Speaker
Description
Hybrid pixel detectors such as Timepix3 provide simultaneous spatial, energetic, and temporal information at the single-pixel level, making them attractive for radiation diagnostics across high-energy physics, fusion, and applied environments. Exploiting this rich, multidimensional data for reliable particle identification (PID), however, remains challenging for conventional approaches based on handcrafted cluster features. We present a deep-learning diagnostic framework that extends the PointNet++ architecture to four dimensions, treating each detector cluster as an unordered point cloud in (x, y, E, t), thereby preserving full pixel-level information without prior feature engineering. The model introduces a learnable weighted metric that quantifies the relative contribution of the spatial, energetic, and temporal coordinates to the discrimination task, while global descriptors (cluster size, total and maximum energy) retain absolute-scale information. A proof-of-concept on experimental data achieves high classification accuracy; the learned metric reveals the dominant role of spatial morphology, with temporal information providing complementary discrimination. Crucially, blind inference on Am-Be field data reveals emergent neutron/gamma separation and identifies proton-recoil candidates without explicit training on these classes, demonstrating the method's robustness and transferability. The approach offers a physically interpretable, scalable, and detector-agnostic tool for next-generation diagnostics, illustrating how AI can enhance measurement capabilities in pixelated radiation detectors.