Speaker
Description
Artificial intelligence and machine learning techniques have gained increasing attention in recent years as powerful tools for advancing data analysis and simulations across various fields of physics. Among these, generative models are notable for their ability to create complex data distributions, with Generative Adversarial Networks (GANs) already showing promise in reducing the computational costs of scientific simulations.
However, diffusion models, which have proven highly effective for generating high-quality text-to-image translations in industry, remain largely underexplored in high-energy heavy-ion physics.
This work represents the first application of diffusion models in this field, opening new possibilities for simulation techniques. We apply denoising diffusion probabilistic models (DDPMs) for full-detector, whole-event simulations in high-energy heavy-ion experiments [1]. Using Geant4-simulated HIJING events with the sPHENIX detector geometry, we compare the performance of DDPMs with that of GANs. Our results demonstrate that DDPMs not only deliver superior fidelity over GANs but also achieve over 100 times faster simulation speeds compared to Geant4, marking a significant advancement in accelerating event simulations for collider physics.
Additionally, we utilize unpaired image-to-image translation models for jet background subtraction in heavy-ion collisions. UVCGAN [2], a CycleGAN-based model, effectively isolates jets from the combinatorial background, with successful applications demonstrated at both the Relativistic Heavy Ion Collider and the Large Hadron Collider.
[1] Y. Go, D. Torbunov, et al., "Effectiveness of denoising diffusion probabilistic models for fast and high-fidelity whole-event simulation in high-energy heavy-ion experiments," Phys. Rev. C 110, 034912 (2023), arXiv:2406.01602
[2] D. Torbunov, et al., "UVCGAN v2: An Improved Cycle-Consistent GAN for Unpaired Image-to-Image Translation," arXiv:2303.16280
