SEDONA-GesaRaT: an AI-Accelerated Radiative Transfer Program for 3-D Supernova Simulations SEDONA ( Kasen et al. 2006 ) ,pytorch ( Paszke et al. 2017 )

Xingzhuo Chen, Ulisses Braga-Neto, Lifan Wang, Daniel Kasen, Zhengwei Liu, Friedrich K. Röpke, Ming Zhong, David J. Jeffery · 2025

Read SEDONA-GesaRaT: an AI-Accelerated Radiative Transfer Program for 3-D Supernova Simulations SEDONA ( Kasen et al. 2006 ) ,pytorch ( Paszke et al. 2017 ) by Xingzhuo Chen, Ulisses Braga-Neto, Lifan Wang, Daniel Kasen, Zhengwei Liu, Friedrich K. Röpke, Ming Zhong, David J. Jeffery (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-11767.txt.

Introduction

Xingzhuo Chen and co-authors speed supernova light simulations with AI in a 2025 paper. Modeling how light fights through an exploding star is slow, above all the atomic physics part. Their code, SEDONA-GesaRaT, swaps that part for trained neural networks.

The networks learned from 119 one-dimensional runs of type Ia supernovae. They solve atom level counts fast and accurately. The team then ran a full three-dimensional explosion model end to end. Polarized light data cubes came out with clean signal. The whole run cost about three thousand core-hours, a budget that once bought only flat or simplified runs. Wide surveys of explosion structure now come within reach.

Catalog record

CREATORXingzhuo Chen, Ulisses Braga-Neto, Lifan Wang, Daniel Kasen, Zhengwei Liu, Friedrich K. Röpke, Ming Zhong, David J. Jeffery
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE38 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-11767

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

SEDONA-GesaRaT: an AI-Accelerated Radiative Transfer Program for 3-D Supernova Simulations SEDONA ( Kasen et al. 2006 ) ,pytorch ( Paszke et al. 2017 ) Xingzhuo Chen, Ulisses Braga-Neto, Lifan Wang, Daniel Kasen, Zhengwei Liu, Friedrich K. Röpke, Ming Zhong, David J. Jeffery 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 ( Seitenzahl et al. 2013 ) 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) ( Chen et al. 2024b ) . 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. I Introduction Radiative transfer (RT) simulations are a critical tool in supernova (SN) research, providing the ability to synthesize observable predictions, including light-curves, spectra, and spectropolarimetry, based on the theoretical structures of supernovae (SNe). Many SN RT codes (e.g., TARDIS Kerzendorf & Sim 2014 , CMFGEN, Hillier & Miller 1998 ) are based on a 1-D spherically symmetric approximation. Reasonable agreement between different RT programs has been shown ( Blondin et al. 2022 ) , and spectral and light-curve modeling on both type Ia supernovae (SNe Ia) and core-collapse supernovae (CCSNe) has been successful in matching observations (e.g., Chen et al. 2024a ; Dessart et al. 2014 ; Vogl et al. 2020 ). However, imaging of nearby SN remnants (e.g., Ferrazzoli et al. 2023 ; Temim et al. 2024 ) and spectropolarimetry observations of SNe (e.g., Cikota et al. 2019 ; Nagao et al. 2024 ) suggest that both SNe Ia and CCSNe have complex 3-D structures, which are also revealed in multidimensional hydrodynamic simulations (e.g., Gamezo et al. 2004 ; Kromer et al. 2013 ; Vartanyan et al. 2024 ). Although several multidimensional RT codes have been developed (e.g., Kasen et al. 2006 ; Kromer & Sim 2009 ) based on the Monte-Carlo method, 3-D time-dependent RT simulation remains a computationally challenging task due to the large number of Monte-Carlo quanta required to achieve high signal-to-noise ratio (S/N) results, and the large number of spatial grid points needed to solve for the plasma excitation and ionization state. To reduce the Monte-Carlo noise in the simulated spectropolarimetry result, Bulla et al. 2015 proposed a virtual packet method. Based on the virtual packet method, Chen et al. 2024b proposed an integral-based technique (IBT), which further increases the computational efficiency. Because a full non-local thermodynamic equilibrium (NLTE) calculation requires a detailed calculation of atomic transition rates to determine the statistical equilibrium of all the atomic levels, most 3-D RT simulations adopt a simple local thermodynamic equilibrium (LTE) or approximated NLTE to strike a balance between computational resources and physical fidelity. In some RT simulations, a level-merging method Höflich 2003 ; Anderson 1985 ; Anderson 1989 has been developed to reduce the number of atomic levels considered in NLTE for an approximate result. Boos et al. 2024 proposed a data-driven method that synthesizes the 3-D NLTE RT spectra from the 3-D LTE simulation and the 1-D NLTE simulation using the extracted SN model along the line-of-sight. SEDONA ( Kasen et al. 2006 ) is a multi-dimensional Monte-Carlo RT code for SN spectral and light-curve simulation. Physics processes, including (1) gamma-ray emission from radioactive elements; (2) gamma-ray bound-free transition, Compton scattering, and photoelectric scattering; (3) optical photon bound-bound, bound-free, free-free transitions, are included in SEDONA simulation. An NLTE solver in SEDONA can be turned on for individual elements to determine the level populations in statistical equilibrium, but the computation cost is increased by a factor of \sim 100 relative to LTE. This is a computation bottleneck for high-precision 3-D SN RT simulations. Inspired by the progress in machine learning accelerated computational fluid dynamics (e.g., ( Kochkov et al. 2021 ) ), we propose an artificial intelligence (AI) method to accelerate the NLTE calculation. In addition to the IBT algorithm ( Chen et al. 2024b ) developed in our previous research, we present SEDONA-GesaRaT (the Gesamtkunstwerk of Radiative Transfer developed on SEDONA), an AI-accelerated RT program that is capable of 3-D NLTE SN RT simulations based on SEDONA ( Kasen et al. 2006 ) . Section II discusses the results of SNe Ia 1-D NLTE RT simulations, which serve as the training and testing data set for the AI. Section III tests the SEDONA-GesaRaT code on 3-D time-dependent RT simulations, and presents the first 3-D RT simulation result with both NLTE processes and spatially-resolved spectropolarimetry data cubes, using the SN Ia explosion model N100 ( Seitenzahl et al. 2013 ) . Section IV gives the conclusion. II NLTE Radiative Transfer Figure 1: The spectropolarimetry time sequence of the 3-D N100 model at the viewing direction \mu=0.5,\ \varphi=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 wavelength flux from 2000\ \rm{\AA} to 10000\ \rm{\AA} , the second panel shows the infrared wavelength flux from 10000\ \rm{\AA} to 50000\ \rm{\AA} , the third panel shows the linear polarization percentage Q/I , the fourth panel shows the linear polarization percentage U/I . The x axes in the third and the fourth panels are in logarithmic scale for better presentation. We downloaded the 119 SN Ia ejecta structures from the Heidelberg Supernova Model Archive (HESMA) 1 1 1 https://hesma.h-its.org/ ( Kromer et al. 2017 ) , then used SEDONA ( Kasen et al. 2006 ) to conduct a 1-D time-dependent RT simulation of all the ejecta structures. Lists of the ejecta structures are shown in Appendix A . In the 1-D time-dependent RT simulation, we use two sets of atomic data libraries: CMF_All and CMF_100. The atomic data library CMF_100 is reduced from the CMFGEN 2 2 2 https://sites.pitt.edu/ \sim hillier/web/CMFGEN.htm ( Hillier & Miller 1998 ) database to limit the data for each ion to no more than the 100 lowest lying levels; the total number of levels is 28,105, and the total number of transitions is 274,128. CMF_All includes all of the atomic levels and transitions from the CMFGEN database, but the Si, S, and Ca elements are kept the same as CMF_100 in order to control the NLTE atomic physics computation time; this amounts to, for all ionic species, 497,424 levels and 35,344,426 bound-bound transitions. The forbidden lines (e.g., Ca II 7291 \rm{\AA} , Ca II

Read the complete text