Elucidating interfacial reaction and diffusion mechanisms with unsupervised characterization of ML-driven molecular simulations

Recent breakthroughs in machine learning interatomic potentials have allowed access to large-scale molecular dynamics (MD) simulations, necessitating automatic and flexible analysis of their dynamics. Conventional techniques, like common neighbor analysis or polyhedral template matching, constrain analysis to crystalline structures or rely entirely on prior knowledge of the system dynamics. To overcome these limitations, we employ an unsupervised clustering algorithm to characterize phase transitions and product formation of large-scale interfacial reaction simulations. By clustering geometric descriptors of local atomic environments within MD trajectories, we reveal dynamic insights previously inaccessible.

We demonstrate this technique by examining high-temperature and pressure phase transitions of silicon carbide and interfacial reactions leading to the formation of the solid-electrolyte interphase in solid-state lithium batteries. Our method not only traces the time evolution of intricate phases—including complex molecules and crystalline, amorphous, or solid-state compounds—but also identifies the formation of crystalline P, Cl co-doped Li2S, differing from the thermodynamic predictions. This automated, data-driven framework provides insight into the intricate dynamic processes of solid-state reactions and phase nucleation in large-scale MD simulations.