A Quantum-Assisted Surrogate for ATLAS Calorimeter Showers

Particle collisions at the Large Hadron Collider enable precision measurements of Standard Model processes and searches for new physics, but the detector simulations needed to interpret them are computationally expensive. This motivates fast generative surrogates that preserve the response and geometry of the target detector. We present a quantum-assisted generator for photon showers in a voxelized ATLAS calorimeter dataset. Our three-stage architecture generates calorimeter layer energies, samples a binary latent representation with a conditional restricted Boltzmann machine (RBM), and maps it to voxel energies with a decoder. We sample the same trained RBM either by Gibbs sampling on a GPU or on a D-Wave quantum annealer, thereby directly testing the transfer of the learned latent sampler to quantum hardware. Across incident energies from 1 to 300~GeV, quantum processing unit (QPU) sampling closely reproduces the principal calorimeter-response and shower-shape distributions obtained with GPU sampling, with remaining discrepancies from Geant4 concentrated primarily in sparse tails. Reliable hardware transfer is enabled by QPU-aware model training and robust sampling and aggregation procedures. These results establish end-to-end QPU-sampled generation, from incident-energy conditioning to complete voxelized output, of ATLAS-specific calorimeter showers.