🔭 astrophysics

Towards independent event horizon imaging of the supermassive black holes in M87 and the Milky Way

This paper presents an independent analysis of EHT data for M87* and Sgr A* using a novel, calibration-resilient imaging framework called GenDIReCT, which combines closure invariants with diffusion-based generative deep learning to reconstruct black hole images without relying on traditional station-based calibration.

Original authors: Nithyanandan Thyagarajan, Samuel Lai, Ivy Wong, Foivos Diakogiannis

Published 2026-08-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Nithyanandan Thyagarajan, Samuel Lai, Ivy Wong, Foivos Diakogiannis

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 our galaxy and in a distant galaxy far beyond, two monsters of gravity hold court: supermassive black holes. These are not empty holes in space, but regions where matter is crushed so densely that not even light can escape. To see them, astronomers cannot simply point a camera at the sky. The black holes themselves are invisible; we can only see the glowing ring of superheated gas swirling around them just before it falls in. To capture this faint, tiny ring, scientists use a technique called very long baseline interferometry. Imagine linking radio telescopes scattered across the entire globe so they act as a single, Earth-sized dish. This massive virtual mirror provides the sharpness needed to resolve details as small as the event horizon, the point of no return around a black hole. However, stitching together signals from telescopes thousands of miles apart is incredibly difficult. The data is often faint, the connections between telescopes are sparse, and the instruments themselves introduce errors that can blur the final picture. For years, the only way to clear up this blur was to rely on complex computer models that made assumptions about what the black hole should look like, a process that risked seeing what researchers expected rather than what was actually there.

A team of researchers from Australia has now developed a new way to see these cosmic objects that does not depend on those assumptions. They created a method that strips away the need to trust the individual performance of each telescope, focusing instead on the relationships between the signals that remain true regardless of instrument errors. To do this, they turned to a mathematical concept known as closure invariants. In simple terms, while the signal received by a single telescope might be distorted by local weather or equipment glitches, a specific combination of signals from three or more telescopes cancels out those distortions automatically. These "closure invariants" act as a reliable fingerprint of the black hole's shape, immune to the noise that usually plagues the data. The team then fed these clean, distortion-free fingerprints into a powerful artificial intelligence system called GenDIReCT. This system was not taught what a black hole looks like; instead, it was trained primarily on non-astronomical images from the CIFAR-10 dataset, learning the general rules of how shapes, edges, and textures fit together in the real world.

When the researchers applied this new framework to real data from the Event Horizon Telescope, the results were striking. They first tested the system on synthetic data, where the true answer was known, and found that the AI successfully reconstructed the correct shapes in almost every case. They then turned their attention to real astronomical targets. Looking at the quasar 3C 279, a bright jet of material shooting out from a black hole, the new method identified three distinct components in the image that matched the structure found by the original team. More impressively, by analyzing data taken over several days, the system detected a specific part of the jet moving away from the center. It measured this motion as a shift of 4.6 microarcseconds over roughly 5.39 days, a speed that corresponds to an apparent velocity of about ten times the speed of light, a phenomenon known as superluminal motion. This finding aligned perfectly with previous results, confirming that the new method could track rapid changes in the sky without needing to be tuned to the specific target.

The team also applied the technique to Centaurus A, the closest radio galaxy to Earth. The AI reconstructed a clear image showing two bright lines of gas running along the base of the jet, matching the orientation and angle seen in earlier studies. These successes suggest that the method is robust enough to handle the messy, incomplete data typical of real-world observations. The researchers are now using this same independent approach to re-examine the famous images of the black holes in M87 and our own Milky Way. By relying on data that is naturally immune to calibration errors and an AI that learns from general shapes rather than specific astronomical theories, they offer a fresh, unbiased perspective on the most extreme environments in the universe. This work does not just confirm what we already know; it provides a new, independent path to verify the most fundamental observations of black holes, ensuring that the pictures we see of these cosmic giants are truly what is there, and not just what we hoped to find.

Technical Summary: Towards Independent Event Horizon Imaging of the Supermassive Black Holes in M87 and the Milky Way

Problem Statement
Imaging supermassive black holes (SMBHs) at event-horizon scales using Very Long Baseline Interferometry (VLBI) presents an inherently difficult inverse problem. Observations are characterized by extremely sparse aperture coverage, low signal-to-noise ratios, and significant calibration uncertainties arising from instrumental and propagation effects at geographically dispersed stations. Consequently, traditional image reconstruction often relies on a range of algorithmic choices, regularisation strategies, and hyperparameters. Variations in these choices have led to alternative reconstructions and interpretations that sometimes diverge from published Event Horizon Telescope (EHT) results. The authors argue that given the scientific significance of horizon-scale imaging, independent analyses based on fundamentally different assumptions and methodologies are essential to validate these findings and mitigate confirmation bias.

Methodology
To address these challenges, the authors propose a framework that is independent in both observables and reconstruction methodology. The approach consists of two primary components:

Closure Invariants: Instead of relying on calibrated visibilities, which are susceptible to station-based gain errors, the framework utilizes closure invariants. These are interferometric observables formed from specific combinations of visibilities that are intrinsically immune to station-based calibration errors. The authors specifically employ copolar closure invariants derived from the general theory of closure invariants, which offer a unified and homogeneous form distinct from traditional closure phases and amplitudes.
GenDIReCT (Generative Deep learning Image Reconstruction with Closure Terms): To reconstruct images from these observables, the authors utilize GenDIReCT, a generative diffusion framework. This algorithm operates in the latent space of images and is conditioned directly on closure invariants.
- Training Strategy: To avoid embedding astrophysical expectations into the learned image priors, the model was trained primarily on non-astronomical images (CIFAR-10 dataset) and simple geometric shapes.
- Validation: Before application to EHT data, the framework underwent a blind, multi-tier validation process. This included testing on synthetic challenge datasets (ngEHT Analysis Challenges) and assessing consistency across different reference stations and imaging fields of view. Metrics included image-domain similarity, closure-invariant residual statistics ( $\chi^2$ ), and the Continuous Ranked Probability Score (CRPS) to probe multimodality and uncertainty.
- Independence: Crucially, no tuning was performed on the target EHT datasets (M87* and Sgr A*) during development. The algorithm and hyperparameters were fixed based on the validation criteria prior to any application to the scientific targets.

Key Results
The paper presents the application of this framework to synthetic data and real EHT observations of non-target sources to demonstrate robustness and consistency:

Synthetic Challenge Data: GenDIReCT was applied to ngEHT Analysis Challenge datasets simulating M87* and Sgr A* at 230 GHz and 345 GHz. The framework successfully recovered structures in all cases except one (Sgr A* at 345 GHz with the EHT 2022 array configuration). In most instances, it achieved comparable or superior results to other analyses based on the challenge's performance metrics.
3C 279 (Quasar): Applied to 2017 EHT data, GenDIReCT identified three distinct components consistent with the EHT Collaboration (EHTC) reference image. By analyzing data from individual observation days, the framework inferred apparent superluminal motion of the southern jet ejecta with a proper motion of $4.6 \pm 1.0$ $\mu$ as over $\approx 5.39$ days, corresponding to a velocity of $\simeq (10 \pm 2)c$ . This result was consistent with EHTC findings.
Centaurus A (Radio Galaxy): Analysis of 2017 EHT data for Centaurus A consistently reconstructed two bright ridge-lines along the sheath and near the base of the approaching jet. The orientation and opening angle of these features were consistent with EHTC reports.
M87 and Sgr A:** The authors state that analyses of the public EHT datasets for M87* and Sgr A* are currently underway. These will represent the first truly independent analysis of these targets using closure invariants and generative diffusion.

Significance and Claims
The paper positions GenDIReCT as a "calibration-resilient framework" for VLBI that offers a complementary avenue for interpreting horizon-scale black hole observations. Its significance lies in:

Calibration Independence: By utilizing closure invariants, the method isolates source-structure information while mitigating one of the dominant sources of systematic uncertainty in VLBI: station-based calibration errors.
Probabilistic Reconstruction: Unlike methods producing a single image, the diffusion-based approach samples the posterior distribution of images. This generates an ensemble of plausible solutions, allowing for the direct investigation of reconstruction degeneracies, multimodal structures, and the statistical significance of recovered features.
Methodological Independence: The framework provides a path to independent verification of EHT results that is distinct not only in the research team but in the fundamental observables and reconstruction algorithms used.
Future Scalability: The authors note that future developments aim to extend GenDIReCT to generalize across arbitrary array configurations and frequencies, reducing the need for retraining as next-generation VLBI arrays (such as the ngEHT and space-VLBI missions) come online.

The work concludes that combining calibration-independent observables with generative probabilistic inference provides an "orthogonal pathway" for characterizing uncertainties and assessing the robustness of scientific conclusions in high-resolution radio astronomy.

Technical Summary: Towards Independent Event Horizon Imaging of the Supermassive Black Holes in M87 and the Milky Way

More like this