Neural Network Engineering: How Machine Learning Decodes the Equation of State of Neutron Stars ✦
A physically informed regularized conditional invertible neural network (cINN) has been developed, directly mapping posterior mass–radius distributions of neutron stars to central energy density and pressure, without explicit multidimensional sampling. Regularization ensures causality and thermodynamic stability, providing physically consistent predictions without direct modeling. The method is shown to accurately recover posterior equation-of-state distributions for data similar to NICER observations, preserving the connection between macro- and micro-properties of dense matter. A systematic optimization of 62,400 simulated mass–radius measurements was carried out to find the most informative targets. It was found that the constraining power strongly depends on the position on the mass–radius plane; the optimal strategy alternates between compact massive stars and extended stars of intermediate mass, reducing the uncertainty of the inferred equation of state to ∼9%–10% relative to current NICER data.
Context
Understanding the equation of state (EoS) of nuclear matter at densities many times higher than that of atomic nuclei is a key challenge in modern astrophysics. Neutron stars, especially those observed as pulsars in the X-ray range by the NICER mission, serve as the only available laboratories for studying such an exotic state of matter. Measured mass–radius distributions, together with data from gravitational waves from mergers like GW170817, recorded by interferometric observatories LIGO and Virgo, provide invaluable information, but their interpretation requires solving the inverse problem: from macroscopic observables to microscopic EoS. The classical approach, based on directly solving the equations of general relativity — the Tolman–Oppenheimer–Volkoff (TOV) equations, derived from the works of Karl Schwarzschild and developed within the framework of Albert Einstein's theory, — necessitates iterating over thousands of parametrizations and enormous computational costs. This work proposes a revolutionary approach that eliminates these limitations.
Methods
The researchers created a conditional invertible neural network (cINN) — a deep learning architecture capable of bijective mapping between complex distributions. The network was trained on 90,000 equations of state generated by the sound speed interpolation method and their corresponding mass–radius curves obtained by numerically solving the TOV equations. The key innovation is physics-informed regularization (PIR), which actively penalizes any predictions during training that violate fundamental constraints: the speed of sound in matter cannot exceed the speed of light (causality, guaranteed by the principle of constancy of the speed of light) and must be positive (thermodynamic stability). Thanks to the network's invertibility, direct penalties are not only imposed on training data but also on random samples from the latent space, turning training into a joint optimization of accuracy and physical consistency.
Results
The model demonstrated high accuracy in reconstructing posterior distributions of central energy density and pressure from synthetic contours simulating NICER observations. For real data, including contours from four pulsars (PSR J0030+0451, J0437-4715, J0614-3329, J0740+6620) and the GW170817 event, cINN successfully identified a jointly allowed EoS region overlapping with Bayesian analysis results by 84.4% (Jaccard index), while covering the most probable area. A systematic search for optimal future targets among 62,400 hypothetical observations revealed a strategy of alternating compact massive stars and extended intermediate-mass stars — this reduces EoS uncertainty by up to 8.5% relative to the current baseline. Remarkably, even with lower measurement precision (errors of 0.3 solar masses and 2 km), the strategy remains effective, yielding a 9.46% reduction. The maximum mass of non-rotating neutron stars, first estimated by Subrahmanyan Chandrasekhar and later refined, is closely related to the stiffness of the EoS; the results are consistent with modern constraints.
Implications
The developed approach marks a paradigm shift in astrophysical inference. Instead of repeatedly solving the forward problem (EoS → mass–radius), the neural network directly inverts observational constraints, delivering physically consistent probabilistic estimates in seconds. This opens up the possibility for rapid data analysis in the era of multi-wavelength and multi-messenger observations, when each new event — whether an X-ray pulsar or a compact object merger — must immediately fit into the global picture of nuclear matter properties.
Future development
The method can be naturally extended to include additional data types: neutron star astroseismology, X-ray polarization observations, neutrino burst data, and future third-generation gravitational-wave detectors. The flexibility of the cINN architecture allows for incorporating more subtle physical constraints, such as lepton number conservation or chemical composition, as well as directly modeling phase transitions (e.g., to quark matter). The platform can also be adapted for joint analysis of neutron star populations, enabling global constraints on the EoS.
Impact
The results will have a direct impact on planning observational campaigns for NICER, the future STROBE-X observatory, and on rapid alert strategies in multi-messenger astronomy.
Next steps
Immediate next steps include testing the model on more realistic simulations that include systematic instrumental errors and developing active learning methods for automatically selecting the next optimal observation in real time.
Key open problems
The research is directly linked to one of the greatest unsolved problems in physics: determining the equation of state of cold nuclear matter at densities unattainable in terrestrial accelerators, and to the question of the nature of phase transitions in quantum chromodynamics (QCD) — in particular, the possible existence of quark matter in the cores of massive neutron stars.
🎯 The speed of sound at the center of a neutron star can approach the speed of light, and in some models even temporarily exceed it under certain conditions — that's why the causality condition serves as such a strict filter for theoretical EoS.
🎬 In Robert Forward's novel "Dragon's Egg," life on the surface of a neutron star is described, where colossal gravity and extreme densities create unimaginable forms of consciousness. Our work may one day allow us to learn how real such worlds are, by refining the properties of matter beneath their 'crust'.
Key numbers
- maximum reduction in EoS uncertainty: 8.5% (baseline model) or 9.46% (with increased errors)
- number of scanned hypothetical targets: 62,400
- training set of EoS: 90,000
- initial EoS band width: 0.5877 dex
- final EoS band width after 6 iterations: 0.5377 dex