Abstract
We present Inspectorch, an open-source Python framework designed to efficiently identify rare and unusual events in large, multidimensional solar observation datasets. It utilizes flow-based models (specifically Normalizing Flows) as flexible, unsupervised density estimators. By learning the multidimensional distribution of solar data, the model assigns a probability score to each sample. The framework flags rare events that receive consistently low probability scores, indicating significant deviation from common features and trends. Because modern solar observatories generate massive volumes of data that cannot be manually inspected, Inspectorch helps researchers focus computational and analytical resources on the most physically relevant and informative phenomena, with successful applications to Hinode/SP, IRIS, MiHI (SST), SDO/AIA, and Solar Orbiter/EUI.
