Ultra Fast Calorimeter Simulation with Generative Machine Learning on FPGAs

Computationally expensive Geant4-based calorimeter simulations are a major bottleneck for LHC analyses and future collider studies, motivating the development of fast, generative machine learning (ML)-based surrogates. While ML simulation methods have been traditionally deployed on CPU/GPU, FPGA implementations offer potentially even faster and lower power inference while making full use of available computational resources at collider experiments. We present a hardware-aware variational autoencoder (VAE) model for fast calorimeter simulation, optimized specifically for field-programmable gate array (FPGA) deployment. Quantization-aware training and other compression techniques are applied to a model designed for the 2022 community CaloChallenge to respect the resource constraints of a single FPGA. We synthesize the decoder of the VAE to high-level synthesis code using the hls4ml library. Synthesis timing reports indicate sub millisecond latency resulting in a considerable speed up compared to traditional GPU implementation with minimal performance drop. This feasibility study demonstrates the potential of utilizing existing FPGA architecture at LHC experiments for efficient calorimeter simulations.