Point-cloud based machine learning for classifying rare fission events

In the region of the neutron deficient pre-actinides, around ¹⁹⁸Pb and ¹⁸⁰Hg, there is an unexpected island of asymmetric fission. To develop a method to determine fission barriers for rare isotopes and to study the transition between symmetric and asymmetric fission in this region, a cocktail beam of heavy isotopes, including ¹⁹⁰Au, ¹⁹²Hg, ¹⁹⁸Pb, and ²⁰⁰Bi, was produced at the National Superconducting Cyclotron Laboratory at Michigan State University and fused with ⁴He target nuclei in the Active Target Time Projection Chamber (AT-TPC). Of the large volume of data produced in the experiment, less than 3% of the events were ones where fission occurred. A heuristic method was developed to classify events as Fission or Non-Fission based on tuned parameters, but that method fails to classify 5% of the events, placing them into an Unlabeled category. This category includes 15% of all fission events. A PointNet model was trained on a subset of the data labeled by the heuristic method and used to generate labels for events in the Unlabeled category. Using a combination of the heuristic method and the machine learning method, 99% of fission events are correctly identified. The method of training the PointNet model with the labeled data, an analysis of the biases in the models, and the effect of physics results will be presented.