Speaker
Description
Hybrid pixel detectors of the Timepix family record, for every pixel above threshold, position, deposited charge (ToT) and arrival time (ToA), and particle identification in mixed fields is usually performed by condensing each cluster into a few engineered features. We ask a narrow, measurable question: on clusters of a few pixels, what does a pixel-level representation contribute beyond a handful of scalar summaries? Each Timepix3 cluster is treated as an unordered set of 4D points (x, y, ToA, ToT) and classified end-to-end by a PointNet++ network extended with a learnable diagonal metric and with three global scalars — number of pixels, total and maximum ToT — that restore the absolute scales removed by per-cluster normalisation. On 14,716 measured clusters (Po-210 alpha particles, Sr-90/Cs-137 electrons and an ambient-background class), the network reaches 90.7 ± 1.5 % accuracy over five seeds, indistinguishable from a gradient-boosted tree on 14 engineered features (90.3 ± 0.4 %) and above a capacity-matched CNN on rasterised clusters (81.3 ± 3.6 %). An ablation at fixed architecture and parameter count, read against the measured run-to-run reproducibility of the pipeline, shows that on the confound-free alpha/electron pair the spatial point cloud carries information beyond the scalars (+0.033 ± 0.013 in macro F1, paired p = 0.005), that the redundant per-pixel energy and time channels cancel this gain (p = 0.032), and that the learned metric weights, although reproducible, do not respond when the information content of an axis is changed from nothing to everything (p = 0.91): on objects with a median of four pixels the neighbourhood has nothing to steer. Applied without retraining to a third detector configuration exposed to an Am-Be field, the same models isolate a non-electron population that is stable across repeated runs and well above the no-source background; its composition is left to a labelled campaign. On few-hit detector objects, set-based networks should be ablated against their own scalar summaries before any claim about pixel-level learning is made.