Using machine learning to explore the long-term evolution of IGR J17091-3624

Reporter：Ming-Chun Hung

Supervisor：Albert Kong, Alvina On, Surojit Saha

Motivation

• Probing Extreme
• Physics near Black hole X-ray binaries (XRBs) exhibit relativistic disk winds, jet ejections, and complex Quasi-Periodic Oscillations (QPOs).
• Classifying these exotic variability states is crucial for understanding accretion disk instabilities and strong-field gravity.
• The Manual Inspection Bottleneck
• Space missions (e.g., RXTE, NICER) have produced decades of massive, high-time-resolution X-ray archives.
• State classification has historically relied on manual visual inspection—a process that is subjective, labor-intensive, and unscalable for massive surveys.
• Research Objective: Automated State Reconstruction
• Build a robust deep learning framework to autonomously recognize fast oscillatory patterns.
• Seamlessly map local temporal features into the global, multi-year accretion state evolution of IGR J17091-3624.

Overview of X-ray Binaries (XRBs)

• System Structure
• A compact object (Black Hole / Neutron Star) accreting matter from a companion star.
• Gravitational energy is converted into intense X-ray emission.
• Classification
• LMXB: Mass transfer via Roche-lobe overflow from a low-mass donor.
• HMXB: Mass transfer via stellar winds from a massive companion.
• Key Observables
• Accretion Disk: Drives transitions between Hard and Soft spectral states.
• Timing Variability: Exhibits broad-band noise, QPOs, and structured flaring.

• Form an accretion disk where viscous dissipation heats plasma to millions of Kelvin.

​

Overview of IGR J17091-3624

• Source Profile
• A transient Galactic black hole candidate / microquasar system.
• Monitored extensively by RXTE during its prominent 2011–2013 outburst.
• GRS 1915+105 Analogue
• The only other known black hole system displaying similar exotic variability patterns to GRS 1915+105.
• Features structured, high-amplitude "heartbeat" oscillations driven by accretion disk instabilities.
• Astrophysical Significance
• Exhibits 9 distinct variability classes (Classes I–IX) spanning diverse timescales.
• Serves as a key laboratory for studying limit-cycle accretion dynamics, disk winds, and QPOs.

• Form an accretion disk where viscous dissipation heats plasma to millions of Kelvin.

​

Target

To identify different QPO light curves from the 2011–2013 long-term data of IGR J17091-3624.

​

Method

To train a 1D CNN neural network classifier using the labels from the latter part of the paper "An atlas of exotic variability in IGR J17091−3624: a comparison with GRS 1915+105" and the RXTE results as input

Problem

The long-term data from RXTE has an exposure time of 16 seconds, whereas the short-term data (QPO light curve)records the exact arrival times of individual photons.

Each data point has a 16-second exposure.

Problem

Each point represents a single photon and its energy. The horizontal axis indicates the time elapsed since the first photon (the original data records the absolute arrival time).

Solution

Since the short-term data records the exact arrival times at the telescope, I can segment the data to count the number of photons arriving within 16-second intervals. Subsequently, I can use this processed data to retrain the neural network to match the long-term data.

Solution

The figure shows the raw RXTE data. The interval between the blue lines is 16 seconds. I summed the energy within these intervals and converted it to mCrab to match the unit of the long-term data.

Solution

I segmented the raw RXTE data into 1-D arrays of 1,024 seconds each (with a bin size of 16 * 64) to serve as input for the neural network. Segments were generated from the same observation using a sliding window approach to increase the dataset size.

Neural Network Architecture： CNN Classifier

Input

1D CNN

1D CNN

1D CNN

1D CNN

Linear

Output

Result：CNN Classifier

Device：RTX 3090

AI Platfrom：Pytorch

Max Iteration：500

Iteration：500

Accuracy: 87.5%

​

Result：CNN Classifier

I applied label smoothing to restrict the model's confidence to a maximum of 96%, which consequently capped the lowest achievable loss at around 0.6.

Problem： Confusion between Class I and Class IX

As shown in the confusion matrix, Class IX has a relatively high probability of being classified as Class IX.

Solution：Confusion between Class I and Class IX

We incorporated FFT into the architecture to see if it could capture subtle differences among QPOs, thereby improving classification accuracy.

Neural Network Architecture：CNN+FFT

Input

1D CNN

1D CNN

1D CNN

1D CNN

Linear

Output

FFT

Linear

Result：CNN+FFT Classifier

Device：RTX 3090

AI Platfrom：Pytorch

Max Iteration：500

Iteration：500

Accuracy: 94%

As shown in the confusion matrix, incorporating FFT significantly reduced the misclassification rate between Class I and Class IX.

Result

CNN Classifier vs CNN+FFT Classifier

• Overall Accuracy Improvement:
• Baseline 1D-CNN: 87.5% → CNN + FFT: 94.0% (+6.5%)
• Mitigating Class I vs. Class IX Confusion:
• Misclassification of Class IX as Class I significantly reduced from 7 samples to 2 samples.
• Spectral Domain Contribution:
• FFT bypass provides orthogonal power spectral signatures, effectively separating quiet, noise-dominated states and boosting recognition of multi-scale oscillations (e.g., Class III, V, VII).

Next Step

Next, I plan to apply a sliding-window-like algorithm to have the neural network classify the 2011–2013 input data into their respective classes.

By using the light curve of this segment as the input to the neural network, we can determine its corresponding class.

Result：Class I

• Initial Rise Detection:
• Successfully captures the early outburst rising phase before MJD 55650, matching literature ground truth.
• Late-Decay Over-Prediction:
• Heavy cluster of false positives detected during the late decay phase (after MJD 55830).
• Waveform Ambiguity:
• Continuous inference misclassifies late-stage Class IX as Class I due to their shared low-variability, Poisson-dominated profiles.

Result：Class II

• Low-Flux Quiescent State:
• Corresponds to the prominent deep intensity dips during the outburst.
• Precise Valley Mapping:
• Accurately maps the entire low-flux valley between MJD 55700 and 55740 with high consistency.
• Robust Pattern Recognition:
• Demonstrates exceptional network reliability in identifying sustained, quiet dip structures.

Result：Class III

• Spike Transition State:
• Characterized by sharp, upward intensity spikes along state transition boundaries.
• Phase-Aligned Detection:
• Accurately identifies occurrences during initial transitions and around the outburst rise near MJD 55760.
• High Ground-Truth Fidelity:
• Closely reproduces the reference temporal distribution.

Result：Class IV

• Heartbeat Oscillations:
• Represents iconic, highly structured periodic "heartbeat" multi-second oscillations.
• Comprehensive Tracking:
• Captures nearly all active intervals across the early outburst and secondary peaks around MJD 55780.
• Strong Morphological Disentanglement:
• Distinct periodic waveform enables near-perfect alignment with literature observations.

Result：Class V

• Dense Large-Amplitude Bursts:
• Features dense, high-amplitude multi-peak oscillatory patterns.
• Dual-Epoch Identification:
• Successfully isolates both active intervals: the early plateau (~MJD 55670) and the post-peak phase (MJD 55750–55800).
• Epoch-Invariant Reliability:
• Robustly identifies complex oscillatory states across different flux levels.

Result：Class VI

• Bright-Phase Flaring:
• Characterized by rapid, high-amplitude flaring dynamics at peak luminosity.
• High-Flux Peak Mapping:
• Reliably maps detections to the brightest outburst intervals between MJD 55750 and 55830.
• Luminosity Correlation:
• Effectively correlates with the most energetic accretion states.

Result：Class VII

• High-Frequency Burst Dynamics:
• Exhibits complex, rapid oscillatory burst sequences.
• Localized Burst Isolation:
• Correctly confines detections around the primary outburst peak near MJD 55760.
• Transient State Capture:
• Effectively isolates short-lived, exotic burst intervals from surrounding states.

Result：Class VIII

• Dipping-and-Flaring Dynamics:
• Rare, highly dynamic state combining deep dips with rapid flaring.
• Peak-Epoch Detection:
• Successfully captures the prominent occurrence near the outburst maximum at MJD 55760.
• Sample Sparsity Impact:
• Sparser detections reflecting the limited training instances available for this rare state.

Result：Class IX

• Late-Decay Dominance:
• Literature ground truth shows complete dominance throughout the late decay phase (after MJD 55830).
• Sparse Continuous Detections:
• Only isolated points detected, as the majority of samples are absorbed into Class I.
• Fundamental Physical Degeneracy:
• Confirms that 16-second binning blurs sub-second timing differences, creating a physical degeneracy between flat quiet states.

Conclusion

• High Classification Accuracy (94%): Integrating 1D CNN with FFT spectral bypass improved classification accuracy from 87.5% to 94%, effectively distinguishing complex QPO/flaring states (e.g., Class IV, V, VIII).
• Mapping Long-term State Transitions: Sliding-window inference successfully reconstructed the accretion state evolution of IGR J17091-3624 across 2011–2012, mapping low-flux quiet states (Class I, II) and flare peaks (Class IV–VIII).

Conclusion

• Class I vs. Class IX Degeneracy: The persistent confusion between Class I and IX is physically driven by their shared steady/low-variability profiles and the loss of sub-second timing features under 16-second binning.
• Future Improvement: Incorporating X-ray hardness ratios (energy spectra) and multi-scale temporal modeling will provide the missing orthogonal information to separate Poisson-dominated quiet states.

Thank You for Listening.�Q&A