average binding energy for the solid and the liquid compounds, the authors need to show the statistical significance of their hypothesis. Their hypothesis is that they can associate their calculated binding energy with the physical state of the ionic liquid. If the population of each category (liquid, solid 1, solid 2, solid 3) was the same, this comparison could be done by comparing the standard deviations and see whether they are overlapping. However, since the population of each category is different, the authors need to perform a t-test to prove the statistical significance of their hypothesis.
Reviewer #2 (Remarks to the Author): In the revised manuscript, authors have addressed all the comments in great detail. It is commendable that they carried out experiments with a not-so-typical ionic liquid to assess the performance of ionic liquid-polymer electrolyte. On the theoretical side, the set of ionic liquids for which quantum calculations were carried has been expanded. Although I still believe that the method concerning selection of features for machine learning has not been described to permit reproduction of results on its own, availability of Github code can alleviate this challenge.
Overall, the manuscripts combines machine learning to identify ionic liquids with suitable ionic conductivity and electrochemical window. Some of the top performing ionic liquids have been mixed with PBDT and promising performance of the solid electrolyte has been reported for Li-ion batteries. I recommend the publication of the manuscript.

Response:
We thank the reviewer for complimenting the revisions we made to the manuscript. The corresponding comments proposed by the reviewer again are very insightful. We hope that the following explanation and revision to this manuscript based on the comments will be satisfactory.  Figure 2), as shown in the table as follow, 4 ILs contains more than two records, but the variations between the records is very high, this uncertainty from the database (experimental values) itself will further increase the fluctuation of our validation shown in Figure 2. there are still many ILs on the list will deserve more investigation in the future, which will be very insightful for other researchers in the field.
Above all, we conclude that the model is important and the distinctive R 2 value is insightful to the field. This also indicates that we can pay more attention to the commonly existing bias of database collection and management for future ML investigations.
Modifications to the manuscript and supplementary information: (1) We added the following description in the main manuscript about the R 2 factor on Pages 13-14, Lines 249-256.
"As shown in Fig. 3d (2) We updated the new plot in the Figure 2 of the manuscript and added more discussion about the R 2 factor as shown above and the following cation and anion types, it has been discussed in previous literature. 3,4 Overall, the inconsistency is highly related to the limitation of this theory for estimating particular cation and anion pairs, for example, the C2mimBF4 is usually overestimated with a "weak" cation paired with a strong anion. We also emphasize in the manuscript in Line 313 about the deviation of the BF4 anion. For the cations, the imidazolium type is also not very accurate, because the description of the top of the valence band for some of the imidazolium-based ILs is not very accurate using the DFT and related approximations, especially for the imidazolium ones with BF4 anions. 4 However, the overall trend of the ECW values is reliable and shows enough accuracy for screening of potential IL in this application. Thanks very much for the reviewer's suggestion about adding the number of studied ILs on the plots. We agree that this is very important for us to see the distribution of the dataset. There are only 47 ionic liquids with measured ECW in IoLiTec. We added the label for each group in the updated plot as below.
Modifications to the manuscript and supplementary information: (1) We added a comment on (Page 10, Lines 182 -184) in the manuscript to emphasize this deviation for ECW.
"We observe that the derivations for some cation and anion types are higher. The explanation for the uncertainty in groups like imidazolium and BF4 is included in Supplementary Note 3." (2) We added the above discussion about the deviation for imidazolium and BF4 type ILs in Supplementary Note 3 and updated the plots in the manuscript.
Comment #3: Similarly, in Figure R5 and Table R2 where the authors show the average binding energy for the solid and the liquid compounds, the authors need to show the statistical significance of their hypothesis. Their hypothesis is that they can associate their calculated binding energy with the physical state of the ionic liquid. If the population of each category (liquid, solid 1, solid 2, solid 3) was the same, this comparison could be done by comparing the standard deviations and see whether they are overlapping. However, since the population of each category is different, the authors need to perform a t-test to prove the statistical significance of their hypothesis.

Response:
We thank the reviewer for these important suggestions. We highly agree that the t-test is a valuable method to further confirm our demonstration. Here, we actually use a combination of one-way ANOVA and the two-sample T-test with equal variance (The equal variance was confirmed with F-test for all T-tests) to confirm our demonstration. The results for the two tests are appended below. We observe a significant difference between the groups based on the p-value 0.0045 < 0.05 for the one-way ANOVA. However, the null hypothesis for the one-way ANOVA is that there are at least two pairs that have significant differences for the means. Thus, as shown in  T-test table, besides the first pair, the other pairs show significant small p-values < 0.05, thus we reject the null hypothesis is that there is no difference between these groups. Thus, we can conclude that there is a significant difference between the liquid group and groups with two models showing "solid" prediction results. We believe that there will be some overlapping between the groups, because the binding energy is not the only one important factor to determine the phase, many other factors will also influence the final state of the ILs. Thus, the ML model will play a very important role to combine every factor to give us more reliable results. Similarly, the number in each group has been added to the updated plot in the brackets as well shown below.
Overall, based on the t-test results, we can conclude that the binding energy is a very critical physical property to classify the phase of ILs. The T-test shows more details and indicates significant differences between the liquid and Solid-2/3, Solid-3/3 except for Solid-1/3. Thus, we can conclude that there is a significant difference between the liquid group and groups with more than 2 models showing solid prediction results." (2) Figure 3a has been updated with the number labels for each group in the main manuscript.

Response to Reviewer #2's Comments
Statement: In the revised manuscript, authors have addressed all the comments in great detail. It is commendable that they carried out experiments with a not-so-typical ionic liquid to assess the performance of ionic liquid-polymer electrolyte. On the theoretical side, the set of ionic liquids for which quantum calculations were carried has been expanded. Although I still believe that the method concerning selection of features for machine learning has not been described to permit reproduction of results on its own, availability of Github code can alleviate this challenge.
Overall, the manuscripts combines machine learning to identify ionic liquids with suitable ionic conductivity and electrochemical window. Some of the top performing ionic liquids have been mixed with PBDT and promising performance of the solid electrolyte has been reported for Li-ion batteries. I recommend the publication of the manuscript.

Response:
We greatly appreciate Reviewer #2's comments again. We added more description about the selection of features for the ML in the Method section of the manuscript, thus ensuring the integrity of the manuscript alone. Additionally, we append the GitHub link again for your reference, which has been available to the public community since our last revision. We expect to see more insight on the website.