SEDONA-GesaRaT: 3D NLTE Supernova Modeling

• SEDONA-GesaRaT is a radiative-transfer framework that replaces computationally expensive NLTE atomic physics with the APNN surrogate, enabling efficient 3D supernova simulations.
• It integrates SEDONA’s Monte Carlo engine with an integral-based spectropolarimetric extraction technique to produce low-noise observer data at reduced core-hour costs.
• The method achieves a reduction from ~50 core-seconds to 0.17 core-seconds per zone, making detailed 3D NLTE spectropolarimetric analyses of Type Ia supernovae feasible.

Searching arXiv for the core SEDONA-GesaRaT paper and closely related SEDONA methodology papers. Searching arXiv for related work on the integral-based technique and SEDONA-based multidimensional radiative transfer. SEDONA-GesaRaT is a SEDONA-based radiative-transfer framework for supernova modeling in which the dominant computational bottleneck of non-local thermodynamic equilibrium atomic physics is replaced by a learned surrogate, the atomic physics neural network (APNN). In the reported implementation, the framework retains SEDONA’s Monte Carlo transport machinery, couples it to the integral-based technique (IBT) for spectropolarimetric extraction, and is demonstrated on the three-dimensional Type Ia supernova model N100, including spatially resolved linear-polarization data cubes. Its central claim is practical rather than terminological: direct $3$-D, time-dependent, NLTE spectropolarimetric simulation becomes feasible at a total cost of about $\sim 3000$ core-hours, where earlier workflows were effectively restricted to either $1$-D NLTE or $3$-D LTE calculations at similar computational budgets (Chen et al., 15 Jul 2025).

1. Definition and scientific setting

SEDONA-GesaRaT is presented as a modified SEDONA framework rather than a wholly separate transport code. The scientific problem it addresses is the long-standing tradeoff in supernova radiative transfer between multidimensional geometry and realistic atomic kinetics. In $3$-D Monte Carlo transport, the cost of packet propagation is already large; in NLTE, the additional requirement to solve the atomic level-population problem zone by zone and timestep by timestep raises the cost sharply. The paper states that, in SEDONA, enabling the traditional NLTE solver for selected elements increases computation cost by roughly a factor of $\sim 100$ relative to LTE. This matters because supernova remnants, spectra, and spectropolarimetry all indicate intrinsically asymmetric ejecta, while $1$-D NLTE and $3$-D LTE each discard one of the two essential ingredients: realistic atomic physics or realistic geometry (Chen et al., 15 Jul 2025).

Within that context, SEDONA-GesaRaT is defined by an internal substitution. The transport solver, packet propagation, and interaction treatment remain those of SEDONA, but the costly atomic-physics stage is replaced by APNN, a neural-network surrogate trained on $119$ one-dimensional SN Ia radiative-transfer simulations. The framework is therefore aimed at accelerating the NLTE material-property evaluation that supplies the transport coefficients, not at replacing Monte Carlo transport itself. This distinction is central to its identity.

2. Framework composition and workflow

The framework consists of three functional layers: the original SEDONA transport engine, APNN as the atomic-physics surrogate, and IBT as the spectropolarimetric extraction method. The resulting workflow is local-condition $\rightarrow$ transport-coefficient prediction $\sim 3000$0 Monte Carlo propagation $\sim 3000$1 low-noise observer extraction.

Component	Function	Basis
SEDONA	Time-dependent multidimensional Monte Carlo radiative transfer	(Chen et al., 15 Jul 2025)
APNN	Predicts $\sim 3000$2, $\sim 3000$3, and $\sim 3000$4 from local state variables	(Chen et al., 15 Jul 2025)
IBT	High-S/N spectropolarimetric extraction using integral packets	(Chen et al., 2024)

Operationally, ejecta structure is supplied on a grid, radioactive decay deposition and packet transport proceed as in SEDONA, and each zone and timestep provide the local radiation estimator $\sim 3000$5, density, and composition. Instead of explicitly solving the rate-equation problem for atomic level populations and then recomputing emissivities and opacities, APNN maps local conditions directly to the absorption or extinction coefficient spectrum $\sim 3000$6, the emissivity spectrum $\sim 3000$7, and the electron-scattering coefficient $\sim 3000$8 or $\sim 3000$9. Monte Carlo transport then uses those coefficients exactly where the original SEDONA solver would have used conventionally computed NLTE quantities (Chen et al., 15 Jul 2025).

A common misconception is that SEDONA-GesaRaT is an AI replacement for radiative transfer in its entirety. The paper is explicit that this is not the case. The machine-learning component is inserted inside the source-function and atomic-property calculation stage, whereas transport itself remains the original SEDONA Monte Carlo scheme. This preserves the transport formalism while relocating the computational savings to the atomic microphysics.

3. APNN and the learned NLTE surrogate

The expensive physical task that APNN bypasses is the determination of atomic level populations under statistical equilibrium. Rather than solving directly for $1$0 from the coupled transition-rate system, the network learns an operational mapping

$1$1

This reframes NLTE atomic physics as a supervised regression problem whose outputs are the transport coefficients actually required by SEDONA (Chen et al., 15 Jul 2025).

The network inputs are a rescaled density, a normalized mean-intensity spectrum, a rescaled integrated radiation energy, and a $1$2-element abundance vector from H to Ni. The appendix states these explicitly as $1$3, $1$4, $1$5, and $1$6. The outputs are not trained in raw form. Because $1$7 and $1$8 span many orders of magnitude, each spectrum is decomposed into a minimum, a maximum, and a normalized spectral shape; electron scattering is predicted through a separate rescaled scalar $1$9. This output factorization is intended to compress dynamic range and separate amplitude from morphology.

Architecturally, APNN is a convolutional neural network with $3$0 convolution layers, $3$1 fully connected layers, and $3$2 trainable parameters. The $3$3-point input spectrum $3$4 passes through four down-sampling blocks, each block containing four sub-blocks of convolution, layer normalization, and ReLU followed by max pooling. The encoded spectral features are concatenated with $3$5, $3$6, and $3$7, passed through the fully connected stack to predict the scale parameters $3$8, and then decoded through four up-sampling blocks plus a convolution layer to reconstruct the spectral shapes $3$9 and $3$0. The design is therefore encoder-decoder-like and tailored to structured frequency-dependent outputs.

Training data are generated from $3$1 one-dimensional SN Ia SEDONA simulations using HESMA ejecta models, evolved from day $3$2 to day $3$3. At each radial zone and timestep, the recorded quantities are elemental abundances, plasma density, $3$4, extinction coefficients, and emissivities. The frequency grid spans $3$5 to $3$6 with logarithmic spacing $3$7, giving $3$8 frequency points. Four atomic-physics targets are used: CMF_100 LTE, CMF_All Ca NLTE, CMF_All Si, Ca NLTE, and CMF_All Si, S, Ca NLTE. The $3$9 models are split into $\sim 100$0 training, $\sim 100$1 validation, and $\sim 100$2 testing cases. Training uses Adam, mean squared error loss, batch size $\sim 100$3, learning rate $\sim 100$4, and $\sim 100$5 epochs; training time is about $\sim 100$6 days on an NVIDIA A100 GPU, with implementation in PyTorch and deployment through LibTorch (Chen et al., 15 Jul 2025).

The paper emphasizes that APNN is not universal. It is trained on SN Ia ejecta only, tied to the specific $\sim 100$7-point frequency grid, and learns only the atomic recipes represented in its training set. This restricts immediate portability, but also defines the domain within which the reported acceleration claims are meaningful.

4. Transport formalism, polarization, and IBT

SEDONA-GesaRaT retains SEDONA’s time-dependent Monte Carlo transport for supernova ejecta. The simulations include gamma-ray emission from radioactive isotopes, gamma-ray interactions through Compton scattering and photoionization or photoelectric processes, and optical, UV, and IR transport with Thomson scattering, bound-bound transitions, bound-free transitions, and free-free transitions. The ejecta are assumed to undergo homologous expansion, while line transport uses the Sobolev approximation and expansion-opacity formalism (Chen et al., 15 Jul 2025).

Several transport coefficients and interaction probabilities are stated explicitly. The comoving-frame Thomson extinction coefficient is

$\sim 100$8

and the Sobolev optical depth for an individual line is

$\sim 100$9

The effective interaction probability is written as

$1$0

and the total bound-bound extinction in a frequency bin is

$1$1

These expressions clarify that the framework accelerates coefficient generation but does not alter the radiative-transfer formalism itself.

For spectropolarimetry, the reported observables are the standard linear-polarization fractions $1$2 and $1$3. The companion methodological paper on IBT defines it as an observer-extraction algorithm based on integral packets rather than event-spawned virtual packets. Implemented in $1$4-D SEDONA, IBT was reported to be faster than the event-based technique by a factor of $1$5–$1$6 and to improve signal-to-noise by a factor of $1$7–$1$8 at fixed Monte Carlo quanta, while also passing spherical-symmetry, mirror-symmetry, and cross-comparison tests against direct counting and event-based extraction (Chen et al., 2024). In SEDONA-GesaRaT, IBT is the mechanism that makes the low-noise spatially resolved polarization products practical once the atomic-physics bottleneck is removed.

5. Validation, N100 application, and computational performance

Validation is reported at two levels. First, after APNN is integrated into SEDONA-GesaRaT, the authors run one-dimensional time-dependent tests on the held-out $1$9-model test set and report that the spectral time series reproduce the original SEDONA results from early to late phases with high accuracy, with predicted and conventional absorption or emission spectra overlapping visually. Second, they compare SEDONA-GesaRaT against original SEDONA in $3$0-D using the simpler CMF_100 LTE setup on the N100 model, because a direct $3$1-D NLTE-to-$3$2-D NLTE comparison with the traditional solver would be prohibitively expensive (Chen et al., 15 Jul 2025).

For the $3$3-D LTE benchmark, the setup is a $3$4 Cartesian grid, $3$5 packets per timestep, evolution from day $3$6 to day $3$7, and execution on a $3$8-core node. The reported wall-clock times are $3$9 hr for original SEDONA and $119$0 hr for SEDONA-GesaRaT. Flux spectra and linear polarization are generally consistent from $119$1–$119$2 Å and from $119$3–$119$4 days, with the main discrepancies occurring in Si II $119$5 Å around day $119$6 and the Ca II IR triplet around day $119$7.

The paper’s principal scientific demonstration is the $119$8-D NLTE application to the delayed-detonation SN Ia model N100. The ejecta are mapped to a $119$9 Cartesian grid with maximum ejecta velocity $\rightarrow$0. The showcased viewing direction is

$\rightarrow$1

The framework computes time-dependent $\rightarrow$2-D spectra, linear spectropolarimetry, and spatially resolved spectropolarimetry data cubes. The resolved image calculation uses $\rightarrow$3 packets per timestep and a $\rightarrow$4 observer image grid spanning

$\rightarrow$5

in both projected axes. The total runtime for this $\rightarrow$6-D NLTE spectropolarimetry simulation is $\rightarrow$7 hours on a $\rightarrow$8-core node, or $\rightarrow$9 core-hours (Chen et al., 15 Jul 2025).

The physical outputs are not limited to global spectra. At $\sim 3000$00–$\sim 3000$01 days after explosion, putting S in NLTE enhances the emission feature of S II $\sim 3000$02 Å and the absorption features of S II $\sim 3000$03 Å and $\sim 3000$04 Å. Putting Si in NLTE enhances the P-Cygni profile of Si II $\sim 3000$05 Å, enhances the absorption feature of Si II $\sim 3000$06 Å, and suppresses the linear polarization signal near $\sim 3000$07 Å. The spatially resolved polarization maps at $\sim 3000$08 Å and $\sim 3000$09 Å are interpreted as tracing Ca II H&K line formation in velocity slices of roughly $\sim 3000$10–$\sim 3000$11 and $\sim 3000$12–$\sim 3000$13, respectively, and their morphology broadly overlaps the Ca density distribution in those slices.

The per-zone timing benchmark summarizes the acceleration more starkly:

Atomic treatment	Core-seconds per zone
CMF_100 LTE	1.19
CMF_All Ca NLTE	57.92
CMF_All Si, Ca NLTE	65.86
CMF_All Si, S, Ca NLTE	72.09
APNN	0.17

These figures imply that APNN is about $\sim 3000$14 times faster than the reduced LTE calculation and roughly $\sim 3000$15–$\sim 3000$16 times faster than the explicit NLTE variants. The paper summarizes this as a reduction from about $\sim 3000$17 core-seconds to $\sim 3000$18 core-seconds per zone, that is, roughly two to three orders of magnitude (Chen et al., 15 Jul 2025).

6. Limits, interpretation, and research role

The principal limitations are sharply defined. APNN is trained only on SN Ia models; application to core-collapse supernovae would require new training data. The learned surrogate is tied to a specific $\sim 3000$19-point frequency grid and to selected atomic recipes rather than a general all-species NLTE treatment. Residual feature-level inaccuracies remain, particularly in S II $\sim 3000$20 Å and parts of the Ca II IR triplet. The transport still relies on the expansion-opacity formalism, whose reliability degrades when a bin contains multiple optically thick lines. More broadly, improvements may require better atomic data libraries, more stable or more complete NLTE solvers, more diverse training sets, and refined network architectures (Chen et al., 15 Jul 2025).

These limits also define the correct interpretation of the framework. SEDONA-GesaRaT is not a proof that NLTE atomic physics has been solved generically by machine learning; it is a domain-specific acceleration strategy for a constrained but astrophysically important class of SN Ia problems. Its significance lies in changing the feasible scale of multidimensional radiative-transfer studies. The paper explicitly connects the reported efficiency to systematic surveys of $\sim 3000$21-D explosion models, parameter studies over viewing angle, composition, and asymmetry, model libraries for interpreting observed spectropolarimetry, and integration with earlier AI-assisted inversion approaches. A plausible implication is that the framework repositions detailed $\sim 3000$22-D NLTE spectropolarimetry from a boutique calculation to a surveyable regime, but only within the bounds of its training domain and transport approximations.

In that sense, SEDONA-GesaRaT occupies a specific place in the SEDONA lineage. It leaves the Monte Carlo transport core intact, incorporates IBT for practical observer extraction, and relocates the dominant computational burden from direct atomic solving to supervised surrogate evaluation. The result is a SEDONA derivative whose novelty is architectural and computational: direct $\sim 3000$23-D NLTE spectropolarimetric simulation of the N100 model, including spatially resolved linear polarization data cubes, at a reported cost of $\sim 3000$24 core-hours (Chen et al., 15 Jul 2025).