SEDONA-GesaRaT: an AI-Accelerated Radiative Transfer Program for 3-D Supernova Simulations

Daniel Kasen; David J. Jeffery; F. K. Roepke; Lifan Wang; Ming Zhong; Ulisses Braga-Neto; Xingzhuo Chen; Zhengwei Liu

REVIEW 4 major objections 6 minor 36 references

SEDONA-GesaRaT: an AI-Accelerated Radiative Transfer Program for 3-D Supernova Simulations

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read By replacing the NLTE atomic solver with neural networks, SEDONA-GesaRaT runs a full 3-D NLTE spectropolarimetric simulation of a type Ia supernova in about 3000 core-hours, a cost previously attached to 1-D NLTE or 3-D LTE runs.

desk verdict CNN surrogate for NLTE atomic physics gives SEDONA a real speedup and a ~3000 core-hour 3D NLTE spectropolarimetry run, but the 3D NLTE result is unvalidated; treat it as a promising demonstration, not a calibrated capability. read the letter →

arxiv 2507.11767 v1 pith:EPYCPDYP submitted 2025-07-15 astro-ph.HE

Xingzhuo Chen , Ulisses Braga-Neto , Lifan Wang , Daniel Kasen , Zhengwei Liu , F. K. Roepke , Ming Zhong , David J. Jeffery This is my paper · ORCID

classification astro-ph.HE

keywords supernovaradiativetransferNLTEneuralnetworksurrogatespectropolarimetryMonteCarlosimulationtypeIamachinelearningacceleration

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Supernova spectropolarimetry probes the 3-D structure of exploding stars, but full non-local thermodynamic equilibrium (NLTE) radiative transfer in three dimensions has been too expensive to run. This paper claims to remove that bottleneck by replacing the slow NLTE atomic level-population solver in the Monte Carlo code SEDONA with a convolutional neural network trained on 119 one-dimensional type Ia supernova simulations. Combined with an integral-based technique for extracting spectra with high signal-to-noise, the resulting code, SEDONA-GesaRaT, reproduces the 3-D NLTE spectropolarimetric time series and resolved polarization maps of the N100 type Ia model for roughly 3000 core-hours. If correct, 3-D NLTE modeling shifts from a single-run luxury to a tool for systematic studies of supernova structure and for comparison with observed polarization.

What carries the argument

The atomic physics neural network (APNN) is a convolutional network with 33 convolution layers and 13,480,395 trainable parameters that takes rescaled plasma density, 28 elemental abundances, the mean intensity array, and integrated radiation energy as inputs, and predicts the maxima, minima, and normalized frequency shapes of the absorption and emission coefficients plus electron-scattering opacity. It is trained once on 119 1-D SEDONA runs spanning four atomic recipes, then replaces the iterative NLTE solver inside the Monte Carlo transport. The second carrying piece is the integral-based technique (IBT), which estimates Stokes flux integrals directly from Monte Carlo packets and boosts spectropolarimetric signal-to-noise by about 30 times for about 30 percent extra compute.

What would settle it

Run the N100 model in 3-D with SEDONA's traditional NLTE solver for Si, S, and Ca at one or two epochs and compare the resulting spectra, fluxes, and Q/U polarization maps directly against SEDONA-GesaRaT at the same viewing direction; substantial disagreement at the Si II 5640 A or Ca II infrared triplet lines would show that the surrogate's 3-D errors exceed what the LTE test suggests. Alternatively, feed 3-D zone data from N100 both into the APNN and into the exact atomic solver and compare the predicted extinction and emission coefficients zone by zone.

Watch

Extended reading notes

Core claim

The paper's central claim is that a neural network emulator of atomic physics, the atomic physics neural network (APNN), can accurately map local density, elemental abundances, and mean radiation intensity to the frequency-resolved extinction and emission coefficients that drive radiative transfer, including the NLTE level populations of Si, S, and Ca. Validated on 1-D SN Ia models, the surrogate reduces the per-zone atomic calculation from tens of core-seconds to 0.17 core-seconds. Applied to the 3-D N100 model, SEDONA-GesaRaT yields spectra and linear polarization maps consistent with SEDONA under the LTE approximation, with noted errors on the Si II 5640 A line and the Ca II infrared triplet, and then produces the first 3-D NLTE spectropolarimetric time series and resolved Q/U polarization data cubes for the model. The roughly 3000 core-hour total cost is presented as opening a regime previously closed to 3-D NLTE simulations.

Load-bearing premise

The neural network, trained only on spherically symmetric 1-D type Ia models, stays accurate when its inputs come from a 3-D anisotropic radiation field, despite the absence of direct 3-D NLTE validation and the presence of small line errors already visible in the 3-D LTE comparison.

Editorial extensions

If this is right

A 3-D NLTE spectropolarimetric simulation of a type Ia model, including a full time series and resolved polarization maps, becomes a routine roughly 3000 core-hour calculation instead of a prohibitive one.
Treating silicon, sulfur, and calcium in NLTE measurably changes predicted spectra and polarization; for example, silicon in NLTE enhances Si II absorption features and suppresses the 6100 A polarization signal, so future fits to observed spectropolarimetry can be done in 3-D NLTE.
The combination of APNN and IBT makes it practical to compute many viewing directions and many explosion models in 3-D, enabling systematic mapping of how supernova internal structure imprints on polarization.
The IBT-based resolved images connect the line-of-sight density structure of elements such as calcium directly to the sign and magnitude of Q and U polarization in velocity slices, tying 3-D hydrodynamic models to imaging data.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The 3000 core-hour figure suggests that campaign-style studies, scanning dozens of SN Ia models over multiple viewing angles with NLTE fidelity, are now feasible on typical university clusters, though the paper itself demonstrates only a single model and a single viewing direction.
Because the APNN replaces a generic atomic physics step, the same trained emulator could in principle be transplanted into other Monte Carlo radiative transfer codes that share SEDONA's input variables, spreading the speedup beyond one code base.
The line-specific errors seen at Si II 5640 A and the Ca II infrared triplet in the 3-D LTE comparison imply that the surrogate's accuracy is line-dependent; a testable improvement is to add 3-D training zones or more diverse ejecta models to harden those specific transitions.
Until a direct 3-D NLTE solver comparison is made, the 3-D NLTE spectra and polarization maps should be read as predictions from the surrogate, with the LTE agreement as partial but not complete support.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Here is the quick take. The genuinely new thing is an atomic-physics neural network (APNN) that replaces the NLTE level-population/opacity/emissivity solve inside SEDONA. Trained on 119 1D SN Ia models with four atomic recipes, it reproduces held-out 1D spectra well, and per-zone cost drops from ~72 core-seconds (Si/S/Ca NLTE) to 0.17 core-seconds—a real two-order-of-magnitude gain. Using it, the authors run the 3D N100 model with Si/S/Ca in NLTE and get spatially resolved polarization maps for ~3000 core-hours. If that holds, it opens a new regime for affordable 3D NLTE SN Ia modeling.

Credit where it is due: the 1D validation is proper—separate train/validation/test splits, and the test-set spectra in Figure 7 look convincing. The paper also flags its own blemishes: the Si II 5640 over-opacity in the 3D LTE test and the increasing validation MSE as more elements are treated in NLTE. That honesty is real.

The soft spot is the absence of any 3D NLTE validation. The surrogate is trained on 1D spherical models; the only 3D check is an LTE comparison, which already shows errors on Si II 5640 near 17 days and on the Ca II IR triplet near 41 days—the same lines that drive the NLTE discussion. Then Figure 2 presents 3D NLTE spectra at 17-18 days with no comparison to SEDONA's traditional NLTE solver. The authors acknowledge in the conclusion that a more complete NLTE solution is needed, but that caveat sits after the abstract has already declared the 3D NLTE result successful.

The 1D-to-3D extrapolation is not argued. No test on anisotropic radiation fields, no convergence study, no exploration of where the APNN might fail. The network uses local quantities, but the distribution of mean-intensity shapes in 3D can be far from the 1D training distribution, and the 50^3 Cartesian grid is a different geometry from spherical shells. The LTE comparison is a useful first step, but it is not the right test for a surrogate that matters most when NLTE is on.

Smaller issues: no code or weights are released, so the 3000 core-hour claim is not independently checkable; and the LTE wall-clock comparison (73h vs 56h) is only 1.3x, so the headline speedup comes from per-zone timing, not a direct 3D NLTE-to-NLTE benchmark.

Bottom line: this is a promising methods paper with an honest 1D validation and a genuinely useful speedup, but the central 3D NLTE result is not yet validated. It should go to peer review—the idea is worth referee time—but the referee should require at least one 3D NLTE validation, e.g., running the traditional solver on a low-resolution or angle-averaged version of N100, plus an explicit discussion of extrapolation limits. Treat the current 3D NLTE spectra as a demonstration, not a benchmark.

Referee Report

4 major / 6 minor

Summary. SEDONA-GesaRaT replaces the atomic physics module of the Monte Carlo radiative transfer code SEDONA with a set of convolutional neural networks (APNNs) that map local density, abundances, and mean intensity to extinction and emission coefficients. The APNNs are trained on 119 one-dimensional SNe Ia ejecta models from HESMA, run with SEDONA under four atomic recipes: CMF 100 LTE, and CMF All with Ca, Si+Ca, and Si+S+Ca in NLTE. The paper reports 1D held-out spectral reproductions, a 3D LTE comparison against SEDONA on the N100 model, and then 3D NLTE spectropolarimetry of N100 at 17-18 and 45 days, including resolved polarization maps, with a total cost of ~3000 core-hours.

Significance. If the claimed accuracy holds, SEDONA-GesaRaT would be a significant methodological advance, making 3D NLTE spectropolarimetry of SNe Ia computationally feasible and enabling systematic studies of ejecta structure. The paper's strengths are its concrete engineering: a well-defined train/validation/test split over a public archive (HESMA), a reproducible training recipe (architecture, optimizer, normalization), direct per-zone timing measurements (Table 4), and the integration of the IBT retrieval method. The empirical speed-up on the per-zone atomic physics step is large and credible. However, the central scientific claim—that the APNN surrogate is accurate enough for 3D NLTE spectropolarimetry—is not directly validated. The only 3D comparison is under LTE and shows known discrepancies at wavelengths that are central to the NLTE analysis. The paper's significance is therefore conditional on additional validation or a clearly quantified error bound.

major comments (4)

[Sec. 3.1-3.2, Figs. 1-3] The only 3D validation of the APNN surrogate is the CMF 100 LTE comparison in Sec. 3.1, which the authors themselves report as overestimating the opacity of Si II 5640 Å near 17 days and of the Ca II infrared triplet near 41 days. The 3D NLTE results in Sec. 3.2 (Figs. 2 and 3) are presented at exactly these epochs and feature these very line features. The assertion that the APNN is 'appropriate for most of 3-D NLTE RT simulations' (Sec. 3.1) is supported only by 1D held-out tests (Appendix B), which do not probe generalization to the anisotropic radiation field and 50^3 Cartesian geometry of the N100 model. A direct 3D NLTE comparison against SEDONA's traditional solver for at least one epoch and viewing direction (or a carefully quantified extrapolation bound from the observed LTE discrepancies) is needed to support the headline claim. Without it, the 3D NLTE spectra and polarization maps in Figs. 2-4 remain unvalidated.
[Sec. 3.3, Abstract] The claimed cost of ~3000 core-hours is the wall-clock time of SEDONA-GesaRaT on the 3D N100 NLTE run, not a comparison against a traditional 3D NLTE calculation. The only end-to-end timing comparison in the paper (Sec. 3.1) shows 73.3 hours for SEDONA versus 56.3 hours for SEDONA-GesaRaT on 48 cores for 3D LTE, a factor of only ~1.3, not the ~300-400x per-zone speedup of Table 4. The abstract's statement that previous codes 'could only finish 1-D NLTE simulation, or 3-D LTE simulation' with similar resources should be backed by an explicit scaling estimate based on the measured per-zone NLTE cost (e.g., 72.09 core-seconds) and the number of zones and timesteps in the N100 simulation. Please provide such an estimate and state its assumptions.
[Appendix A, Fig. 6; Sec. 4] The left panel of Fig. 6 shows that the validating-set MSE increases monotonically as more elements are included in the NLTE treatment, yet Appendix A states that in the most complex recipe 'the performance does not decrease compared to the other three simpler recipes,' and Sec. 4 states that 'the 1-D RT simulated spectral time sequence accuracy is not reduced.' These statements are inconsistent with the reported MSE trend and are not backed by any quantitative spectral accuracy metric on the held-out 1D testing set. Please report the numerical MSE values (or an equivalent spectral error metric such as RMS flux deviation or line equivalent-width error) for each of the four APNNs on the 19-model testing set, and connect these numbers to the claimed fidelity of the 3D NLTE results.
[Sec. 2 and Appendix B] The paper does not describe how the APNN is embedded in the iterative Monte Carlo radiative transfer loop: how many transport/population iterations are performed per timestep, whether the mean intensity J_nu fed to the APNN is taken from a previous iteration or iterated to consistency, and whether the training data were generated with converged J_nu. Because NLTE level populations depend nonlinearly on the local radiation field, a surrogate that is accurate on fixed single-zone inputs may be biased in a coupled time-dependent run if the feedback loop is not converged. Please specify the iteration protocol and include a convergence test (e.g., sensitivity of the output spectra to the number of iterations).

minor comments (6)

[Sec. 4 vs Sec. 3.1, Fig. 1] The text refers to 'S II 5640 Å' in Sec. 4 and the abstract region of the conclusion, while Sec. 3.1 and Fig. 1 call the same feature 'Si II 5640 Å'; resolve this line-identification inconsistency.
[Sec. 4, Table 4] Sec. 4 states that the computation time is reduced 'from ~50 core-seconds to 0.17 core-seconds,' but Table 4 lists 1.19 core-seconds for CMF 100 LTE and 57.92-72.09 core-seconds for the NLTE recipes; specify which recipe the ~50 core-seconds refers to.
[Abstract and Sec. 1] The abstract and introduction describe the code as performing '3-D NLTE radiative transfer calculation' without clarifying that only Si, S, and Ca are treated in NLTE, while all other elements remain in LTE; please state this limitation early to avoid overclaiming.
[Fig. 1 and Fig. 7 captions] The color-coding of the lines is described inconsistently: Fig. 1 says green is SEDONA and red is SEDONA-GesaRaT, while the Fig. 7 text implies the opposite; make the color assignments consistent across all figure captions.
[Abstract, Sec. 3.3] The claim that 'previous codes could only finish 1-D NLTE simulation, or 3-D LTE simulation' under-represents existing approximate 3D NLTE efforts, including the level-merging and data-driven approaches cited in Sec. 1; please rephrase to acknowledge these methods and clarify the distinction.
[Sec. 2 vs Sec. 3.3] The training simulations in Sec. 2 use 5x10^4 energy packets per timestep, while the 3D simulations in Sec. 3.3 use 4x10^6 packets per timestep; discuss whether the lower packet number in the training data could imprint Monte Carlo noise into the APNN targets and whether any noise-averaging was applied.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the APNN is a standard emulator validated on held-out SEDONA models, and the 3D NLTE extrapolation is a generalizability risk rather than a definitional reduction.

full rationale

The central machine-learning component, APNN, is trained on the output of SEDONA's traditional level-population solver and then compared with that same solver on a deliberately held-out testing set (19 models not used in training or validation). This is the standard and non-circular protocol for emulator validation: the testing models were not fitted, so agreement there is independent evidence of interpolation within the 1D training distribution. The paper explicitly states the split: 80 models in training, 20 in validation, and 19 in testing, with the testing set used only to compare APNN results with traditional methods. The 3D NLTE results, by contrast, are an extrapolation of a 1D-trained surrogate to a 3D model with anisotropic radiation fields and a 50x50x50 Cartesian grid. The paper provides no direct 3D NLTE validation and even acknowledges accuracy issues on Si II 5640 and the Ca II infrared triplet in the 3D LTE comparison, as well as increasing MSE when more elements are treated in NLTE. That is a serious correctness and generalization risk, but it is not circularity: the 3D N100 output is not a fitted target of the APNN, and no equation or construction forces the 3D spectra to equal the training data. The self-citation to Chen et al. (2024b) for the integral-based technique is a normal method dependency; the present paper applies IBT and shows its use in figures, and the citation does not function as an unverified premise that the derivation reduces to. Overall, the paper's 1D emulator claim is self-contained and properly benchmarked, while the 3D NLTE claim rests on an unvalidated extrapolation that belongs in the correctness column, not the circularity column.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the APNN as a fitted surrogate, with no released code or weights, and on several domain assumptions about the representativeness of 1D training data for 3D applications. The atomic data reduction is a further modeling choice that affects the NLTE results.

free parameters (4)

APNN neural network weights = 13,480,395 trained parameters
The surrogate predictions of absorption and emission coefficients are entirely determined by these fitted weights, trained on SEDONA 1D output. No independent physical constraints are imposed.
Input/output normalization ranges = Hand-chosen, e.g., rho_new=(log10(rho)+23)/4, E_new=(log10(sum J_nu dnu)-10.5)/3.5
These scaling constants are selected by the authors to keep quantities in the network's dynamic range; the central claim depends on them only through training performance.
Neural network hyperparameters = Learning rate 1e-4, batch size 64, 200 epochs, 33 conv layers, 5 FC layers
Chosen empirically for training stability; not derived from physics.
Atomic data library reduction = CMF All has 497,424 levels and 35,344,426 transitions; Si, S, Ca kept at CMF 100 levels
The choice of which elements to keep at 100 levels is a modeling decision that affects the accuracy of NLTE results, especially for the species highlighted in the 3D runs.

assumptions (5)

standard math Monte Carlo radiative transfer with Sobolev and expansion opacity approximations is a valid description of SN atmospheres.
Invoked in Section 2 (Eqs. 1-4); these are standard approximations in SN RT codes, though they limit fidelity on optically thick lines.
domain assumption The 119 1D HESMA SN Ia models span the physical conditions (density, composition, radiation field) of the 3D N100 model and of general SNe Ia.
Section 2 states the training set 'covered the range of physical conditions relevant for SNe Ia'; this is asserted without a quantitative coverage analysis.
domain assumption NLTE level populations and resulting opacities/emissivities depend only on local density, composition, and mean intensity J_nu.
This is inherent in the standard NLTE rate equations used by SEDONA, but the APNN is trained on 1D spherical (angle-averaged) radiation fields and applied to 3D fields; non-local and angle-dependent effects are neglected.
domain assumption The N100 model mapped to a 50^3 grid with max velocity 28600 km/s adequately represents the ejecta structure.
Section 3 states the mapping without resolution convergence tests.
domain assumption Forbidden lines are not important for the epochs and wavelengths studied.
Section 2 notes forbidden lines (e.g., Ca II 7291, 7324) are excluded from the atomic libraries; this may affect late-time spectra.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SEDONA-GesaRaT: an AI-Accelerated Radiative Transfer Program for 3-D Supernova Simulations." pith.science (2026). https://pith.science/paper/EPYCPDYP

@misc{pith2026250711767,
  author       = {Pith},
  title        = {Pith review of: SEDONA-GesaRaT: an AI-Accelerated Radiative Transfer Program for 3-D Supernova Simulations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EPYCPDYP}},
  note         = {Machine review of arXiv:2507.11767}
}

abstract

We present SEDONA-GesaRaT, a rapid code for supernova radiative transfer simulation developed based on the Monte-Carlo radiative transfer code SEDONA. We use a set of atomic physics neural networks (APNN), an artificial intelligence (AI) solver for the non-local thermodynamic equilibrium (NLTE) atomic physics level population calculation, which is trained and validated on 119 1-D type Ia supernova (SN Ia) radiative transfer simulation results showing great computation speed and accuracy. SEDONA-GesaRaT has been applied to the 3-D SN Ia explosion model N100 to perform a 3-D NLTE radiative transfer calculation. The spatially resolved linear polarization data cubes of the N100 model are successfully retrieved with a high signal-to-noise ratio using the integral-based technique (IBT). The overall computation cost of a 3-D NLTE spectropolarimetry simulation using SEDONA-GesaRaT is only $\sim$3000 core-hours, while the previous codes could only finish 1-D NLTE simulation, or 3-D local thermodynamic equilibrium (LTE) simulation, with similar computation resources. The excellent computing efficiency allows SEDONA-GesaRaT for future large-scale simulations that systematically study the internal structures of supernovae.

Figures

Figures reproduced from arXiv: 2507.11767 by the authors.

**Figure 1.** The spectropolarimetry time sequence of the 3-D N100 model at the viewing direction µ = 0.5, φ = 3.4455. The green line is calculated by SEDONA which use traditional method to solve the source function, while the red line is calculated by SEDONA-GesaRaT using APNN to solve the source function, both solutions use the CMF 100 LTE recipe for atomic physics. From left to right, the first panel shows the optical waveleng… view at source ↗

**Figure 2.** The spectropolarimetry of the N100 model between 17 and 18 days after the explosion, in the viewing direction µ = 0.5, φ = 3.4455. Upper panel is the spectral flux, middle panel is the linear polarization percentage Q/I, lower panel is the linear polarization percentage U/I. The identified spectral lines are marked with vertical lines and labeled with the element names and the blueshift velocities [PITH_FULL_IMAGE:… view at source ↗

**Figure 3.** Same as [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Left and middle panel show the linear polarization maps of the Q and U components at 17-18 days after the explosion. Upper panel is at the wavelength 3697 ˚A and lower panel is at the wavelength 3800 ˚A. The maps are split into four regions following the value symbol. …

**Figure 5.** Figure 5: The illustrative neural network structure for APNN. We use adam (Kingma & Ba 2014) optimization algorithm to train APNN, and we use the mean square error (MSE) as a loss function. The training batch size is 64, and the learning rate is 0.0001. The neural network is tra…

**Figure 6.** Figure 6: Left Panel: The MSE value on the validating data set of the APNNs trained on different atomic physics libraries and approximations. Middle and Right Panel: A comparison of the absorption and emission coefficients calculated from APNN (orange thin line) and from traditi…

**Figure 7.** Figure 7: A comparison between the SEDONA spectral time series (green line) and SEDONA-GesaRaT spectral time series (red line). The 1-D SN Ia explosion model names are shown in the title of each panel. The time after the explosion is labeled on the left of each spectrum. Atomic …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 15 canonical work pages

[1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter doi edition editor eprint howpublished institution journal key month number organization pages publisher school series title misctitle type volume year version url label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts ...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION format.url url empty "" new.block "" url * "" * if FUNCTION format.eprint eprint empty "" archivePrefix empty "" archivePrefix "arXiv" = new.block " " eprint * " " * new.block " " eprint * " " * if if if FUNCTION format.doi doi empty "" " " doi * " " * if FUNCTION format.pid doi empty eprint empty ur...
[3]

&E( cֳnm=檖 <z f Yb7q RhQq61 T* 1 /_ _Я_? H(zQ 5UTɑ9(d< R ϟG tԉ ߐ

thebibliography [1] 20pt to REFERENCES 6pt =0pt -12pt 10pt plus 3pt =0pt =0pt =1pt plus 1pt =0pt =0pt -12pt =13pt plus 1pt =20pt =13pt plus 1pt \@M =10000 =-1.0em =0pt =0pt 0pt =0pt =1.0em @enumiv\@empty 10000 10000 `\.\@m \@noitemerr \@latex@warning Empty `thebibliography' environment \@ifnextchar \@reference \@latexerr Missing key on reference command E...

2021
[4]

Anderson , L. S. 1985, , 298, 848, 10.1086/163668

work page doi:10.1086/163668 1985
[5]

1989, , 339, 558, 10.1086/167317

---. 1989, , 339, 558, 10.1086/167317

doi:10.1086/167317 1989
[6]

P., et al

Blondin , S., Blinnikov , S., Callan , F. P., et al. 2022, , 668, A163, 10.1051/0004-6361/202244134

doi:10.1051/0004-6361/202244134 2022
[7]

J., Deems, S., Furlani, T

Boerner, T. J., Deems, S., Furlani, T. R., Knuth, S. L., & Towns, J. 2023, in Practice and Experience in Advanced Research Computing, 173--176

work page 2023
[8]

Non-LTE Synthetic Observables of a Multidimensional Model of Type Ia Supernovae

Boos , S. J., Dessart , L., Shen , K. J., & Townsley , D. M. 2024, arXiv e-prints, arXiv:2410.22276, 10.48550/arXiv.2410.22276

work page Pith review arXiv doi:10.48550/arxiv.2410.22276 2024

Show all 36 references

[9]

A., & Kromer , M

Bulla , M., Sim , S. A., & Kromer , M. 2015, , 450, 967, 10.1093/mnras/stv657

2015 doi
[10]

2020, , 250, 12, 10.3847/1538-4365/ab9a3b

Chen , X., Hu , L., & Wang , L. 2020, , 250, 12, 10.3847/1538-4365/ab9a3b

2020 doi
[11]

Chen , X., Wang , L., Hu , L., & Brown , P. J. 2024 a , , 962, 125, 10.3847/1538-4357/ad0a33

2024 doi
[12]

2024 b , arXiv e-prints, arXiv:2409.07729, 10.48550/arXiv.2409.07729

Chen , X., Wang , L., & Kasen , D. 2024 b , arXiv e-prints, arXiv:2409.07729, 10.48550/arXiv.2409.07729

2024 doi arXiv
[13]

2019, , 490, 578, 10.1093/mnras/stz2322

Cikota , A., Patat , F., Wang , L., et al. 2019, , 490, 578, 10.1093/mnras/stz2322

2019 doi
[14]

P., Young , P

Del Zanna , G., Dere , K. P., Young , P. R., & Landi , E. 2021, , 909, 38, 10.3847/1538-4357/abd8ce

2021 doi
[15]

J., Blondin , S., & Khokhlov , A

Dessart , L., Hillier , D. J., Blondin , S., & Khokhlov , A. 2014, , 441, 3249, 10.1093/mnras/stu789

2014 doi
[16]

2023, , 945, 52, 10.3847/1538-4357/acb496

Ferrazzoli , R., Slane , P., Prokhorov , D., et al. 2023, , 945, 52, 10.3847/1538-4357/acb496

2023 doi
[17]

N., Khokhlov , A

Gamezo , V. N., Khokhlov , A. M., & Oran , E. S. 2004, , 92, 211102, 10.1103/PhysRevLett.92.211102

2004 doi
[18]

J., & Miller , D

Hillier , D. J., & Miller , D. L. 1998, , 496, 407, 10.1086/305350

1998 doi
[19]

2003, in Astronomical Society of the Pacific Conference Series, Vol

H \"o flich , P. 2003, in Astronomical Society of the Pacific Conference Series, Vol. 288, Stellar Atmosphere Modeling, ed. I. Hubeny , D. Mihalas , & K. Werner , 185

2003
[20]

Jeffery , D. J. 1991, , 375, 264, 10.1086/170187

1991 doi
[21]

H., Lasher , G., Chan , K

Karp , A. H., Lasher , G., Chan , K. L., & Salpeter , E. E. 1977, , 214, 161, 10.1086/155241

1977 doi
[22]

C., & Nugent , P

Kasen , D., Thomas , R. C., & Nugent , P. 2006, , 651, 366, 10.1086/506190

2006 doi
[23]

E., & Sim , S

Kerzendorf , W. E., & Sim , S. A. 2014, , 440, 387, 10.1093/mnras/stu055

2014 doi
[24]

P., & Ba , J

Kingma , D. P., & Ba , J. 2014, arXiv e-prints, arXiv:1412.6980, 10.48550/arXiv.1412.6980

2014 doi arXiv
[25]

A., Alieva , A., et al

Kochkov , D., Smith , J. A., Alieva , A., et al. 2021, Proceedings of the National Academy of Science, 118, e2101784118, 10.1073/pnas.2101784118

2021 doi
[26]

Kromer , M., Ohlmann , S., & R \"o pke , F. K. 2017, , 88, 312, 10.48550/arXiv.1706.09879

2017 doi arXiv
[27]

Kromer , M., & Sim , S. A. 2009, , 398, 1809, 10.1111/j.1365-2966.2009.15256.x

2009
[28]

2013, , 429, 2287, 10.1093/mnras/sts498

Kromer , M., Fink , M., Stanishev , V., et al. 2013, , 429, 2287, 10.1093/mnras/sts498

2013 doi
[29]

L., & Bell , B

Kurucz , R. L., & Bell , B. 1995, Atomic line list

1995
[30]

A., & Lucy , L

Mazzali , P. A., & Lucy , L. B. 1993, , 279, 447

1993
[31]

2024, , 681, A11, 10.1051/0004-6361/202346715

Nagao , T., Patat , F., Cikota , A., et al. 2024, , 681, A11, 10.1051/0004-6361/202346715

2024 doi
[32]

Paszke, A., Gross, S., Chintala, S., et al. 2017

2017
[33]

R., Ciaraldi-Schoolmann , F., R \"o pke , F

Seitenzahl , I. R., Ciaraldi-Schoolmann , F., R \"o pke , F. K., et al. 2013, , 429, 1156, 10.1093/mnras/sts402

2013 doi
[34]

M., Kavanagh , P

Temim , T., Laming , J. M., Kavanagh , P. J., et al. 2024, , 968, L18, 10.3847/2041-8213/ad50d1

2024 doi
[35]

Vartanyan , D., Tsang , B. T. H., Kasen , D., et al. 2024, arXiv e-prints, arXiv:2411.03434, 10.48550/arXiv.2411.03434

2024 doi arXiv
[36]

E., Sim , S

Vogl , C., Kerzendorf , W. E., Sim , S. A., et al. 2020, , 633, A88, 10.1051/0004-6361/201936137

2020 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter doi edition editor eprint howpublished institution journal key month number organization pages publisher school series title misctitle type volume year version url label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts ...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION format.url url empty "" new.block "" url * "" * if FUNCTION format.eprint eprint empty "" archivePrefix empty "" archivePrefix "arXiv" = new.block " " eprint * " " * new.block " " eprint * " " * if if if FUNCTION format.doi doi empty "" " " doi * " " * if FUNCTION format.pid doi empty eprint empty ur...

[3] [3]

&E( cֳnm=檖 <z f Yb7q RhQq61 T* 1 /_ _Я_? H(zQ 5UTɑ9(d< R ϟG tԉ ߐ

thebibliography [1] 20pt to REFERENCES 6pt =0pt -12pt 10pt plus 3pt =0pt =0pt =1pt plus 1pt =0pt =0pt -12pt =13pt plus 1pt =20pt =13pt plus 1pt \@M =10000 =-1.0em =0pt =0pt 0pt =0pt =1.0em @enumiv\@empty 10000 10000 `\.\@m \@noitemerr \@latex@warning Empty `thebibliography' environment \@ifnextchar \@reference \@latexerr Missing key on reference command E...

2021

[4] [4]

Anderson , L. S. 1985, , 298, 848, 10.1086/163668

work page doi:10.1086/163668 1985

[5] [5]

1989, , 339, 558, 10.1086/167317

---. 1989, , 339, 558, 10.1086/167317

doi:10.1086/167317 1989

[6] [6]

P., et al

Blondin , S., Blinnikov , S., Callan , F. P., et al. 2022, , 668, A163, 10.1051/0004-6361/202244134

doi:10.1051/0004-6361/202244134 2022

[7] [7]

J., Deems, S., Furlani, T

Boerner, T. J., Deems, S., Furlani, T. R., Knuth, S. L., & Towns, J. 2023, in Practice and Experience in Advanced Research Computing, 173--176

work page 2023

[8] [8]

Non-LTE Synthetic Observables of a Multidimensional Model of Type Ia Supernovae

Boos , S. J., Dessart , L., Shen , K. J., & Townsley , D. M. 2024, arXiv e-prints, arXiv:2410.22276, 10.48550/arXiv.2410.22276

work page Pith review arXiv doi:10.48550/arxiv.2410.22276 2024