arxiv: 2607.01591 · v1 · pith:3NEGP5RXnew · submitted 2026-07-02 · ⚛️ physics.comp-ph

Verification and Performance Assessment of NuDEAL, a GPU-Accelerated Deterministic Transport Framework on Unstructured Meshes

Kyung Min Kim , Jaeuk Im , Han Gyu Lee , Yeon Sang Jung This is my paper

Pith reviewed 2026-07-03 02:38 UTC · model grok-4.3

classification ⚛️ physics.comp-ph

keywords deterministic neutron transportGPU accelerationunstructured meshesmethod of characteristicsdiscontinuous Galerkinfinite element methodreactor core simulationNuDEAL framework

0

0 comments

The pith

GPU-accelerated deterministic solvers on unstructured meshes deliver accuracy and scalability for whole-core reactor simulations.

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

The paper presents the NuDEAL framework containing three GPU-accelerated solvers for the multigroup transport equation on unstructured meshes: MOC/HFEM, DGMOC, and DFEM-SN. These methods are implemented with GPU optimizations such as memory alignment and sequential sweeps, then validated on the C5G7 benchmark before application to the ABTR, Empire microreactor, and MSRE problems. DFEM-SN reaches eigenvalue errors below 50 pcm while the other two methods match the speed of large CPU clusters on a single GPU. The work shows that deterministic transport on unstructured meshes can now support practical whole-core calculations for heterogeneous advanced reactors. The unified framework also sets up extensions to transient and multiphysics modeling on large GPU systems.

Core claim

NuDEAL unifies three complementary deterministic solvers—MOC/HFEM, DGMOC, and DFEM-SN—on unstructured meshes, all accelerated on GPUs through memory alignment, compressed-flux storage, and sequential azimuthal sweeps. Validation on C5G7 and application to ABTR, Empire, and MSRE show DFEM-SN achieving eigenvalue errors below 50 pcm while MOC/HFEM and DGMOC deliver single-GPU runtimes comparable to large CPU clusters, establishing that deterministic GPU solvers can enable practical whole-core simulations for heterogeneous advanced reactors.

What carries the argument

The NuDEAL unified framework that implements and selectively deploys MOC/HFEM, DGMOC, and DFEM-SN solvers, each optimized for GPU execution via memory alignment, compressed-flux storage, and sequential azimuthal sweeps to solve the multigroup transport equation consistently on unstructured meshes.

If this is right

DFEM-SN supplies the highest accuracy among the three solvers with eigenvalue errors below 50 pcm.
MOC/HFEM and DGMOC achieve the best computational efficiency, with single-GPU times matching large CPU clusters.
The framework permits choosing among the three solvers to trade accuracy against memory and runtime cost for a given geometry.
Whole-core simulations become practical for heterogeneous advanced reactors that require unstructured meshes.
The same code base provides a direct path to transient and multiphysics extensions on large-scale GPU hardware.

Where Pith is reading between the lines

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

The same GPU optimizations could be applied to deterministic transport in other fields such as radiative transfer or charged-particle problems.
Coupling the framework to existing multiphysics codes would allow testing whether the reported speedups persist under coupled iteration.
Users facing memory-limited problems could systematically compare the three solvers on the same mesh to quantify the accuracy-efficiency frontier.
Extension to multi-GPU or distributed GPU clusters would test whether the observed single-GPU scaling continues linearly with problem size.

Load-bearing premise

The benchmark results are taken as sufficient proof that the solvers contain no significant implementation or numerical errors and that the tested problems adequately represent real advanced reactor applications.

What would settle it

A run of the C5G7 or MSRE problem on the same hardware showing eigenvalue error above 50 pcm for DFEM-SN or single-GPU runtimes substantially exceeding the reported CPU-cluster equivalents for MOC/HFEM or DGMOC.

Figures

Figures reproduced from arXiv: 2607.01591 by Han Gyu Lee, Jaeuk Im, Kyung Min Kim, Yeon Sang Jung.

Figure 1. Figure 1: 32-bit (top) and 128-bit aligned (bottom) memory layout. [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗

Figure 2. Figure 2: Data serialization example for RB sweep in HFEM solution procedure. [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗

Figure 3. Figure 3: GPU thread access pattern to the response matrix and operand vectors during RB sweep. [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗

Figure 4. Figure 4: DoF-wide GPU parallelism for DFEM-SN sweep. [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗

Figure 5. Figure 5: Benchmark problem 2D configurations. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗

Figure 6. Figure 6: ABTR 2D core element-wise relative power error distribution. [PITH_FULL_IMAGE:figures/full_fig_p024_6.png] view at source ↗

Figure 7. Figure 7: Empire fuel assembly mesh in the radial (left) and axial (right) views. [PITH_FULL_IMAGE:figures/full_fig_p025_7.png] view at source ↗

Figure 8. Figure 8: Empire 1/12 core configuration in the radial (left) and axial (right) views. [PITH_FULL_IMAGE:figures/full_fig_p027_8.png] view at source ↗

Figure 9. Figure 9: Empire 2D core element-wise relative power error distribution. [PITH_FULL_IMAGE:figures/full_fig_p027_9.png] view at source ↗

Figure 10. Figure 10: Empire 3D 1/12 core element-wise relative power error distribution. [PITH_FULL_IMAGE:figures/full_fig_p028_10.png] view at source ↗

Figure 11. Figure 11: MSRE 2D mesh for fuel channel lattice. (a) Base. (b) Refined 1. (c) Refined 2. (d) Refined 3 [PITH_FULL_IMAGE:figures/full_fig_p029_11.png] view at source ↗

Figure 12. Figure 12: MSRE 2D mesh for outer structure. 29 [PITH_FULL_IMAGE:figures/full_fig_p029_12.png] view at source ↗

Figure 13. Figure 13: MSRE 2D core channel-wise relative power error distribution for the baseline mesh. [PITH_FULL_IMAGE:figures/full_fig_p031_13.png] view at source ↗

Figure 14. Figure 14: MSRE 2D core channel-wise relative power error distribution for the most refined mesh. [PITH_FULL_IMAGE:figures/full_fig_p031_14.png] view at source ↗

Figure 15. Figure 15: MSRE 3D octant core element-wise power error distribution. [PITH_FULL_IMAGE:figures/full_fig_p033_15.png] view at source ↗

read the original abstract

High-fidelity neutronic analyses of advanced reactors require deterministic transport solvers capable of handling complex unstructured geometries while maintaining computational efficiency. This work presents the development and verification of three GPU-accelerated deterministic solvers implemented within a unified framework, Neutronics using Deterministic Finite Element Algorithm (NuDEAL): the planar Method of Characteristics coupled with the Hybrid Finite Element Method (MOC/HFEM), the Discontinuous Galerkin Method of Characteristics (DGMOC), and the Discontinuous Finite Element discrete ordinate method (DFEM-SN). These solvers provide complementary capabilities for consistently solving the multigroup transport equation and can be selectively employed to balance accuracy, computational cost, and memory requirements for a given problem. All methods emphasize efficient GPU execution by leveraging memory alignment, compressed-flux storage, and sequential azimuthal sweeps. The solvers are validated on the C5G7 benchmark and applied to advanced reactor problems, including the ABTR, Empire microreactor, and MSRE. DFEM-SN achieved the highest accuracy, with eigenvalue errors below 50 pcm, while MOC/HFEM and DGMOC provided superior efficiency, with single-GPU runtimes comparable to those of large CPU clusters. The results demonstrate that deterministic GPU solvers on unstructured meshes can deliver both accuracy and scalability, enabling practical whole-core simulations for heterogeneous advanced reactors. The unified NuDEAL framework establishes a foundation for future extensions toward transient and multiphysics analyses on large-scale GPU architectures.

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

NuDEAL gives a working unified GPU framework for three transport methods with solid benchmark numbers, but the whole-core scalability claim extrapolates from smaller test cases.

read the letter

The main thing to know is that this paper delivers a practical GPU implementation of deterministic transport on unstructured meshes. It unifies MOC/HFEM, DGMOC, and DFEM-SN in one framework, adds GPU-specific optimizations such as memory alignment, compressed flux storage, and sequential sweeps, and reports concrete verification on C5G7 plus runs on ABTR, Empire, and MSRE. DFEM-SN hits eigenvalue errors below 50 pcm while the others show single-GPU times that match large CPU clusters. That is useful engineering output for reactor analysis.

The integration and the side-by-side performance data are the real additions. The methods themselves are established, but putting them together with these GPU tweaks and showing they can be swapped for accuracy versus speed is a reasonable contribution. The verification step uses standard benchmarks and reports direct error metrics, which is the right way to do it.

The soft spot is scale. Empire is a microreactor and MSRE an experiment; both are orders of magnitude smaller than commercial whole-core models. The abstract's claim that this enables practical whole-core simulations for heterogeneous advanced reactors therefore rests on extrapolation rather than direct evidence from large meshes. If the full paper contains bigger cases or explicit scaling studies, that would address the concern; otherwise the stress-test point stands.

This work is for nuclear computational physicists and engineers who need GPU transport tools on unstructured geometries. A reader looking for implementation details and benchmark comparisons would get value from the numbers and the framework description.

It deserves peer review. The verification data and performance claims are concrete enough to justify referee time, even if the scalability discussion needs tightening.

Referee Report

1 major / 0 minor

Summary. The paper presents NuDEAL, a unified GPU-accelerated framework implementing three deterministic transport solvers (planar MOC/HFEM, DGMOC, and DFEM-SN) on unstructured meshes for solving the multigroup transport equation. It reports verification on the C5G7 benchmark (eigenvalue errors below 50 pcm for DFEM-SN) and applications to ABTR, Empire microreactor, and MSRE problems, with single-GPU runtimes claimed comparable to large CPU clusters, concluding that the methods enable practical whole-core simulations for heterogeneous advanced reactors.

Significance. If the reported accuracy and performance metrics hold under independent verification, the work demonstrates concrete GPU acceleration for unstructured-mesh deterministic transport, providing a foundation for efficient high-fidelity neutronic analysis of advanced reactors and extensions to multiphysics.

major comments (1)

[Abstract] Abstract and results discussion: the central claim that the solvers 'enable practical whole-core simulations for heterogeneous advanced reactors' is load-bearing but rests on extrapolation; the tested cases (C5G7 plus ABTR, Empire microreactor, MSRE) are orders of magnitude smaller than commercial whole-core models, with no reported scaling studies, larger-domain timings, or heterogeneity metrics to support the assertion.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address the single major comment below and agree that revisions are warranted to moderate the central claim.

read point-by-point responses

Referee: [Abstract] Abstract and results discussion: the central claim that the solvers 'enable practical whole-core simulations for heterogeneous advanced reactors' is load-bearing but rests on extrapolation; the tested cases (C5G7 plus ABTR, Empire microreactor, MSRE) are orders of magnitude smaller than commercial whole-core models, with no reported scaling studies, larger-domain timings, or heterogeneity metrics to support the assertion.

Authors: We agree with the referee that the tested problems are substantially smaller than commercial whole-core models and that the manuscript contains no explicit strong-scaling studies, larger-domain timings, or quantitative heterogeneity metrics to directly support the assertion. The claim was intended to reflect the observed single-GPU performance relative to large CPU clusters on the reported benchmarks together with the unstructured-mesh capability, but we recognize that this constitutes an extrapolation. We will revise the abstract, introduction, and conclusions to replace the phrasing 'enabling practical whole-core simulations' with 'providing a foundation toward practical whole-core simulations' and will add an explicit statement noting the absence of scaling studies to larger domains and the need for future work in that direction. These changes will be made without altering the reported verification and timing results. revision: yes

Circularity Check

0 steps flagged

No circularity: claims rest on direct benchmark runs and runtime measurements

full rationale

The paper implements three transport solvers (MOC/HFEM, DGMOC, DFEM-SN) inside NuDEAL, verifies them on the external C5G7 benchmark, and reports wall-clock times and eigenvalue errors on ABTR, Empire, and MSRE. All performance numbers and accuracy statements are obtained by executing the code on those meshes and comparing against reference solutions or CPU baselines; no parameter is fitted to a subset of the target data and then re-labeled as a prediction, no equation is defined in terms of its own output, and no load-bearing uniqueness theorem is imported from prior self-citations. The central assertion that the solvers are accurate and scalable therefore reduces to the measured results on the reported cases rather than to any self-referential construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available, so no specific free parameters, axioms, or invented entities can be extracted; the work rests on standard deterministic transport methods and GPU programming techniques from prior literature.

pith-pipeline@v0.9.1-grok · 5805 in / 1047 out tokens · 29604 ms · 2026-07-03T02:38:48.281799+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

39 extracted references · 39 canonical work pages

[1]

Benchmark of

Allen, Kenneth and Knight, Travis and Bays, Samuel , year = 2011, journal =. Benchmark of. doi:10.1016/j.pnucene.2011.01.007 , abstract =

work page doi:10.1016/j.pnucene.2011.01.007 2011
[2]

Blacker, T. D. and Bohnhoff, W. J. and Edwards, T. L. , year = 1994, journal =

work page 1994
[3]

Botsch, M and Steinberg, S and Bischoff, S and Kobbelt, L , year = 1999, journal =

work page 1999
[4]

Boyd, William and Shaner, Samuel and Li, Lulu and Forget, Benoit and Smith, Kord , year = 2014, journal =. The

work page 2014
[5]

Cho, Jin Young and Kim, Kang Seog and Lee, Chung Chan and Zee, Sung Quun and Joo, Han Gyu , year = 2007, journal =. Axial

work page 2007
[6]

Optimization of Neutron Tracking Algorithms for

Choi, Namjae and Kim, Kyung Min and Joo, Han Gyu , year = 2021, month = nov, journal =. Optimization of Neutron Tracking Algorithms for. doi:10.1016/j.anucene.2021.108508 , urldate =

work page doi:10.1016/j.anucene.2021.108508 2021
[7]

Progress in Nuclear Energy , volume =

Practical Acceleration of Direct Whole-Core Calculation Employing Graphics Processing Units , author =. Progress in Nuclear Energy , volume =. doi:10.1016/j.pnucene.2021.103631 , urldate =

work page doi:10.1016/j.pnucene.2021.103631 2021
[8]

DeHart, Mark D and Ortensi, Javier and Laboure, Vincent M , year = 2020, journal =

work page 2020
[9]

Nuclear Science and Engineering , volume =

Variational Nodal Methods for Neutron Transport , author =. Nuclear Science and Engineering , volume =. doi:10.13182/NSE85-A27436 , file =. https://doi.org/10.13182/NSE85-A27436 , pages =

work page doi:10.13182/nse85-a27436
[10]

doi:10.1016/j.jcp.2011.04.010 , urldate =

Gong, Chunye and Liu, Jie and Chi, Lihua and Huang, Haowei and Fang, Jingyue and Gong, Zhenghu , year = 2011, month = jul, journal =. doi:10.1016/j.jcp.2011.04.010 , urldate =

work page doi:10.1016/j.jcp.2011.04.010 2011
[11]

and Brown, P

Hanebutte, U. and Brown, P. N. , year = 1999, number =

work page 1999
[12]

Multiphysics

Im, Jaeuk and Jeong, Myung Jin and Choi, Namjae and Kim, Kyung Min and Cho, Hyoung Kyu and Joo, Han Gyu , year = 2023, journal =. Multiphysics

work page 2023
[13]

Annals of Nuclear Energy , volume =

Practical Numerical Reactor Employing Direct Whole Core Neutron Transport and Subchannel Thermal/Hydraulic Solvers , author =. Annals of Nuclear Energy , volume =. doi:10.1016/j.anucene.2013.06.031 , urldate =

work page doi:10.1016/j.anucene.2013.06.031 2013
[14]

Kim, Kyung Min and Lee, Han Gyu and Im, Jaeuk and Shim, Hyung Jin , year = 2024, volume =. Joint

work page 2024
[15]

Benchmark

Kim, Taek K , year = 2020, journal =. Benchmark

work page 2020
[16]

International

Kim, Kyung Min and Lee, Han Gyu and Yoon, Jooil and Joo, Han Gyu , year = 2023, month = aug, address =. International

work page 2023
[17]

, year = 2013, journal =

Kochunas, Brendan M. , year = 2013, journal =. A

work page 2013
[18]

Kunen, A and Loffeld, J and Black, A and Chen, R and Nowak, P and Bailey, T and Brown, P and Rennich, S and Maginot, P , year = 2019, langid =. Porting. International

work page 2019
[19]

Lewis, E. E. , year = 2004, journal =. Much Ado about Nothing:

work page 2004
[20]

Performance Analysis of a Two-Step Calculation Procedure Based on

Lim, Changhyun and Kwon, Sung Joon and Yoon, Jooil , year = 2025, journal =. Performance Analysis of a Two-Step Calculation Procedure Based on

work page 2025
[21]

Argonne National Laboratory , number =

Development. Argonne National Laboratory , number =

work page
[22]

and Smith, M

Palmiotti, G. and Smith, M. and Rabiti, C. and Leclere, M. and Kaushik, D. and Siegel, A. and Smith, B. and Lewis, E. E. , year = 2007, abstract =. Joint

work page 2007
[23]

and Gaston, Derek R

Permann, Cody J. and Gaston, Derek R. and Andr. SoftwareX , volume =. doi:10.1016/j.softx.2020.100430 , urldate =

work page doi:10.1016/j.softx.2020.100430 2020
[24]

and Gibson, Marc A

Poston, David I. and Gibson, Marc A. and Godfroy, Thomas and McClure, Patrick R. , year = 2020, journal =

work page 2020
[25]

and Jung, Yeon Sang and Kumar, Shikhar and Wang, Yaqi and Hanophy, Joshua T

Prince, Zachary M. and Jung, Yeon Sang and Kumar, Shikhar and Wang, Yaqi and Hanophy, Joshua T. and Laboure, Vincent M. and Lee, Changho , year = 2022, journal =. Performance

work page 2022
[26]

and Thomas, J

Raviart, P.A. and Thomas, J. M. , year = 1977, journal =. Primal

work page 1977
[27]

Shen, Dan and Fratoni, Massimiliano and Ilas, Germina and Powers, Jeffrey , year = 2006, journal =

work page 2006
[28]

Smith, Michael A and Lewis, E. E. and Na, Byung-Chan , year = 2005, journal =. Benchmark on Deterministic Transport Calculations without Spatial Homogenization (

work page 2005
[29]

Smith, M. A. and Lewis, E. E. and Shemon, E. R. , year = 2014, journal =

work page 2014
[30]

A Discrete-Ordinates Variational Nodal Method for Heterogeneous Neutron

Sun, Qizheng and Liu, Xiaojing and Chai, Xiang and He, Hui and Zhang, Tengfei , year = 2024, journal =. A Discrete-Ordinates Variational Nodal Method for Heterogeneous Neutron. doi:10.1016/j.camwa.2024.06.032 , abstract =

work page doi:10.1016/j.camwa.2024.06.032 2024
[31]

Vladimirov, V. S. , year = 1961, journal =. Mathematical

work page 1961
[32]

Vladimirov, V S , langid =. A

work page
[33]

and Park, Hansol and Calvin, Olin W

Wang, Yaqi and Prince, Zachary M. and Park, Hansol and Calvin, Olin W. and Choi, Namjae and Jung, Yeon Sang and Schunert, Sebastian and Kumar, Shikhar and Hanophy, Joshua T. and Labour. Griffin:. Annals of Nuclear Energy , volume =. doi:10.1016/j.anucene.2024.110917 , urldate =

work page doi:10.1016/j.anucene.2024.110917 2024
[34]

Rattlesnake:

Wang, Yaqi and Schunert, Sebastian and Ortensi, Javier and Laboure, Vincent and DeHart, Mark and Prince, Zachary and Kong, Fande and Harter, Jackson and Balestra, Paolo and Gleicher, Frederick , year = 2021, month = jul, journal =. Rattlesnake:. doi:10.1080/00295450.2020.1843348 , urldate =

work page doi:10.1080/00295450.2020.1843348 2021
[35]

Rattlesnake

Wang, Yaqi and Schunert, Sebastian and Laboure, Vincent , year = 2018, month = apr, number =. Rattlesnake. doi:10.2172/1466687 , urldate =

work page doi:10.2172/1466687 2018
[36]

Accelerating a Three-Dimensional

Zhang, ZhiZhu and Wang, Kan and Li, Qing , year = 2013, month = dec, journal =. Accelerating a Three-Dimensional. doi:10.1016/j.anucene.2013.06.039 , urldate =

work page doi:10.1016/j.anucene.2013.06.039 2013
[37]

doi:10.1016/j.net.2025.103597 , urldate =

Zhang, Ao and Hong, Ser Gi and Jeong, Seungil and Chen, Jingen , year = 2025, month = sep, journal =. doi:10.1016/j.net.2025.103597 , urldate =

work page doi:10.1016/j.net.2025.103597 2025
[38]

Quadratic Axial Expansion Function with Sub-Plane Acceleration Scheme for the High-Fidelity Transport Code

Zhang, Guangchun and Yang, Won Sik , year = 2020, journal =. Quadratic Axial Expansion Function with Sub-Plane Acceleration Scheme for the High-Fidelity Transport Code

work page 2020
[39]

Nuclear Engineering and Technology , volume =

Variational Nodal Methods for Neutron Transport: 40 Years in Review , author =. Nuclear Engineering and Technology , volume =. doi:10.1016/j.net.2022.04.012 , abstract =

work page doi:10.1016/j.net.2022.04.012 2022