Transformers for Point Cloud Completion in the AT-TPC

In many experiments with the Active Target Time Projection Chamber, there are breaks in reconstructed particle tracks from nuclear reactions. These breaks in the particle tracks can be the result of a variety of reasons: over biased pads, disconnected pads, or a physical hole in the pad plane. Broken tracks reduce the effectiveness of fitting these particle tracks to extract kinematic information. We propose the use of SnowflakeNet, a deep learning architecture designed for point cloud completion, to fill in the breaks in the tracks for more accurate analyses. The transformer-based architecture of SnowflakeNet enables itself to learn from the context from the point cloud as a sequence and use that “knowledge” in completing the tracks. The model is trained on artificially broken simulated ²²Mg + α and simulated ¹⁶O + α reactions and applied to data from two experiments: ¹⁴C + p at Argonne National Laboratory and ²²Mg + α at the Facility for Rare Isotope Beams. Results will be presented from these experiments.