Nanoscale and condensed matter systems evolve on multiple length-and time-scales, and rare events such as local phase transformation, ion segregation, defect migration, interface reconstruction, and grain boundary sliding can have a profound influence on material properties. We demonstrate how outlier detection indices can be used to identify rare events in machinelearning based, high-dimensional molecular dynamics (MD) simulations. Designed to order data-points from typical to untypical, the indices enable one to capture atomic events that are hard to detect otherwise. We demonstrate the approach with a nanosecond MD simulation of a grain boundary in a metal halide perovskite that is extensively studied for solar energy and optoelectronic applications. The method captures the initial grain boundary sliding and a spontaneous fluctuation half a nanosecond later, both events giving rise to persistent deep electronic trap states that impact charge carrier lifetime and transport and material performance. The approach offers a generalizable and simple method for identifying outlier events in complex condensed matter, molecular, and nanoscale systems.