Ionic polymer electrolytes (IPEs) containing the non-flammable ions embedded with the mechanically supporting polymers with predetermined ionic pathway have received considerable attention toward reviving clean energy storage and conversion devices, such as batteries,1-6 fuel cells,7 supercapacitors,8 mechanical actuators9 and reverse osmosis membranes.10 As promising candidates for safe and environmentally friendly electrolyte materials, ionic liquids (ILs) are room temperature (RT) molten salts with low vapor pressure, high thermal stability, wide electrochemical window and high ionic conductivity.1,5,11 In recent years, liquid crystalline polymers have shown the capability to effectively reduce the interfacial resistance, meanwhile raising unique ion conduction mechanisms in lithium metal batteries (LMBs).1,12 Li metal anode coupled with high-energy-density cathodes, for example Li-air and Li-sulfur batteries, usually require highly conductive, thermal-stable and electrochemical-stable electrolytes to suppress inhomogeneous Li dendrites, overcome the side reactions and break the tradeoff between conductivity and modulus in the composite electrolytes.13,14 To alleviate these issues synergistically, IPEs have shown the capability to block dendrites through the robust polymer matrix, meanwhile guarantee the extreme safety by avoiding organic plasticizer in LMBs.13,15-17
As critical components in IPEs, it is desirable to develop a method to screen suitable ILs from a large population of ion pairs to develop successful IPEs for LMBs. ML has been widely discussed to predict properties and learn the rules underlying datasets, thus efficiently simplifying the material-discovery process.18-23 Here, we describe a machine learning (ML) workflow embedded with the quantum chemistry calculation and graph convolutional neural network (GCN) to discover potential ILs with high ionic conductivity and sufficient electrochemical window. Driven by the structure-property relationships, previous researchers have developed diverse statistical methods and regression models to predict physical properties, for example, melting point24, viscosity25 and ionic conductivity26 based on the structure descriptors of the ILs.27 However, the reported high accuracy usually originates from the overfitting of the dataset.28,29 Among the training datasets, the sample size of unique ILs is extremely limited.27,28 The investigated datasets usually contain the datapoints of the same ILs at varying temperatures, these replicated datapoints will increase the appeal accuracy of the reported models artificially. 25,30 Thus, it is still challenging to predict the accurate properties of new ILs without enough labeled datapoints and preparation of featured factors for the explanatory variables. However, to overcome the data scarcity issue, we can specifically focus on the design of statistical regression and classification workflow to discover new applicable ILs instead of predicting the absolute physical properties of the IL pairs, thus will relief the issue of data scarcity in practical application of ML in material science. In addition, this work also demonstrates the efficiency to use GCN for the graph classification task based on the graph-to-property relationship of ILs.
Based on the screening results from ML, we experimentally investigate some of the filtered ILs combined with a liquid crystalline polyelectrolyte Poly 2,2′-disulfonyl-4,4′-benzidine terephthalamide (PBDT) and a predetermined Li salt to develop a series of IPEs. As reported previously, Li metal anode shows highly reversible cycling-performance with a stable solid-electrolyte interphase (SEI) formed by electrochemically reduction of the bis(fluorosulfonyl)imide (FSI-) anion.31,32 Thus, we employ LiFSI dissolved in ILs as an ion exchange medium to load Li ions in the IPEs. By incorporating these IPEs into batteries with Li metal as an anode, we can experimentally confirm the properties of these developed IPEs, including thermal stability, ionic conductivity, Li+ transference number, electrochemical window and Li dendrite suppression.
Machine Learning-guided Screening Of Ionic Liquids
This machine learning workflow requires two main steps: (1) Unsupervised learning, followed by (2) Supervised learning to target promising ILs. As shown in Fig. 1, we obtain 76 cations and 31 anions from the web scrapping of IoLiTech website. The permutation of cations and anions forms an ionic liquid pool containing 2356 unique ILs, but only less than 13% of the ILs showing measured properties, for example melting point, viscosity, conductivity and electrochemical window. Three open-source platforms, including RDKit, Psi4 and Pytorch Geometrics (PyG), have been employed to generate the molecular descriptors of the raw dataset. RDKit is a powerful tool to calculate the molecular structure and three dimensional (3D) descriptors of the molecules.33 Psi4 is an open-source ab initio electronic structure program for high-throughput quantum chemistry.34 We employ the self-consistence field (SCF) method along with the basis set of 6-311 + G** to optimize the geometric structure and then calcualte the energy, the highest occupied molecular orbital energy (EHOMO), the lowest unoccupied molecular orbital energy (ELUMO) and the molecular dipole moment for the cations and anions seperately. In this work, we combine 60 molecular structure descriptors from RDKit and 14 calculated variables from Psi4 for cations, anions in the final dataset. PyG is a library based on Pytorch to build and train Graph Neural Networks (GNN) for a wide range of prediction tasks. For the unsupervised learning, we employ boxplots, pair plots and hierarchical clustering to summarize the basic rules underlying the dataset. In terms of the supervised learning, we employ both statistical regression and classification to promote the screening efficiency. Among the 2356 ILs, we initially use the classification method to predict the state (solid/liquid) of the ion pairs at RT. Then we employ both statistical regression and classification to evaluate the conductivity of the obtained 1240 ILs (liquid at RT). As we expected, it is particularly challenging to obtain the absolute conductivity of the ILs with high accuracy through regression, especially for the ions with small samples in the training model. To improve the screening efficiency, we utilize a complementary classification method, which will classify the ILs into two categories, based on predetermined and tunable threshold of conductivity, thus will alleviate the low accuracy arises from the limited datapoints. The ensemble learning is based on the combination of algorithms including support vector machine (SVM), random forest (RF), XGBoosting (XGB) and GCN. The last screening step is determined by the computed electrochemical window value (ECW) according to Eq. 1–3 through Psi4. The ECW of an IL is determined by both cations and anions. The cathodic limit (VCL) is the maximum value of the cathodic potential determined by the ELUMO for cations and anions. Similarly, the anodic limit (VAL) is the minimum value of the anodic potential determined by the EHOMO for cations and anions.35 According to this machine learning workflow shown in Fig. 1, we finally obtain 40 ILs in the recommendation list. We will expand the discussion of the unsupervised learning and the supervised learning in the following sections.
\({V}_{CL}=\text{max}\left(-\frac{{E}_{LUMO}[+]}{e},-\frac{{E}_{LUMO}[-]}{e}\right)\) Eq. (1)
\({V}_{AL}=\text{min}\left(-\frac{{E}_{HOMO}[+]}{e},-\frac{{E}_{HOMO]}[-]}{e}\right)\) Equation (2)
\(\text{E}\text{C}\text{W}= {V}_{AL}- {V}_{CL}\) Eq. (3)
Unsupervised Learning
To investigate the features and the underlying correlations in the dataset, we employ the unsupervised learning as a feature detector. Figure 2a,b display the boxplots of the ILs with known σ in the dataset, which are classified by the cation and anion types correspondingly. In terms of the cations, the average σ follows the order of ammonium > sulfonium > pyrrolidinium > imidazolium > pyridinium > piperidinium > phosphonium. Accordingly, the rule for the anions is nitrate > thiocyanate > DCA > BF4 > triflate > sulfate > imide > sulfonate > acetate > halogen > phosph. Figure 2c,d display the rankings for the measured ECW of the ILs correspondingly. We observe that the ECW boxplots showing large variations, but the ammonium-based cations and imide-based anions commonly displaying promissing ECWs. In order to understand the corelations between different features, as shown in Fig. 2e, we observe that the σ is highly corelated with the viscosity of ILs, which can be elucidated by Nernst-Einstein equation and Stokes-Einstein equation included in Supplementary Note 1.36 However, as shown in Fig. 2f, there is no strong corelation between σ and ECW. Thus, we can conclude that σ and ECW of ILs are two independent features. As we familar, both σ and ECW are essential properties for electrolytes applied for LMBs. Zhang30 has employed hierachical clustering to predict fast inorganic Li ion conductors. Here, we also employ this algorithm to detect the clusters of ILs with high conductivity and wide ECW. The key features used for the hierachical clustering include the computed ECW and the top 10 factors obtained from the XGBoosting model for conductivity classification (Supplementary Note 2). The clustering criterior is a combination of the σ and the ECW of the ILs. The final results are shown in Fig. 2g. Additionally, in order to validate the effectiveness of the clustering, we also plot the finally screened ILs based on the supervised learning in Fig. 2h, we can clearly observe that the results based on the two learning protocols are highly overlaped, which indicates the high efficiency of the unsupervised learning as compared to the supervised learning introduced in the following section.
Supervised Learning
Building on the unsupervised learning, we initially propose a complementary supervised learning to filter ILs with liquid phase at RT based on the ensemble learning of SVM, RF, XGB and GCN. The classification results are shown in the Supplementary Note 3. Figure 3a indicates the corelation between the radius of gyration for anion and the heavy atom molecular weight for the classified solid and liquid groups, we observe that the ILs with high sphericity tend to be in solid state, whereras those with high asymmetry of anions will usually turn to be liquid state. We also include the top 10 importance features of XGB model in Fig. 3b. As we expected, the phase of ILs are highly dependent on the gyometic strucutres of anions. To further validate the predicted results, we employ quantum chemistry calculation to calculate the interaction energy between cations and anions. We randomly select 20 ILs from the predicted solid and liquid groups, seperately. The interaction energies of the selected ion pairs are calculated based on Eq. 4. As shown in Fig. 3c, the interaction energies for the solid ILs are generally lower compared to the liquid ILs. The detailed calculation results are included in Supplementary Table 1. As we know, the interaction energy of the cation and anion is the energy required to combine cation and anion. The lower of the interaction energies means easier for the cations and anions to stay in tightly associated pairs. In addition, we also employ GCN model to investigate this classification. The workflow of the GCN model is shown in Fig. 3d. The details about the GCN model used in this work are also included in the Supplementary Note 4. The efficiencies of employed models are included in Table 1. We observe that the GCN model offers similar performance only based on the chemcial geometric descriptors without any additional calculation results, which shows high promise for this algotihrm for future materials genome projects. Additionally, we also apply the other three models to predict the absolute σ of the liquid (1084) ILs pairs at RT. The boxplots shown in Fig. 3e further indicate the consistency between the regression and the classification results. The median value of the predicted conductivity for cluster with σ < 5 is 1.9 mS cm− 1. Meanwhile, the median value for cluster with σ ≥ 5 is 8.9 mS cm− 1. Finally, we propose 40 ILs with σ ≥ 5 mS cm− 1 along with an satisfactory ECW > 3.5 V in the final recommendation list for development of IPE in LMBs. The full list is included in Supplementary Table 2. The bubble plots shown in Fig. 3f, g indicate the distribution of the filtered ILs based on the cation and anion types, correspondingly. In terms of the cations, we observe that the imidazolium cations display the highest average ionic conductivity. In terms of the anions, BF4, sulfate and triflate show both promising ECW and σ. Base on this prliminary rules, we select four widely discussed ILs from the bubbles and conduct experimetanl validation in the following section.
\({\text{E}}_{interaction energy}= {\text{E}}_{\left[+\right]\left[-\right]}- {\text{E}}_{[+]}-{\text{E}}_{[-]}\) Eq. (4)
Table 1
The performance of the classification tasks based on different algorithms.
| Algorithm | R2 (Solid or Liquid) | R2(σ ≥ 5 or σ < 5) |
| Support Vector Machine (SVM) | 0.80 | 0.82 |
| Random Forest (RF) | 0.85 | 0.81 |
| XGBoosting (XGB) | 0.86 | 0.81 |
| Graph Convolutional Neural Network (GCN) | 0.83 | NA |
*The R2 for SVM, RF and XGB algorithms are the avergae of the 5-fold cross-validation accuracy results.
Electrochemical Performance Of Ipes Based On The Filtered Il
To verify the efficiency of this screening process, we develop a series of IPEs based on 4 hydrophilic ILs on the final recommendation list and validate the cycling behavior and electrochemical performance of the developed IPEs coupled with the Li metal anodes. The selected ILs include 1-ethyl-3-methylimidazolium triflate (C2mimTFO), 1-ethyl-3-methylimidazolium tetrafluoroborate (C2mimBF4), 1-ethyl-3-methylimidazolium ethyl sulfate (C2mimEthylSulfate) and diethylmethylammonium triflate (DemaTFO). All these ILs are hydrophilic and showing high ionic conductivity at RT, we can use solvent-casting method to prepare ultrathin (~ 50 µm) polymer electrolytes as reported previously.1,6,37 The fabricated films are transparent and mechanically strong with 5/10wt% of PBDT. After ion exchange with the ionic liquid electrolyte (ILE) with LiFSI dissolved in N-propyl-N-methylpyrrolidinium bis(fluorosulfonyl)imide (C3mpyrFSI), we obtain large area and flexible IPE membranes. The solid-state membrane electrolytes with diameter of 9 mm shown as insets in Fig. 4a-c are assembled in LMBs. Initially, we perform cyclic voltammetry (CV) on Li|IPEs|SUS to evaluate the Li plating (negative scan) and stripping (positive scan) behavior of the selected ILs combined with PBDT. We observe that upon scanning in the negative direction, this electrolyte displays excellent cathodic stability for Li-metal cycling. The achieved high current density offers promise for these IPEs as conductive electrolytes in LMBs. Among the four ILs, only DemaTFO displays no Li deposition in the cathodic deposition, which indicates that the ion exchange protocol fails to incorporate Li ions in DemaTFO based IPEs. Thus, we conclude that the ion association behavior and varying interaction energies are essential factors to determine the final chemical composition in the IPEs. In the following discussion, we will only focus on the other three IPEs. For the positive direction, as shown in the enlarged curves in the insets, we observe that C2mimTFO based IPEs show the highest anodic stability up to 5 V vs Li|Li+, which confirms the high electrochemical stability of IPEs based on C2mimTFO. However, both BF4- and Ethyl Sulfate display fluctuate current starting from 4V for BF4 and 3.5V for Ethyl Sulfate, which are lower compared to the computed ECWs. As reported previously, the calculated anodic limits for BF4- are usually overestimated using the vacuum calculation model compared to the experimental results.35 The reasons for the inconsistency originate from many factors, for example the assumption of the calculation model and the complicated ion association in the real system. For comparison, we include the CV results for the three neat ILs in Supplementary Note 5. Surprisingly, the cathodic limits for the neat ILs are much lower compared to the IPEs. We propose that the high Li+ concentration and PBDT in the IPEs can improve the cathodic stability of the composite electrolytes, which is analogy to the “water in salt” electrolytes proposed by Suo et al.38 The solvation of the Li+, cations and anions will change dramatically with the adjustment of the relative ion concentrations. We have demonstrated that the rigid-rod polyelectrolyte PBDT backbone will selectively absorb cations and anions during the ion exchange process, which will promote the Li transference number (tLi+) of the system. Here, we estimate the tLi+ based on the steady-state current of the Li symmetric cell assembled with IPEs as electrolyte and separator in Fig. 4d. The insets show the corresponding impedance spectra of the cells before and after the polarization. The tLi+ of Bruce-Vincent analysis is defined in Supplementary Note 6.39 The determined tLi+ in these IPEs (0.4–0.5) are much higher compared to the pure ILE with LiFSI|C3mpyrFSI (tLi+ = 0.18)40. Among these three IPEs, C2mimTFO based IPEs show higher tLi+= 0.5 compared to C2mimBF4 (0.4) and Dema Ethyl Sulfate (0.4) for the varying interaction between PBDT chains and the anions in ILs. We are conducting in depth quantum chemistry calculations to measure the interaction energies between anions and PBDT polymer chains, thus will offer more clear idea about the determining factors in the ion exchange process. Figure 4e shows ionic conductivities of the IPEs as a function of temperature. The exceedingly high σ at RT (> 1 mS cm-1) originates from the fibrillar and nanocrystalline conducting phase formed in the composite structure of IL and liquid-crystalline polymers reported previously.1,6 Meanwhile, all samples show stable Li stripping and plating at varying current grades in the Li||Li cell cycling process. As shown in Fig. 4f, C2mimTFO and C2mim Ethyl Sulfate display stable cycling at high current density (J) at 6 mA cm-2 at 80°C, which is promising, since most organic cells reported cannot sustain any stable performance at this high temperature without any safety concerns.
In this section, we mainly extend the investigation of IPEs based on C2mimTFO at varying temperature and cell configurations. In Fig. 5a, we initially test the symmetric cell performance at varying J at RT. The critical J is 2.0 mA cm− 2 at RT. As shown in Fig. 5b, the cells can maintain at least 800 hours at 1 mA cm− 2 at RT, which is rarely seen by other single-layer polymer electrolytes without any supporting separators or organic plastiziers.41 In addition, we prepare the Li|IPEs|Cu cell to investigate the plating and striping of Li at the Cu anode based on these IPEs. Figure 5b shows the voltage-cycle profile for the cells at RT. The secondary axis shows the corresponding columbic efficiency (CE) values, the average value of which is > 98%. The reported average CE for Li||Cu cell using regular organic electrolytes is ~ 90%, which indicates the highly reversible Li deposition on Cu surface of these IPEs.42–44 The SEM images of the deposited Li on current collector is included in the Supplementary Note 7. We observe no Li dendrites formed using these IPEs coupled with Li metal anodes.
For practical application, we also report the full cell performance based on the commercial LiFePO4 cathode with high loading (10.3 mg cm− 2). Figure 6a shows the long-term cycling of the full cell at RT at 0.5 C rate, the cell shows 96% capacity retention at 350 cycle. The voltage capacity files for the selected main cycles are shown in Fig. 6b. The major population of the CE is between 100.1% – 100.2%, which is slightly higher than the theoretical maximum value (100%) because of the thermal fluctuation at the ambient temperature. Thus indicates the high reversibility of the Li metal cells based on this IPEs at RT. Figure 6c shows the long-term cycling of the full cell at 50°C at 2 C rate, the cell shows not only high average CE (> 99.9), but also 80% capacity retention at 350th cycle, which is promising to satisfy the fast charge/discharge requirements for widely used portable devices. The voltage capacity files for the selected main cycles are shown in Fig. 6d. At last, we also investigate the fast charge/discharge capability and thermal stability of this IPEs at 80°C, which is a dangerous temperature for regular organic cells. Figure 6e shows the cycling performance of the IPE at increasing C rate from 0.5 C (0.53mA/cm2) to 5 C (5.3 mA cm-2)) at 80°C. Here, we observe that the cell can cycle without short circuit at super high J up to 5C (8.3 mA cm− 2), which show promise for this IPEs as next-generation solid-state electrolytes for fast charge LMBs at mid-high temperatures. The corresponding voltage capacity curves for the selected cycles are shown in Fig. 6f.
In summary, we have described a machine learning-guided screening protocol to filter promising ILs with high ionic conductivity and wide electrochemical window for preparation of IPEs in LMBs. In addition, we further confirm the rigid-rod liquid crystalline polyelectrolyte PBDT as an essential polymer matrix to develop a series of solid-state polymer electrolytes with extremely high CE and excellent fast charge and discharge performance at high temperature. This platform shows immense potential to serve as an efficient method to quickly focus on the essential ionic liquids for specific applications. More importantly, this work provides novel insights into strategies to overcome the data scarcity issues and realize the efficient utilization of ML in material design and optimization. Through investigation of the golden rules, we could fabricate IPEs with tunable variations in mechanical, structural and transport properties for a large array of applications in versatile functional devices, including capacitors, mechanical actuators, water reverse osmosis membranes and so on.