Machine learning approach to ionizing particle recognition using hybrid active pixel detectors

AbstraktAbstract

Timepix detectors record frames containing characteristic patterns corresponding to particles of ionizing radiation passing through a layer of semiconductive material. Since their inception by the Medipix collaboration at CERN, attempts have been made at reconstruction of particle trajectories and recognition of particle species based on the observed patterns. This thesis proposes novel approaches to both problems inspired by the recent works and methods used in computer vision applications. The presented algorithms for trajectory reconstruction are robust and combine local optimization with energies available in the Time-over-Threshold operation mode to achieve subpixel precision. Unlike alternate approaches, the proposed methods have been found to successfully detect and separate up to 10 overlapping patterns. A supervised machine learning algorithm is presented for particle species classification based on the energy loss feature model. Cross-validated evaluation of the classifier with heavy ion dataset indicates that the proposed model is viable and can accurately determine particle species. An analysis of recent Timepix data from the MoEDAL Experiment at LHCb, CERN has been conducted to demonstrate usage of the presented methods in research applications.

Timepix detectors record frames containing characteristic patterns corresponding to particles of ionizing radiation passing through a layer of semiconductive material. Since their inception by the Medipix collaboration at CERN, attempts have been made at reconstruction of particle trajectories and recognition of particle species based on the observed patterns. This thesis proposes novel approaches to both problems inspired by the recent works and methods used in computer vision applications. The presented algorithms for trajectory reconstruction are robust and combine local optimization with energies available in the Time-over-Threshold operation mode to achieve subpixel precision. Unlike alternate approaches, the proposed methods have been found to successfully detect and separate up to 10 overlapping patterns. A supervised machine learning algorithm is presented for particle species classification based on the energy loss feature model. Cross-validated evaluation of the classifier with heavy ion dataset indicates that the proposed model is viable and can accurately determine particle species. An analysis of recent Timepix data from the MoEDAL Experiment at LHCb, CERN has been conducted to demonstrate usage of the presented methods in research applications.