Speaker
Description
Foundation models such as the GPT models, BERT, and DALL-E have shown impressive performance in text and image domains. Such models are built through large-scale training on self-supervised tasks. Similarly, foundation models built for physics-native data structures show potential for applications in nuclear physics experiments. This talk presents work toward developing a multi-purpose deep learning model for time projection chamber (TPC) detector systems that can be fine-tuned for various tasks, including event identification, particle or track identification, and regression. Time-projection chambers are widely used across various subfields of nuclear physics experiments to provide three-dimensional “images” of particle reactions or decays. By treating TPC data as 4-dimensional point clouds, we explore multiple self-supervised pre-training tasks, including a point-cloud shuffling task, to build the backbone of our foundation model. To evaluate these backbone models, we share representation learning results, which include cross-detector transferability, where a model pre-trained on data from one detector retains useful “knowledge” when transferred to another detector. Last, we demonstrate the usefulness of these models when tuned to downstream physics tasks which are evaluated using experimental data withheld in the pretraining phase. Our models are developed using data from two TPCs: the Active-Target Time Projection Chamber (AT-TPC) and the Gaseous Detector with Germanium Tagging (GADGET II) at the Facility for Rare Isotope Beams at Michigan State University and demonstrate broader applicability to nuclear physics TPCs. This work also demonstrates the broader impacts of foundation models to nuclear physics facilities beyond TPCs.
This work is supported in part by NSF grants OAC-2311263, OAC-1836650, PHY-2012865 and the Davidson College RISE program.