Particle collisions at the Large Hadron Collider (LHC) enable tests of the Standard Model and open avenues for discovering new physics. However, the quest for higher fidelity in LHC simulations demands immense computational resources, with projections indicating a need for millions of CPU-years annually during the High Luminosity LHC (HL-LHC) era. Simulating a single LHC event in Geant4 can require on the order of 1000 CPU seconds, particularly dominated by calorimeter subdetector modeling.
In this talk, I will present a novel approach to address these challenges: a conditioned quantum-assisted deep generative model, which we call Calo4pQVAE. Our framework combines a conditioned variational autoencoder (VAE) with a conditioned Restricted Boltzmann Machine (RBM) in the latent space. We engineer the RBM to leverage qubits and couplers on D-Wave's Zephyr-structured Advantage2 quantum annealer, enabling efficient sampling from the latent space.
A key aspect of our method is a new scheme for conditioning the quantum-assisted RBM via flux biases, effectively bridging the flexibility of classical RBMs as universal discrete distribution approximators with the potential speedup and scalability offered by quantum annealing. We further introduce an adaptive mapping procedure to estimate the effective inverse temperature in quantum annealers, ensuring both faster convergence and improved robustness.
To validate Calo4pQVAE, we benchmark our approach using Dataset 2 of the CaloChallenge.
Zoom: https://stanford.zoom.us/j/93308701034?pwd=f51uCvoka42KZpseDu0jrlGGldeU6f.1&from=addon
