Leveraging differentiable programming in the inverse problem of neutron stars

Thibeau Wouters, Peter T. H. Pang, Hauke Koehn, Henrik Rose, Rahul Somasundaram, Ingo Tews, Tim Dietrich, Chris Van Den Broeck · 2025

Read Leveraging differentiable programming in the inverse problem of neutron stars by Thibeau Wouters, Peter T. H. Pang, Hauke Koehn, Henrik Rose, Rahul Somasundaram, Ingo Tews, Tim Dietrich, Chris Van Den Broeck (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-2504-15893.txt.

Introduction

Dead star cores pack stuff past all earth norms, and their size marks are the sole probe of that stuff's law. Thibeau Wouters and co-authors run the probe back in a 2025 study with sums that can be taught by slope. Slope fed samplers read the full doubt of the law in less than an hour on one card.

No schooled stand in tool is need, and wide dial sets scale clean. A twin slope hunt reads a law straight off a size curve. The frame shows where core lore and star lore clash or kiss. One such read sets the depth at which a stock law tale breaks. Twin law dials that ape each other are shown to bend the hunt.

Catalog record

CREATORThibeau Wouters, Peter T. H. Pang, Hauke Koehn, Henrik Rose, Rahul Somasundaram, Ingo Tews, Tim Dietrich, Chris Van Den Broeck
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE51 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-15893

Browse the full library · About the Tower of Babel

Full text

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

Leveraging differentiable programming in the inverse problem of neutron stars Thibeau Wouters, Peter T. H. Pang, Hauke Koehn, Henrik Rose, Rahul Somasundaram, Ingo Tews, Tim Dietrich, Chris Van Den Broeck Abstract \Acp NS probe the high-density regime of the nuclear EOS ( EOS ). However, inferring the EOS from observations of NS is a computationally challenging task. In this work, we efficiently solve this inverse problem by leveraging differential programming in two ways. First, we enable full Bayesian inference in under one hour of wall time on a GPU by using gradient-based samplers, without requiring pre-trained machine learning emulators. Moreover, we demonstrate efficient scaling to high-dimensional parameter spaces. Second, we introduce a novel gradient-based optimization scheme that recovers the EOS of a given NS mass-radius curve. We demonstrate how our framework can reveal consistencies or tensions between nuclear physics and astrophysics. First, we show how the breakdown density of a metamodel description of the EOS can be determined from NS observations. Second, we demonstrate how degeneracies in EOS modeling using nuclear empirical parameters can influence the inverse problem during gradient-based optimization. Looking ahead, our approach opens up new theoretical studies of the relation between NS properties and the EOS , while effectively tackling the data analysis challenges brought by future detectors. Preprint: LA-UR-25-23486 I Introduction The nuclear EOS remains uncertain and an open problem in physics [ 121 , 152 , 29 , 34 ] . Nuclear experiments only constrain its lower density behavior, i.e., around the nuclear saturation density, {\rho_{\rm{sat}}\equiv 2.7\times 10^{14}\ \rm{g}\ \rm{cm}^{-3}} (or equivalently, a baryon number density of {n_{\rm{sat}}\equiv$0.16\text{\,}{\mathrm{fm}}^{-3}$} ). For instance, experiments measuring the neutron-skin thickness of atomic nuclei constrain the symmetry energy at nuclear saturation [ 17 , 210 , 7 , 178 , 8 , 80 ] , while heavy-ion collisions uniquely probe isospin-symmetric matter above the nuclear saturation density [ 44 , 123 , 100 , 203 ] . In addition, nuclear theory calculations provide constraints on isospin-asymmetric matter up to two times the nuclear saturation density [ 208 , 52 , 12 ] . Matter at the most extreme densities, up to 6-8n_{\rm{sat}} , is only realized inside NS , the remnants of core-collapse supernovae [ 118 ] . Their macroscopic properties, such as mass, radius, quadrupole moment, and tidal deformability, are determined through the TOV ( TOV ) equations, which require the EOS for closure. Consequently, measurements of NS properties can then be used to infer the EOS [ 129 ] , which is referred to as the inverse problem of NS . Several such measurements are now available. For instance, precise radio timing of binary pulsar systems determines the masses of NS [ 94 , 46 ] . Furthermore, pulse-profile modeling of X-ray emissions from hotspots of millisecond pulsars provides joint estimates on the mass and radius of NS [ 222 ] , and has been achieved with the NICER ( NICER ) [ 76 , 23 , 182 , 183 , 141 , 184 , 142 ] . Additionally, the tidal deformability of NS leaves a measurable imprint on GW emitted during the inspiral and merger of BNS ( BNS ) systems. This way, GW170817 [ 2 , 4 ] , the first BNS merger observed by Advanced LIGO [ 1 ] and Advanced Virgo [ 6 ] , provided the first estimates on NS radii derived from GW [ 3 ] . Constraining the EOS with input from nuclear physics and astrophysics is commonly done with Bayesian inference [ 172 , 168 , 10 , 173 , 37 , 31 , 143 , 207 , 125 , 122 , 100 , 21 , 166 , 156 , 20 , 58 , 155 , 60 , 26 , 187 , 27 , 189 , 19 , 98 , 167 , 105 , 229 ] (see Ref. [ 110 ] for a recently compiled overview of EOS constraints). However, current inference methods remain computationally expensive, particularly when directly sampling a set of EOS parameters, as this requires solving the TOV equations on the fly for millions of samples. This computational burden will hinder our ability to extract precise EOS constraints from NS data delivered by future instruments, such as upcoming X-ray telescopes (e.g., eXTP and STROBE-X [ 232 , 221 , 176 ] ) and the next-generation of ground-based GW detectors (i.e., the Einstein Telescope [ 165 , 134 , 61 , 62 ] and Cosmic Explorer [ 63 ] ). For example, next-generation GW observatories are projected to detect approximately \mathcal{O}(10^{5}) BNS mergers annually [ 134 , 161 ] . These detectors will measure NS radii potentially with an unprecedented accuracy of 10-100 meters [ 35 , 164 , 102 , 101 , 219 ] . This increase in both the quantity and quality of observations highlights the need for more efficient inference methods for the inverse problem of NS . Ideally, such methods should be capable of solving the TOV equations on the fly to ensure accurate and robust EOS constraints. One possible solution, thoroughly explored in past works, is to resort to machine-learning surrogates. For instance, previous works have accelerated the TOV solver by replacing it with machine learning-based approximations using regression methods [ 181 , 103 ] , neural networks [ 66 , 209 , 211 , 179 , 135 , 47 , 201 ] , support vector machines [ 66 ] , Gaussian processes and reduced basis methods [ 179 ] , and dynamic mode decomposition [ 115 ] (see Ref. [ 233 ] for a review). Other works, rather than relying on traditional sampling techniques, use machine learning models to directly recover the EOS from NS data, e.g., with neural networks [ 71 , 72 , 146 , 213 , 112 , 73 , 197 , 199 , 64 , 91 , 67 , 32 , 198 , 113 , 228 , 33 , 217 ] , normalizing flows [ 147 , 25 , 96 ] , transformers [ 85 ] , principal component analysis [ 159 ] , and symbolic regression [ 104 , 160 , 158 , 15 ] . However, training these surrogate models can take on the order of hours [ 139 ] to weeks [ 96 ] (depending on the model complexity and hardware used for training) and requires additional modeling efforts. Moreover, we are often interested in comparing inferences with different parameterizations or prior assumptions. Since emulators must be retrained whenever these change, their computational speed-up comes at the cost of reduced flexibility in inference. In this work, we propose an alternative solution to efficiently solve the inverse problem of NS without resorting to machine-learning emulators. In particular, we leverage differential programming and hardware acceleration with GPU to infer the EOS from NS properties in two ways. First, we make Bayesian inference of the EOS from NS data in high-dimensional EOS parameter spaces computationally efficient. Second, we introduce a gradient-based optimization scheme that can recover the parameters of the EOS underpinning a given NS mass-radius curve. While the former method can take observational uncertainties into account, the latter can be used to improve our understanding of how accurate knowledge of NS propagates into constraints on the EOS

Read the complete text