15–18 Dec 2026
Duke University
America/New_York timezone
Artificial Intelligence for the Electron Ion Collider (2026)

4D PointNet++ on Timepix3 clusters: measuring what pixel-level information adds to particle identification

Not scheduled
20m
LSRC B101 Love Auditorium (Duke University)

LSRC B101 Love Auditorium

Duke University

Durham, NC 27710
AI/ML for ePIC and Beyond AI/ML for ePIC and Beyond

Speaker

Francesco Cordella

Description

Hybrid pixel detectors such as Timepix3 record position, energy (ToT) and time of arrival (ToA) for every hit, yet particle identification (PID) still compresses each cluster into a few engineered scalars before any classifier sees it. We treat each cluster as an unordered point cloud in (x, y, E, t) and extend PointNet++ to four dimensions, with a learnable diagonal metric and three absolute-scale descriptors (pixel count, total and maximum ToT) injected at the classification head. The 166k-parameter network is trained with five seeds on 14,716 experimental clusters from two chips: Po-210 alphas and Sr-90/Cs-137 electrons on silicon, and a third class on CdTe.
On unseen test data it reaches 90.7 ± 1.5% accuracy, matching a mature 14-feature boosted decision tree (90.3 ± 0.4%) and 9 percentage points above a parameter-matched CNN on rasterised cluster images. The main contribution is the measurement protocol. A 35-training ablation is read against a run-to-run reproducibility floor measured on identical inputs (±0.04 macro F1): removing the three global scalars is the only contrast that clears it (−0.20 macro F1), while on the confound-free alpha/electron pair the spatial point cloud beats its own scalars by +0.033 macro F1 (p = 0.005) and the per-pixel energy and time channels cancel exactly that gain (p = 0.032): without a time-walk correction, in-cluster time restates the amplitude. A charge balance identifies the third class as ambient-gamma background carried by the CdTe chip, turning a limitation into a result. A parsing accident became a falsification test: a dead-versus-alive ToA channel leaves the learned metric weights unchanged (p = 0.91), because with a 4-pixel median cluster and 16 requested neighbours there is no neighbourhood to steer, a caution for any set-based network on few-hit objects. Blind inference on Am-Be data from a third, unseen detector isolates a recoil-proton-like population (38.5 ± 4.2%).
Paired seeds, a measured floor, a confound audit and a metric falsification test transfer unchanged to sparse hit clouds in silicon trackers and PID detectors at the EIC and beyond, and to Timepix4's sub-200 ps timing. Data, code and checkpoints are public (doi 10.5281/zenodo.21643310); the full study is submitted to JINST.

Author

Co-authors

Dr Antonella Tamburrino (INAIL) Dr Danilo Pacella (ENEA / INFN) Dr Fabio Panza Dr Gerardo Claps (ENEA / INFN) Dr Veronica De Leo (ENEA / INFN)

Presentation materials

There are no materials yet.