Ionic polymer electrolytes (IPEs), incorporating non-flammable ions within mechanically supportive polymers, are gaining traction for next-generation energy storage devices. Ionic liquids (ILs), with their inherent properties of high ionic conductivity, wide electrochemical windows, low vapor pressure, and high thermal stability, are particularly promising candidates for IPEs designed for safe and high-energy-density lithium metal batteries (LMBs). LMBs, particularly those employing high-energy-density cathodes like Li-air and Li-sulfur, demand electrolytes that possess high conductivity, thermal stability, and electrochemical stability to mitigate issues such as Li dendrite formation and side reactions. IPEs offer a compelling solution by providing a robust polymer matrix that effectively blocks dendrites while eliminating flammable organic plasticizers. The challenge lies in efficiently screening suitable ILs from a vast pool of candidates. Traditional methods are time-consuming and resource-intensive. Machine learning (ML) techniques offer a powerful alternative, enabling rapid screening and prediction of material properties. While ML has been applied to predict IL properties, challenges remain, particularly in addressing data scarcity and overfitting issues. Existing models often rely on limited datasets or overfit to repeated data points at varying temperatures, hindering accurate prediction for novel ILs. This work addresses these challenges by employing a novel ML workflow that combines unsupervised and supervised learning, utilizing quantum chemistry calculations and graph convolutional neural networks (GCNNs) for efficient screening and subsequent experimental validation.

Previous research has explored various statistical methods and regression models to predict IL properties like melting point, viscosity, and ionic conductivity using structural descriptors. However, these methods often suffer from overfitting due to limited datasets and the inclusion of repeated data points, compromising the accuracy of predictions for new ILs. The use of liquid crystalline polymers to improve ion conduction mechanisms in LMBs has also been investigated, highlighting the potential for enhancing performance. Studies on the use of ionic liquids in energy storage devices and other applications, including batteries, fuel cells, supercapacitors, actuators and membranes showcase their versatile properties and potential. Several publications address the use of machine learning in material discovery and the prediction of material properties in general, and in particular, some studies have focused on applying machine learning to predict the properties of ionic liquids and their suitability for use in battery applications. However, these existing studies often suffer from limitations related to data scarcity and overfitting which were addressed in this study by developing a robust workflow combining unsupervised and supervised learning techniques.

The proposed ML workflow consists of two primary stages: unsupervised and supervised learning. The unsupervised learning phase employs web scraping to gather data on 74 cations and 30 anions from the IoLiTec website, creating a pool of 2220 unique ILs. Only a small fraction of these ILs have experimentally measured properties. Open-source tools, RDKit, Psi4, and PyTorch Geometric (PyG), are utilized to generate molecular descriptors. RDKit calculates 3D molecular descriptors, while Psi4, an open-source ab initio electronic structure program, employs the self-consistent field (SCF) method with Hartree-Fork theory and a 6-311+G** basis set to optimize geometry and calculate electronic properties (energy, HOMO, LUMO, dipole moments) for cations and anions individually. These descriptors are combined to create the final dataset. Unsupervised learning methods—boxplots, pair plots, and hierarchical clustering—analyze the dataset, revealing underlying correlations between properties. The supervised learning phase uses statistical regression and classification algorithms (SVM, Random Forest, XGBoosting, and GCNN) to predict IL properties. First, a classification model predicts whether an IL is liquid or solid at room temperature (RT). Subsequently, liquid ILs are classified into those with ionic conductivity (σ) above and below 5 mS cm⁻¹ (a threshold chosen to ensure high IPE conductivity). Regression models predict the absolute σ values. Finally, the electrochemical window (ECW) is calculated using the HOMO/LUMO energies, and ILs with ECW > 4 V are selected. This multi-step approach significantly improves screening efficiency. Based on the ML recommendations, a series of IPEs are experimentally fabricated using selected ILs, combined with a rigid-rod polyelectrolyte (PBDT) and LiFSI salt. The fabrication process involves solvent evaporation to create precursor membranes followed by ion exchange with a concentrated LiFSI/C3mpyrFSI solution. The resulting IPEs are characterized for ionic conductivity, Li⁺ transference number, electrochemical window, and dendrite suppression using various electrochemical techniques (cyclic voltammetry, impedance spectroscopy, and battery cycling tests). The mechanical properties of the IPEs are also investigated using stress-strain and dynamic mechanical analysis.

The ML workflow successfully identified 49 promising ILs with high ionic conductivity (σ ≥ 5 mS cm⁻¹) and wide electrochemical windows (ECW > 4 V). Unsupervised learning revealed that ammonium-based cations and imide-based anions tend to exhibit better ECWs, while pair plots showed no correlation between σ and ECW, indicating that both are independent factors. The calculated ECWs showed good agreement with experimental values from IoLiTec, with a mean absolute error (MAE) < 1.1 V. Hierarchical clustering effectively identified ILs with both high conductivity and wide ECW. Supervised learning further refined the selection of ILs, with GCNN demonstrating comparable performance to other algorithms. The model's accuracy was validated using the NIST ILThermo database, resulting in an R² of 0.82 and an MAE of 1.8 mS cm⁻¹. Experimental validation using five selected ILs demonstrated that the fabricated IPEs were transparent, mechanically robust (>200 MPa), and exhibited high Li⁺ transference numbers (0.4–0.5). The Li|IPEs|Li cells exhibited ultrahigh critical current density (6 mA cm⁻²) at 80 °C. Li|IPEs|LiFePO₄ cells (with 10.3 mg cm⁻² cathode loading) demonstrated outstanding capacity retention (>96% at 0.5C over 350 cycles, >80% at 2C), fast charge/discharge capability (146 mAh g⁻¹ at 3C), and excellent coulombic efficiency (>99.92%). These results significantly outperform many previously reported single-layer polymer electrolytes for LMBs without flammable organic components. The symmetric cell performance at varying current density was investigated, showing a critical current density of 2.0 mA cm⁻² at room temperature and stable operation at 1 mA cm⁻² for at least 800 h without short circuit. The Li|IPEs|Cu cell showed high coulombic efficiency (>98%) and no Li dendrite formation. The full-cell performance at varying temperatures and C-rates revealed high capacity retention and efficiency.

The superior performance of the developed IPEs confirms the effectiveness of the ML-guided screening protocol. The high critical current density, excellent capacity retention, and fast charge/discharge capabilities are attributed to the synergistic effects of the selected ILs, the rigid-rod PBDT polymer matrix, and the LiFSI salt. The PBDT backbone provides mechanical integrity and promotes nanoscale structuring, facilitating fast Li⁺ transport. The combination of high Li⁺ concentration, decomposition products of FSI, and the liquid crystalline PBDT contributes to enhanced cathodic stability, similar to water-in-salt electrolytes. The high Li⁺ transference numbers suggest selective absorption of cations and anions by the PBDT during ion exchange. The exceptional results demonstrate the potential of this ML-guided approach for accelerating the discovery of high-performance electrolytes for LMBs. The use of commercially available ILs from IoLiTec enhances the practicality and reproducibility of the findings.

This study presents a highly efficient machine learning-guided screening protocol for identifying promising ionic liquids for use in ionic polymer electrolytes designed for lithium metal batteries. The combination of unsupervised and supervised learning techniques, coupled with quantum chemical calculations and graph convolutional neural networks, successfully identified ILs that, upon experimental validation, resulted in IPEs demonstrating superior electrochemical performance. The results significantly advance the field of solid-state electrolytes and highlight the power of integrating machine learning and experimental validation for materials discovery. Future research could focus on exploring a wider range of ILs and polymer matrices, investigating the detailed interactions between the components in IPEs, and optimizing the ion exchange process for even better performance.

The accuracy of the ML model depends on the quality and quantity of the training data. While efforts were made to mitigate overfitting, there's always a risk of limited generalizability to ILs outside the scope of the training dataset. The experimental validation was performed on a limited number of ILs; further experiments with additional ILs could broaden the understanding of the structure-property relationships. The calculated electrochemical window using the HOMO/LUMO theory, while providing a useful screening metric, does not capture all the complexities of the electrochemical behavior in real battery systems. Future studies focusing on these limitations could further improve the efficacy of the overall workflow.