(368b) A Machine Learning Study on the Result of Polarizable Molecular Dynamics of Ionic Liquid-Based Solid Polymer Electrolytes for Li+-Ion Batteries By Graph Dynamical Networks

Solid polymer electrolytes (SPEs) are promising candidates for safe and stable lithium-ion secondary battery electrolytes. Poly(ethylene oxide) (PEO) has been widely used as a polymer host, and room temperature ionic liquids (RTILs) have been added to significantly increase lithium ion mobility in SPEs. Our study focused on exploring the role of each component and geometrical condition in the electrolytes from simple ionic liquid electrolytes to multicomponent complex SPEs: LiTFSI/PYR₁₄TFSI binary bulk, LiTFSI/PYR₁₄TFSI/PEO ternary bulk, and LiTFSI/PYR₁₄TFSI mixture in a PEO-branched cylindrical nanopore. Molecular dynamics (MD) simulations of these systems were performed previously [1] under the APPLE&P polarizable force field.

An unsupervised learning method called Graph Dynamical Networks (GDyNets) [2] was suggested for learning atomic scale dynamics from MD trajectories of arbitrary condensed phases. GDyNets can classify local configurational states around a target atom using graph convolutional neural networks and obtain transition probabilities between the states based on the Koopman model [3].

In this study, the MD trajectory of each electrolyte was trained with GDyNets where lithium ions were determined as target atoms. For each trajectory, we measured several state-wise and transition-wise properties of all Li⁺ ions from the trained model. Population of each classified state was calculated. State-weighted radial and spatial distribution functions around the Li⁺ ions identified the local structure of each state. The Koopman matrix was evaluated after the lag time was determined. From the eigenvector decomposition of the Koopman matrix, the slowest transitions were identified, and their timescales were evaluated. The Chapman-Kolmogorov test was performed to check whether the trained model is Markovian. Conductivity of each transition revealed that which transitions have major contributions to the overall conductivity. In addition to properties of all Li⁺ ions, we analyzed individual ion trajectories with the same model, and ions with similar behavior were clustered into a group. Our study contributes to elucidate lithium-ion acceleration conditions in SPEs and give a systematic method to analyze multicomponent MD trajectories with machine learning techniques.

References

[1] Kim, Seulwoo, et al. "Molecular dynamics study on lithium‐ion transport in PEO branched nanopores with PYR₁₄TFSI ionic liquid." Battery Energy (2022): 20210013.

[2] Xie, Tian, et al. "Graph dynamical networks for unsupervised learning of atomic scale dynamics in materials." Nature Communications 10.1 (2019): 1-9.

[3] Koopman, Bernard O. "Hamiltonian systems and transformation in Hilbert space." Proceedings of the National Academy of Sciences of the United States of America 17.5 (1931): 315.