⚛️ nuclear theory

NS-UNO: Neutron Star EoS Inference from an Unconstrained Number of Observations

This paper introduces NS-UNO, a flexible Neural Posterior Estimation framework that combines hierarchical DeepSets and conditional normalizing flows to infer the neutron star equation of state from an unconstrained number of mass-radius observations, demonstrating robust performance across varying dataset sizes, precisions, and equation-of-state families.

Original authors: Valéria Carvalho, Márcio Ferreira, Michał Bejger, Constança Providência

Published 2026-09-01

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Valéria Carvalho, Márcio Ferreira, Michał Bejger, Constança Providência

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Deep inside the universe, matter is crushed to densities that cannot be recreated on Earth. In the cores of neutron stars, a single teaspoon of material would weigh as much as a mountain, yet we cannot touch it or measure it directly. To understand what happens under such extreme pressure, scientists study the "equation of state." This is simply a set of rules describing how matter behaves when squeezed, dictating how heavy a star can get before it collapses and how big it can be for a given weight. For decades, researchers have tried to pin down these rules by combining theoretical models with observations from telescopes and gravitational wave detectors. However, a major hurdle has always remained: the number of observations is constantly changing. As new telescopes come online, the amount of data grows, and the precision of each measurement improves. Traditional methods for analyzing this data struggle to adapt when the number of sources varies, often requiring scientists to start their calculations over from scratch every time a new star is discovered.

A team of researchers has now introduced a new approach called NS-UNO, designed to handle an unlimited and changing number of observations without losing the detailed information contained in each one. Instead of treating every new star as a single point of data, this system learns from the full range of possibilities for each measurement, much like looking at a cloud of points that represents the uncertainty of a single observation. The researchers trained a sophisticated computer model on a vast library of theoretical star behaviors, teaching it to recognize patterns in how mass and radius relate to the underlying physics of dense matter. The model was exposed to two very different ways of describing matter: one based on fixed mathematical formulas and another that is completely flexible and data-driven. By learning from both, the system became robust enough to infer the properties of matter even when faced with star types it had never seen before.

The results show that this new framework can accurately reconstruct the rules governing dense matter using sets of observations that vary in size. When the researchers tested the model, it successfully identified the correct physical rules for stars, even when the input data came from different theoretical families. The accuracy of the reconstruction depended heavily on the variety of the stars being observed. If the data only included low-mass stars, the model could only determine the rules for lower densities. If the data included only heavy stars, it could only see the high-density rules. The most accurate picture emerged when the observations covered a wide range of masses, from light to heavy, because this provided a complete map of how matter behaves under different levels of compression. The study found that having more observations was helpful, but having observations that spanned a broader range of masses was far more valuable for understanding the full picture.

The researchers also tested how well their model would handle the messy, complex data coming from real-world instruments like NICER and the gravitational wave event GW170817. While the model was trained on simplified, idealized data, it responded sensibly to the real observations, assigning higher likelihood to star models that matched the current measurements and lower likelihood to those that did not. This suggests the system is ready to be used as a flexible tool for future astronomy, capable of digesting the growing flood of data from next-generation telescopes without needing to be reprogrammed for every new discovery. The work demonstrates that it is possible to build a single, adaptable system that learns from a variable number of sources, preserving the rich detail of each measurement to reveal the hidden nature of the universe's densest matter.

Technical Summary: NS-UNO

Problem Statement
Understanding the Equation of State (EoS) of dense matter in neutron stars (NS) is a central challenge in nuclear physics and astrophysics. While multimessenger observations (gravitational waves and X-ray pulse profiles) have significantly constrained the EoS, traditional Bayesian inference methods relying on Markov Chain Monte Carlo (MCMC) or nested sampling face scalability issues. These methods require recomputing the posterior distribution whenever new observations are added or different EoS parametrizations are considered, which becomes computationally prohibitive as future observatories (e.g., Einstein Telescope, Cosmic Explorer) increase the volume and diversity of data. Furthermore, existing Neural Posterior Estimation (NPE) approaches for NS EoS inference have typically assumed a fixed number of observations, each represented by a single point estimate. They are not designed to handle hierarchical observational datasets where both the number of sources and the number of posterior samples per source vary.

Methodology
The authors introduce NS-UNO (Neutron Star EoS inference from an Unconstrained Number of Observations), a Simulation-Based Inference (SBI) framework using Neural Posterior Estimation. The architecture is designed to accept an arbitrary number of NS observations, where each observation is represented by a set of posterior samples (e.g., mass-radius pairs) rather than a single point.

Hierarchical DeepSets Architecture:
- Level 1 (Within-observation pooling): Each observation (a set of $N_S$ samples) is processed by an embedding network ( $\phi_1$ ). The embedded samples are aggregated using masked mean pooling and masked second-moment pooling to capture first- and second-order statistics. This creates a permutation-invariant representation for each observation.
- Level 2 (Observation aggregation): The representations of individual observations are projected into a higher-dimensional space ( $\phi_2$ ) and aggregated using an attention pooling mechanism. This assigns scalar scores to each observation, generating a weighted average context vector. This step ensures the model is permutation-invariant regarding the order of observations and produces a fixed-dimensional context vector ( $c_i$ ) regardless of the number of observations ( $N_{obs}$ ).
Conditional Normalizing Flow (CNF):
- The fixed-dimensional context vector conditions a CNF, which learns the posterior distribution of the EoS parameters ( $\theta$ ).
- The EoS is represented non-parametrically by the pressure values at 20 fixed baryon densities.
- The flow uses rational quadratic spline coupling transforms to map a base distribution to the target posterior.
- Loss Function: The training minimizes the negative log-likelihood of the posterior density, augmented with a monotonicity regularization term ( $L_{mono}$ ) to enforce physical consistency (pressure must not decrease with increasing density).
Training Data:
- The model is trained on a merged dataset of Piecewise-Polytropic (PT) and Gaussian Process (GP) EoS ensembles. This combination exposes the network to both parametric and non-parametric descriptions of dense matter.
- Mock Observations: Synthetic datasets are generated with variable numbers of observations ( $N_{obs} \in [5, 40]$ ) and variable mass distributions. Each observation is represented by 300 synthetic samples drawn from a bivariate Gaussian distribution centered on the true TOV solution, with randomly sampled uncertainties.

Key Results

Reconstruction Accuracy: The model achieves accurate posterior reconstructions for both PT and GP test sets. The median absolute relative residual is 7.28% for GP and 11.86% for PT over the full density range. Accuracy is highest within the density regimes directly constrained by the observations (up to the maximum central density of the observed stars) and degrades at higher densities where data is absent.
Calibration: The inferred posteriors are well-calibrated, with Coverage Calibration Error (CCE) typically below 10% across the density range.
Impact of Observational Coverage:
- Mass Range vs. Number: The mass range covered by observations has a significantly larger impact on reconstruction accuracy than the sheer number of observations. Observations spanning a broad mass range (e.g., $1.0 - M_{max} M_\odot$ ) provide complementary constraints on different density regimes, whereas increasing the number of observations within a narrow mass range yields diminishing returns.
- Uncertainty Robustness: The model remains robust to variations in the number and precision of observations.
Generalization: The framework successfully generalizes to EoS families not seen during training, including various Relativistic Mean Field (RMF) models (e.g., SLy4, DD2, NL3) and hyperonic extensions (DD2Y, SFHoY). While reconstruction quality varies for models with extreme $M(R)$ relations, the framework captures the underlying relationship between observables and the EoS.
Application to Real Data: The authors qualitatively applied the trained model to current NICER and GW170817 observations. While the model was not trained on the specific non-Gaussian posteriors of real data, it successfully assigned different posterior densities to candidate theoretical EoSs in a manner consistent with their compatibility with current observational constraints.

Significance and Claims
The paper claims that NS-UNO is, to the best of the authors' knowledge, the first NPE framework capable of inferring the NS EoS from an unconstrained number of observations while preserving the posterior information (samples) associated with each measurement.

Scalability: By utilizing a hierarchical DeepSets architecture, the framework naturally accommodates the variable size and diversity of future multimessenger datasets without requiring retraining or recomputation of the posterior for new data configurations.
Model Independence: By training on both PT and GP ensembles, the framework reduces dependence on specific EoS parametrizations.
Methodological Proof of Concept: The authors position this work as a methodological proof of concept. They explicitly note limitations, such as the use of simplified Gaussian uncertainties in training (which differ from the complex structures of current NICER/LVK posteriors) and the inability to reliably reconstruct EoSs with phase transitions given the current training data. They state that quantitative application to real data will require training on more realistic uncertainty models in future work.

In summary, NS-UNO provides a flexible, scalable, and physically consistent approach to NS EoS inference, designed to meet the challenges of the increasingly diverse and numerous observational datasets expected from next-generation multimessenger astronomy.

Technical Summary: NS-UNO

More like this