Choose timezone
Your profile timezone:
Nuclear Physics Seminars at BNL
Foundational Models for Nuclear and Particle Physics
by
→
US/Eastern
Description
The rapid growth and success of large language models, utilizing self-supervised pretraining at scale, has inspired research into the potential use of foundation models (FM) in many scientific domains. However, the application of experimental nuclear and particle physics data to a FM is challenging as the data is sparse, irregular, and context and geometry dependent, unlike the dense structure of language. To address this question, a FM trained on simulated space points in the sPHENIX Time Projection Chamber was developed with the goal of characterizing model performance with model size and the performance of downstream reconstruction tasks. Track finding, particle identification, and noise tagging were studied using architectures with up to 188 million parameters. Similarly to behavior observed in other large language models, the performance of the model scales with the model size; furthermore, the largest model trained consistently surpasses the reconstruction metrics of other AI/ML models in the literature in all downstream tasks studied. This is due to the utilization of a frozen FM backbone augmented by lightweight task-specific adapters. This talk will discuss the model architecture, downstream reconstruction, and implications and possible future paths for reconstruction algorithms utilizing FMs in high energy nuclear and particle physics.
Organised by
Takao Sakaguchi
