(Invited) Scalable and Generalizable Machine Learning Force Fields for Modeling Complex Battery Materials

Le résumé fourni par la source

The rapid advancements in rechargeable Li-ion battery (LIB) technology has revolutionized several key industries such as automotive and consumer electronics . However, new battery chemistries are needed to improve the power density, safety, reliability, and lifetime of LIBs. Existing classical force fields are not accurate enough to predict bulk properties of LIB materials without time-consuming and customized parametrization. To move towards accurate and reliable modeling of battery chemistries, we developed a machine-learned force field (MLFF) using a charge recursive neural network (QRNN) architecture, which includes both long-range interactions and global charge redistribution. The MLFF is trained to model a large chemical space of industrially relevant liquid electrolyte chemistries and enables large scale molecular dynamics (MD) simulations of realistic electrolyte formulations. In this presentation, I will demonstrate our generalized active learning framework and sampling workflow used to generate accurate training data for liquid and amorphous systems. I will discuss large scale benchmarks carried out to evaluate the performance of MLFF against experimental data for key physical and transport properties of liquid electrolytes including density, viscosity and ionic conductivity. Our results indicate that MLFF outperforms the classical force fields in terms of quantitative agreement with experimental data across a broad range of electrolyte chemistries. I will also discuss the novel molecular level insights into the unique Li+ cation solvation structures predicted by MLFF and their validation using experimental nuclear magnetic resonance (NMR) spectroscopy. Finally, I will briefly discuss our recent work to develop a message passing neural network architecture to train universal MLFF for inorganic materials covering up to 94 elements from the periodic table allowing for efficient design of materials for cathodes, coatings and and solid-state electrolytes for applications in next-generation batteries.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.