Machine learning models for single-particle classification with Timepix 3 detector

Semiconductor hybrid pixel detectors with Timepix 3 ASIC chips developed by the Medipix collaboration at CERN enable high-resolution, single-particle detection, measuring deposited energy and time of arrival per pixel. This study introduces machine learning models—BiLSTM, fully connected neural networks, and CatBoost decision trees—for classifying individual particle events based on morphological and spectral features. These models were trained on an extensive, labeled experimental dataset and validated on a separate test set. Classification was performed across five particle categories: electrons & photons, alpha particles, heavy nuclei, low-energy protons, and high-energy protons. The models demonstrated high accuracy, with BiLSTM requiring no preprocessing but slower inference, while CatBoost achieved the fastest predictions. Additionally, an unsupervised Gaussian Mixture Model was used to explore the separation between electrons and photons. Feature importance analysis revealed LET, energy distribution, and cluster morphology as key discriminators. Future work will integrate these models into the TraX Engine platform developed by ADVACAM.

Text supplied in this site's publication record. Consult the full paper for methods, experimental conditions and limitations.