Sparse Tensor Computation for Active-Target Time Projection Chamber Data

The Active-Target Time Projection Chamber (AT-TPC) is used for imaging reactions of rare isotopes. In this project, we use data from the ¹⁶O + α experiment conducted in the AT-TPC at the Facility for Rare Isotope Beams (FRIB). The reactions in the experiment produced events with varying numbers of tracks, or reaction products. Four- and five-track events are of interest for understanding the production of carbon in stars. This project utilizes supervised deep learning to identify and select these types of events. A PointNet model implemented using the Minkowski Engine (ME) library was trained to determine multiplicity in AT-TPC events, with a focus on selecting four- and five-track events. The ME provides significant computational advantages for sparse tensor computation, omitting trivial calculations that typical networks evaluate. The ME PointNet architecture is used to train a classification model for track-counting in the ¹⁶O + α experiment. We achieved a 0.93 average F1-score for selecting events with 4 and 5 reaction products in the AT-TPC and an accuracy of 0.85 across all multiplicities.