pith. sign in

arxiv: 2606.26089 · v1 · pith:UTRAV7UFnew · submitted 2026-06-24 · ✦ hep-ex · hep-ph

Parnassus: A GPU-enabled, Python-based Package for Fast Particle Detector Simulation and Reconstruction

Abdelrahman Elabd , Eilam Gross , Dmitrii Kobylianskii , Benjamin Nachman This is my paper

Pith reviewed 2026-06-25 19:28 UTC · model grok-4.3

classification ✦ hep-ex hep-ph
keywords detector simulationfast simulationmachine learningPyTorchGPU computingCMS detectorDelphesparticle physics
0
0 comments X

The pith

Parnassus supplies a Python PyTorch framework with interchangeable neural and parametric models for fast GPU detector simulation and reconstruction.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents Parnassus as a public software package that supplies GPU-compatible PyTorch backends for simulating and reconstructing particle detector responses. Neural models emulate full Geant4 chains while parametric models replicate selected Delphes-style responses, and both types share one process-agnostic API so the same code works on any generated event sample. Initial releases include flow-matching and parametric versions for the CMS detector plus parametric and neural versions for ATLAS and ALEPH. Native links to Pythia event generation and FastJet clustering are provided, and no ROOT installation is required. The approach matters because it moves detector-response modeling into ordinary Python workflows that run on GPUs.

Core claim

Parnassus is a Python/PyTorch, GPU-compatible framework for fast detector simulation and reconstruction that supplies interchangeable detector models: neural models emulate Geant4-based chains and parametric models supply PyTorch implementations of Delphes-style responses. All models share the same process-agnostic and detector-agnostic API so a user selects a detector card once and applies the identical tool to new physics processes without retraining. The release includes two CMS models, PyTorch versions of the ATLAS and ALEPH Delphes cards, and a flow-matching neural model of ALEPH, together with direct interfaces to Pythia and FastJet.

What carries the argument

The process-agnostic and detector-agnostic API that lets a user select any released detector card (neural or parametric) and apply the same simulation and reconstruction steps to events from any generator.

If this is right

  • The same API applies unchanged to Standard Model and beyond-Standard-Model processes once a detector card is chosen.
  • GPU execution is available for both the flow-matching neural backends and the parametric smearing models.
  • No ROOT dependency is needed, removing a common installation barrier for Python-only analysis chains.
  • Users can swap between a neural CMS model and a parametric CMS model inside one script without changing any other code.
  • Pre-built Delphes-style cards for ATLAS and ALEPH extend the same workflow to e+e- environments.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The shared API could allow analysts to test the impact of switching from a full simulation to a fast emulation on the same downstream analysis code.
  • Community additions of new neural models for other detectors would expand the set of experiments that can use the framework without custom coding.
  • Hybrid runs that route some particles through a neural model and others through a parametric model become possible inside the same event loop.
  • Direct Pythia and FastJet hooks lower the barrier for quick generator-level studies that still include detector effects.

Load-bearing premise

The released code and models actually implement the described neural emulation and parametric responses and can be used through the stated process-agnostic API without further user validation or modification.

What would settle it

Install the package, load one of the released CMS models, generate events with the built-in Pythia interface, run the detector response, and check whether the output object distributions agree with a reference Geant4 sample to within the emulation tolerance stated in the model documentation.

Figures

Figures reproduced from arXiv: 2606.26089 by Abdelrahman Elabd, Benjamin Nachman, Dmitrii Kobylianskii, Eilam Gross.

Figure 1
Figure 1. Figure 1: Parnassus generation pipeline. Truth-level events are routed through either the neural backend (conditional flow matching) or the parametric backend (TorchDelphes), both producing particle-flow ob￾jects. Physics-based event post-processing (e.g. jet clustering) computes derived quantities and writes the final output. 4 [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Example event display for H → ZZ → 4ℓ: truth particles (left) and Parnassus PFlow output (right) in the η–ϕ plane. Marker size scales with pT; dashed circles indicate anti-kt jets (R = 0.5). Particle classes: charged hadrons (blue circles), electrons (red triangles), muons (orange squares), neutral hadrons (purple diamonds), photons (green stars). pT > 10 GeV. Since no lepton cleaning is applied, isolated … view at source ↗
Figure 3
Figure 3. Figure 3: PFlow-level comparison of SM (H→ZZ(∗)→4ℓ, 50,000 events) and BSM (H → aa, a → gg or µ +µ −, 50,000 events) Higgs processes. All PFlow candidates, including leptons, are clustered into anti-kt objects with R = 0.5 and pT > 10 GeV. The distributions illustrate differences in hadron-like activity and visible resonance structure between the two processes. line and Python API for generating detector-level parti… view at source ↗
read the original abstract

We present the public software release of Parnassus, a Python/PyTorch, GPU-compatible framework for fast detector simulation and reconstruction in particle and nuclear physics. Parnassus provides a user-friendly framework with interchangeable detector models: neural models can emulate computationally expensive Geant4-based detector simulation and reconstruction chains, while parametric models provide PyTorch implementations of selected Delphes-style detector responses. This initial release includes two models of the CMS detector: one based on a flow-matching neural network architecture and one based on a PyTorch implementation of the Delphes CMS card (parametric bias and smearing). PyTorch versions of the ATLAS and ALEPH Delphes cards are also available, together with a flow-matching neural model of the ALEPH detector that extends the framework to the e+e- LEP environment. All detector-specific backends share the same process-agnostic and detector-agnostic API: users select a detector card - analogous to choosing a detector card in Delphes - and the same tool can be applied to new physics processes without retraining the released detector model. There are native interfaces to the event generator Pythia and the event clustering package FastJet. Unlike previous C++/ROOT-based tools, Parnassus provides GPU-capable PyTorch detector-response backends and requires no ROOT installation. We describe the installation, command-line and Python API, configuration system, and demonstrate the framework on Standard Model and BSM processes.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript presents Parnassus, a Python/PyTorch, GPU-compatible framework for fast detector simulation and reconstruction. It offers interchangeable detector models: neural flow-matching networks to emulate Geant4-based simulation and reconstruction chains, and parametric PyTorch implementations of selected Delphes-style detector responses. The initial release includes two CMS detector models (one neural, one parametric), PyTorch versions of ATLAS and ALEPH Delphes cards, and a flow-matching neural model of ALEPH. All share a process-agnostic and detector-agnostic API; native interfaces to Pythia and FastJet are provided, with no ROOT dependency required. The paper describes installation, APIs, configuration, and demonstrates use on SM and BSM processes.

Significance. If the implementation is correct and the neural emulations are shown to be accurate and generalizable, Parnassus could provide a useful modern alternative to Geant4 and Delphes by enabling GPU-accelerated, Python-native detector response modeling without ROOT, potentially simplifying workflows for new physics studies.

major comments (2)
  1. [Abstract] Abstract: The central claim that neural models emulate the full Geant4-based detector simulation and reconstruction chains and that the released models (including the CMS flow-matching network) can be applied to arbitrary new physics processes without retraining is unsupported, as the manuscript supplies no performance numbers, validation plots, efficiency curves, or direct comparisons to Geant4 or Delphes.
  2. [Abstract] Abstract and demonstration section: No out-of-distribution tests are reported for the neural models on BSM processes with kinematics outside the training distribution (e.g., high-pT exotics or unusual multiplicities), which is required to substantiate the process-agnostic API claim given that flow-matching networks learn a specific transport map.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful review and constructive comments on the Parnassus manuscript. We address each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that neural models emulate the full Geant4-based detector simulation and reconstruction chains and that the released models (including the CMS flow-matching network) can be applied to arbitrary new physics processes without retraining is unsupported, as the manuscript supplies no performance numbers, validation plots, efficiency curves, or direct comparisons to Geant4 or Delphes.

    Authors: We agree that the manuscript provides no quantitative performance numbers, validation plots, efficiency curves, or direct Geant4/Delphes comparisons. This paper is a software release note focused on the framework, API, and installation rather than model validation. The abstract describes the intended use of the released models. We will revise the abstract to remove unsupported claims about full emulation and add a brief limitations paragraph noting that the neural models are pre-trained examples whose accuracy must be assessed by users for their specific processes. revision: yes

  2. Referee: [Abstract] Abstract and demonstration section: No out-of-distribution tests are reported for the neural models on BSM processes with kinematics outside the training distribution (e.g., high-pT exotics or unusual multiplicities), which is required to substantiate the process-agnostic API claim given that flow-matching networks learn a specific transport map.

    Authors: We agree that no out-of-distribution tests are reported. Flow-matching networks learn transport maps specific to their training distributions, so performance on significantly different BSM kinematics is not guaranteed. We will revise the text to clarify that the process-agnostic and detector-agnostic API refers to the shared interface (users select a card and apply the same code), while explicitly noting the limitation that neural model accuracy depends on kinematic overlap with training data. A short discussion of this point will be added to the demonstration section. revision: yes

Circularity Check

0 steps flagged

No circularity: software release paper with no derivations or fitted predictions

full rationale

The paper is a software package announcement describing Parnassus, a PyTorch-based framework for detector simulation. It presents no mathematical derivations, equations, fitted parameters, or predictions that could reduce to inputs by construction. The core claims concern code implementation, API design, and availability of neural and parametric models, with no self-referential equations or load-bearing self-citations of uniqueness theorems. The reader's assessment of 0.0 circularity is confirmed by inspection of the abstract and full-text description, which contain only implementation details and usage instructions. No steps qualify under any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The paper is a software release description and introduces no free parameters, axioms, or invented physical entities.

pith-pipeline@v0.9.1-grok · 5804 in / 1165 out tokens · 38458 ms · 2026-06-25T19:28:18.510234+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

14 extracted references · 2 canonical work pages

  1. [1]

    de Favereau, et al., Delphes 3: A modular framework for fast simulation of a generic collider experiment, JHEP 02 (2014) 057.arXiv:1307.6346

    J. de Favereau, et al., Delphes 3: A modular framework for fast simulation of a generic collider experiment, JHEP 02 (2014) 057.arXiv:1307.6346

    Pith/arXiv arXiv 2014

  2. [2]

    Dreyer, E

    E. Dreyer, E. Gross, D. Kobylianskii, V. Mikuni, B. Nachman, Conditional deep gen- erative models for simultaneous simulation and reconstruction of entire events (2025). arXiv:2503.19981

    arXiv 2025

  3. [3]

    Y. Lo, D. Kobylianskii, B. Nachman, An ai-based detector simulation and reconstruction model for the aleph experiment at lep (2026).arXiv:2604.11834

    Pith/arXiv arXiv 2026

  4. [4]

    Cacciari, G

    M. Cacciari, G. P. Salam, G. Soyez, Fastjet user manual, European Physical Journal C 72 (2012) 1896.arXiv:1111.6097

    Pith/arXiv arXiv 2012

  5. [5]

    Pivarski, et al., Uproot: Root i/o in pure python and numpy

    J. Pivarski, et al., Uproot: Root i/o in pure python and numpy. URLhttps://github.com/scikit-hep/uproot5

  6. [6]

    Agostinelli, et al., Geant4—a simulation toolkit, Nuclear Instruments and Methods in Physics Research Section A 506 (2003) 250

    S. Agostinelli, et al., Geant4—a simulation toolkit, Nuclear Instruments and Methods in Physics Research Section A 506 (2003) 250. 25

    2003

  7. [7]

    Lipman, R

    Y. Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, Flow matching for generative modeling, in: International Conference on Learning Representations, 2023.arXiv: 2210.02747

    Pith/arXiv arXiv 2023

  8. [8]

    X. Liu, C. Gong, Q. Liu, Flow straight and fast: Learning to generate and transfer data with rectified flow, in: International Conference on Learning Representations, 2023. arXiv:2209.03003

    Pith/arXiv arXiv 2023

  9. [9]

    R. Brun, F. Rademakers, Root—an object oriented data analysis framework, Nuclear Instruments and Methods in Physics Research Section A 389 (1997) 81–86.doi:10. 1016/S0168-9002(97)00048-X

    1997

  10. [10]

    Paszke, et al., Pytorch: An imperative style, high-performance deep learning library, in: Advances in Neural Information Processing Systems, 2019.arXiv:1912.01703

    A. Paszke, et al., Pytorch: An imperative style, high-performance deep learning library, in: Advances in Neural Information Processing Systems, 2019.arXiv:1912.01703

    Pith/arXiv arXiv 2019

  11. [11]

    Dreyer, E

    E. Dreyer, E. Gross, D. Kobylianskii, V. Mikuni, B. Nachman, N. Soybelman, Par- nassus: An automated approach to accurate, precise, and fast detector simulation and reconstruction (2024).arXiv:2406.01620

    arXiv 2024

  12. [12]

    Buckley, P

    A. Buckley, P. Ilten, D. Konstantinov, L. Lönnblad, J. Monk, W. Pokorski, T. Przedzin- ski, A. Verbytskyi, The HepMC3 Event Record Library for Monte Carlo Event Genera- tors, Computer Physics Communications 260 (2021) 107310, arXiv:1912.08005 [hep-ph]. doi:10.1016/j.cpc.2020.107310. URLhttp://arxiv.org/abs/1912.08005

    work page doi:10.1016/j.cpc.2020.107310 2021

  13. [13]

    Bierlich, S

    C. Bierlich, S. Chakraborty, N. Desai, L. Gellersen, I. Helenius, P. Ilten, L. Lönnblad, S. Mrenna, S. Prestel, C. T. Preuss, T. Sjöstrand, P. Skands, M. Utheim, R. Verheyen, A comprehensive guide to the physics and usage of PYTHIA 8.3, SciPost Physics Code- bases (2022) 008doi:10.21468/SciPostPhysCodeb.8. URLhttps://scipost.org/SciPostPhysCodeb.8

    work page doi:10.21468/scipostphyscodeb.8 2022

  14. [14]

    URLhttps://atlas.web.cern.ch/Atlas/GROUPS/PHYSICS/PLOTS/SIMU-2026-09/ 26

    ATLAS Collaboration, Combined fast simulation and reconstruction in ATLAS with PARNASSUS, ATLAS Public Plot PLOT-SIMU-2026-09, accessed: 2026-06-23 (2026). URLhttps://atlas.web.cern.ch/Atlas/GROUPS/PHYSICS/PLOTS/SIMU-2026-09/ 26

    2026