Speaker
Description
Detailed simulation of particle interactions in calorimeters represents a major computational bottleneck for high-energy physics experiments, particularly in the upcoming High-Luminosity LHC (HL-LHC) era. While Generative Adversarial Networks (e.g., CaloGAN) have demonstrated the potential of ML-based fast simulation, they often suffer from mode collapse and limited precision in modeling complex correlations. Diffusion Models (DMs) have emerged as a robust alternative, offering superior generation fidelity and training stability, though typically at a prohibitive computational cost due to their iterative sampling nature.
In this work, we present a novel diffusion-based framework for calorimeter simulation that outperforms GAN-based baselines while successfully addressing the inference speed challenge. We introduce an optimized inference pipeline—leveraging advanced solvers and acceleration techniques—that achieves a speedup factor of $10\text{--}100\times$ compared to standard diffusion samplers, with negligible loss in physical accuracy. While the computational complexity (FLOPS) remains higher than that of single-pass GANs, our approach strikes a critical balance, delivering high-fidelity generation where traditional methods fall short. We demonstrate that our model surpasses CaloGAN on key metrics, improving the PRD AUC (Energy/Physics) from 0.97 to 0.98, while significantly enhancing the visual and statistical quality of generated shower shapes. This study establishes a viable pathway for integrating high-precision diffusion models into production-level fast simulation pipelines.