Accelerating Diffusion Models in Particle Physics

Our research investigates the optimization of sampling techniques applied to particle physics datasets, focusing on numerical methods and comprehensive data analysis. The unique challenges encompassed by particle physics data, such as continuous coordinates, stochastic dimensionality, and permutation invariance, distinguish it from stan- dard datasets. The prevalent deep generative models, mainly designed for image data, fall short for these datasets. Our approach centers on a novel neural network simulation called Fast Point Cloud Diffusion (FPCD). The core aim is to boost the efficiency and speed of diffusion models, assessed via minimized Wasserstein Distances between authentic and generated data distributions and reduced sampling time. We propose a stochastic differential equation solver implementing Euler-Maruyama scheme with predictor-corrector method, resulting in an 8.4x acceleration in performance compared to the baseline ODE solver from the SciPy library. The research uses resources of National Energy Research Scientific Computing Center, , a DOE Office of Science User Facility supported by the Office of Science of the U.S. Department of Energy".