PointNetArrow: Learning Temporal Dynamics of Reactions in Time Projection Chambers

We investigate the efficacy of building a pre-trained model as a foundation for various machine learning tasks across different experiments at the Active-Target Time Projection Chamber (AT-TPC). In pursuit of this, we developed PointNetArrow, a deep neural network trained in a self-supervised fashion to determine the time evolution of AT-TPC events. PointNetArrow takes in 3D point cloud events, preprocessed into stacks of sequential frames. The model is trained on a surrogate task of predicting the direction of time of these events. Successful prediction of this `arrow of time' indicates the model's ability to decipher both low-level visual cues (such as shapes, volume, and charge) and high-level cues (like spatial relationships and temporal movements). The PointNetArrow model was trained using simulated data from a ²²Mg +α experiment. This pretrained model was then evaluated on a track counting from simulated ¹⁶O +α data. Results from both the surrogate task and the downstream task will be presented.