Track Completion in TPCs using Transformers

The Active-Target Time Project Chamber (AT-TPC) is a charged particle detector used at various facilities around the world, including the Facility for Rare Isotope Beams (FRIB). A gas fills the volume of the detector and serves as both the detection medium and as the target for nuclear collisions. The AT-TPC produces “images” of nuclear reactions, recording the trajectories of particles within the chamber as a series of points. These point clouds often have “broken” regions in tracks due to experimental conditions such as missing sensor pads, or over-biased pad regions.

We train a point-cloud based transformer to reconstruct broken tracks from the ¹⁶O+α and ²²Mg+α experiments conducted at FRIB in 2021 and 2020 respectively. To achieve this, we simulated ¹⁶O+α and ²²Mg+α tracks that were then artificially broken by removing 25% of the original points either from the center of the event, from a randomly chosen region of the event, or from throughout the event. We compared the AdaPoinTr and SnowflakeNet point cloud transformer architectures, with our best results generating 94.7% of points within 1% of their original, known location for simulated ¹⁶O+α data, and 87.3% for simulated ²²Mg+α data.