🔭 high-energy astrophysics

The hunt for new pulsating ultraluminous X-ray sources: a clustering approach

This study employs an unsupervised clustering AI approach on XMM-Newton data to identify a sample of 85 new candidate pulsating ultraluminous X-ray sources that share multi-dimensional properties with known pulsating ULXs, highlighting the method's predictive power while underscoring the need for higher-statistics observations to confirm their pulsations.

Original authors: Nicolò Oreste Pinciroli Vago, Roberta Amato, Matteo Imbrogno, GianLuca Israel, Andrea Belfiore, Konstantinos Kovlakas, Piero Fraternali, Mario Pasquato

Published 2025-07-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nicolò Oreste Pinciroli Vago, Roberta Amato, Matteo Imbrogno, GianLuca Israel, Andrea Belfiore, Konstantinos Kovlakas, Piero Fraternali, Mario Pasquato

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

The Big Picture: Hunting for Cosmic Heartbeats

Imagine the universe is a giant, noisy party. Most of the guests (stars and black holes) are just chatting or dancing quietly. But then, there's a special group of guests called Ultraluminous X-ray Sources (ULXs). These are the "superstars" of the party, glowing so brightly in X-rays that they seem to be breaking the laws of physics by eating more food (matter) than they should be able to handle.

For a long time, astronomers thought these superstars were all Black Holes (the heavy, silent eaters). But then, they discovered a few of them were actually Neutron Stars (the dense, pulsing leftovers of exploded stars) that were eating so fast they were glowing brighter than a black hole. These are called PULXs (Pulsating ULXs). They have a "heartbeat"—a rhythmic pulse of light.

The Problem: Finding these heartbeats is like trying to hear a specific drumbeat in a rock concert. You need a lot of data (time and attention) to hear it. Many of these sources are too faint, or we haven't looked at them long enough, so we missed their pulses. We have a list of 2,000 "superstars," but we've only confirmed about 6 of them as having a heartbeat.

The Goal: The authors of this paper wanted to find the other Neutron Stars hiding in the crowd. They didn't want to just listen harder; they wanted to use Artificial Intelligence (AI) to spot the "look-alikes."

The Method: The "Bodyguard" and the "Clustering Party"

The researchers treated the data like a massive guest list for a party. They had 640 different "stars" with about 1,800 different "observations" (snapshots of them over time).

1. The Training (Teaching the AI)

They took the 6 confirmed PULXs (the ones we know have heartbeats) and showed them to the AI. They said, "These are the VIPs. Learn what they look like."

2. The Clustering (Sorting the Guests)

Instead of asking the AI to guess "Is this a PULX? Yes/No," they used a technique called Clustering.

The Analogy: Imagine you have a huge pile of mixed-up socks. You don't know which ones are pairs yet. You use an algorithm to sort them into two piles: Pile A and Pile B.
The AI looked at all the data points (how bright they are, how fast they change, how "hard" their light is) and tried to group them into two distinct clusters.
The goal was to make sure Pile A contained almost all the known VIPs (the confirmed PULXs). If the AI did this well, then Pile B would contain the "unknowns," and Pile A would also contain the new candidates that look exactly like the VIPs, even if we haven't heard their heartbeat yet.

3. The Secret Ingredient: The "Peak Flash"

The researchers tried many different ways to sort the socks. They found that the most important thing to look at was the Maximum Brightness (the brightest flash the star ever made).

The Metaphor: Imagine trying to identify a famous singer in a crowd. You might look at their height, their shoes, or their voice. But the researchers found that if you know the loudest note they ever sang, you can identify them much better than by looking at their shoes.
By including the "peak brightness" in the sorting, the AI became incredibly good at separating the VIPs from the rest.

The Results: Finding the Hidden Twins

The AI did its job and split the guests into two groups.

Group 1 (The VIPs): This group held all 6 known PULXs (100% success rate).
Group 2 (The New Candidates): This group held 85 new stars that the AI said, "Hey, these guys look exactly like the VIPs in every way we can measure, even though we haven't heard their heartbeat yet."

The Catch: The AI is a very good guesser, but it's not a magician.

The researchers took these 85 new candidates and tried to listen for the heartbeat using traditional, very sensitive methods.
Result: They didn't find any new heartbeats yet.
Why? It's likely that the "signal" (the heartbeat) is just too quiet or the "noise" (the background static) is too loud. The stars are there, they look like PULXs, but we need to look at them for a longer time or with better equipment to hear the pulse.

Why This Matters (The "So What?")

This paper is a great example of AI as a detective.

Efficiency: Instead of manually checking 2,000 stars one by one (which takes forever), the AI scanned them all in a multi-dimensional space (looking at brightness, color, variability, etc., all at once) and pointed the astronomers to the 85 most promising suspects.
The Future: These 85 candidates are now the top priority for future telescopes. If we point our most powerful X-ray eyes at them, we might finally hear the heartbeats of dozens of new Neutron Stars.
The Lesson: Just because we can't hear the pulse yet doesn't mean the star isn't a PULX. The AI told us, "Trust me, these guys belong in the VIP section."

Summary in One Sentence

The authors used a smart computer program to sort through thousands of bright X-ray stars, finding 85 new suspects that look exactly like the known "pulsing" neutron stars, giving astronomers a targeted "hit list" of where to look next to confirm their existence.

1. Problem Statement

Ultraluminous X-ray sources (ULXs) are extragalactic sources with luminosities exceeding the Eddington limit of a stellar-mass black hole ( $L_X \gtrsim 10^{39}$ erg s $^{-1}$ ). While initially thought to be intermediate-mass black holes, the discovery of coherent pulsations in several ULXs (PULXs) confirmed that at least a fraction are neutron stars (NSs) accreting at super-Eddington rates.

However, the detection of pulsations is statistically limited by:

Low photon counts: Many observations lack the $\sim10,000$ photons required for standard timing analysis.
Low pulsed fractions: Typical PULXs have low pulsed fractions ( $\sim5-15\%$ ).
Variable spin rates: Rapid spin-up requires specialized search techniques.

Consequently, only $\sim6$ confirmed PULXs exist out of $\sim2,000$ known ULXs. The authors hypothesize that many non-pulsating ULXs share physical characteristics with PULXs but remain undetected due to unfavorable observational conditions (e.g., low statistics). The challenge is to identify these "hidden" candidates using multi-dimensional data without relying solely on direct pulsation detection.

2. Methodology

The authors employed an unsupervised Machine Learning (ML) approach using Gaussian Mixture Models (GMMs) to cluster ULX observations based on their spectral and temporal properties.

Data Source

Dataset: An updated catalog of ULXs observed by XMM-Newton, cross-matched with the Walton et al. (2022) catalog and 4XMM-DR13.
Sample Size: 640 unique sources with a total of 1,769 observations.
Labels: 95 observations from 6 sources were flagged as known PULXs (used as a reference set).
Features: The model utilized 10 parameters, including:
- Peak flux ( $F_{peak}$ ) and luminosity ( $L_{peak}$ ) in the 0.2–12 keV band.
- Fluxes in specific energy bands (EPIC-pn/MOS).
- Hardness Ratio ( $HR_{Hard}$ ).
- Variability flags ( $VAR\_FLAG$ ).
- Distance and host galaxy properties.

The Clustering Pipeline

Preprocessing: Data underwent scaling (MinMax, Standard, Quantile, Power) and optional logarithmic transformation and Principal Component Analysis (PCA).
Clustering Algorithm: A Bayesian Gaussian Mixture Model with two components was fitted.
- Training: The GMM was trained only on observations of ULXs with unknown compact object nature (excluding known PULXs to prevent bias).
- Prediction: The trained model assigned a probability of belonging to the "PULX-like" cluster ( $C_P$ ) to all observations, including the known PULXs.
Optimization Metrics:
- PULXs Ratio (PR): The fraction of known PULX observations assigned to the PULX cluster. The goal was $PR \geq 0.99$ .
- Uncertain Ratio (UR): The fraction of unknown ULXs assigned to the other cluster ( $C_U$ ). Maximizing UR ensures the model distinguishes PULXs from non-PULXs effectively.
Explainability: A Decision Tree (DT) was trained on the GMM outputs to extract interpretable physical rules defining the PULX cluster.

3. Key Results

Optimal Configuration

The best performance was achieved when:

$F_{peak}$ (Maximum Observed Flux) was included as a feature.
$L_{peak}$ was excluded (it had negligible impact).
No Logarithmic transformation or PCA was applied.
MinMaxScaler was used for scaling.
Covariance type: "Full".

Under these conditions, the model achieved a PR of 0.99 (capturing all known PULXs) and a high UR of 0.79, indicating a robust separation between PULX-like and non-PULX-like sources.

Candidate Identification

New Candidates: The algorithm identified 85 unique sources (355 observations) as new candidate PULXs.
Characteristics: The Decision Tree revealed that candidates are primarily defined by:
1. High broadband flux ( $\gtrsim 4 \times 10^{-13}$ erg cm $^{-2}$ s $^{-1}$ ).
2. High maximum observed flux ( $F_{peak} \gtrsim 1 \times 10^{-12}$ erg cm $^{-2}$ s $^{-1}$ ).
3. Presence of intra-observational variability ( $VAR\_FLAG \geq 0.5$ ).
Distribution: Most candidates are located within 2–40 Mpc. The inclusion of $F_{peak}$ was crucial for correctly classifying low-flux observations of known PULXs that would otherwise be missed.

Validation

Timing Analysis: A preliminary search for pulsations in the 85 candidates using the HENaccelsearch tool yielded no new detections. This is attributed to the low photon statistics in the archival data, reinforcing the need for targeted high-statistics follow-up.
QPO Correlation: 19 out of 22 observations with reported Quasi-Periodic Oscillations (QPOs) were classified into the PULX cluster, supporting the hypothesis that QPOs are a signature of accreting NSs.

4. Significance and Contributions

AI-Driven Discovery: This work demonstrates the power of unsupervised ML to identify astrophysical populations that are statistically difficult to detect via traditional methods. It shifts the paradigm from "searching for signals" to "identifying sources with similar multi-dimensional properties."
Target List for Future Observations: The paper provides a curated list of 85 high-probability candidates. These sources are prioritized for future deep XMM-Newton or Chandra observations to acquire the necessary photon statistics to detect pulsations.
Insight into ULX Physics: The results suggest that the "PULX" population is likely much larger than currently confirmed (potentially $\sim25\%$ of ULXs with sufficient data). The reliance on $F_{peak}$ highlights that many PULXs are only detectable as pulsars during their brightest outbursts.
Methodological Framework: The pipeline is designed as a "living" project. It can be updated with new observations and improved spectral modeling (e.g., using multi-component models instead of simple power laws) to refine the candidate list.

5. Limitations and Future Work

Data Quality: The fluxes used are derived from standard 4XMM power-law fits, which may not perfectly represent ULX spectra (often requiring double-component models).
Contamination: While the sample is likely pure, distant sources ( $>100$ Mpc) may suffer from contamination by background AGNs, though these are unlikely to be high-flux candidates.
Next Steps: Future work involves re-running the analysis with physically motivated spectral fits, incorporating uncertainty estimates (bootstrapping), and expanding the search to include lower-luminosity X-ray binaries ( $10^{38}$ erg s $^{-1}$ ) to test the continuity between standard X-ray pulsars and PULXs.

In conclusion, this study successfully utilizes clustering to bridge the gap between known PULXs and the broader ULX population, providing a statistically robust roadmap for the next generation of pulsation discoveries.