Employing deep-learning techniques for the conservative-to-primitive recovery in binary neutron star simulations
Ranjith Mudimadugula UP, Federico Schianchi UP,UIB, Anna Neuweiler UP, Thibeau Wouters GRASP,Nikhef, Henrique Gieg UP, Tim Dietrich UP,AEI · 2025
Read Employing deep-learning techniques for the conservative-to-primitive recovery in binary neutron star simulations by Ranjith Mudimadugula UP, Federico Schianchi UP,UIB, Anna Neuweiler UP, Thibeau Wouters GRASP,Nikhef, Henrique Gieg UP, Tim Dietrich UP,AEI (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-2503-08289.txt.
Introduction
Ranjith Mudimadugula and co-authors test, in twenty twenty-five, neural networks inside simulations of merging neutron stars. Neutron star mergers shed light on how heavy elements form and on the state of ultra-dense matter. Simulations are costly and need big computers. One costly step converts conserved variables, the tracked bookkeeping values, into physical values such as pressure and density. The team asks if a neural network can do that step.
They ran what they report as the first merger simulations that use such a method for this step. The runs stayed stable and matched classic methods in accuracy. Total cost was comparable. The authors call it a proof of principle, and say more speed gains are needed before the method pays off.
Catalog record
| CREATOR | Ranjith Mudimadugula UP, Federico Schianchi UP,UIB, Anna Neuweiler UP, Thibeau Wouters GRASP,Nikhef, Henrique Gieg UP, Tim Dietrich UP,AEI |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 40 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-08289
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Employing deep-learning techniques for the conservative-to-primitive recovery in binary neutron star simulations Ranjith Mudimadugula UP, Federico Schianchi UP,UIB, Anna Neuweiler UP, Thibeau Wouters GRASP,Nikhef, Henrique Gieg UP, Tim Dietrich UP,AEI Abstract The detection of GW170817, together with its electromagnetic counterparts, has proven that binary neutron star mergers are of central importance to the field of nuclear astrophysics, e.g., through a better understanding of the formation of elements and novel constraints on the supranuclear dense equation of state governing the matter inside neutron stars. Essential for understanding the binary coalescence are numerical-relativity simulations, which typically come with high computational costs requiring high-performance computing facilities. In this work, we build on recent studies to investigate whether novel techniques, such as neural networks, can be employed in the conversion of conservative variables to primitive hydrodynamical variables, such as pressure and density. In this regard, we perform – to the best of our knowledge – the first binary neutron star merger simulations in which such methods are employed. We show that this method results in stable simulations, reaching accuracies similar to traditional methods with an overall comparable computational cost. These simulations serve as a proof of principle that, in the future, deep learning techniques could be used within numerical-relativity simulations. However, further improvements are necessary to offer a computational advantage compared to traditional methods. ∎ Journal: Eur. Phys. J. A 1 Introduction The multi-messenger detection of GW170817 Abbott, 2017 , GRB170817A Goldstein, 2017 , and AT2017gfo Coulter, 2017 has been a breakthrough in the field of multi-messenger astronomy by incorporating gravitational waves as an observational window. This groundbreaking event, associated with the merger of two neutron stars, has already provided significant insights into the evolution of the Universe Abbott, 2017a ; Hotokezaka et al., 2019 ; Nakar & Piran, 2021 ; Coughlin et al., 2020 ; Dietrich et al., 2020 , the properties of matter at supranuclear densities Abbott, 2017 ; Dietrich et al., 2020 ; Annala et al., 2018 ; Bauswein et al., 2017 ; Fattoyev et al., 2018 ; Ruiz et al., 2018 ; Radice et al., 2018 ; De et al., 2018 ; Most et al., 2018 ; Capano et al., 2020 ; Abbott, 2019 ; Abbott, 2018 ; Coughlin, 2018 ; Coughlin et al., 2019 ; Huth, 2022 , and the processes responsible for the formation of heavy elements Cowperthwaite, 2017 ; Smartt, 2017 ; Kasliwal, 2017 ; Kasen et al., 2017 ; Watson, 2019 ; Rosswog et al., 2018 ; Abbott, 2017b . To extract information from such high-energetic astrophysical observations, one has to compare the measured data with predictions of the binary evolution. Regarding gravitational-wave (GW) astronomy, this can be done by correlating the measured strain signal with existing GW models, maximizing their agreement Veitch, 2015 . For an interpretation of possible electromagnetic (EM) counterparts, one needs to relate the properties of the observed signals with the predicted light curves and spectra caused by the material outflow from two neutron stars during the coalescence Radice et al., 2018a ; Waxman et al., 2018 . Because of the complexity of the merger process, an accurate description of the system’s dynamics requires us to solve Einstein’s field equations together with the equations of general relativistic hydrodynamics. For this reason, we have to perform numerical-relativity (NR) simulations to understand the binary neutron star (BNS) merger dynamics. In general, such simulations also allow us to extract the emitted GW signal and to validate/calibrate existing GW approximants, e.g., Dietrich et al., 2017 ; Dietrich & Hinderer, 2017 ; Kawaguchi et al., 2018 ; Abac et al., 2024 ; Gamba, 2023 . In addition, they also allow us to understand the properties of the outflowing material Radice et al., 2018a ; Hotokezaka et al., 2013 ; Nedora et al., 2022 . Regarding the latter, one requires a proper description of microphysical processes. This implies that one has to use a nuclear-physics motivated equation of state (EOS) for the description of the supranuclear dense material, incorporate a description of neutrino radiation and of magnetic fields; cf. e.g., Rosswog & Liebendoerfer, 2003 ; Liebling et al., 2010 ; Sekiguchi et al., 2011 ; Abdikamalov et al., 2012 ; M“”osta et al., 2014 ; Galeazzi et al., 2013 ; Giacomazzo et al., 2015 ; Kiuchi et al., 2015 ; Foucart, 2018 ; Kiuchi et al., 2018 ; Foucart et al., 2018 ; Palenzuela et al., 2018 ; Foucart et al., 2020 ; Radice et al., 2022 ; Schianchi et al., 2024 for such efforts. Unfortunately, even by employing approximations for the microphysical description, NR simulations come with high computational costs, and single simulations require between tens of thousands to several millions of CPU hours, depending on the incorporated physical descriptions and the employed resolution. Hence, we can typically only investigate the last few orbits and the first tens to hundreds of milliseconds after the merger; cf. Kiuchi et al., 2023 ; Ng et al., 2024 for recent state-of-the-art simulations covering up to one second. This large computational footprint motivates the continuous effort within the NR community to increase the efficiency of NR simulations. While there has been progress, e.g., through the application of new discretization schemes, e.g., Bugner et al., 2016 ; Kidder, 2017 ; Deppe et al., 2022 ; Adhikari et al., 2025 , or the usage of GPUs, e.g., Shankar et al., 2023 ; Zhu et al., 2024 , some works also investigated possibilities to speed up individual routines such as the conservative-to-primitive recovery Galeazzi et al., 2013 ; Kastaun et al., 2021 ; Dieselhorst et al., 2021 . The conversion from the evolved conservative variables to the primitive variables is a critical ingredient within the often employed Valencia formulation of general-relativistic hydrodynamics Font, 2008 ; Anton et al., 2006 . In this article, we want to build on the previous works of Refs. Dieselhorst et al., 2021 ; Kacmaz et al., 2024 ; Wouters, 2024 and present, to our knowledge, the first general-relativistic simulations in which deep-learning techniques are employed to speed up the conservative-to-primitive recovery during BNS simulations. The article is structured as follows. In Sec. 2 we will review the basic methods and also discuss the constructed neural networks (NNs). In Sec. 3 we perform validation tests of our new method and in Sec. 4 we present results of BNS simulations. Throughout this article, we will use geometric units G=c=M_{\odot}=1 for all our calculations, which means that one code unit for lengths refers to \simeq 1.48\rm km , for times refer to \simeq 5\mu s , and for masses refer to 1M_{\odot} . 2 Numerical Methods 2.1 General relativistic hydrodynamics simulations The evolution equations to describe general-relativistic hydrodynamics for a perfect fluid, neglecting the influence of magnetic fields or neutrino radiation, can be derived from the conservation laws of the baryon number and the energy-momentum: \nabla_{\mu}\left(\rho u^{\mu}\right)=0,\hskip 14.22636pt\nabla_{\mu}T^{\mu\nu}=0. (1) To write the evolution system into the form of a balance law \frac{\partial\textbf{q}}{\partial t}+\frac{\partial\textbf{F}^{i}}{\partial x^{i}}=\textbf{s}, (2) one typically transforms the primitive variables \textbf{w}=\left(\rho,v^{i},\epsilon,p\right) , corresponding to the rest-mass energy density \rho , the fluid velocity v^{i} , the specific internal energy \epsilon , and the pressure p , to the following conservative variables q : \displaystyle D= \displaystyle\sqrt{\gamma}\rho W, (3) \displaystyle S_{j}= \displaystyle\sqrt{\gamma}\rho