arxiv: 2604.24949 · v1 · submitted 2026-04-27 · 🌌 astro-ph.HE · nucl-th

Recognition: unknown

A Physics Informed Bayesian Neural Network for the Neutron Star Equation of State

J.D. Baker , C.A. Bertulani , R.V. Lobato

Authors on Pith no claims yet

Pith reviewed 2026-05-08 01:34 UTC · model grok-4.3

classification 🌌 astro-ph.HE nucl-th

keywords neutron starsequation of stateBayesian neural networksphysics-informed machine learningNICER observationstidal deformabilityhadronic modelsTolman-Oppenheimer-Volkoff equation

0 comments

The pith

A Bayesian neural network learns neutron-star equations of state from hadronic models and physical constraints.

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

This paper introduces a physics-informed Bayesian neural network to infer the equation of state for neutron stars. The network is trained on a representative set of hadronic equations of state while enforcing constraints like causality and monotonicity. It then propagates uncertainties to predict mass-radius relations and tidal deformabilities that align with observational data from NICER and maximum mass constraints. The approach offers a non-parametric way to connect microphysical uncertainties to stellar observables.

Core claim

The central discovery is a Bayesian neural-network framework that, trained on hadronic EoS ensembles with soft constraints at saturation density and perturbative QCD, learns P(ε) via stochastic variational inference, enforces monotonicity and causality, matches to SLy4 crust, and evolves through TOV-plus-tidal solver to yield posterior predictions consistent with NICER radii and 2.0 solar mass limits, specifically R_{1.4}=12.1^{+1.4}_{-0.9} km, Λ_{1.4}=580^{+520}_{-240}, and M_max ≃ 2.11±0.05 M⊙ at 90% CI.

What carries the argument

Physics-informed Bayesian neural network performing stochastic variational inference on the pressure-energy density relation P(ε), with penalties for monotonicity and causality.

If this is right

The resulting posteriors for radius and tidal deformability at canonical mass can be compared to gravitational wave observations.
The framework provides a flexible, non-parametric mapping from EoS uncertainties to neutron-star observables.
Accepted core EoSs are matched to SLy4 crust and solved via unified Tolman-Oppenheimer-Volkoff equations.
The method allows a posteriori assessment against current gravitational-wave constraints on tidal deformability.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

This neural-network approach could reduce computational cost compared to traditional parametric EoS sampling in future analyses.
If the training ensemble is representative, the model may help identify inconsistencies between theoretical priors and new observations.
Extending the network to include phase transitions or exotic matter could test for deviations from hadronic models.
Integration with real-time gravitational wave data analysis might enable rapid EoS updates.

Load-bearing premise

The chosen ensemble of hadronic EoSs plus the soft constraints at saturation density and from perturbative QCD suffice to produce an unbiased posterior without the network memorizing the training set.

What would settle it

A precise measurement of a 1.4 solar mass neutron star radius falling well outside the range 12.1^{+1.4}_{-0.9} km, or a maximum mass significantly below 2.0 solar masses, would falsify the inferred posterior distribution.

Figures

Figures reproduced from arXiv: 2604.24949 by C.A. Bertulani, J.D. Baker, R.V. Lobato.

**Figure 1.** Figure 1: FIG. 1. ELBO convergence history for the PI-BNN training view at source ↗

**Figure 2.** Figure 2: FIG. 2. Microphysical predictions of the filtered PI-BNN posterior. Left: pressure as a function of energy density, where the view at source ↗

**Figure 3.** Figure 3: FIG. 3. Joint observable-plane predictions derived from the PI-BNN posterior. Left: the mass-radius ( view at source ↗

read the original abstract

We present a physics-informed Bayesian neural-network framework to infer neutron-star equations of state from theoretical priors and to propagate the associated uncertainties to stellar observables. Trained on a large and representative ensemble of hadronic EoSs, the model learns $P(\epsilon)$ via stochastic variational inference, incorporating soft constraints at saturation density and from perturbative QCD, together with penalties enforcing monotonicity and causality. The accepted core EoSs are matched to an SLy4 crust and evolved through a unified Tolman-Oppenheimer-Volkoff-plus-tidal solver to generate posterior predictions in the mass-radius ($M$-$R$) and mass-tidal-deformability ($M$-$\Lambda$) planes. The inferred posterior is consistent with NICER radius measurements and the observed $2.0\,M_\odot$ maximum-mass constraint, yielding $R_{1.4}=12.1^{+1.4}_{-0.9}\,\mathrm{km}$, $\Lambda_{1.4}=580^{+520}_{-240}$, and $M_{\mathrm{max}}\simeq 2.11\pm0.05\,M_\odot$ (90\% CI). The resulting canonical tidal deformability can be assessed \emph{a posteriori} against current gravitational-wave constraints. Overall, this framework provides a flexible, non-parametric mapping from microphysical EoS uncertainties to neutron-star observables.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This BNN framework for neutron star EoS tries a physics-informed variational route to propagate uncertainties but the quoted posteriors rest on untested assumptions about the training ensemble.

read the letter

The key point here is that this Bayesian neural network approach offers a way to map EoS uncertainties to neutron star observables while baking in physical constraints, but the reported results rest on assumptions about the training ensemble that aren't fully tested in the abstract. The new element is the use of stochastic variational inference on a physics-informed BNN that includes soft constraints at saturation density and from pQCD, along with monotonicity and causality penalties. They then match to SLy4 crust and run a unified TOV and tidal solver to get the M-R and M-Lambda posteriors. This specific pipeline appears distinct from prior work. It does well in producing a flexible, non-parametric framework that gives posteriors consistent with current observations, like the NICER radii and 2 solar mass stars, with R_1.4 at 12.1 km plus or minus about 1 km and max mass 2.11. Where it is soft is the missing details on training and validation splits, how the constraint penalties are balanced, and whether the credible intervals cover all uncertainty sources. The risk that the model overfits the finite hadronic EoS training set instead of learning something general is worth checking, as the stress test notes. If the ensemble isn't diverse enough, the outputs could just echo the training statistics. The separation from the TOV solver helps, but the acceptance of core EoSs might still tie back to the training priors. This paper is for astrophysicists and nuclear physicists working on dense matter and neutron star modeling. Someone looking for new tools to handle EoS uncertainties in light of upcoming data would find value, provided the methods hold up. It should go to peer review because the framework is relevant and the central idea is sound enough to warrant checking the implementation and robustness.

Referee Report

3 major / 3 minor

Summary. The manuscript presents a physics-informed Bayesian neural network (BNN) trained via stochastic variational inference on an ensemble of hadronic equations of state (EoSs). The model learns P(ε) while incorporating soft constraints at nuclear saturation density and from perturbative QCD, plus penalties for monotonicity and causality. Accepted core EoSs are matched to an SLy4 crust and integrated through a unified TOV-plus-tidal solver to produce posterior distributions in the M-R and M-Λ planes. The resulting posteriors are reported as consistent with NICER radius measurements and the 2 M⊙ maximum-mass constraint, with specific values R_{1.4}=12.1^{+1.4}_{-0.9} km, Λ_{1.4}=580^{+520}_{-240}, and M_max ≃ 2.11±0.05 M⊙ (90% CI).

Significance. If the BNN produces a genuinely generalizable posterior rather than an interpolation of the training ensemble, the framework offers a flexible non-parametric route to propagate microphysical EoS uncertainties into observable predictions. This could be useful for joint analyses of NICER, gravitational-wave, and future multi-messenger data. The reported consistency with existing constraints is a positive indication, but the overall significance hinges on demonstrating robustness to training-set choice and hyperparameter variation.

major comments (3)

[§3] §3 (Model Architecture and Training): No details are provided on the training/validation split of the hadronic EoS ensemble, the size or diversity metrics of the ensemble, or any cross-validation procedure. This is load-bearing for the central claim because the quoted credible intervals (e.g., for R_{1.4} and Λ_{1.4}) may not fully account for epistemic uncertainty if the variational approximation collapses or the network memorizes training-set features.
[§3.3] §3.3 (Loss Function and Constraints): The balancing of the soft saturation-density and pQCD constraint penalties against the variational objective is not quantified or ablated. Different relative weights could shift the accepted core EoS distribution and therefore the final M_max and radius posteriors, undermining the claim that the reported values are robust.
[Results] Results section (posterior propagation): The credible intervals for R_{1.4}, Λ_{1.4}, and M_max are presented without explicit decomposition into contributions from the BNN variational parameters, hyperparameter choices, and the TOV integration. This makes it impossible to assess whether the intervals capture all relevant uncertainties or primarily reflect the statistics of the finite training ensemble after SLy4 crust matching.

minor comments (3)

[Abstract] The abstract states that the model is 'trained on a large and representative ensemble' but provides no numerical size, composition, or diversity measure for that ensemble.
[§3] Notation for the neural-network output (P(ε)) and the precise form of the monotonicity/causality penalty terms should be defined more explicitly before the results are presented.
[Figures] Figure captions for the M-R and M-Λ posterior plots should include the exact number of accepted EoS samples used to generate the contours.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed report. We address each major comment below and have revised the manuscript accordingly to strengthen the presentation of training details, constraint robustness, and uncertainty quantification.

read point-by-point responses

Referee: §3 (Model Architecture and Training): No details are provided on the training/validation split of the hadronic EoS ensemble, the size or diversity metrics of the ensemble, or any cross-validation procedure. This is load-bearing for the central claim because the quoted credible intervals (e.g., for R_{1.4} and Λ_{1.4}) may not fully account for epistemic uncertainty if the variational approximation collapses or the network memorizes training-set features.

Authors: We agree that these procedural details are required to evaluate generalization and epistemic uncertainty. The original text described the ensemble only as 'large and representative.' In the revised §3 we now specify: an ensemble of 5000 hadronic EoS models spanning a broad range of nuclear saturation parameters; an 80/20 train/validation split; quantitative diversity metrics (e.g., coverage of K_0, J, and L); and 5-fold cross-validation results showing stable ELBO convergence and no evidence of memorization. These additions allow readers to assess whether the reported credible intervals properly reflect epistemic uncertainty. revision: yes
Referee: §3.3 (Loss Function and Constraints): The balancing of the soft saturation-density and pQCD constraint penalties against the variational objective is not quantified or ablated. Different relative weights could shift the accepted core EoS distribution and therefore the final M_max and radius posteriors, undermining the claim that the reported values are robust.

Authors: We acknowledge that an explicit ablation of the constraint weights was missing. In the revised §3.3 we have added a sensitivity study in which the saturation-density and pQCD penalty coefficients are varied over two orders of magnitude. The resulting shifts in the 90 % credible intervals for R_{1.4} and M_max remain smaller than the quoted uncertainties, and we now report the relative magnitudes of each loss term at convergence. This demonstrates that the quoted posteriors are robust to reasonable choices of the balancing hyperparameters. revision: yes
Referee: Results section (posterior propagation): The credible intervals for R_{1.4}, Λ_{1.4}, and M_max are presented without explicit decomposition into contributions from the BNN variational parameters, hyperparameter choices, and the TOV integration. This makes it impossible to assess whether the intervals capture all relevant uncertainties or primarily reflect the statistics of the finite training ensemble after SLy4 crust matching.

Authors: We agree that a clearer decomposition improves interpretability. The revised Results section now includes a dedicated paragraph that separates the sources: (i) epistemic uncertainty from the variational posterior (sampled via 1000 draws), (ii) sensitivity to hyperparameter choices (tested by varying network depth, learning rate, and prior scale, producing <10 % variation in the intervals), and (iii) the deterministic TOV integration (numerical contribution negligible). We also compare the BNN-generated distribution against the raw training-ensemble statistics to show that the network generalizes rather than merely interpolates. The SLy4 crust matching is fixed and therefore does not contribute additional variance. revision: yes

Circularity Check

0 steps flagged

No circularity detected; derivation uses external training ensemble and independent propagation to observables

full rationale

The paper trains a BNN on an external ensemble of hadronic EoSs via stochastic variational inference, adds soft saturation and pQCD constraints plus monotonicity/causality penalties, matches to SLy4 crust, and evolves through a separate TOV solver to produce M-R and M-Λ posteriors. These posteriors are then compared for consistency against NICER and 2 M⊙ observations. No quoted step shows a prediction reducing to a fitted input by construction, no self-citation load-bearing the central claim, and no ansatz or uniqueness imported from the authors' prior work. The framework remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The central claim rests on the representativeness of the hadronic EoS training ensemble, the adequacy of the chosen soft constraints, and the assumption that the SLy4 crust can be matched without introducing systematic bias; no new particles or forces are postulated.

free parameters (2)

Neural-network architecture and variational parameters
Number of layers, hidden units, and variational distribution parameters are chosen to fit the training ensemble and constraint penalties.
Constraint penalty strengths
Weights on monotonicity, causality, saturation-density, and pQCD soft constraints are set by hand or cross-validation.

axioms (3)

domain assumption Pressure must be a monotonically increasing function of energy density
Enforced as a penalty during training; standard for stable stellar models.
domain assumption Speed of sound must remain below the speed of light
Causality constraint applied as a soft penalty.
domain assumption SLy4 crust model can be matched to the learned core EoS
Standard nuclear-physics choice; no independent justification given in abstract.

pith-pipeline@v0.9.0 · 5545 in / 1448 out tokens · 58034 ms · 2026-05-08T01:34:14.468741+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

54 extracted references · 38 canonical work pages · 1 internal anchor

[1]

Training proceeds by minimizing a loss function (such as Mean Squared Error) over a training dataset using gradient descent algorithms, yielding a single optimal set of weights ˆθ

The Deterministic Limitation In a standard, deterministic ANN, the network parametersθ={W l, bl}are treated as fixed point estimates. Training proceeds by minimizing a loss function (such as Mean Squared Error) over a training dataset using gradient descent algorithms, yielding a single optimal set of weights ˆθ. While a well-trained deterministic ANN can...
[2]

Letf θ(ϵ) represent the neural network parametrized by weights and biasesθ

BNN Architecture To reconstruct the equation of state (EoS) with robust uncertainty quantification, we model the mapping from energy densityϵto pressurePusing a Bayesian Neural Network (BNN). Letf θ(ϵ) represent the neural network parametrized by weights and biasesθ. Unlike standard deterministic networks, we treatθas random variables to capture epistemic...
[3]

We apply a Gaussian penalty centered at ϵsat = 150 MeV/fm3 with a target pressureP sat = 1.5 MeV/fm3 and varianceσ 2 sat

Nuclear Saturation Density: The EoS must pass through the well-constrained empirical saturation point. We apply a Gaussian penalty centered at ϵsat = 150 MeV/fm3 with a target pressureP sat = 1.5 MeV/fm3 and varianceσ 2 sat
[4]

The target pressure is set toP pQCD = 3,333 MeV/fm 3, so that the anchor respects the approach toward the conformal limitP≈ϵ/3

Perturbative QCD (pQCD) Limit: To ensure strict theoretical control over the perturbative expansion, we place the high-density anchor at ϵpQCD = 10,000 MeV/fm 3, corresponding to a baryon density of approximatelyn B ≈40n 0, following the regime identified by Komoltsev and Kurkela as safely perturbative [39]. The target pressure is set toP pQCD = 3,333 MeV...
[5]

Monotonicity Constraint: Monotonicity is not imposed as a hard rejection rule during training; instead, we add a soft derivative penalty on an auxiliary density grid throughpyro.factor, which suppresses negative-slope regions while preserving posterior flexibility. Writing the dense auxiliary grid asϵ mono, the penalty is Lmono =λ mono X ϵi∈ϵmono ReLU δ−c...
[6]

We therefore use post-training checks only as a permissive numerical consistency step, not as the primary mechanism for enforcing physical constraints

Post-Training Consistency Screening Because thermodynamic stability and causality are already encouraged directly in the ELBO, samples drawn from the optimized variational posteriorq ∗ ϕ(θ) are typically physically admissible before any downstream screening. We therefore use post-training checks only as a permissive numerical consistency step, not as the ...
[7]

To model this region, we adopt the tabulated SLy4 equation of state

Crust-Core Stitching The physics of low-mass neutron stars (M≲1.0M ⊙) is dominated by the outer and inner crust, a regime not directly constrained by the BNN training, which is restricted to supranuclear densities. To model this region, we adopt the tabulated SLy4 equation of state
[8]

for the crust. Rather than using stitching as the primary mechanism for imposing monotonicity, we use it to preserve the already softened monotonic behavior of the sampled core EoSs while maintaining an ordered connection to the SLy4 crust. Let (ϵ crust,max, Pcrust,max) denote the terminal point of the SLy4 crust. For each sampled core EoSf (k) θ (ϵ) draw...
[9]

Tidal Deformability and Love Number In addition to the TOV background, we solve the linear even-parity tidal perturbation equations for each stellar model using the same interpolated EoS and adaptive integrator. The unified solver evolves the background variables together with a first-order radial metric response variabley(r) (introduced by Hinderer [40–4...

2000
[10]

Oertel, M

M. Oertel, M. Hempel, T. Kl¨ ahn, and S. Typel, Reviews of Modern Physics89, 015007 (2017), 1610.03361

work page arXiv 2017
[11]

Komoltsev, R

O. Komoltsev, R. Somasundaram, T. Gorda, A. Kurkela, J. Margueron, and I. Tews, Physical Review D109, 094030 (2024), 2312.14127

work page arXiv 2024
[12]

GW170817: Observation of Gravitational Waves from a Binary Neutron Star Inspiral

B. Abbottet al.(LIGO Scientific, Virgo), Phys. Rev. Lett.119, 161101 (2017), 1710.05832

work page internal anchor Pith review arXiv 2017
[13]

GW170817: Measurements of Neutron Star Radii and Equation of State

B. Abbottet al.(LIGO Scientific, Virgo), Physical Review Letters121, 161101 (2018), 1805.11581

work page Pith review arXiv 2018
[14]

T. E. Rileyet al., The Astrophysical Journal887, L21 (2019), 1912.05702

work page arXiv 2019
[15]

T. E. Rileyet al., The Astrophysical Journal Letters918, L27 (2021), 2105.06980

work page arXiv 2021
[16]

Choudhury et al., Astrophys

D. Choudhuryet al., The Astrophysical Journal Letters 971, L20 (2024), 2407.06789

work page arXiv 2024
[17]

Annala, T

E. Annalaet al., Phys. Rev. Lett.120, 172703 (2018), 1711.02644

work page arXiv 2018
[18]

A Massive Pulsar in a Compact Relativistic Binary

J. Antoniadiset al., Science340, 6131 (2013), 1304.6875

work page Pith review arXiv 2013
[19]

H. T. Cromartieet al.(NANOGrav), Nature Astron.4, 72 (2019), 1904.06759

work page arXiv 2019
[20]

Fonseca et al., Astrophys

E. Fonsecaet al., Astrophys. J. Lett.915, L12 (2021), 2104.00880

work page arXiv 2021
[21]

A. W. Steiner, J. M. Lattimer, and E. F. Brown, The Astrophysical Journal722, 33 (2010), 1005.0811

work page Pith review arXiv 2010
[22]

Landry, R

P. Landry, R. Essick, and K. Chatziioannou, Physical Review D101, 123007 (2020), 2003.04880

work page arXiv 2020
[23]

Raaijmakerset al., The Astrophysical Journal Letters 918, L29 (2021), 2105.06981

G. Raaijmakerset al., The Astrophysical Journal Letters 918, L29 (2021), 2105.06981

work page arXiv 2021
[24]

Zhouet al., Progress in Particle and Nuclear Physics 135, 104084 (2024), 2303.15136

K. Zhouet al., Progress in Particle and Nuclear Physics 135, 104084 (2024), 2303.15136

work page arXiv 2024
[25]

Papigkiotiset al., Physical Review D111, 083056 (2025), 2501.18544

G. Papigkiotiset al., Physical Review D111, 083056 (2025), 2501.18544

work page arXiv 2025
[26]

Liodis, E

I. Liodis, E. Smyrniotis, and N. Stergioulas, Physical Review D109, 104008 (2024), 2309.03991

work page arXiv 2024
[27]

Fujimoto, K

Y. Fujimotoet al., Physical Review D110, 034035 (2024), 2401.12688

work page arXiv 2024
[28]

Stergioulas, inCompact Objects in the Universe, edited by E

N. Stergioulas, inCompact Objects in the Universe, edited by E. Papantonopoulos and N. Mavromatos (Cham, 2024) pp. 329–356

2024
[29]

Nget al., Classical and Quantum Gravity42, 205008 (2025), 2507.03232

S. Nget al., Classical and Quantum Gravity42, 205008 (2025), 2507.03232

work page arXiv 2025
[30]

J.-J. Li, Y. Tian, and A. Sedrakian, Physical Review C 111, 055804 (2025), 2502.20000

work page arXiv 2025
[31]

Donget al., Physical Review D112, 063043 (2025), 2509.03069

X. Donget al., Physical Review D112, 063043 (2025), 2509.03069

work page arXiv 2025
[32]

Li and A

J.-J. Li and A. Sedrakian, Physical Review C112, 015802 (2025), 2505.00911

work page arXiv 2025
[33]

J.-J. Li, Y. Tian, and A. Sedrakian, Physics Letters B 865, 139501 (2025), 2412.16513

work page arXiv 2025
[34]

Papigkiotis, G

G. Papigkiotis, G. Vardakas, and N. Stergioulas, Physical Review D113, 023018 (2026), 2508.05850

work page arXiv 2026
[35]

Typelet al.(CompOSE Core Team),58, 221 (2022), 2203.03209

S. Typelet al.(CompOSE Core Team),58, 221 (2022), 2203.03209

work page arXiv 2022
[36]

M. D. Hoffmanet al., Journal of Machine Learning Research14, 1303 (2013)

2013
[37]

Douchin and P

F. Douchin and P. Haensel, Astronomy & Astrophysics 380, 151 (2001), astro-ph/0111092

work page arXiv 2001
[38]

R. C. Tolman, Physical Review55, 364 (1939)

1939
[39]

J. R. Oppenheimer and G. M. Volkoff, Physical Review 55, 374 (1939)

1939
[40]

J. S. Read, B. D. Lackey, B. J. Owen, and J. L. Friedman, Physical Review D79, 124032 (2009), 0812.2163

work page arXiv 2009
[41]

Lindblom, Physical Review D82, 103011 (2010), 1009.0738

L. Lindblom, Physical Review D82, 103011 (2010), 1009.0738

work page arXiv 2010
[42]

E. V. Chimanski, R. V. Lobato, A. R. Goncalves, and C. A. Bertulani, Particles6, 198 (2023), 2205.01174

work page arXiv 2023
[43]

Chatziioannou, H

K. Chatziioannou, H. T. Cromartie, S. Gandolfi, I. Tews, D. Radice, A. W. Steiner, and A. L. Watts, Reviews of Modern Physics97, 045007 (2025), 2407.11153

work page arXiv 2025
[44]

J. M. Lattimer and Y. Lim, The Astrophysical Journal 771, 51 (2013), 1203.4286

work page arXiv 2013
[45]

Margueron, R

J. Margueron, R. Hoffmann Casali, and F. Gulminelli, Physical Review C97, 025805 (2018), 1708.06894

work page arXiv 2018
[46]

Ligo algorithm library - lalsuite,

LIGO Scientific Collaboration, “Ligo algorithm library - lalsuite,” (2018)

2018
[47]

Binghamet al., Journal of Machine Learning Research 20, 1 (2019)

E. Binghamet al., Journal of Machine Learning Research 20, 1 (2019)

2019
[48]

Komoltsev and A

O. Komoltsev and A. Kurkela, Physical Review Letters 128, 202701 (2022), 2111.05350

work page arXiv 2022
[49]

Tidal Love numbers of neutron stars

T. Hinderer, The Astrophysical Journal677, 1216 (2008), 0711.2420

work page Pith review arXiv 2008
[50]

Tidal deformability of neutron stars with realistic equations of state and their gravitational wave signatures in binary inspiral

T. Hinderer, B. D. Lackey, R. N. Lang, and J. S. Read, Physical Review D81, 123016 (2010), 0911.3535

work page Pith review arXiv 2010
[51]

Postnikov, M

S. Postnikov, M. Prakash, and J. M. Lattimer, Physical Review D82, 024016 (2010), 1004.5098

work page arXiv 2010
[52]

PyTorch: An Imperative Style, High-Performance Deep Learning Library

A. Paszkeet al., (2019), 10.48550/arXiv.1912.01703, 1912.01703

work page Pith review doi:10.48550/arxiv.1912.01703 2019
[53]

Annala, T

E. Annalaet al., Nature Physics , 1 (2020), 1903.09121

work page arXiv 2020
[54]

This behavior is consistent with the adopted crust-core stitching procedure and indicates that low-mass configurations are not artificially truncated

The low-mass regime (M≲0.5M ⊙) exhibits a sharp expansion in radius, characteristic of stellar structures dominated by the nuclear envelope. This behavior is consistent with the adopted crust-core stitching procedure and indicates that low-mass configurations are not artificially truncated. For the sake of simplicity we truncated the plot at 0.5M ⊙ since ...