🔭 astrophysics

Extracting the X-ray reverberation response functions from the AGN light curves using an autoencoder

This study demonstrates that a variational autoencoder trained on simulated X-ray light curves can directly extract reverberation response functions from AGN observations, successfully revealing the variable height of the corona in IRAS 13224-3809 and its correlation with luminosity.

Original authors: Sanhanat Deesamutara, Poemwai Chainakun, Tirawut Worrakitpoonpon, Kamonwan Khanthasombat, Wasutep Luangtip, Jiachen Jiang, Francisco Pozo Nuñez, Andrew J. Young

Published 2026-04-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sanhanat Deesamutara, Poemwai Chainakun, Tirawut Worrakitpoonpon, Kamonwan Khanthasombat, Wasutep Luangtip, Jiachen Jiang, Francisco Pozo Nuñez, Andrew J. Young

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: Listening to the Echoes of a Black Hole

Imagine a supermassive black hole at the center of a galaxy. It's not just a vacuum cleaner; it's a cosmic lighthouse. As gas swirls around it, it forms a hot, glowing disk (the accretion disk). Above this disk, there is a super-hot, energetic cloud of particles called the corona (think of it as a glowing, flickering candle flame hovering over a pond).

Sometimes, this "candle" flashes. The light from the flash hits the "pond" (the disk) and bounces back. This bouncing light is called a reflection. Because the light has to travel from the candle to the pond and back, there is a tiny delay. This delay is called reverberation.

By measuring how long that delay takes, astronomers can figure out how high the candle is hovering above the pond. If the candle is low, the echo comes back quickly. If it's high, the echo takes longer.

The Problem: The "Static" is Too Loud

For decades, scientists have tried to measure these echoes. But real data is messy. It's like trying to hear a whisper in a crowded stadium. The light curves (graphs showing how bright the black hole gets over time) are full of random noise, gaps in data, and chaotic fluctuations.

Traditionally, scientists tried to analyze this data using complex math in the "Fourier domain" (a way of looking at data based on frequencies, like breaking a song down into its individual notes). It works, but it's indirect and complicated. They wanted a way to look at the raw "whisper" (the light curve) and directly pull out the "echo shape" (the response function) without all the heavy math gymnastics.

The Solution: A Machine Learning "Echo Finder"

The authors of this paper decided to use a type of Artificial Intelligence called a Variational Autoencoder (VAE).

The Analogy: The Master Chef and the Apprentice
Imagine you are trying to teach an apprentice how to recognize a specific type of soup just by looking at the steam rising from the bowl.

The Training: You don't show the apprentice real soup yet. Instead, you generate thousands of fake soups on a computer. You make some with the candle low, some high, some with extra noise, some with gaps. You show the apprentice the "steam" (the light curve) and the "recipe" (the actual echo shape).
The Learning: The AI (the apprentice) learns the patterns. It realizes, "Ah, when the steam looks this way, the candle must be 5 meters high. When it looks that way, it's 15 meters high."
The Test: Once the apprentice is trained, you show it a bowl of real soup (real telescope data). The apprentice looks at the steam and says, "I bet the candle is 9 meters high."

What They Did

Simulated the Universe: They used a supercomputer to create thousands of fake black hole light curves. They made them look as realistic as possible, adding in random noise and "glitches" that real telescopes see.
Trained the AI: They fed these fake curves into the VAE. The AI learned to reverse-engineer the data: "If I see this specific pattern of brightness changes, what does the echo shape look like?"
The Real Deal: They took the trained AI and pointed it at real data from a famous black hole called IRAS 13224–3809. This black hole is known for being very active and changing its brightness a lot.

The Results: The Candle is Dancing

The AI worked! It successfully looked at the messy, real data and pulled out the "echo shapes."

Here is what they discovered:

The Corona Moves: The "candle" (corona) isn't stuck in one place. It moves up and down!
The Connection: When the black hole gets brighter (more luminous), the corona moves higher (from about 3 to 20 times the size of the black hole itself). When it dims, the corona drops lower.
Validation: This matches what other scientists found using different, older methods. It confirms that the black hole's atmosphere is dynamic and changes with its mood (brightness).

Why This Matters

This paper is a breakthrough because it proves we can use Machine Learning to look at raw time-based data and instantly understand the geometry of a black hole.

Old Way: Like trying to solve a puzzle by looking at the shadow it casts on the wall.
New Way: Like using a smart camera that can look at the object and tell you exactly what it is, even if it's in the dark.

The Limitations (The "But...")

The authors are honest about the limits:

The "Fuzzy" Details: The AI is great at seeing the general shape of the echo (is the candle high or low?), but it sometimes misses tiny, specific details right at the peak of the echo.
Simplification: They made some assumptions to make the math work (like fixing the spin of the black hole). In reality, things might be even more complex.
Noise Sensitivity: If the data is too noisy (like trying to hear a whisper in a hurricane), the AI gets confused. It needs the data to be reasonably clear (Signal-to-Noise ratio > 50).

The Bottom Line

This study is like giving astronomers a new pair of glasses. Instead of squinting at complex graphs, they can now use AI to "see" the shape of the echoes directly. It confirms that the corona around black holes is a dynamic, moving structure that reacts to how much energy the black hole is eating. It's a powerful new tool for mapping the invisible architecture of the universe's most extreme objects.

1. Problem Statement

Active Galactic Nuclei (AGN) exhibit X-ray variability driven by processes in the accretion disk and the hot corona. A key phenomenon is X-ray reverberation, where X-rays from the corona illuminate the disk, causing a time-delayed reflection. Traditionally, the geometry of this system (specifically the coronal height) is inferred by analyzing lag-frequency spectra or power spectral densities (PSD) in the Fourier domain.

However, this Fourier-based approach has limitations:

It relies on statistical averaging and does not directly recover the physical response function (the impulse response of the disk) from a single light curve.
It assumes stationarity, which may not hold for all AGN states.
There is a lack of methods to extract the response function directly in the time domain using machine learning.

The authors aim to develop a machine learning framework to directly extract X-ray reverberation response functions from AGN light curves, bypassing the Fourier domain, and to apply this to real observational data to constrain coronal geometry.

2. Methodology

A. Data Simulation and Training Sets

The authors generated synthetic AGN light curves to train and test their models. They progressively increased the realism of the simulations:

REV Model: The simplest model. $s(t) = ba(t) + R a(t) \otimes \psi(t)$ , where $a(t)$ is a red-noise driving signal, $\psi(t)$ is the disk response function (generated via the kynxilrev ray-tracing code), and $R$ is the reflection fraction.
REV+PROP Model: Adds a "top-hat" response function ($TH(t)$) to simulate disk-propagating fluctuations (hard lags) on longer timescales.
REV+PROP+NG Model: Adds uncorrelated Gaussian noise ( $n(t)$ ) at various Signal-to-Noise Ratios (SNR).
REV+PROP+NG+NP Model: The most realistic model, incorporating Poisson noise (simulating CCD readout) and Gaussian noise. This model was used for the final application to real data.

Parameters:

Coronal Heights ( $h$ ): Varied from 2.3 to 30 $r_g$ (gravitational radii).
Reflection Fraction ( $R$ ): Varied to account for dilution effects.
Target: The goal was to predict the response function $\psi(t)$ given the input light curve $s(t)$ .

B. Machine Learning Architecture

The study employs Variational Autoencoders (VAE), a type of unsupervised/supervised deep learning architecture, compared against a standard Autoencoder (AE).

Architecture: Based on Multilayer Perceptrons (MLP).
- Encoder: Maps the input light curve to a latent space defined by a mean ( $\mu$ ) and variance ( $\sigma^2$ ).
- Reparameterization: Samples a latent vector $z = \mu + \epsilon\sigma$ (where $\epsilon$ is Gaussian noise) to ensure a continuous latent space.
- Decoder: Reconstructs the output, which is the predicted response function.
Loss Function: A combination of Binary Cross-Entropy (BCE) to measure reconstruction accuracy and Kullback-Leibler (KL) divergence to regularize the latent space distribution.
Input/Output: Input is a binned light curve (1s or 5s bins); Output is a response function vector of length 10,000.

C. Real Data Application

The trained VAE model was deployed on XMM-Newton observations of the Narrow-line Seyfert 1 galaxy IRAS 13224–3809.

Preprocessing: Light curves were background-subtracted, binned, and gaps were filled using Piecewise Cubic Hermite Interpolating Polynomials (PCHIP).
Selection: Only observations with SNR $\gtrsim 50$ (calculated via power spectral density) were used.
Inference: The model predicted the response function for each observation. These predictions were compared against a library of kynxilrev simulated response functions.
Metric: The best-matched coronal height was determined using zero-lag cross-correlation. Heights yielding a correlation $\ge 0.8$ of the maximum score were used to calculate the mean height and standard deviation (the "80% rule").

3. Key Contributions

First Time-Domain Extraction: This is the first study to use a VAE to directly extract X-ray reverberation response functions from AGN light curves in the time domain, rather than inferring them indirectly via Fourier lags.
Robustness to Noise: The study systematically tested the model's performance under varying noise conditions (Gaussian and Poisson), establishing that the method is reliable for SNR $\gtrsim 50$ .
Real-World Application: Successfully applied the model to 13 observations of IRAS 13224–3809, providing direct estimates of response functions and coronal heights without relying on complex spectral fitting for every epoch.
Validation: Demonstrated that the VAE can recover the shape of the response function (including double-peak features for specific heights) even in the presence of disk-propagating fluctuations and noise.

4. Results

Model Performance:
- For noise-free data, both AE and VAE achieved high accuracy (Cross-Correlation > 0.9).
- Performance degraded as SNR decreased. The model maintained high accuracy (CC $\gtrsim 0.9$ ) for SNR $\ge 50$ .
- The VAE outperformed the standard AE in handling the stochastic nature of the data and generalizing to the test set.
IRAS 13224–3809 Analysis:
- The model successfully estimated response functions for 13 observations.
- Coronal Height ( $h$ ): The estimated heights ranged from 3 $r_g$ to 20 $r_g$ .
- Luminosity Correlation: A positive correlation was found between the coronal height and the source X-ray luminosity ( $L_X$ ). The Spearman correlation coefficient was 0.68 ( $p$ -value: 0.011).
- Comparison: These results are consistent with previous studies (e.g., Alston et al. 2020; Caballero-García et al. 2020) that used Fourier lag-frequency analysis, validating the VAE approach.
- Discrepancies: Some individual height estimates differed slightly from previous works, likely due to the VAE using only soft-band (0.3–1 keV) light curves and a fixed photon index ( $\Gamma=2$ ), whereas previous works used broader energy bands and variable $\Gamma$ .

5. Significance and Future Outlook

Independent Tool: This method provides an independent, time-domain tool to map the disk-corona geometry, complementing traditional Fourier-based techniques.
Physical Insight: The direct recovery of the response function shape allows for a more intuitive understanding of the reverberation physics, such as identifying specific features like double peaks associated with specific coronal heights.
Limitations:
- The current model struggles to resolve fine details around the response peak (e.g., specific sub-features) and focuses on the overall shape.
- It assumes a "lamp-post" geometry and fixed parameters (spin, inclination, density), which may introduce degeneracies.
- It does not yet account for a "warm corona" or Broad Line Region (BLR) contamination, which can mimic reverberation delays.
Future Work: The authors suggest improving the model by training on more diverse synthetic data (including varying $\Gamma$ and spin), implementing Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs) for better feature extraction, and incorporating multi-band data to break parameter degeneracies.

In conclusion, this paper demonstrates that Variational Autoencoders are a powerful and viable method for extracting physical response functions from AGN light curves, offering a new pathway to study the dynamic geometry of supermassive black hole environments.