Speaker
Description
Future Higgs and electroweak studies at the FCC-ee will require Monte Carlo samples far larger than their already unprecedented datasets, while detector designs are still actively being optimized. Thus, full detector simulation and reconstruction is a critical computational bottleneck. We apply Parnassus, a generative machine-learning-based surrogate that merges simulation and reconstruction into a single point-cloud-to-point-cloud mapping, to the CLD detector concept proposed for FCC-ee. Using $e^+e^- \to Z \to q\bar{q}$ events at $\sqrt s = 91.2$ GeV, fully simulated in Geant4 and reconstructed using CLDConfig within Key4hep, we train Parnassus to emulate two distinct particle flow reconstructions: the established Pandora particle flow algorithm and a hit-level machine-learning particle flow algorithm (HitPF).
In both cases, Parnassus reproduces the full simulation at the level of individual reconstructed particles and jets, including particle multiplicities, kinematics, and substructure at the percent-to-sub-percent-level fidelity, significantly outperforming Delphes fast simulation. As a stringent test of the correlations relevant to downstream physics, we also evaluate flavor tagging performance against full simulation. Since Parnassus learns the response of whichever reconstruction chain it is trained on, it offers a reconstruction-agnostic fast surrogate for detector optimization studies and high-statistics analyses at future colliders.
