First-hand information for everyone
This paper announces Parnassus, a new open-source software package that turns particle-level events from event generators into detector-level reconstructed particles. It is written in Python and PyTorch so it can run on graphics processing units (GPUs). The goal is to provide a fast surrogate for slow, full detector simulation chains based on Geant4, while keeping outputs that look like standard reconstruction results.
Parnassus offers two kinds of detector backends. One is a neural backend based on conditional flow matching, a generative neural-network technique trained on full simulation and reconstruction data. The other is a parametric backend called TorchDelphes, a PyTorch re-implementation of selected Delphes fast-simulation cards. The first public release includes neural models for two detectors (CMS and ALEPH) and TorchDelphes cards for CMS, ATLAS and ALEPH. The package also connects natively to the Pythia event generator and to FastJet for jet clustering.
At a high level, Parnassus maps ‘‘truth’’ particles (the particles produced by an event generator) to detector-level particle-flow objects (the reconstructed particles experiments use). The neural models are conditioned on the truth-level particle properties, and they are sampled by integrating a learned vector field; the user can change the number of integration steps to trade speed for quality. The parametric TorchDelphes path simulates tracking efficiency, momentum smearing, calorimeter response and similar effects in PyTorch. Both backends produce the same output format, so downstream analysis steps such as jet clustering work identically.
Why this matters: full Geant4-based simulation is computationally expensive and often dominates computing budgets for large experiments. The authors present Parnassus as a faster surrogate that is ‘‘orders of magnitude’’ cheaper to run than full simulation while preserving the usual reconstruction outputs. Because it is Python-native and GPU-capable, Parnassus can be plugged into modern analysis workflows without needing the C++, ROOT, or Delphes build steps that other tools require.
Important caveats and current limits are clearly described. Neural models in this release have a fixed maximum number of constituents inherited from their training sets (for example, the CMS neural model supports up to 400 truth or particle-flow particles and the ALEPH model up to 128). Public training data that include realistic pileup (many low-energy overlapping collisions) are not available, so the neural backends must model pileup unconditionally; this can produce ‘‘fake’’ particles and is an area of ongoing work. The parametric backend handles pileup by merging minimum-bias interactions in a Delphes-like way. The package is process-agnostic within the detector model’s validated phase space, but users should be cautious when applying the released models far outside the conditions they were trained and validated for.