Unsupervised outlier detection in heavy-ion collisions

Abstract We present different methods of unsupervised learning which can be used for outlier detection in high energy nuclear collisions. This method is particular interest heavy ion collisions where a direct comparison experimental data to model simulations often ambiguous and it not easy determine whether an observation due new physics, incomplete understanding the known physics or artefact. The UrQMD generate bulk background events as well variants may result from misidentified centrality detector malfunctions. presented here generalized novel effects. To detect outliers, dimensional reduction algorithms are implemented, speciftically Principle Component Analysis (PCA) Autoencoders (AEN). find that mainly reconstruction error good measure distinguish outliers background. performance compared using ROC curve. It shown number reduced (encoded) dimensions describe single event contributes significantly task. best suited separate requires reconstructing at same time small parameters.