Inspectorch: Efficient rare event exploration in solar observations

Modern solar observatories resolve the solar atmosphere in unprecedented detail, enabling studies of its activity on spatial scales below 0.1 arcsec and temporal cadences of a few seconds. However, the large volume of data collected by our telescopes cannot be fully analyzed with conventional methods. Popular machine learning methods identify general trends from observations but tend to overlook unusual events due to their low frequency of occurrence. We study the applicability of unsupervised probabilistic methods for efficiently identifying rare events in multidimensional solar observations, thereby optimizing our computational resources to the study of these extreme phenomena. We introduce an open-source framework that utilizes flow-based models --- flexible density estimators capable of learning the multidimensional distribution of high-dimensional datasets. Once optimized, it assigns a probability to each sample, allowing us to identify unusual events. We illustrate the potential of this approach by applying it to observations from the Hinode Spectro-Polarimeter (Hinode/SP), the Interface Region Imaging Spectrograph (IRIS), the Microlensed Hyperspectral Imager (MiHI) at the Swedish 1-m Solar Telescope (SST), the Atmospheric Imaging Assembly (AIA) on board the Solar Dynamics Observatory (SDO), and the Extreme Ultraviolet Imager (EUI) on board Solar Orbiter. On those datasets, we find that the algorithm assigns consistently lower probabilities to spectra that exhibit unusual features. For example, it identifies profiles with very strong Doppler shifts, uncommon broadening, complex profiles where the magnetic field polarity changes with height, and temporal dynamics associated with small-scale reconnection events, among others. demonstrates that density estimation using flow-based models offers a powerful approach to identifying rare events in large solar datasets. We find that the framework is robust across different data regimes and requires limited parameter tuning. The resulting probabilistic anomaly scores allow computational resources to be focused on the most informative and physically relevant events. The Python package is publicly available.