Pith. sign in

REVIEW 2 major objections 5 minor 44 references

A diffusion model trained on 60,000 simulated bunch crossings reproduces the particle- and detector-level structure of incoherent pair creation backgrounds at the FCC-ee with percent-level accuracy and about a 10,000x speedup.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 15:12 UTC pith:2EOQPVLK

load-bearing objection A solid, well-engineered fast-surrogate paper for FCC-ee backgrounds, but the detector-level classifier test is too weak to back the 'statistically consistent' claim. the 2 major comments →

arxiv 2607.18512 v1 pith:2EOQPVLK submitted 2026-07-20 hep-ph hep-exphysics.data-an

De^+e^-ffusion: Capturing the Beam-Beam Physics of e^+e^- Collisions with Diffusion Models

Antonio Chahine , Mariarosaria D'Alfonso , Jan Eysermans , Emmett Forrestel , Loukas Gouskos , Lindsey Gray , Katie Kudela , Haoyun Liu
show 5 more authors
Benedikt Maier Dimitrios Ntounis Christoph Paus Umar Sohail Qureshi Caterina Vernieri
This is my paper
classification hep-ph hep-exphysics.data-an
keywords incoherent pair creationbeam-induced backgroundsdiffusion modelfast simulationFCC-eegenerative surrogateGuineaPig++set-valued generative model
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper introduces De+e−ffusion, a denoising diffusion model that generates entire e+e− pair-production events from beam-beam interactions at the FCC-ee. Trained on just 60,000 bunch crossings from the GuineaPig++ generator, the model reproduces the one-dimensional kinematic, angular, and spatial distributions of all three pair-production channels to about a percent, including the process-dependent longitudinal profile. After full detector simulation, a transformer-based classifier cannot clearly distinguish generated from reference events (AUC 0.553 ± 0.016, against a null noise floor of 0.522 ± 0.024). The model generates events about 10,000 times faster than GuineaPig++, making samples of 10^5 bunch crossings feasible for detector design studies. The central modeling choices are process conditioning, where each particle's QED subprocess is supplied as a never-noised input, and inverse-frequency loss reweighting that prevents rare channels from being washed out.

Core claim

The central claim is that a permutation-equivariant denoising diffusion model, conditioned on the QED subprocess (Breit–Wheeler, Bethe–Heitler, or Landau–Lifshitz) of each particle, can act as a fast surrogate for the incoherent pair creation background at FCC-ee. The authors demonstrate percent-level agreement with GuineaPig++ on every one-dimensional particle-level marginal examined, including the bimodal longitudinal production profile, and detector-level agreement within statistical consistency after full detector simulation. They attribute the success to three choices: diffusing in an unsquashed velocity variable to avoid the lightlike boundary, keeping the subprocess label as a time-in

What carries the argument

The central object is a denoising diffusion probabilistic model operating on a set (particle cloud) rather than a fixed vector. The continuous state (log-energy, unsquashed velocity u = atanh(β), and position) is diffused with a standard variance-preserving Gaussian kernel, while the charge label is diffused with a discrete uniform-flip kernel; both are denoised by a shared transformer encoder with two output heads. The QED subprocess label is never noised—it is supplied as conditioning at every reverse step, which preserves the multi-modal longitudinal structure. An inverse-frequency weighting of the noise-prediction loss, plus a soft charge-balance regularizer, keeps rare channels and glob

Load-bearing premise

The model's fidelity is judged only against a modified GuineaPig++ simulation of a single FCC-ee configuration, and the statistical test used to certify detector-level agreement has a noise floor nearly as large as the measured effect—so if that simulator is wrong or the test lacks power, the claim weakens.

What would settle it

Generate ten times more events from both the model and GuineaPig++, rerun the transformer classifier on the detector-level hit collections, and compare the AUC to the null distribution. If the AUC rises decisively above the null (say above 0.58 with small uncertainty), the claim that the surrogate is statistically consistent at the detector level is falsified; if it stays within the null, the current consistency claim is supported.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Design studies that need O(10^5) bunch crossings become feasible: the surrogate's ~10^4 speedup turns a CPU-years-scale campaign into a GPU-hours-scale one.
  • Because process conditioning preserves per-channel z profiles, detector occupancy studies can be done per subprocess, not just for the sum.
  • The classifier two-sample result, AUC 0.553 vs a null floor of 0.522, implies residual detector-level differences are small at current sample sizes, so the model can be used to iterate on vertex-detector geometry.
  • Training cost (~9,000 seconds) is less than a single GuineaPig++ bunch crossing, so retraining on updated lattice parameters remains practical.
  • Optional subsampling of the reverse trajectory (T' = 250 steps) gives a further 4x speedup with quantified tail degradation, enabling quick scans.

Where Pith is reading between the lines

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

  • The paper does not claim extrapolation to unseen machine configurations; a natural test would be to condition on bunch sizes and crossing angle and check interpolation across the FCC-ee design space.
  • The classifier null experiment shows the 'statistically consistent' result sits only about 1.3 standard deviations above the null mean; a stronger test with more events could either confirm the small residual separation or show it was statistical noise.
  • The same set-valued diffusion recipe could be applied to beamstrahlung photon backgrounds, closing the full beam-background loop, or to other rare-process event generation where multi-modal per-channel structure matters.
  • Because the ground truth is a modified GuineaPig++ fork, the surrogate inherits any deficiencies of that fork; validating against an independent particle-in-cell simulation would test whether the learned distributions generalize beyond the training simulator.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper introduces De+e−ffusion, a process-conditioned denoising diffusion probabilistic model trained on 6e4 GuineaPig++ bunch crossings at the FCC-ee Z pole. The model generates variable-size, permutation-invariant particle clouds for incoherent pair creation, jointly diffusing seven continuous per-particle features and a binary charge label while conditioning on the QED subprocess (Breit–Wheeler, Bethe–Heitler, Landau–Lifshitz). Multiplicity and event composition are handled non-parametrically by sampling from the training set. Validation is carried out at particle level via one-dimensional marginal comparisons and per-process longitudinal profiles, at the fiducial-volume level, and at detector level by propagating events through a Geant4 simulation of the CLD vertex detector and training a transformer two-sample classifier. The reported detector-level classifier AUC is 0.553 ± 0.016 against a same-sample null of 0.522 ± 0.024. The paper claims approximately percent-level particle-level fidelity and detector-level statistical consistency, with a wall-clock speedup of nearly four orders of magnitude relative to GuineaPig++.

Significance. If the validation claims hold, this is a practically valuable fast-simulation surrogate for FCC-ee beam-background studies. The paper is refreshingly explicit that the model is a surrogate for the GuineaPig++ simulator rather than a first-principles calculation; training and validating against the same generator is the correct protocol for this benchmark. Strengths include a detailed and reproducible architecture description, release of the training/sampling code, process conditioning to protect rare channels, inverse-frequency reweighting, and a detector-level Geant4-based validation that goes beyond simple marginal checks. The main weakness is that the headline detector-level 'statistical consistency' claim rests on a classifier test whose statistical power is not demonstrated, and the 'percent-level fidelity' claim is broader than the fiducial-volume results support.

major comments (2)
  1. [§V.D and §VI] The central detector-level claim is not established by the quoted classifier test. The paper reports AUC = 0.553 ± 0.016 for De+e−ffusion versus GuineaPig++, while the same-sample null experiment gives AUC_null = 0.522 ± 0.024. The separation is about 1.1σ, the individual signal runs ([0.540, 0.576]) substantially overlap the null runs ([0.505, 0.546]), and only five seeds and five null partitions are used. With this power, an AUC in the null range cannot be read as evidence that the two detector-level distributions are consistent: absence of a detectable difference is not a demonstrated equivalence. The statement in §VI that the model is 'statistically consistent' with GuineaPig++ at detector level therefore overstates the evidence. Please provide a calibrated permutation p-value, a power analysis (for example, how large a known perturbation must be before the classifier detects it at t
  2. [§V.C vs §VI] Section VI states that the model 'achieves approximately percent-level fidelity on every one-dimensional particle-level marginal we have examined.' This is directly contradicted by §V.C, where the fiducial-volume comparison is explicitly said not to reach percent-level agreement, with high-E, high-|p|, and high-pT tails under-predicted by approximately 10–15%. The fiducial subset is exactly the population most relevant for detector occupancy, so this is not a cosmetic wording issue. The manuscript should either restrict the percent-level statement to the inclusive particle-level marginals of Fig. 3 or provide a quantitative fidelity metric for the fiducial subset that supports whatever claim is made.
minor comments (5)
  1. [Abstract and §V.E] The abstract says the model is 'nearly four orders of magnitude faster than Geant4,' but Fig. 8 and §V.E compare generation time with GuineaPig++ (about 15000 s/event), not with the Geant4 simulation time. The surrogate replaces the generator but still requires Geant4 detector simulation afterwards. Please compare against the correct baseline or rephrase.
  2. [§V.D and Fig. 7] Figure 7 is labeled 'ROCmax{AUC,1−AUC}' while the text reports 'AUC.' If the classifier orientation is corrected by taking the maximum, the reported values are not raw ROC AUCs and should be defined explicitly in the text; otherwise the comparison to AUC = 0.5 and to the null distribution is ambiguous.
  3. [§IV.G, step 3] The sampling description says 'Sample k_{t−1} ∼ Cat(σ(ν_θ)) at masked slots.' This appears to be a typo for unmasked slots, since padded slots should not receive sampled charge labels. Please correct the wording.
  4. [§V.D] The classifier checkpoint is selected by highest validation AUC and the quoted uncertainty is the standard deviation over five runs. This can understate variability due to checkpoint selection and the small number of runs; please clarify whether the same held-out test set is used across seeds and consider reporting the mean and standard error as well.
  5. [Appendix A] The non-parametric handling of multiplicity and per-event composition is disclosed clearly, but the abstract and introduction could state more prominently that event multiplicity and process composition are drawn from cached training events rather than generated by the model, since this affects the meaning of 'generative surrogate.'

Circularity Check

1 steps flagged

Only circularity is the non-parametric multiplicity/composition resampling presented as a validation; the central kinematic and detector-level claims rest on genuine held-out comparisons.

specific steps
  1. self definitional [Section III.C, Section IVG, and Appendix A (Figs. 9–10)]
    "We do not learn the multiplicity distribution p(K) directly. Instead we cache, using the training dataset: The empirical multiplicity histogram {K_n}... and the per-event process composition f_n... At sampling time ... a real event n⋆ is drawn uniformly at random and supplies both K_n⋆ and f_n⋆ to the model. This non-parametric handling of multiplicity and composition is exact by construction up to the resolution of the empirical sample. ... In this appendix we verify that the sampling procedure accurately reproduces both the total and per-process multiplicity distributions."

    The generated event multiplicity and per-process composition are not outputs of the diffusion model: sampling directly resamples the empirical histogram and per-event composition cached from the training set. The Appendix's agreement of total and per-process multiplicities with GuineaPig++ is therefore a comparison of the resampled input with itself, forced by construction up to finite-sample noise. Including multiplicity among the 'one-dimensional particle-level marginals' claimed in Section VI overstates what this check establishes. This is peripheral to the kinematic, angular, spatial, and detector-level claims, which are held-out evaluations of learned continuous features.

full rationale

D e+e−ffusion is an emulator benchmarked against its own GuineaPig++ training simulator, which is the correct protocol for a fast-surrogate study and is not itself circular: the kinematic/angular/spatial comparisons (Figs. 3–6) and the classifier two-sample test are held-out evaluations of learned degrees of freedom, and the results are presented with visible limitations (e.g., AUC 0.553 vs null 0.522, and acknowledged tail mismodeling). The one step that reduces by construction is the multiplicity/composition handling: the empirical multiplicity histogram and per-event process fractions are cached from training and resampled at generation time, so the Appendix A 'verification' is tautological for event size and channel composition. This is a real but minor circularity and does not affect the central phase-space fidelity claims. The same-group GuineaPig++ fork (Ref. [6]) is a validation-source limitation, not a derivation loop, and the statistical overlap of the classifier AUC with its null baseline is a power/interpretation concern, not circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 7 axioms · 0 invented entities

The central claim is conditional on the modified GuineaPig++ fork as ground truth; no real-data check exists since FCC-ee is not built. Free parameters are confined to ML hyperparameters and loss weights, none tuned to a physics target: benchmarks are computed on held-out samples. No new physical entities are postulated; the process labels are existing QED channels.

free parameters (5)
  • Inverse-frequency reweighting exponent α_w = 0.5
    Eqs. (22)-(23) set α_w=1/2 to balance rare-channel gradient weight against variance; hand-chosen and directly controls BW/BH fidelity.
  • Auxiliary loss weights λ_pdg, λ_bal = 0.35, 1.1e-2
    Eq. (25); hand-tuned so that the denoising loss dominates the gradient; affects charge-prediction balance.
  • Velocity unsquashing offset ε = 1e-6
    Eq. (7); arbitrary small offset from the lightlike boundary, chosen to keep atanh well-defined.
  • Cosine schedule offset s = 2.27e-3
    Eq. (11); standard cosine schedule parameter from Ref [20], not physics-derived.
  • Empirical standardization (µ, σ) = Training-set means/std (not quoted)
    Per-feature affine map from Sec. III.B; data-derived preprocessing, bijective, does not alter the modeled distribution.
axioms (7)
  • domain assumption GuineaPig++ (modified fork [6]) is a faithful model of IPC physics at FCC-ee.
    All training data and validation targets come from this code; no real-beam data exist to check it. The surrogate's fidelity is bounded by the simulator's fidelity. Invoked in Sec. III.A and throughout Sec. V.
  • domain assumption The three QED subprocesses BW, BH, LL exhaust incoherent pair creation.
    Sec. II.A lists these as the IPC channels; the model conditions on these labels and cannot generate any other process. A missing channel would be structurally absent from the surrogate.
  • domain assumption Equivalent-photon approximation and beamstrahlung parameter Υ describe the interacting photon fluxes.
    Sec. II.A uses the Yokoya-Chen formalism [2]; standard for beam-beam physics but an approximation, especially the real/virtual photon separation.
  • standard math Denoising score matching equals the (weighted) variational bound for the continuous block; D3PM x0-prediction is a consistent discrete kernel.
    Eqs. (17)-(19) rely on Refs [10,11,23,24] and [21,22]; standard diffusion theory assumed without proof here.
  • standard math The atanh map is bijective from R^3\{0} to the open unit ball.
    Eqs. (7)-(8) require this bijection to keep sampled velocities subluminal; standard calculus with the ε offset.
  • domain assumption Geant4/ddsim/CLD simulation correctly maps generator particles to detector hits.
    Detector-level validation in Sec. V.D applies the same simulation chain to both samples, so shared detector biases cancel in the comparison, but absolute occupancy predictions carry them.
  • domain assumption Fiduciality condition Eq. (37) approximates production at the origin and uniform B field.
    Used to select particles reaching the CLD innermost layer (Sec. V.C); an approximation that shapes the tail-fidelity comparison.

pith-pipeline@v1.3.0-alltime-deepseek · 214 in / 17383 out tokens · 267699 ms · 2026-08-01T15:12:05.697386+00:00 · methodology

0 comments
read the original abstract

Beam-induced backgrounds at high-luminosity $e^+e^-$ colliders, such as the FCC-ee, are dominated by incoherent pair creation (IPC), and require computationally expensive simulations with dedicated Monte Carlo (MC) event generators. Reliable detector and machine-detector interface studies necessitate event samples that are several orders of magnitude larger than what is practically attainable with existing MC. To alleviate this bottleneck, we present D$e^+e^-$ffusion, a denoising diffusion probabilistic model that operates as a permutation-equivariant, set-valued surrogate for fast IPC simulation. Trained on a small GuineaPig++ sample, D$e^+e^-$ffusion faithfully reproduces the marginal and joint kinematic, angular, and positional distributions of all three IPC production processes. In addition, we assess the fidelity at the detector level by propagating both Geant4 and D$e^+e^-$ffusion events through a Geant4 simulation of the CLD vertex detector and by training a transformer-based two-sample classifier; the classifier achieves an area under the ROC curve of $0.553 \pm 0.016$. The trained model generates events nearly four orders of magnitude faster than Geant4, paving the way for a fast-simulation surrogate for FCC-ee design studies.

Figures

Figures reproduced from arXiv: 2607.18512 by Antonio Chahine, Benedikt Maier, Caterina Vernieri, Christoph Paus, Dimitrios Ntounis, Emmett Forrestel, Haoyun Liu, Jan Eysermans, Katie Kudela, Lindsey Gray, Loukas Gouskos, Mariarosaria D'Alfonso, Umar Sohail Qureshi.

Figure 1
Figure 1. Figure 1: FIG. 1: Representative Feynman diagrams for incoherent [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2: Architecture of the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3: Particle-level one-dimensional marginal distributions of [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4: Longitudinal [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5: Particle-level one-dimensional marginal distributions of [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6: Detector-level comparison of hits in the CLD vertex detector following [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8: Per-event generation wall-clock time for [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9: Particle-level per-event multiplicity split by QED subprocess: Breit–Wheeler (left), Bethe–Heitler (center), [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: FIG. 10: Total per-event particle-level multiplicity for [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗

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

44 extracted references · 20 linked inside Pith

  1. [1]

    [20], as: αt = f(t) f(0) , f(t) = cos 2 t/T+s 1 +s π 2 ,(11) withsa small offset (we uses= 2.27×10 −3)

    Continuous Block LetT= 10 3 andlet{ αt}T t=0 bethecumulative-product schedule of the cosineβ-schedule from Ref. [20], as: αt = f(t) f(0) , f(t) = cos 2 t/T+s 1 +s π 2 ,(11) withsa small offset (we uses= 2.27×10 −3). The associ- ated per-step noise schedule isβt = 1− αt/αt−1, clamped to[0,0.999]. The forward kernel acting on the continu- ous block is the s...

  2. [2]

    With γt ∈[0,1]the per-step flip probability, q(kt|kt−1) = (1−γ t)δkt,kt−1 +γ t 1 C ,(14) whereC= 2is the number of categories

    Discrete Block: Uniform-Flip Kernel for Charge The charge labelki ∈ {0,1}is diffused with a discrete uniform-flip kernel of the D3PM family [21, 22]. With γt ∈[0,1]the per-step flip probability, q(kt|kt−1) = (1−γ t)δkt,kt−1 +γ t 1 C ,(14) whereC= 2is the number of categories. The corresponding stationary distribution is uniform on {0, . . . , C−1}, mirror...

  3. [3]

    process conditioning

    Process Labels The subprocessπ i ∈ {0,1,2}is held fixed: q(πt|π0) =δ πt,π0 , t= 0, . . . , T.(16) This is the central modeling choice of the present work and we refer to it as “process conditioning”. At every re- verse step the denoiser is told the ground-truth process of every particle. We motivate this choice further in Sec- tion V, where we observe tha...

  4. [4]

    and full inverse-frequency weighting (αw = 1); the square-root choice keeps the gradient variance of the rarest class manageable while still upweighting it by roughly a factor of 7. The raw weights are normalized so that their training-set mean is unity and then clipped to[w min, wmax] = [0.2,5.0]to bound the worst-case influ- ence of any single particle,...

  5. [5]

    Computeε θ, νθ =M θ(xt, t, kt, π, m)

  6. [6]

    (18), then clip to∥xt−1∥∞ ≤10

    Apply Eq. (18), then clip to∥xt−1∥∞ ≤10

  7. [7]

    The decoded event is obtained by inverting the standard- ization, applying Eq

    Samplek t−1 ∼Cat(σ(ν θ))at masked slots. The decoded event is obtained by inverting the standard- ization, applying Eq. (8) and Eq. (6), and writing: Egen = (11(1−2k i), Ei,β i,r i, πi) i<K .(34) Wherelowerfidelitysamplesareacceptable, wealsoallow the user to optionally subsample the reverse trajectory toT ′ < Tsteps with linearly-spaced time steps for fa...

  8. [8]

    Schulte,Beam-beam simulations with GUINEA-PIG, Tech

    D. Schulte,Beam-beam simulations with GUINEA-PIG, Tech. Rep. CERN-PS-99-014-LP (CERN, 1998)

    1998

  9. [9]

    Benediktet al.,FCC Feasibility Study Report, Tech

    M. Benediktet al.,FCC Feasibility Study Report, Tech. Rep. CERN-FCC-PHYS-2025-001 (CERN, 2025)

    2025

  10. [10]

    Yokoya and P

    K. Yokoya and P. Chen, Beam-beam phenomena in linear colliders, Lect. Notes Phys.400, 415 (1992)

    1992

  11. [11]

    Breit and J

    G. Breit and J. A. Wheeler, Collision of two light quanta, Phys. Rev.46, 1087 (1934)

    1934

  12. [12]

    H.BetheandW.Heitler,Onthestoppingoffastparticles and on the creation of positive electrons, Proc. Roy. Soc. Lond. A146, 83 (1934)

    1934

  13. [13]

    L. D. Landau and E. M. Lifshitz, On the production of electrons and positrons by a collision of two particles, Phys. Z. Sowjetunion6, 244 (1934)

    1934

  14. [14]

    Eysermans, GUINEA-PIG, GitLab repository, fork of CLIC Software / GUINEA-PIG, accessed: 2026-06-24

    J. Eysermans, GUINEA-PIG, GitLab repository, fork of CLIC Software / GUINEA-PIG, accessed: 2026-06-24

    2026

  15. [15]

    Rimbault, P

    C. Rimbault, P. Bambade, O. Dadoun, G. Le Meur, F. Touze, M. C. del Alabau, and D. Schulte, GUINEA PIG++: An upgraded version of the linear collider beam-beaminteractionsimulationcodeGUINEAPIG,in Proceedings of the 2007 Particle Accelerator Conference (PAC07), Conf. Proc. C, Vol. 070625 (2007) p. 2728

    2007

  16. [16]

    Mikuni, B

    V. Mikuni, B. Nachman, and M. Pettee, Fast point cloud generation with diffusion models in high energy physics, Phys. Rev. D108, 036025 (2023), arXiv:2304.01266

    Pith/arXiv arXiv 2023

  17. [17]

    Dreyer, E

    E. Dreyer, E. Gross, D. Kobylianskii, V. Mikuni, B. Nachman, and N. Soybelman, Parnassus: An auto- mated approach to accurate, precise, and fast detector simulation and reconstruction, (2024), arXiv:2406.01620 [physics.data-an]

    Pith/arXiv arXiv 2024

  18. [18]

    Ganguli, Deep unsupervised learning using nonequi- librium thermodynamics, inProceedings of the 32nd In- ternational Conference on Machine Learning (ICML) (2015) arXiv:1503.03585

    J.Sohl-Dickstein, E.A.Weiss, N.Maheswaranathan,and S. Ganguli, Deep unsupervised learning using nonequi- librium thermodynamics, inProceedings of the 32nd In- ternational Conference on Machine Learning (ICML) (2015) arXiv:1503.03585

    Pith/arXiv arXiv 2015

  19. [19]

    J. Ho, A. Jain, and P. Abbeel, Denoising diffusion prob- abilistic models, inAdvances in Neural Information Pro- cessing Systems (NeurIPS)(2020) arXiv:2006.11239

    Pith/arXiv arXiv 2020

  20. [20]

    Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, Score-based generative modeling through stochastic differential equations, inInternational Conference on Learning Representations (ICLR)(2021) arXiv:2011.13456

    Pith/arXiv arXiv 2021

  21. [21]

    Mikuni and B

    V. Mikuni and B. Nachman, Score-based generative mod- els for calorimeter shower simulation, Phys. Rev. D106, 092009 (2022), arXiv:2206.11898

    Pith/arXiv arXiv 2022

  22. [22]

    Buhmann, S

    E. Buhmann, S. Diefenbacher, E. Eren, F. Gaede, G. Kasieczka, A. Korol, A. Krause, K. Krueger, and P. McKeown, CaloClouds: Fast geometry-independent highly-granular calorimeter simulation, JINST18, P11025, arXiv:2305.04847

    Pith/arXiv arXiv

  23. [23]

    Golling, PC-JeDi: Diffusion for particle cloud generation in high energy physics, SciPost Phys.16, 018 (2024), arXiv:2303.05376

    M.Leigh, D.Sengupta, G.Quetant, J.A.Raine, K.Zoch, and T. Golling, PC-JeDi: Diffusion for particle cloud generation in high energy physics, SciPost Phys.16, 018 (2024), arXiv:2303.05376

    Pith/arXiv arXiv 2024

  24. [24]

    D. P. Kingma, T. Salimans, B. Poole, and J. Ho, Variational diffusion models, inAdvances in Neu- ral Information Processing Systems (NeurIPS)(2021) arXiv:2107.00630

    Pith/arXiv arXiv 2021

  25. [25]

    Kasieczkaet al., Diffusion-model-based pile-up sub- traction, (2024), arXiv:2401.13537 [hep-ph]

    G. Kasieczkaet al., Diffusion-model-based pile-up sub- traction, (2024), arXiv:2401.13537 [hep-ph]

    Pith/arXiv arXiv 2024

  26. [26]

    Agostinelliet al., GEANT4: A simulation toolkit, Nucl

    S. Agostinelliet al., GEANT4: A simulation toolkit, Nucl. Instrum. Meth. A506, 250 (2003)

    2003

  27. [27]

    Bacchettaet al., CLD - a detector concept for the FCC-ee, (2019), arXiv:1911.12230 [physics.ins-det]

    N. Bacchettaet al., CLD - a detector concept for the FCC-ee, (2019), arXiv:1911.12230 [physics.ins-det]

    Pith/arXiv arXiv 2019

  28. [28]

    A. Q. Nichol and P. Dhariwal, Improved denoising dif- 18 fusion probabilistic models, inProceedings of the 38th International Conference on Machine Learning (ICML) (2021) arXiv:2102.09672

    Pith/arXiv arXiv 2021

  29. [29]

    Hoogeboom, D

    E. Hoogeboom, D. Nielsen, P. Jaini, P. Forré, and M. Welling, Argmax flows and multinomial diffusion: Learning categorical distributions, inAdvances in Neu- ral Information Processing Systems (NeurIPS)(2021) arXiv:2102.05379

    Pith/arXiv arXiv 2021

  30. [30]

    Austin, D

    J. Austin, D. D. Johnson, J. Ho, D. Tarlow, and R.vanden Berg, Structureddenoisingdiffusion modelsin discrete state-spaces, inAdvances in Neural Information Processing Systems (NeurIPS)(2021) arXiv:2107.03006

    Pith/arXiv arXiv 2021

  31. [31]

    Vincent, A connection between score matching and denoising autoencoders, Neural Computation23, 1661 (2011)

    P. Vincent, A connection between score matching and denoising autoencoders, Neural Computation23, 1661 (2011)

    2011

  32. [32]

    251, 03026 (2021)

    Gaede, Frank, Ganis, Gerardo, Hegner, Benedikt, Helsens, Clement, Madlener, Thomas, Sailer, Andre, Stewart, Graeme A, Volkl, Valentin, and Wang, Joseph, Edm4hep and podio - the event data model of the key4hep project and its implementation, EPJ Web Conf. 251, 03026 (2021)

    2021

  33. [33]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, At- tentionisallyouneed,inAdvances in Neural Information Processing Systems (NIPS)(2017) arXiv:1706.03762

    Pith/arXiv arXiv 2017

  34. [34]

    Xiong, Y

    R. Xiong, Y. Yang, D. He, K. Zheng, S. Zheng, C. Xing, H. Zhang, Y. Lan, L. Wang, and T.-Y. Liu, On layer normalization in the Transformer architecture, inPro- ceedings of the 37th International Conference on Machine Learning (ICML)(2020) arXiv:2002.04745

    Pith/arXiv arXiv 2020

  35. [35]

    Loshchilov and F

    I. Loshchilov and F. Hutter, Decoupled weight decay regularization, inInternational Conference on Learning Representations (ICLR)(2019) arXiv:1711.05101

    Pith/arXiv arXiv 2019

  36. [36]

    L. N. Smith and N. Topin, Super-convergence: Very fast training of neural networks using large learning rates, (2017), arXiv:1708.07120 [cs.LG]

    Pith/arXiv arXiv 2017

  37. [37]

    Petrič, M

    M. Petrič, M. Frank, F. Gaede, S. Lu, N. Nikiforou, and A. Sailer, Detector simulations with dd4hep, Journal of Physics: Conference Series898, 042015 (2017)

    2017

  38. [38]

    Frank, F

    M. Frank, F. Gaede, C. Grefe, and P. Mato, Dd4hep: A detector description toolkit for high energy physics ex- periments, Journal of Physics: Conference Series513, 022010 (2014)

    2014

  39. [39]

    Ganis, C

    G. Ganis, C. Helsens, and V. Völkl (Key4hep), Key4hep, a framework for future HEP experiments and its use in FCC, Eur. Phys. J. Plus137, 149 (2022), arXiv:2111.09874 [hep-ex]

    Pith/arXiv arXiv 2022

  40. [41]

    J.-L. Vay, M. Acciarri, A. Almgren, L. D. Amorim, I. An- driyash, J. R. Angus, T. Antoun, D. Belkin, D. Biz- zozero, A. Blelly, S. E. Clark, E. Dammak, L. Fedeli, A. Formenti, M. Garten, L. Ge, L. Giacomel, K. Gott, R. E. Groenewald, D. Grote, J. Gu, C. Harrison, M. Haseeb, A. Huebl, R. Jambunathan, I. Kara-Mostefa, H. Klion, P. Kumar, R. Lehe, B. Loring,...

    2018

  41. [42]

    Gessner, R

    P.Kicsiny, A.Formenti, S. Gessner, R. Lehe, J.Park,and J.-L.Vay,WarpX:amassivelyparallelPICframeworkfor FCC-ee, Talk at the 9th FCC Physics Workshop, Garch- ing, Germany (2026),https://warpx.readthedocs.io

    2026

  42. [43]

    ONNX Runtime developers, Onnx runtime,https:// onnxruntime.ai/(2021), version: x.y.z

    2021

  43. [44]

    Acostaet al., SubMIT: A Physics Analysis Facil- ity at MIT, Comput

    P. Acostaet al., SubMIT: A Physics Analysis Facil- ity at MIT, Comput. Softw. Big Sci.10, 2 (2026), arXiv:2506.01958 [cs.DC]

    Pith/arXiv arXiv 2026

  44. [64]

    The checkpoint with the highest validation AUC is retained for the final test, and the AUC and ROC curve quoted below are evaluated on the held-out test split. On the held-out test set, the classifier achievesAUC = 0.553±0.016, where the uncertainty reflects the standard deviation across five independent training runs with dif- ferent random seeds control...