Machine Learning-Guided Screening of Advantageous Solvents for Solid Polymer Electrolytes in Lithium Metal Batteries

Technical Summary: Machine Learning-Guided Screening of Advantageous Solvents for Solid Polymer Electrolytes in Lithium Metal Batteries

Problem Statement
Solid polymer electrolytes (SPEs) are critical for advancing safe, high-energy lithium metal batteries due to their favorable processability and stable electrode interfaces. However, they often suffer from lower ionic conductivity and mechanical strength compared to inorganic counterparts. While strategies like incorporating inorganic fillers exist, the introduction of organic solvents remains a simpler and widely adopted method to enhance lithium salt dissociation and stabilize interfaces. A significant challenge lies in the complexity of solvent selection: trace residual solvents (either deliberately added or inadvertently retained from synthesis) significantly impact electrolyte and interface properties. Currently, selecting optimal solvents relies heavily on empirical trial-and-error or empirical solvent blending, which is costly and lacks a unified theoretical framework. Furthermore, existing machine learning approaches for material discovery often remain purely theoretical, lacking experimental validation, or rely heavily on data quality without establishing robust links between electronic structures and macroscopic performance.

Methodology
The authors propose an integrated paradigm combining high-throughput Density Functional Theory (DFT) calculations, machine learning (ML), and experimental validation to screen and design advantageous solvents for SPEs.

1. High-Throughput DFT Screening: A dataset of approximately 10,000 organic solvents containing C, H, O, N, and F was extracted from the ChemSpider database. High-precision DFT calculations were performed to determine key electronic descriptors (HOMO and LUMO energy levels) and macroscopic properties: polarizability ($\alpha$), dipole moment ($\mu$), dielectric constant ($\varepsilon$), and donor number (DN). The dielectric constant was calculated via the Clausius-Mossotti equation, and DN was determined by interaction energies with the Lewis acid SbCl$_5$.
2. Machine Learning Modeling: To establish predictive models linking electronic structures to macroscopic properties, the authors employed a multi-level ML strategy:
   • XGBoost and SISSO: Used for feature importance analysis and initial screening. These interpretable models confirmed that electronic structure attributes (specifically HOMO-LUMO gaps) contribute over 65% to predicting dielectric constants.
   • Crystal Graph Convolutional Neural Network (CGCNN): Adapted for molecular systems under periodic boundary conditions to capture complex non-linear interactions. This model achieved high predictive accuracy ($R^2$ = 0.91 for $\mu$, 0.97 for $\varepsilon$, and 0.98 for the HOMO-LUMO gap).
   • Validation: The models underwent rigorous validation, including 50-fold cross-validation, multiple-seed tests, and y-scrambling, confirming that the models learned genuine structure-property relationships rather than memorizing data.
3. Screening Criteria and Candidate Selection: Based on the ML insights, quantitative screening criteria were established: HOMO < –6 eV, LUMO > –1 eV, HOMO-LUMO gap > 5 eV, $\varepsilon$ between 20–50, and $\mu$ between 3–6 D. These parameters aim to balance oxidative/reductive stability, salt dissociation, and ion-solvent interactions.
4. Experimental Validation: Two top candidates, 2,2,2-trifluoro-N,N-dimethylacetamide (TFDMA) and the newly proposed N-methoxy-N-methyl-2,2,2-trifluoroacetamide (TFOMA), were selected. Trace amounts of these solvents were incorporated into a poly(vinylidene fluoride-co-hexafluoropropylene) (PVDF-HFP) matrix. The resulting SPEs were characterized for electrochemical stability, ionic conductivity, transference numbers, and cycling performance in full cells with LiFePO$_4$ and LiNi$_{0.9}$Co$_{0.05}$Mn$_{0.05}$O$_2$ (NCM91) cathodes.

Key Results

• Structure-Property Correlations: The study confirmed a universal correlation where larger dipole moments and relative polarizabilities generally correspond to higher dielectric constants. A non-linear dependence was observed between the HOMO-LUMO gap and macroscopic properties; properties decrease sharply as the gap increases from 0.5 to 2 eV, then stabilize.
• Model Performance: The CGCNN model successfully captured the "black box" nature of deep learning, with latent embeddings showing strong correlations to electronic structure descriptors. SHAP analysis confirmed the dominant role of frontier orbitals in determining dielectric constants and dipole moments.
• Candidate Identification: TFOMA and TFDMA were identified as superior candidates. TFOMA, featuring an additional methoxy group, was predicted to enhance ionic pathways.
• Experimental Performance:
  • Stability: PVDF-HFP@TFOMA exhibited a wide electrochemical stability window of 4.5 V, outperforming TFDMA (4.25 V).
  • Conductivity: At 30°C, TFOMA-based SPEs achieved an ionic conductivity of 5.5 $\times$ 10$^{-4}$ S cm$^{-1}$, significantly higher than TFDMA (1.0 $\times$ 10$^{-4}$ S cm$^{-1}$). The activation energy for Li$^+$ migration was lower for TFOMA (0.29 eV) compared to TFDMA (0.45 eV).
  • Transference Number: TFOMA-based electrolytes demonstrated a high Li$^+$ transference number of 0.78.
  • Cycling Stability: In LiFePO$_4$ cells, the TFOMA system retained 86.7% capacity after 500 cycles at 2C. In high-voltage NCM91 cells, it retained 98.7% capacity after 200 cycles at 2C. These results surpassed those of TFDMA and other common solvents like DMF, NMP, and DMSO.
  • Mechanism: Molecular dynamics simulations suggested that the methoxy group in TFOMA facilitates the formation of localized high-concentration structures and creates new ion channels, reducing Li$^+$-anion Coulombic interactions.

Significance and Claims
The paper claims to establish a universal, data-driven paradigm for solvent screening in solid polymer electrolytes. By integrating high-throughput DFT, advanced machine learning algorithms, and experimental validation, the authors demonstrate a pathway to rationally design electrolytes that balance ion transport, electrochemical stability, and cycling durability. The study highlights that trace residual solvents, when properly selected, are not merely impurities but can be engineered to enhance battery performance. The successful identification and validation of TFOMA as a superior solvent for PVDF-HFP matrices underscores the efficacy of this integrated approach. The authors posit that this methodology can be extended to other advanced electrolyte designs, accelerating the development of safer, high-energy lithium metal batteries. The work bridges the gap between theoretical electronic structure predictions and practical battery performance, offering a robust framework for next-generation materials discovery.