Neural Mode

The neural generator uses flow-based generative models to simulate detector response. It runs a multi-stage pipeline:

Event model -- generates global event-level features
Particle model -- generates per-particle detector-level quantities
Impact model (optional) -- generates impact parameters (d0, z0)

Input particles are processed as padded sequences of shape (batch_size, max_particles, n_features).

Configuration

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

generator:
  type: "neural"
  name: "cms_2011_flow_v00"
  num_steps: 50
  batch_size: 2000
  device: "cpu"

Fields

Field	Type	Default	Description
`type`	string	--	Must be `"neural"`
`name`	string	--	Model identifier from the registry
`num_steps`	integer	`50`	Number of ODE integration steps for flow sampling
`batch_size`	integer	`2000`	Number of events per batch
`device`	string	`"cpu"`	Computation device: `"cpu"`, `"cuda"`, or `"mps"`

Choosing `num_steps`

Higher values produce better-quality samples at the cost of longer inference. 50 is the default and recommended. With values below ~20 output quality degrades noticeably. Inference time scales roughly linearly with num_steps.

Available Models

Name	Description
`cms_2011_flow_v00`	CMS 2011 era flow-based model. Includes event-level, particle-level, and optional impact parameter models. Uses model-defined max particles, variable transformations, and tunable sampling steps.
`aleph_flow_v00`	ALEPH flow-based model for LEP-era \(e^+e^-\) collisions. Includes event-level and particle-level models. Selects truth particles above a 0.5 GeV pT cut (defined in the model metadata).

Example

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

The -n CLI flag overrides the config value:

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

Output Collections

Collection	Description
`Truth`	Truth-level input particles (PT, Eta, Phi, vertex, ClassID, PID)
`PFlow`	Generated detector-level particles (PT, Eta, Phi, vertex, ClassID, PID); impact parameters D0, Z0, ErrorD0, ErrorZ0 included if the impact model ran
`Electrons` / `Muons`	Lepton kinematics (always present); impact parameters if available; isolation fields added by isolation pipeline
`Event`	Per-event scalars: EventNumber, TruthHT, PFlowHT, TruthMET, PFlowMET
`<JetName>`	One collection per configured clustering pipeline (only if cluster pipeline defined)

See Output Reference for all branch names.