arxiv: 2605.12165 · v1 · submitted 2026-05-12 · ⚛️ physics.ins-det · cs.LG· physics.comp-ph

Recognition: 2 theorem links

· Lean Theorem

Machine Learning for neutron source distributions

Jingjing Li, Jose Ignacio Robledo, Klaus Lieutenant, Norberto Schmidt, Paul Zakalek, Stefan Kesselheim

Pith reviewed 2026-05-13 04:19 UTC · model grok-4.3

classification ⚛️ physics.ins-det cs.LGphysics.comp-ph

keywords neutron source distributiongenerative modelsmachine learningMonte Carlo simulationvariational autoencodernormalizing flowgenerative adversarial networkdenoising diffusion

0 comments

The pith

Probabilistic generative models trained on Monte Carlo neutron lists can sample source distributions efficiently without storing the original data.

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

The paper proposes training probabilistic generative models on a Monte Carlo particle list to capture neutron source distributions. Once trained, the models generate new samples independently of the list, in a rapid and memory-efficient way. The authors test variational autoencoders, normalizing flows, generative adversarial networks, and denoising diffusion models, then compare their performance to traditional estimation methods. They discuss the advantages and disadvantages of each approach. A reader would care because this could simplify large-scale neutron transport calculations that currently depend on bulky precomputed particle lists.

Core claim

The authors establish that source distributions can be modeled through the use of probabilistic generative models trained on a Monte Carlo particle list; after training the model becomes independent of the original list and supports further sampling in an efficient, rapid, and memory-costless manner, while offering an alternative to existing source distribution estimations.

What carries the argument

Probabilistic generative models (variational autoencoder, normalizing flow, generative adversarial network, denoising diffusion model) trained to reproduce the neutron source distribution from a finite Monte Carlo particle list.

If this is right

After training, particle sampling proceeds without reference to the original Monte Carlo list and requires far less memory.
Different generative models can be evaluated and ranked against conventional source estimation techniques.
The learned models open routes to further refinements in neutron source modeling for simulation work.
Trade-offs between model types (speed, fidelity, training cost) become quantifiable for practical use.

Where Pith is reading between the lines

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

The approach could be inserted into existing Monte Carlo workflows to replace repeated loading of large particle files during iterative design studies.
Because sampling is cheap, it may support on-the-fly generation of many source variants for uncertainty propagation in shielding calculations.
The same training pipeline might extend directly to other particle types or to time-dependent sources if the underlying Monte Carlo data are available.

Load-bearing premise

The trained generative models can faithfully reproduce the full statistical properties and physical correlations of the neutron source without bias or loss of rare events when given only a finite Monte Carlo particle list.

What would settle it

Generate a large set of particles from one of the trained models, feed them into a standard neutron transport code, and check whether the resulting flux or detector signals deviate systematically from those obtained with the original Monte Carlo list, especially in the tails of the distributions.

Figures

Figures reproduced from arXiv: 2605.12165 by Jingjing Li, Jose Ignacio Robledo, Klaus Lieutenant, Norberto Schmidt, Paul Zakalek, Stefan Kesselheim.

**Figure 1.** Figure 1: Left: Schematic of the Target-Moderator-Assembly used for the PHITS simulations. Copied from [16] under the terms of the Creative Commons Attribution License 4.0. Right: Schematic of experimental setup adopted from [15] with accordance to terms in Creative Commons Attribution License 4.0. as neutron source representations in Monte Carlo workflows, their comparative performance across ML model architectures… view at source ↗

**Figure 2.** Figure 2: Conceptual schematic of a Normalizing Flow model. Blue blocks represent the invertible transformations of data fi(x), which gradually shape the phase-space volume to that of a multivariate gaussian of the same dimensions as the input dimension. Invertability allows the probability flow to go from the multivariate gaussian space, where it is easy to sample, to the phase-space. In our study, we employ NFs wi… view at source ↗

**Figure 3.** Figure 3: Conceptual schematic of a VAE model. Blue boxes represent layers that are fully connected by learnable parameters (purple). Latent space is generated by learning the mean (orange) and standard deviation (yellow) vectors of dimension dlatent. data generated by the Generator as belonging or not to the original dataset, i.e. it classifies as True if it cannot distinguish it from the original dataset and False… view at source ↗

**Figure 4.** Figure 4: Conceptual schematic of a GAN model. The generator transforms a multivariate Gaussian random vector into an estimated sample, which then is compared to and original sample from the dataset by the discriminator in the attempt to discriminate them. If it cannot, the generator model has learned well the task, and can be used to generate new data that follow the original distribution. that the ordering of vari… view at source ↗

**Figure 5.** Figure 5: Conceptual schematic of a DM model. Starting from random noise, the denoising steps remove noise gradually in T steps, until an estimate of the phase-space variables of a neutron is obtained. generated samples remain within the unit interval [0,1]. This is achieved through a mirroring technique that folds and reflects values outside the unit interval back into it. A total of T = 200 time steps was used to … view at source ↗

**Figure 6.** Figure 6: Histograms for azimuthal angle, weight, time and energy variables for TDR dataset. 1D Histograms in the diagonal show the marginal distributions for a sample from the original data (green) and for the AI model (orange). 2D histograms below the diagonal correspond to the data, and those above the diagonal to the NF AI model [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗

**Figure 7.** Figure 7: Comparison of simulated and measured spectra of the J¨ulich platform experiments. the results of simulations using different simulation source estimates are shown. One of the results from the experiments conducted at the J¨ulich platform was the time-of-flight spectrum measured by a 3He detector, which was then converted to a wavelength spectrum following the procedure described in [15]. A comparison of th… view at source ↗

**Figure 8.** Figure 8: Comparison of the results of the Monte Carlo simulations on VITESS, using the MCPL as input file (orange), the KDSource (pink), and the NF model (blue). Marginal distributions of the estimated phase-space variables are shown at two different points of the simulation: at the beginning just after the source module (t0, white background) and at the end of the simulation just before the detector (t1, green bac… view at source ↗

**Figure 9.** Figure 9: Pairplot of variables in the TDR dataset of the HBS project used to train PGMs. Individual histograms are shown on the diagonal and 2D histograms on the off-diagonals [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗

**Figure 10.** Figure 10: Pairplot of all transformed variables in the benchmark dataset of the J¨ulich platform simulations. Individual histograms are shown on the diagonal and 2D histograms on the off-diagonals. In this case, the weight was constant and not estimated by the PGM. Acknowledgments The authors would like to acknowledge the High Brilliance Source project, in particular Ulrich R¨ucker for the measurements performed in… view at source ↗

read the original abstract

In light of the recent advancements in machine learning, we propose a novel approach to neutron source distribution estimation through the utilisation of probabilistic generative models. The estimation is based on a Monte Carlo particle list, which is only required during the training stage of the machine learning model. Once the source distribution has been learned, the model is independent of the original particle list, allowing for further sampling in an efficient, rapid, and memory-costless manner. The performance of various generative models is evaluated, including a variational autoencoder, a normalizing flow, a generative adversarial network, and a denoising diffusion model. These approaches are then compared to existing source distribution estimations, and the advantages and disadvantages of each approach are discussed. The results demonstrate that source distributions can be modeled through the use of probabilistic generative models, which paves the way for further advancements in this field.

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.

Desk Editor's Note private letter to a colleague

The paper applies generative models to compress MC neutron lists into fast samplers, but validation details are missing so the practical gain stays unproven.

read the letter

The colleague should know that this work trains VAE, normalizing flow, GAN, and diffusion models on Monte Carlo particle lists to learn neutron source distributions, then samples from the trained model without storing the original list. The goal is faster, lower-memory source sampling for repeated transport calculations once training is done. That matches a real workflow pain point in neutronics codes. What is new is the targeted use of these models for source estimation in the physics-ins-det setting; the techniques themselves are standard, but the paper frames their application here as a fresh route and runs a side-by-side comparison to conventional estimators while noting pros and cons for each. The setup is described cleanly: the particle list is needed only for training, after which sampling becomes independent and cheap. That part is straightforward and useful to see laid out. The soft spots sit in the evidence. The abstract states that performance was evaluated and advantages discussed, yet the text supplies no numbers on distribution fidelity, conservation of total strength, or downstream observables. Generative models are known to drop rare events or break correlations in high-dimensional phase space, and neutron sources are sensitive to exactly those tails and joint properties. Without metrics such as marginal distances, energy spectra checks, or transport-level tests, the claim that the models faithfully reproduce the source rests on assertion rather than demonstration. The stress-test concern about mode collapse or poor tail coverage therefore lands and is not answered by what is shown. This paper is for neutronics practitioners who already run Monte Carlo source sampling and are looking for lighter alternatives. A reader working on simulation optimization or code development would pick up the model comparison and the efficiency framing. It deserves peer review because the idea is practical, the approach is direct, and referees can require the quantitative checks that would turn the proposal into a usable result.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes using probabilistic generative models (variational autoencoder, normalizing flow, generative adversarial network, and denoising diffusion model) trained on Monte Carlo particle lists to estimate neutron source distributions. After training, the models allow efficient, rapid, memory-independent sampling of the source without retaining the original particle list. The abstract states that model performance is evaluated, compared against existing estimators, and that advantages/disadvantages are discussed, concluding that source distributions can be modeled this way to advance the field.

Significance. If supported by quantitative validation, the work could enable more efficient handling of complex neutron sources in transport simulations by decoupling sampling from large particle lists. The direct, non-circular application of standard generative modeling to simulation data is a methodological strength that aligns with needs in nuclear instrumentation and detector design.

major comments (2)

[Abstract] Abstract: the claim that 'the performance of various generative models is evaluated' and 'compared to existing source distribution estimations' is unsupported, as the manuscript supplies no quantitative metrics (e.g., distribution distances, conservation of total source strength, or downstream transport observables), validation procedures, error analysis, or baseline comparisons. This directly undermines the central assertion of successful demonstration.
[Results] Results section: the assumption that models trained on a finite MC particle list faithfully recover the full joint distribution (position, energy, direction, time) and all physical correlations without bias or loss of rare-event tails is not tested or reported. Neutron-source applications are sensitive to precisely these tails and correlations, yet no statistics addressing mode coverage or tail fidelity are provided.

minor comments (1)

[Abstract] Abstract and introduction: the specific neutron-physics application domain (e.g., reactor, shielding, or detector context) could be stated more explicitly to clarify the target use case.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed comments, which have helped us identify areas where the manuscript requires strengthening. We address each major comment below and have made substantial revisions to include the requested quantitative validations, statistical analyses, and comparisons.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that 'the performance of various generative models is evaluated' and 'compared to existing source distribution estimations' is unsupported, as the manuscript supplies no quantitative metrics (e.g., distribution distances, conservation of total source strength, or downstream transport observables), validation procedures, error analysis, or baseline comparisons. This directly undermines the central assertion of successful demonstration.

Authors: We agree that the original abstract overstates the quantitative aspects of the evaluation and comparison. The submitted manuscript focused on introducing the generative modeling approach with primarily qualitative demonstrations and discussions of advantages/disadvantages, without the specific metrics, validation procedures, or baseline comparisons noted. In the revised manuscript, we have added quantitative metrics including Kullback-Leibler divergence and Wasserstein distance for distribution similarity, explicit checks for conservation of total source strength, and limited downstream transport observables from sampled sources. We now include error analysis, cross-validation procedures, and direct comparisons to baseline methods such as histograms and kernel density estimation. These additions are detailed in a new subsection of the Results and support the revised abstract claims. revision: yes
Referee: [Results] Results section: the assumption that models trained on a finite MC particle list faithfully recover the full joint distribution (position, energy, direction, time) and all physical correlations without bias or loss of rare-event tails is not tested or reported. Neutron-source applications are sensitive to precisely these tails and correlations, yet no statistics addressing mode coverage or tail fidelity are provided.

Authors: This is a valid and important criticism. The original manuscript did not include explicit tests for faithful recovery of the joint distribution, mode coverage, or tail fidelity, nor did it address potential biases from finite Monte Carlo samples. We have revised the Results section to add these analyses: we now report quantitative measures of mode coverage (via dimensionality reduction visualizations and coverage statistics), tail fidelity (quantile-quantile plots and tail probability comparisons for energy and position), and correlation preservation (using correlation matrices and mutual information scores between variables). We also discuss limitations arising from finite training data and the risk of under-representing rare events, providing a more cautious interpretation of the results. revision: yes

Circularity Check

0 steps flagged

No circularity detected in derivation chain

full rationale

The paper applies standard probabilistic generative models (VAE, normalizing flows, GAN, diffusion) to a finite Monte Carlo particle list for learning neutron source distributions. Training occurs once on the list; subsequent sampling is independent of the original data. No equations, derivations, or claims reduce the central result to a fitted parameter, self-definition, or self-citation chain. The method is a direct, non-circular transfer of existing ML techniques to simulation output, with no load-bearing uniqueness theorems or ansatzes imported from prior author work. The reader's assessment of score 0.0 is confirmed by the absence of any quoted reduction of outputs to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that standard generative models are sufficiently expressive to capture the high-dimensional, physics-constrained distribution of neutron sources when trained on Monte Carlo particle lists.

axioms (1)

domain assumption Probabilistic generative models can accurately approximate the distribution of neutron sources from Monte Carlo particle lists
This assumption is required for the post-training sampling to be useful; it is invoked implicitly when the abstract states that the models learn the source distribution.

pith-pipeline@v0.9.0 · 5452 in / 1282 out tokens · 108686 ms · 2026-05-13T04:19:51.461056+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
We propose a novel approach to neutron source distribution estimation through the utilisation of probabilistic generative models... The performance of various generative models is evaluated, including a variational autoencoder, a normalizing flow, a generative adversarial network, and a denoising diffusion model.
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear
The results demonstrate that source distributions can be modeled through the use of probabilistic generative models

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages · 1 internal anchor

[1]

Forster R and Godfrey T 2006 Mcnp-a general monte carlo code for neutron and photon transportMonte-Carlo Methods and Applications in Neutronics, Photonics and Statistical Physics: Proceedings of the Joint Los Alamos National Laboratory-Commissariat ` a l’Energie Atomique Meeting Held at Cadarache Castle, Provence, France April 22–26, 1985(Springer) pp 33–55

work page 2006
[2]

Niita K, Sato T, Iwase H, Nose H, Nakashima H and Sihver L 2006Radiation measurements 411080–1090

work page
[3]

Romano P K, Horelik N E, Herman B R, Nelson A G, Forget B and Smith K 2015Annals of Nuclear Energy8290–97

work page
[4]

Kittelmann T, Klinkby E, Knudsen E B, Willendrup P, Cai X X and Kanaki K 2017 Computer Physics Communications21817–42

work page 2017
[5]

Zsigmond G, Lieutenant K and Mezei F 2002Neutron News1311–14

work page
[6]

Willendrup P K and Lefmann K 2020Journal of Neutron Research221–16 URL https://content.iospress.com/articles/journal-of-neutron-research/jnr190108

work page
[7]

Lin J Y, Smith H L, Granroth G E, Abernathy D L, Lumsden M D, Winn B, Aczel A A, Aivazis M and Fultz B 2016Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment81086–99

work page
[8]

Pan Z Y, Yang N, Tang M, Shen P and Cai X X 2024Computer Physics Communications295 109004 ISSN 0010-4655 URL https://www.sciencedirect.com/science/article/pii/S0010465523003491

work page
[9]

Willendrup P K and Lefmann K 2020Journal of Neutron Research221–16 16 IOP PublishingJournalvv(yyyy) aaaaaa Authoret al

work page
[10]

Schmidt N S, Abbate O I, Prieto Z M, Robledo J I, Dami´ an J M, Marquez A A and Dawidowski J 2022Annals of Nuclear Energy177109309

work page
[11]

Manring, C A and Hawari, A I 2021EPJ Web Conf.24720004 URL https://doi.org/10.1051/epjconf/202124720004

work page doi:10.1051/epjconf/202124720004
[12]

Forget B and Alhajri A 2022Annals of Nuclear Energy170108974

work page
[13]

Sarrut D, Krah N and L´ etang J M 2019Physics in Medicine & Biology64215004 URL https://dx.doi.org/10.1088/1361-6560/ab3fc1

work page doi:10.1088/1361-6560/ab3fc1
[14]

Br¨ uckel, Thomas, Gutberlet, Thomas, Baggemann, Johannes, Chen, Junyang, Claudio-Weber, Tania, Ding, Qi, El-Barbari, Monia, Li, Jingjing, Lieutenant, Klaus, Mauerhofer, Eric, R¨ ucker, Ulrich, Schmidt, Norberto, Schwab, Alexander, Voigt, J¨ org, Zakalek, Paul, Bessler, Yannick, Hanslik, Romuald, Achten, Richard, L¨ ochte, Fynn, Strothmann, Mathias, Felde...

work page doi:10.1051/epjconf/202328602003
[15]

Schmidt, Norberto Sebasti´ an, Schwab, Alexander, Li, Jingjing, R¨ ucker, Ulrich, Zakalek, Paul, Mauerhofer, Eric, Dawidowski, Javier and Gutberlet, Thomas 2025Eur. Phys. J. Plus140 114 URLhttps://doi.org/10.1140/epjp/s13360-025-06046-0

work page doi:10.1140/epjp/s13360-025-06046-0
[16]

Zakalek P, Achten R, Baggemann J, Beßler Y, Beule F, Br¨ uckel T, Chen J, Ding Q, El-Barbari M, Engels R, Felden O, Gebel R, Grigoryev K, Gutberlet T, Hanslik R, Kamerdzhiev V, K¨ ammerling P, Kleines H, Li J, Lieutenant K, L¨ ochte F, Mauerhofer E, Paulin M A, Pechenizkiy I, R¨ ucker U, Schmidt N, Schwab A, Steffens A, Ott F, Valdau Y, Vezhlev E and Voig...

work page doi:10.1051/epjconf/202328602004
[17]

Br¨ uckel T and Gutberlet T (eds) 2023Technical Design Report HBS Volume 2 – Target Stations and Moderators(Schriften des Forschungszentrums J¨ ulich Reihe Allgemeines / Generalvol 9-2) (J¨ ulich: Forschungszentrum J¨ ulich GmbH Zentralbibliothek, Verlag) ISBN 978-3-95806-710-3 URLhttps://juser.fz-juelich.de/record/1016731

work page arXiv
[18]

29805003 URLhttps://doi.org/10.1051/epjconf/202429805003

Zakalek P, Baggemann J, Li J, R¨ ucker U, Gutberlet T and Br¨ uckel T 2024EPJ Web Conf. 29805003 URLhttps://doi.org/10.1051/epjconf/202429805003

work page doi:10.1051/epjconf/202429805003
[19]

Bond-Taylor S, Leach A, Long Y and Willcocks C G 2022IEEE Transactions on Pattern Analysis and Machine Intelligence447327–7347

work page
[20]

Dinh L, Krueger D and Bengio Y 2014arXiv: LearningURL https://api.semanticscholar.org/CorpusID:13995862

work page
[21]

Kingma D P, Salimans T and Welling M 2016ArXivabs/1606.04934URL https://api.semanticscholar.org/CorpusID:11514441

work page arXiv
[22]

Coccaro A, Letizia M, Reyes-Gonz´ alez H and Torre R 2024Symmetry16ISSN 2073-8994 URLhttps://www.mdpi.com/2073-8994/16/8/942

work page 2073
[23]

Singh A and Ogunfunmi T 2021Entropy2455

work page
[24]

Berahmand K, Daneshfar F, Salehi E S, Li Y and Xu Y 2024Artificial Intelligence Review57 28

work page
[25]

Rivera M 2024 How to train your vae2024 IEEE International Conference on Image Processing (ICIP)pp 3882–3888

work page 2024
[26]

Goodfellow I J, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, Courville A and Bengio Y 2014 Generative adversarial networks (Preprint1406.2661) URL https://arxiv.org/abs/1406.2661

work page internal anchor Pith review Pith/arXiv arXiv 2014
[27]

Karras T, Laine S and Aila T 2019 A style-based generator architecture for generative adversarial networksIEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019(Computer Vision Foundation / IEEE) pp 4401–4410 URL http://openaccess.thecvf.com/content_CVPR_2019/html/Karras_A_Style-Based_ Generator_Archit...

work page 2019
[28]

Sohl-Dickstein J, Weiss E, Maheswaranathan N and Ganguli S 2015 Deep unsupervised learning using nonequilibrium thermodynamicsInternational conference on machine learning (pmlr) pp 2256–2265

work page 2015
[29]

Paszke A, Gross S, Massa F, Lerer A, Bradbury J, Chanan G, Killeen T, Lin Z, Gimelshein N, Antiga Let al.2019Advances in neural information processing systems32

work page
[30]

Krause D 2019Journal of large-scale research facilities JLSRF5A135–A135

work page
[31]

Akiba T, Sano S, Yanase T, Ohta T and Koyama M 2019 Optuna: A next-generation hyperparameter optimization frameworkProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data miningpp 2623–2631

work page 2019
[32]

Gretton A, Borgwardt K M, Rasch M J, Sch¨ olkopf B and Smola A 2012The journal of machine learning research13723–773 18

work page