Generative AI Training Tools for Particle Physics Simulation

Detector simulations are a critical step of designing, operating, and analyzing particle physics data. State-ofthe- art simulations are computationally expensive, as they carefully track particles and produced secondary particles through complex instruments. There has been immense progress in automated fast simulation from generative Artificial Intelligence (genAI), which can be used as a surrogate for computational expensive simulations. The goal of this proposal is to adapt genAI strategies for fast simulation to address the challenges of small particle physics experiments and industrial applications of particle transport through dense materials. Gen- Phys will leverage sparse diffusion models to train a foundation model that can be automatically fine-tuned and used as a fast simulation surrogate in software such as Geant4. Phase I Workplan The Phase I effort will prototype the GenPhys training, fine-tuning, and Geant4 integration. The effectiveness and usefulness of the GenPhys tools will be demonstrated using an extension of the CaloChallenge datasets based on the Par04 example in Geant4. Particle Physics simulation is used in various fields; including high energy physics, nuclear physics, medical physics, and space science. Therefore, it has broad applicability in basic and applied research at government agencies (e.g., Department of Energy and NASA), academia, and industry (e.g., radiology, radiotherapy, and private space companies).