⚛️ general relativity

Scalable and sequential inference of the neutron star equation of state with the Einstein Telescope

This paper introduces a GPU-accelerated hybrid sequential Monte Carlo algorithm that enables efficient, real-time hierarchical inference of the neutron star equation of state by adaptively updating results from thousands of binary neutron star mergers without restarting from scratch, making it scalable for future gravitational-wave observatories like the Einstein Telescope.

Original authors: Thibeau Wouters, Thomas C. K. Ng, Hauke Koehn, Fabian Gittins, Peter T. H. Pang, Tim Dietrich, Chris Van Den Broeck

Published 2026-10-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Thibeau Wouters, Thomas C. K. Ng, Hauke Koehn, Fabian Gittins, Peter T. H. Pang, Tim Dietrich, Chris Van Den Broeck

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 in the heart of the universe, matter exists in a state so dense and strange that it defies our everyday understanding. When a star collapses under its own gravity, it can become a neutron star, a city-sized sphere containing more mass than our entire Sun, packed so tightly that a single teaspoon of its material would weigh a billion tons. Inside these cosmic objects, the rules of physics are pushed to their absolute limit. To understand what happens to matter under such crushing pressure, scientists study the "equation of state," a description of how matter behaves at these extreme densities. For decades, researchers have tried to piece together this description using radio waves from pulsars and X-rays from hot surfaces, but a new tool has emerged: gravitational waves. These are ripples in the fabric of space-time itself, created when two neutron stars collide. As they spiral together and smash into one another, they leave a subtle imprint on these waves, revealing the internal structure of the stars. The next generation of detectors, such as the Einstein Telescope, promises to hear these collisions not just a few times a year, but tens of thousands of times, turning gravitational waves into a dominant way to study the building blocks of the universe.

However, a massive new challenge looms on the horizon. While future telescopes will detect an unprecedented flood of these cosmic crashes, the computer methods used to analyze the data are not ready for such volume. Traditionally, scientists analyze each event one by one and then combine the results to learn about the population as a whole. This process is computationally expensive, and if a new event arrives, the old methods often require starting the entire calculation from scratch. With tens of thousands of events expected annually, restarting the analysis for every new discovery would take far too long, making real-time understanding impossible. The problem is not just the sheer number of signals, but the need to update our knowledge continuously as new data arrives, without wasting time re-doing work that has already been done.

To solve this, a team of researchers has developed a new way to process this incoming stream of information. They created a method that allows computers to learn from a growing list of events by building directly on what they already know, rather than starting over. Imagine a student who, instead of re-reading an entire textbook every time they learn a new fact, simply updates their notes with the new information. The researchers combined two advanced statistical techniques to achieve this. First, they used a powerful type of artificial intelligence to summarize the details of each individual collision into a compact form that is easy to work with. Second, they employed a sequential algorithm that can add new data to the existing picture in small, manageable batches. This approach allows the computer to adaptively decide when a new group of events is significant enough to change the overall understanding of the neutron star equation of state, and when the new data is too quiet to warrant a major update.

The team tested this system by simulating a month of observations from the future Einstein Telescope. They generated data for roughly 1,500 binary neutron star mergers, mimicking the signals the telescope would detect. Running their new algorithm on a single high-performance graphics processor, the system processed the entire month's worth of data in just a few hours. In contrast, traditional methods would struggle to handle even a fraction of this volume in a reasonable timeframe. The researchers found that the system worked efficiently by grouping events: when a new batch of signals was very informative, the computer performed a detailed update; when the signals were less impactful, it simply added them to a waiting queue until enough accumulated to justify a new calculation. This adaptive strategy meant the computer spent its energy only where it was needed most.

The results showed that the method could successfully reconstruct the properties of the neutron star matter. By the end of the simulated month, the algorithm had narrowed down the possible behaviors of the dense matter to a very tight range, providing constraints on the tidal deformability—the measure of how easily the stars can be squashed—of about one percent. While the simulated data contained some noise that caused slight deviations in the final answer, the researchers confirmed that in a cleaner environment, the recovery of the true physical properties was even more accurate. They also projected that with current hardware, the same system could process a full year of detections, involving around 30,000 events, in just a few days. This demonstrates that the bottleneck of analyzing massive catalogs is not a fundamental limit of physics, but a solvable engineering problem.

This work establishes a practical path forward for the era of next-generation gravitational-wave astronomy. By proving that sequential inference is possible and efficient, the researchers have shown that the scientific community can keep pace with the deluge of data expected from the Einstein Telescope. The method is not limited to neutron stars; it can be applied to other questions about the universe, such as the distribution of black holes or the expansion of the cosmos itself. As the detectors come online and begin their daily hunt for cosmic collisions, this new approach ensures that every new discovery will be immediately integrated into our growing understanding of the universe, turning a flood of data into a clear and continuous stream of knowledge.

Technical Summary: Scalable and Sequential Inference of the Neutron Star Equation of State with the Einstein Telescope

Problem Statement
Future gravitational-wave (GW) observatories, specifically the Einstein Telescope (ET) and Cosmic Explorer, are projected to detect tens of thousands of binary neutron star (BNS) mergers annually. This unprecedented data volume presents a critical bottleneck for current hierarchical inference methods used to constrain the neutron star equation of state (EOS). Existing approaches typically require restarting the inference process from a wide prior distribution every time a new event is added to the catalog. This "restart-from-scratch" methodology is computationally prohibitive at the scale of next-generation detectors, where daily detection rates will reach tens to hundreds. Furthermore, standard hierarchical inference involves solving the Tolman-Oppenheimer-Volkoff (TOV) equations repeatedly for every EOS draw, a process that is inherently expensive.

Methodology
The authors propose a hybrid Sequential Monte Carlo (SMC) algorithm designed for scalable, real-time hierarchical inference. The method integrates three primary technical components:

Hybrid Tempering Strategy: The core innovation is a combination of data tempering and likelihood tempering.
- Data Tempering: Instead of processing events one by one (which risks a sharp drop in Effective Sample Size, or ESS, if an event is highly informative) or all at once, the algorithm accumulates new events in a "queue." It continuously reweights the current particle set with new single-event likelihoods.
- Adaptive Batching: The algorithm monitors the ESS of the weighted particle set. If the ESS remains above a user-defined threshold (indicating the new data has not significantly altered the posterior), the event is simply added to the queue.
- Likelihood Tempering Update: Once the ESS drops below the threshold (indicating the batch of accumulated events is sufficiently informative), the algorithm triggers an "update step." It uses likelihood tempering (raising the likelihood to an inverse temperature $\beta_t$ ) to bridge the gap between the current partial posterior (conditioned on $n$ events) and the new partial posterior (conditioned on $n+q$ events). This avoids the need to restart from the prior.
GPU Acceleration and Normalizing Flows:
- TOV Solving: Rather than using machine learning surrogates to approximate TOV solutions (which introduces training costs and reduces flexibility), the authors utilize a GPU-native TOV solver.
- Hyperlikelihood Evaluation: To handle the integration over single-event posteriors required for the hierarchical likelihood, the authors employ normalizing flows (neural density estimators). These flows are trained on the posterior samples of individual events to create tractable hyperlikelihoods that can be evaluated at arbitrary points in the parameter space. This allows for high-throughput evaluation on GPUs, even with thousands of events.
EOS Parametrization: The inference samples a flexible EOS parametrization combining a nucleonic metamodel (for densities up to a break point) and an agnostic speed-of-sound extension scheme (for higher densities), totaling 26 dimensions.

Key Results
The authors validated their method using a simulated dataset representing one month of observations by the ET (2L-misaligned configuration).

Efficiency: The algorithm processed 1,579 detected BNS sources (1,555 after convergence checks) in approximately 2 hours on a single NVIDIA H100 GPU.
Scalability Projection: Based on the runtime scaling observed (where cost becomes memory-bound after ~~400 events), the authors project that processing a full year of realistic ET detections (~~30,000 events) would take only a few days on current hardware.
Inference Quality: The method successfully constrained the EOS and the mass-tidal deformability relationship. While the recovered EOS showed some underestimation of pressure at low densities and overestimation at intermediate densities (partially attributed to the specific fiducial EOS features and detector noise), the tidal deformability curve remained within the 95% credible band.
Noise Sensitivity: Comparisons with zero-noise injections showed that the algorithm performs more smoothly without noise, requiring fewer update steps and yielding a median tidal deformability curve that aligns more closely with the injected truth. This highlights that noise, even in large catalogs, introduces biases that the inference must navigate.

Significance and Claims
The paper establishes Sequential Monte Carlo as a practical and necessary route for real-time hierarchical inference in the era of third-generation GW detectors. The authors claim that their hybrid approach solves the dual challenges of scalability (handling the growing catalog size) and sequentiality (updating posteriors without restarting). By reusing previous inference results and adaptively determining when to update the posterior based on the information content of new data batches, the method enables the analysis of massive datasets that would otherwise be computationally infeasible.

The work demonstrates that with current hardware, the community can move from analyzing individual events to continuously refining population properties (such as the EOS, mass distributions, and cosmology) as data arrives. The authors note that while their current implementation uses simplified settings for single-event analysis (e.g., TaylorF2 waveforms, 20 Hz cutoff), the hierarchical framework itself makes no simplifications in how single-event posteriors are incorporated, suggesting the method is robust for future application to more realistic, long-duration signals and advanced waveform models.

Technical Summary: Scalable and Sequential Inference of the Neutron Star Equation of State with the Einstein Telescope

More like this