Breath of the Abyss: Neural Networks Listen to the Magnetic Sighs of Black Holes ✛ ✦

Original: "Identifying Observational Signatures of Flux Eruption Events in Supermassive Black Hole Accretion Flows with Machine Learning"
arXiv:2606.28603 · 2026-06-26 · CC BY 4.0 · 3 min
Neural networks trained on simulations have detected signatures of magnetic flux eruptions near the event horizon: the black hole’s shadow 'inflates', polarization flares up, and radiation, contrary to expectations, dims. These 'sighs' are already within reach of the Event Horizon Telescope.

A black hole is not a frozen void, but a living magnetic engine. Its 'heart'—the accretion disk, pulsates: inhale, matter spirals into the funnel, exhale — magnetic reconnection tears the field lines, and plasma shoots away from the horizon. These events, flux eruption events (FEE), long hid in numerical simulations, eluding telescopes. But now astrophysicists have armed themselves with a 'stethoscope': a convolutional neural network, trained on libraries of GRMHD models, learned to recognize the characteristic patterns of such sighs and translated them into quantities accessible to the interferometer of the Event Horizon Telescope.

About 14% of MAD disk images contain FEE, and these events occur more frequently in black holes with prograde spin — the magnetic field literally winds onto the rotating funnel of spacetime.

Trained on manually labeled images, ResNet-18 achieved near-perfect accuracy (95%), distinguishing details invisible to the human eye. But physicists need not just a detector, but an interpreter. So they trained a random forest — on the same data, but on the output labels of the neural network, and only on parameters measurable in real observations: shadow angular size, fluxes, polarization degree. The cascade worked: it turned out that during an eruption, the shadow noticeably swells (FWHM increases), and linear polarization spikes sharply. But the most unexpected signal is silence: the total flux often does not flare up, but drops. Thus, FEE are not the progenitors of the millimeter flares observed in Sagittarius A*.

Another surprise: the rotation of the polarization vector in the Q–U plane slows down during an eruption. Previously, it was thought that FEE generate polarization loops, but the data refute this hypothesis — the magnetic exhale calms rather than twists the radiation.

These findings are not just a game with data. They chart a course for future observational campaigns. In the images of M87* and Sgr A*, fingerprints of FEE may already be lurking, but they can only be extracted by next-generation interferometers — ngEHT and the space observatory BHEX, whose resolution will rival that of light bending right at the horizon. The next step is to catch the breath in real time: dynamic neural networks trained on time series will link eruptions to the launching of relativistic jets and corona pulsations. And then, perhaps, these magnetic sighs will reveal the answer to the accursed question: does information vanish in a singularity? And does a black hole have 'hair'—contrary to the theorem that John Wheeler forged in debate, and Roger Penrose refined mathematically. Just imagine: one such 'sigh' of M87* lasts 11 days — in that time, a light beam could cross the Solar System from edge to edge. A cosmic arrhythmia that we can now eavesdrop on.

For the supermassive black hole at the center of galaxy M87, one such 'sigh' lasts about 11 days — an instant by astronomical standards, but an eternity for a physicist dreaming to peek beyond the horizon.

If you’re familiar with 'Interstellar', imagine: the astronauts by Gargantua see not only the gravitational lens, but also the accretion disk suddenly warped by a spasm — a magnetic eruption rips it open, hurling plasma at near-light speeds. Alas, at such distance they'd be incinerated before they'd notice.

\phi = \frac{\sqrt{4\pi}}{2\sqrt{\dot{M}_{\bullet,2000}}} \int_{\theta}\int_{\phi} |B^r|_{r=r_H} \sqrt{-g}\, d\phi\, d\theta
Dimensionless magnetic flux threading the horizon; values around 50 correspond to the MAD regime, where the field dominates over accretion.
\mathcal{L} = -\frac{1}{N} \sum_{i=1}^N \left[ y_i \log(p_i) + (1-y_i) \log(1-p_i) \right] w_i
Weighted cross-entropy used in training the neural network to handle class imbalance: rare eruptions accounted for far fewer images than the quiescent disk.
Scientists
Stephen HawkingJacob BekensteinJohn Archibald WheelerFritz ZwickyKarl SchwarzschildSubrahmanyan Chandrasekhar
Concepts
black hole light echo black hole accretion EHT event selection event horizon EEG seizure detection Bekenstein-Hawking entropy quasi periodic eruptions maximum likelihood iron k alpha reverberation Machine Learning accretion disk wind maximum likelihood estimation model comparison prediction error seyfert galaxy symbolic regression neural network regression reverberation mapping astrophysical signal detection