Point cloud-based regression models for the Active-Target Time Projection Chamber

The Active-Target Time Projection Chamber (AT-TPC) is a detector system that produces high resolution, 3D “images” of low energy nuclear reactions. In traditional analyses of data produced in the AT-TPC, extracting kinematic information from this data is time consuming. We propose the use of a point-cloud based deep learning architecture, PointNet, to accomplish this task quickly with equivalent accuracy to the traditional methods. For this work, we focused on predicting the energy loss of reaction products of a simulated ²²Mg + α experiment at the Facility for Rare Isotope Beams.

The AT-TPC records the trajectories of charged particles as they travel through the gas, losing energy in this process. We train a PointNet model to predict this energy loss for each point in each track. The architecture's shared multi-layer perceptron layers process individual points in a permutation invariant manner, while global feature aggregation captures broader characteristics of the events. The PointNet model is trained as a point-wise regression task, with the inputs being the spatial coordinates and charge amplitude recorded in the detector electronics. Performance metrics such as mean absolute error will be presented on a subset of data that was not used in training.