Speaker
Description
High precision calorimeter simulation is essential for particle reconstruction, and precision measurements at the LHC and its High-Luminosity LHC upgrade. In precision channels such as H→γγ, accurate photon shower modeling directly impacts the diphoton mass resolution and signal sensitivity. Achieving the required precision demands very large Monte Carlo samples to constraint systematic uncertainties and backgrounds. As detector granularity and event complexity increase, full simulation with Geant4 becomes a major computational bottleneck, limiting the Monte Carlo statistics required for precision Standard Model studies and rare process searches.
We present a fast simulation framework designed to preserve key physics features of calorimeter showers with accurate energy response, realistic longitudinal and transverse development, and inter-layer correlations while dramatically reducing inference cost. The method combines: (i) an average velocity field integrator enabling sampling in one or a few evaluations; (ii) a data-driven generative prior constructed in shower space; and (iii) physics-guided loss terms that impose inductive biases on observables such as energy conservation and shower shape moments. These constraints act only during training, preserving strictly end-to-end generation at inference.
With one or a few evaluations, the model achieves shower quality competitive with state-of-the-art flow and diffusion approaches on public high granularity calorimeter datasets, providing a computationally efficient solution for large scale collider simulation workflows.