Parametric Mode

The parametric generator uses torch_delphes, a PyTorch-based fast detector simulation that reproduces Delphes-like smearing and efficiencies. It applies particle propagation, tracking efficiency, momentum smearing, and calorimeter simulation.

Unlike neural mode, the parametric pipeline processes all particles independently as a flat (N_particles, n_features) tensor -- particles are not grouped into padded sequences.

Configuration

Set generator.type to "parametric" in your config file:

generator:
  type: "parametric"
  name: "cms"
  seed: 42

Fields

Available Detector Cards

Example

uv run parnassus run \
  -c src/parnassus/configs/parametric_config.yaml \
  -i input.hepmc \
  -ne 100 \
  -bs 10 \
  -o output.root

The --random_seed CLI flag overrides the config value:

uv run parnassus run \
  -c src/parnassus/configs/parametric_config.yaml \
  -i input.hepmc \
  -ne 100 \
  -bs 10 \
  -o output.root \
  --random_seed 123

Output Collections

See Output Reference for all branch names.

Pile-Up Merging

The parametric pipeline supports Delphes-style pile-up merging. When enabled, pile-up particles from a pre-generated MinBias file are merged with hard-scatter particles before the detector simulation card runs. This means the calorimeter correctly sums energy deposits from both hard-scatter and pile-up sources, matching the physics of a real detector.

How it works

1. For each hard-scatter event, the number of pile-up interactions is sampled from a Poisson distribution with the configured mean.
2. For each pile-up interaction, a MinBias event is randomly selected from the .pileup file.
3. Each pile-up event receives independent vertex smearing (z and t offsets from a truncated Gaussian) and a random azimuthal rotation.
4. Optionally, the hard-scatter primary vertex is also smeared (enabled by default), matching the C++ Delphes PileUpMerger behaviour.
5. The combined particle list is passed through the detector card.

The truth output contains only hard-scatter particles (with smeared vertex positions if enabled). Detector-level outputs (PFlow, Track, Tower) include contributions from both hard-scatter and pile-up.

Configuration

Add a pileup section under generator:

generator:
  type: "parametric"
  name: "cms"
  seed: 42
  pileup:
    file_path: "MinBias.pileup"    # Path to Delphes .pileup binary file
    mean_pileup: 50                # Average PU interactions per bunch crossing
    max_z_spread: 0.25             # Vertex z truncation in meters (default)
    max_t_spread: 800e-12          # Vertex t truncation in seconds (default)
    sigma_z: 0.053                 # Vertex z Gaussian sigma in meters (default)
    sigma_t: 160e-12               # Vertex t Gaussian sigma in seconds (default)
    smear_hs_vertex: true          # Also smear hard-scatter vertex (default)

Only file_path and mean_pileup are required; the remaining fields have defaults that match the standard CMS Delphes card.

See Configuration Reference for the full field reference.

MinBias file format

The .pileup file uses the Delphes binary XDR (big-endian) format. These files can be generated using the Delphes hepmc2pileup or root2pileup conversion tools. Each file contains a set of pre-generated minimum-bias events that are randomly sampled during pile-up merging.

Vertex smearing defaults

The default sigma and truncation values reproduce the standard CMS Delphes card vertex distribution: