Speaker
Description
We present denoising diffusion probabilistic models (DDPMs) as high-fidelity, AI-based generative surrogates for producing full-detector, whole-event simulations in heavy-ion experiments [1]. Trained on HIJING minimum-bias data propagated through the sPHENIX detector geometry with Geant4, DDPMs achieve roughly a hundredfold speedup over standard Geant4 simulations and exhibit superior fidelity compared to GANs. This capability enables the rapid generation of large-scale datasets, essential for high-statistics analyses and for embedding rare high-pT signals, such as jets, into complex backgrounds.
In addition, we introduce a generative AI model for jet background subtraction in heavy-ion collisions. While earlier approaches mainly relied on supervised regression techniques, this work represents the first self-supervised application. We trained UVCGAN [2], a Cycle-Consistent Generative Adversarial Network (CycleGAN), using simulated sPHENIX data to transform calorimeter data from heavy-ion collisions into their proton-proton counterparts, and vice versa, without requiring paired samples. This model effectively separates jets from the underlying event background while preserving global jet kinematics and internal jet structure.
[1] Y. Go and D. Torbunov et al, Effectiveness of denoising diffusion probabilistic models for fast and high-fidelity whole-event simulation in high-energy heavy-ion experiments, https://link.aps.org/doi/10.1103/PhysRevC.110.034912, https://arxiv.org/abs/2406.01602
[2] D. Torbunov et al, UVCGAN v2: An Improved Cycle-Consistent GAN for Unpaired Image-to-Image Translation, https://arxiv.org/abs/2303.16280
