pith. sign in

arxiv: 2606.26268 · v1 · pith:WLNKNW6Onew · submitted 2026-06-24 · 🌌 astro-ph.IM · astro-ph.EP· astro-ph.GA· astro-ph.HE· astro-ph.SR

A nonrelativistic radiative transfer module for Idefix

Nicolas Scepi , Geoffroy Lesur This is my paper

Pith reviewed 2026-06-26 01:12 UTC · model grok-4.3

classification 🌌 astro-ph.IM astro-ph.EPastro-ph.GAastro-ph.HEastro-ph.SR
keywords radiative magnetohydrodynamicsRMHDM1 approximationreduced speed of lightperformance portabilityKokkosexascale computingIDEFIX
0
0 comments X

The pith

IDEFIX gains a radiative transfer module that reaches 7×10^8 cell updates per second per node on AMD accelerators at 1.6 times the cost of MHD.

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

The paper adds a nonrelativistic radiative transfer module to the public MHD code IDEFIX to enable radiation magnetohydrodynamics simulations. It employs the M1 approximation in a split explicit-implicit scheme together with a reduced speed of light to relax strict timestep limits. Built on the Kokkos library, the module runs on a range of accelerated hardware and supports multiple geometries, dimensions, and opacity choices. Tests on AMD MI250X nodes show peak throughput of 700 million cell updates per second per node with only modest overhead compared to pure MHD runs. This makes coupled radiation-fluid calculations practical for astrophysical problems where radiation and matter interact strongly.

Core claim

The radiative transfer module for IDEFIX is built on the M1 approximation and uses a split explicit-implicit time integration with a reduced speed of light approximation. It includes several radiation Riemann solvers, supports Cartesian, cylindrical and spherical coordinates in one to three dimensions, and allows built-in or user-supplied opacities. The implicit step is handled by direct matrix inversion. On AMD MI250X nodes the module attains up to 7×10^8 cell updates per second per node while incurring a computational cost only 1.6 times that of a pure magnetohydrodynamic simulation, demonstrating that the code is fast, robust and portable across current and future accelerated architecture

What carries the argument

M1 approximation for radiative transfer, implemented through a split explicit-implicit scheme and reduced speed of light approximation with direct matrix inversion in the implicit step.

Load-bearing premise

The reduced speed of light approximation leaves the physical results unchanged in the regimes the simulations target.

What would settle it

A side-by-side comparison of the same setup run once with the reduced speed of light and once with the physical speed of light, checking whether radiation energy density, momentum exchange, or flow morphology differ beyond numerical noise.

Figures

Figures reproduced from arXiv: 2606.26268 by Geoffroy Lesur, Nicolas Scepi.

Figure 1
Figure 1. Figure 1: First test of an optically thin shock tube. The blue and green lines show the results of simulations with 28 radial cells using a PLM reconstruction and LimO3 reconstruction, respectively. The black line shows the reference solution using PLM reconstruction with 217 radial cells. Left panels: Full solution. Right panels: Zoomed-in view of the left-facing shock at x ≈ 11.2. this behavior is quite general to… view at source ↗
Figure 3
Figure 3. Figure 3: L1 norm error as a function of the number of cells for the first optically thin Riemann problem (top) and the second optically thin Rie￾mann problem (bottom) computed from a reference solution with 217 radial cells. Blue and green points show simulations with PLM and LimO3 reconstructions, respectively. boundary condition as in Melon Fuksman & Mignone (2019) to avoid the implementation of a special boundar… view at source ↗
Figure 4
Figure 4. Figure 4: Top: Color map of the radiation energy density for the free streaming beam with a resolution of 300 × 300 with a LimO3 f￾preserving reconstruction scheme. Bottom: Vertical cut of the radiation energy density at x = 1 and x = 4.35. More precisely, we set Er, f = R(Er,c), (36) F i r, f = R(F i r,c ) if ||R(F i r,c )|| R(Er,c) ≤ 1, (37) F i r, f = R(F i r,c ) R(Er,c) ||R(F i r,c )|| otherwise. (38) We see tha… view at source ↗
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Color map of the radiation energy density without emission source terms (top) and with emission source terms (middle) for the shadow test run. Bottom: Vertical slice of the radiation energy density at x = 0.5 for the run with emission (yellow line and dots) and without emission (blue line and dots). Our units are unitv = c, unitρ = 1.67 × 10−24 g cm−3 and unitL = 1.496 × 1013 cm. To ensure an optically thi… view at source ↗
Figure 8
Figure 8. Figure 8: Gas temperature (solid lines) and radiation temperature (dashed lines) as a function of x for the subcritical radiation shock (top) and supercritical shock (bottom) for different choices of reduced speed of light ˆc/c = 1, 10−2 , 10−3 , and 10−4 as blue, salmon, yellow, and brown lines, respectively. We see in [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Top left and center left: Radiative energy and radiative flux as a function of x for the vertical diffusion test. Empty circles show the cell-centered values. Crosses show the values of radiative flux from the Riemann solver and radiative energy reconstructed from the Riemann fluxes of energy and flux. The black line shows the analytical solution. Top right and center right: Difference between our simulati… view at source ↗
Figure 10
Figure 10. Figure 10: Top: Color map of the gas temperature as a function of r and z/r for the stellar irradiation test. Middle and bottom: Comparison of the temperature for the Monte Carlo and M1 solution in the midplane as a function of r (middle) and at r = 2 AU as a function of θ (bottom). of 2 AU as a function of θ. We retrieve the typical structure of an irradiated disk, with hot upper layers and a cold disk midplane whe… view at source ↗
Figure 11
Figure 11. Figure 11: Temperature at r = 2 AU as a function of θ for four differ￾ent dust-to-gas ratios of 10−2 , 10−1 , 1, and 10, shown as brown, yellow, salmon, and blue lines, respectively. Dashed lines show the results from the Monte Carlo simulations, and solid lines show the results from the M1 simulations. outward. We compare our solution with a multifrequency multi￾dimensional Monte Carlo simulation done with RADMC3D … view at source ↗
Figure 12
Figure 12. Figure 12: Top: Performance-in-cell update per second per cell on AdAs￾tra’s Mi250x and Mi300 as a function of nodes. Bottom: Weak scaling efficiency. 5. Conclusion In this paper, we have presented the new radiative transfer mod￾ule of Idefix. We used the M1 approximation, effectively treating radiation as a fluid, which is evolved using a high-order finite￾volume Godunov method. We solved for the hyperbolic part of… view at source ↗
read the original abstract

Radiation magnetohydrodynamic (RMHD) simulations are essential for comparisons with observations, particularly in the regime where fluids and radiation are dynamically coupled. Although computationally expensive, RMHD is becoming increasingly accessible with the advent of exascale computing. However, only a few public RMHD codes are currently able to fully exploit the diversity of modern accelerated architectures. We present a nonrelativistic radiative transfer module for the public magnetohydrodynamic code IDEFIX; it is built on the Kokkos library to ensure performance portability. Our goal is to provide a user-friendly RMHD code capable of running efficiently on current and future exascale supercomputers. The radiative transfer module is based on the M1 approximation and implemented using a split explicit-implicit scheme. A reduced speed of light approximation is employed to alleviate the timestep constraint imposed by radiation. The module supports several radiation Riemann solvers and Cartesian, cylindrical, and spherical geometries in one, two, and three dimensions. The implicit step relies on a simple matrix inversion, ensuring both robustness and high performance. Users can choose between built-in opacity models or supply tabulated opacities and custom user-defined functions. The radiative module of IDEFIX demonstrates excellent performance on accelerated architectures, including the AMD MI250X and MI300 partitions of the AdAstra supercomputer. On MI250X nodes, it achieves up to $7.\times 10^8$ cell updates per second per node, at a computational cost only 1.6 times higher than a pure magnetohydrodynamic simulation. These results establish IDEFIX as a fast, robust, and portable RMHD code suitable for the wider community.

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 / 2 minor

Summary. The manuscript presents a nonrelativistic radiative transfer module for the public Idefix MHD code, built on the Kokkos library for performance portability. The module implements the M1 closure with a split explicit-implicit scheme and reduced speed of light approximation to relax the radiation CFL constraint; it supports multiple Riemann solvers, 1D-3D Cartesian/cylindrical/spherical geometries, and user-supplied or built-in opacities. The central results are benchmark timings on AMD MI250X/MI300 partitions of AdAstra, reporting up to 7×10^8 cell updates per second per node at 1.6× the cost of pure MHD.

Significance. If validated, the work would deliver a publicly available, portable RMHD capability optimized for current and future accelerated architectures, addressing a documented gap in exascale-ready public codes. The explicit use of Kokkos and the reported performance numbers (cell-update rate and modest overhead) constitute concrete, reproducible strengths that could be directly useful to the community once accuracy is demonstrated.

major comments (2)
  1. [Abstract, §2] Abstract and §2: the claim that the module is 'suitable for comparisons with observations' and that the reduced speed of light 'does not significantly alter the physical results' is load-bearing for the scientific utility asserted in the introduction, yet no controlled RSL-versus-full-c comparison, truncation-error estimate, or test in the optically thick/dynamically coupled regime is presented anywhere in the manuscript.
  2. [§4] §4 (or equivalent numerical-results section): the performance benchmarks are reported without accompanying accuracy metrics, convergence studies, or comparisons against analytic solutions or other established RMHD codes, so the cell-update rates cannot be translated into a claim of scientific readiness.
minor comments (2)
  1. [§3] The description of the implicit matrix inversion (likely §3) would benefit from an explicit statement of the matrix size and conditioning for the M1 system.
  2. [Figure captions] Figure captions for the performance plots should include the exact problem setup, grid size, and number of MPI ranks/GPUs used.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive report. The comments correctly identify that the current manuscript emphasizes implementation and performance but does not yet provide the validation needed to fully support the broader scientific-utility claims. We will revise the manuscript to address both points.

read point-by-point responses

  1. Referee: [Abstract, §2] Abstract and §2: the claim that the module is 'suitable for comparisons with observations' and that the reduced speed of light 'does not significantly alter the physical results' is load-bearing for the scientific utility asserted in the introduction, yet no controlled RSL-versus-full-c comparison, truncation-error estimate, or test in the optically thick/dynamically coupled regime is presented anywhere in the manuscript.

    Authors: We agree that the manuscript does not contain controlled RSL-versus-full-c comparisons or dedicated truncation-error tests in the optically thick regime. The RSL approximation is standard in the non-relativistic RMHD literature, but the absence of such tests means the claim cannot be substantiated from the present results. We will (i) qualify the abstract and introduction statements, (ii) add a short subsection in the numerical-results section that reports a controlled RSL test (e.g., radiation diffusion or a simple shock-tube problem) together with an estimate of the truncation error relative to the full-c limit, and (iii) cite the relevant literature on RSL validity. These additions will be included in the revised version. revision: yes

  2. Referee: [§4] §4 (or equivalent numerical-results section): the performance benchmarks are reported without accompanying accuracy metrics, convergence studies, or comparisons against analytic solutions or other established RMHD codes, so the cell-update rates cannot be translated into a claim of scientific readiness.

    Authors: The manuscript’s primary focus is the Kokkos-based implementation and the measured cell-update rates on AMD GPUs. We acknowledge that performance numbers alone do not demonstrate scientific readiness without accompanying accuracy metrics. We will expand the results section to include (i) convergence studies for the M1 solver on standard test problems, (ii) direct comparisons against analytic solutions (e.g., radiation shock tubes, diffusion tests), and (iii) where possible, side-by-side results with other public RMHD codes. These additions will allow the reported performance to be interpreted in the context of verified accuracy. revision: yes

Circularity Check

0 steps flagged

No circularity: implementation and benchmark paper with no derivations or fitted predictions

full rationale

The paper reports software implementation of an M1-based radiative module with reduced-speed-of-light approximation, split explicit-implicit scheme, and direct performance benchmarks (e.g., 7e8 cell updates/s on MI250X). No equations derive physical predictions from inputs, no parameters are fitted then relabeled as predictions, and no self-citations justify load-bearing uniqueness theorems or ansatzes. All claims rest on code measurements and stated design choices, making the derivation chain self-contained with no reductions by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

As a software implementation paper, no free parameters, physical axioms, or new invented entities are introduced beyond standard numerical methods.

pith-pipeline@v0.9.1-grok · 5841 in / 1105 out tokens · 29183 ms · 2026-06-26T01:12:32.133096+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

35 extracted references

  1. [1]

    & Sekora, M

    Colella, P. & Sekora, M. D. 2008, Journal of Computational Physics, 227, 7069

    2008

  2. [2]

    & Woodward, P

    Colella, P. & Woodward, P. R. 1984, Journal of Computational Physics, 54, 174 Commerçon, B., Teyssier, R., Audit, E., Hennebelle, P., & Chabrier, G. 2011, A&A, 529, A35

    1984

  3. [3]

    Dexter, J., Scepi, N., & Begelman, M. C. 2021, ApJ, 919, L20

    2021

  4. [4]

    Draine, B. T. 2003, ApJ, 598, 1017

    2003

  5. [5]

    1994, ApJ, 424, 275

    Ensman, L. 1994, ApJ, 424, 275

    1994

  6. [6]

    P., Turner, N

    Flock, M., Nelson, R. P., Turner, N. J., et al. 2017, ApJ, 850, 131

    2017

  7. [7]

    2018, MNRAS, 475, 4186

    Foucart, F. 2018, MNRAS, 475, 4186

    2018

  8. [8]

    2023, Living Reviews in Computational Astrophysics, 9, 1

    Foucart, F. 2023, Living Reviews in Computational Astrophysics, 9, 1

    2023

  9. [9]

    Gnedin, N. Y . & Abel, T. 2001, New A, 6, 437 González, M., Audit, E., & Huynh, P. 2007, A&A, 464, 429

    2001

  10. [10]

    Hayes, J. C. & Norman, M. L. 2003, ApJS, 147, 197

    2003

  11. [11]

    Hirose, S., Blaes, O., & Krolik, J. H. 2009, ApJ, 704, 781

    2009

  12. [12]

    2025, ApJ, 988, 43

    Jiang, Y .-F., Blaes, O., Kaul, I., & Zhang, L. 2025, ApJ, 988, 43

    2025

  13. [13]

    M., & Davis, S

    Jiang, Y .-F., Stone, J. M., & Davis, S. W. 2012, ApJS, 199, 14

    2012

  14. [14]

    M., & Davis, S

    Jiang, Y .-F., Stone, J. M., & Davis, S. W. 2019, ApJ, 880, 67

    2019

  15. [15]

    & Puls, J

    Kudritzki, R.-P. & Puls, J. 2000, Annual Review of Astronomy and Astrophysics, 38, 613

    2000

  16. [16]

    Lesur, G. R. J., Baghdadi, S., Wafflard-Fernandez, G., et al. 2023, A&A, 677, A9

    2023

  17. [17]

    LeVeque, R. J. 2002, Finite volume methods for hyperbolic problems, V ol. 31 (Cambridge university press)

    2002

  18. [18]

    Levermore, C. D. 1984, J. Quant. Spectr. Rad. Transf., 31, 149

    1984

  19. [19]

    Liska, M. T. P., Musoke, G., Tchekhovskoy, A., Porth, O., & Beloborodov, A. M. 2022, ApJ, 935, L1

    2022

  20. [20]

    C., Tchekhovskoy, A., Sadowski, A., & Narayan, R

    McKinney, J. C., Tchekhovskoy, A., Sadowski, A., & Narayan, R. 2014, MN- RAS, 441, 3177 Melon Fuksman, D., Flock, M., Klahr, H., Mattia, G., & Muley, D. 2025, A&A, 701, A97 Melon Fuksman, J. D., Klahr, H., Flock, M., & Mignone, A. 2021, ApJ, 906, 78 Melon Fuksman, J. D. & Mignone, A. 2019, ApJS, 242, 20

    2014

  21. [21]

    B., & Bruenn, S

    Mezzacappa, A., Endeve, E., Messer, O. B., & Bruenn, S. W. 2020, Living Re- views in Computational Astrophysics, 6, 4

    2020

  22. [22]

    & Mihalas, B

    Mihalas, D. & Mihalas, B. W. 2013, Foundations of radiation hydrodynamics (Courier Corporation)

    2013

  23. [23]

    M., Ryan, B

    Miller, J. M., Ryan, B. R., Dolence, J. C., et al. 2019, Phys. Rev. D, 100, 023008

    2019

  24. [24]

    M., Sprouse, T

    Miller, J. M., Sprouse, T. M., Fryer, C. L., et al. 2020, ApJ, 902, 66

    2020

  25. [25]

    Minerbo, G. N. 1978, J. Quant. Spectr. Rad. Transf., 20, 541

    1978

  26. [26]

    Murray, N., Quataert, E., & Thompson, T. A. 2010, ApJ, 709, 191

    2010

  27. [27]

    2004, A&A, 417, 793

    Pascucci, I., Wolf, S., Steinacker, J., et al. 2004, A&A, 417, 793

    2004

  28. [28]

    M., & Drew, J

    Proga, D., Stone, J. M., & Drew, J. E. 1998, MNRAS, 295, 595

    1998

  29. [29]

    2001, A&A, 380, 776

    Richling, S., Meinköhn, E., Kryzhevoi, N., & Kanschat, G. 2001, A&A, 380, 776

    2001

  30. [30]

    2001, Combustion Theory and Modelling, 5, 261

    Ripoll, J.-F., Dubroca, B., & E., D. 2001, Combustion Theory and Modelling, 5, 261

    2001

  31. [31]

    2013, MNRAS, 436, 2188

    Rosdahl, J., Blaizot, J., Aubert, D., Stranex, T., & Teyssier, R. 2013, MNRAS, 436, 2188

    2013

  32. [32]

    & Teyssier, R

    Rosdahl, J. & Teyssier, R. 2015, MNRAS, 449, 4380

    2015

  33. [33]

    C., & Pickrel, D

    Roth, N., Anninos, P., Fragile, P. C., & Pickrel, D. 2025, ApJ, 981, 144 S˛ adowski, A., Narayan, R., Tchekhovskoy, A., & Zhu, Y . 2013, MNRAS, 429, 3533

    2025

  34. [34]

    Skinner, M. A. & Ostriker, E. C. 2013, ApJS, 206, 21

    2013

  35. [35]

    A., Quataert, E., & Murray, N

    Thompson, T. A., Quataert, E., & Murray, N. 2005, ApJ, 630, 167 ˇCada, M. & Torrilhon, M. 2009, Journal of Computational Physics, 228, 4118 Article number, page 12

    2005