Point-cloud machine learning methods for analysis of TPC data

Machine learning techniques that operate directly on point cloud data were investigated for events in the Active-Target Time Projection Chamber at the Facility for Rare Isotope Beams at Michigan State University. PointNet++ was used for event classification and track identification in the ²²Mg +⁴He experiment that ran at NSCL¹ and on simulated data for the upcoming ¹⁰Be + ⁴He experiment at NSCL. Accuracy as high as $98\%$ was achieved for the event classification method. Point-wise convolutions were also examined for both data cleaning and simulating detector response tasks. Results are compared with other machine learning methods such as Convolutional Neural Networks and traditional analysis methods.

¹First Direct Measurement of ²²Mg(α,p)²⁵Al and Implications for X-Ray Burst Model-Observation Comparisons. J.S. Randhawa et al. Phys. Rev. Lett. 125, 202701