Unsupervised Learning to Build Pretrained Models for the AT-TPC

Using PointNet, a trained unsupervised model whose latent space can be used for tasks such as event selection and track selection in the Active Target Time Projection Chamber (AT-TPC) was developed. The AT-TPC is a charged particle tracking detector that is used to study rare isotopes and is located at the Facility for Rare Isotope Beams at Michigan State University.

PointNet is a machine learning architecture that is specially developed for point clouds. The model was made by first voxelating each event, translating each voxel to a different location on the grid, then constructing the model by training to unscramble the events. It will be used to investigate the latent representations for event and track identification using the point and global feature layer. Preliminary results will be presented and discussed.