Title: CaloTrilogy: Advancing One-Step, End-to-End, Physics-Guided Shower Generation for Contemporary Calorimeters

Abstract:

As current and upcoming particle colliders demand increasingly precise calorimeter simulations, the associated computational burden is escalating sharply. This trend has spurred the creation of machine-learning surrogates to replace traditional Monte Carlo engines like Geant4. While flow matching and diffusion-based generative models have emerged as top contenders for high-dimensional fast simulation due to their superior sample quality, they generally suffer from inefficiencies. Specifically, these models often require approximately ${\cal O}(100)$ function evaluations during inference and depend on auxiliary networks to enforce constraints on global observables, which disrupts the efficiency of streamlined end-to-end generation.

To address these challenges, we present a unified framework that optimizes the trade-off between computational speed, shower quality, and adherence to physical principles. Our approach integrates three core components: (i) an average velocity field integrator that allows for sampling in just one or a few steps; (ii) a data-driven generative prior defined directly in shower space, bypassing the need for random noise initialization; and (iii) physics-informed loss terms that embed inductive biases regarding key observables directly into the training process. Importantly, these components function as regularizers during training, ensuring that end-to-end inference remains uncompromised and incurs no extra computational cost.

Evaluated against several public high-granularity calorimeter datasets, the model delivers shower quality comparable to leading flow and diffusion methods, despite requiring only one or a few evaluation steps. The results reveal an inter-layer shower structure that aligns with underlying physical laws, positioning this method as a robust candidate for next-generation fast simulation pipelines.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC