Unsupervised characterization of reactions in ML-driven molecular simulations

Machine learning interatomic potentials now enable large-scale reactive molecular dynamics (MD) simulations that capture complex reactions and phase transformations. However, the scale of these simulations renders manual analysis impractical, demanding generalized and automated methods to characterize system dynamics. Conventional structural analysis techniques, such as common neighbor analysis or polyhedral template matching, are constrained to crystalline order. Alternatively, supervised learning approaches rely on prior assumptions and chemical intuition.

To overcome these limitations, we introduce a generalized, unsupervised framework for identifying complex interfacial reactions and phase transitions. The method employs density-based clustering of geometric descriptors of local atomic environments. We demonstrate its versatility across diverse systems, including high-temperature phase transitions in silicon carbide, solid-electrolyte interphase formation in solid-state lithium batteries [1], and hydrogen turnover on Pt(111) catalytic surfaces. Beyond recovering known processes, our framework reveals previously unreported phases, such as Li₂S_0.72P_0.14Cl_0.14 [1] in solid-state lithium batteries. This scalable, data-driven approach offers new insight into the intricate reaction mechanisms at complex interfaces.

[1] Ding, J., Zichi, L., Carli, M., Wang, M., Musaelian, A., Xie, Y., & Kozinsky, B. (2025). Coupled reaction and diffusion governing interface evolution in solid-state batteries. arXiv preprint arXiv:2506.10944.