Evaluating the Generalizability of Inorganic Pretrained General-Purpose Graph Neural Network Potential in Simulations of Liquid Electrolytes

Lithium-ion batteries (LIBs) are crucial to modern technology, powering a wide range of devices from mobile electronics to electric vehicles. Despite their widespread use, challenges such as safety, charging speed, and temperature resilience continue to limit their broader application. Electrolyte engineering, which directly impacts charge transport and the overall stability of the battery system through the formation of the solid electrolyte interphase (SEI), is key to overcoming these challenges. Theoretical approaches are illuminating in their predictive ability prior to trial-and-error experimental methods. However, recent theoretical studies based on density functional theory (DFT), classical potentials, and machine learning potentials have encountered challenges with the complexity of electrolytes due to the vast combinations of solvents, salts, and additives. The DFT method faces scalability limitations and requires high computational costs to investigate diverse chemical systems. Classical force fields lack transferability and overall accuracy. Machine learning interatomic potentials (MLIPs) offer a promising alternative, extending the length and time scales of simulations while maintaining DFT-level accuracy. Nonetheless, bespoke descriptor-based MLIPs require extensive training sets to capture interactions among numerous species and are limited by the number of elements they can handle.<br/> In this presentation, we address computational challenges in electrolyte simulations by employing a pretrained general-purpose graph neural network interatomic potential (GNN-IP) to explore its applicability in liquid electrolyte systems. Specifically, we assess the generalizability and limitations of the SevenNet-0 model (version 11July2024), which was trained on inorganic crystals from the Materials Project database. We conduct molecular dynamics (MD) simulations on 20 solvents and two salts across various compositional combinations. Our results demonstrate that SevenNet-0 accurately describes the first Li-ion solvation shell, including bond lengths and angles in carbonate solvents, and effectively extends to Na and K ions with accurate cation-oxygen distances. Although the model exhibits systematic shifts in density predictions, leading to an underestimation of diffusivity, scaling densities to experimental values yields diffusivity predictions comparable to experimental results. Through Uniform Manifold Approximation and Projection (UMAP) analysis, we suggest that SevenNet-0 infers organic solvent chemical moieties and Li-ion solvation shells from similar inorganic local structures and other cation-oxygen/fluorine interactions. However, SevenNet-0 shows limitations in accurately predicting energy barriers between cis-cis and cis-trans conformers in linear solvents, as well as in atomic forces for specific chemical moieties with 5-atom rings and fluorine doping. Fine-tuning with a small set of organic structures led to improvements in these areas. This study provides valuable insights into the use of inorganic pretrained models for out-of-distribution systems, emphasizing the importance of carefully assessing relevant properties for target applications. Our findings offer a promising pathway for advancing electrolyte engineering in complex battery systems, potentially accelerating the development of next-generation high-performance batteries.