Unsupervised learning of representative local atomic arrangements in molecular dynamics data

Abstract: Molecular dynamics (MD) simulations present a data-mining challenge, given that they can generate a considerable amount of data but often rely on limited or biased human interpretation to examine their information content. By not asking the right questions of MD data we may miss critical information hidden within it. We combine dimensionality reduction (UMAP) and unsupervised hierarchical clustering (HDBSCAN) to quantitatively characterize the coordination environment of chemical species within MD data. By focusing on local coordination, we significantly reduce the amount of data to be analyzed by extracting all distinct molecular formulas within a given coordination sphere. We then efficiently combine UMAP and HDBSCAN with alignment or shape-matching algorithms to classify these formulas into distinct structural isomer families. The outcome is a quantitative mapping of the multiple coordination environments present in the MD data. The method was employed to reveal details of cation coordination in electrolytes based on molecular liquids.