A Wandering 35,000-Solar-Mass Black Hole Fed by a Gravitational Wake

The paper reports the discovery of a wandering intermediate-mass black hole (IMBH) of approximately 35,000 solar masses that is actively accreting gas through a gravitational wake, providing the first direct observational evidence for this predicted accretion channel. The object, designated UGCA320-IMBH, is located off-nucleus in the nearby edge-on dwarf irregular galaxy UGCA 320 (also known as DDO 161), at a distance of 6.03 Mpc. The host galaxy has a stellar mass of 1.3×10 8 M⊙, an H I mass of 9.7×10 8 M⊙, and a star-formation rate of log SFR(UV) = −1.4 M⊙ yr−1.

The black-hole nature of the source is supported by four lines of evidence:

1. Broad Balmer emission lines detected in VLT/MUSE observations from 2021, with the broad Hα component having a FWHM of 854±6.02 km/s, corresponding to a black hole mass of 3.5 × 10 4 M⊙.

2. A spatially coincident, unresolved continuum counterpart in HST ACS F814W imaging at 2.8 pc resolution, ruling out Balmer-dominated supernova remnants.

3. AGN-like stochastic optical variability over 14 years, with an amplitude of 0.15 mag, smaller than typical luminous blue variables.

4. A UV-to-optical SED well described by an AGN power-law continuum partially attenuated by optically thick gas (τb ≫ 1 at λ ≈ 3646 Å) with modest dust extinction, which cannot be reproduced by stellar atmosphere models.

The key discovery is the spectroscopic identification of three distinct gas components predicted by the Bondi–Hoyle–Lyttleton (BHL) accretion framework:

• A blueshifted (−27 km/s), low-density upstream flow with electron density 40 cm−3, traced by [O III] and [S II] emission.

• A redshifted (+27 km/s), dense downstream wake with densities exceeding 10 6 cm−3, traced by permitted Fe II and Ca II emission.

• Dense absorbing clumps at a blueshifted velocity of −50 km/s, with hydrogen column density exceeding 5×10 22 cm−2, located well within the BHL capture radius of 0.3 pc.

Multi-epoch spectroscopy reveals changing-look behavior in the broad Hα and Hβ lines over timescales of a few years: prominent in April 2021, nearly disappeared in April/June 2025, partially reappeared in July 2025, and faded again by April 2026. The absorbing clumps are inferred to be at a Keplerian radius of 0.06 pc, within the BHL capture radius, indicating they are embedded in the accretion flow.

The paper concludes that this discovery establishes a previously unobserved channel for the growth of wandering intermediate-mass black holes, which may represent an important pathway for the early growth of seed black holes before they sink into galactic nuclei and contribute to supermassive black hole assembly.

Improvement 1: Multi-Component Spectral Line Decomposition Engine

• What to improve: Current AI spectral fitting tools often assume single-Gaussian profiles or simple disk/wind models.

• Improved AI capability: An AI system that automatically identifies and separates multiple kinematically distinct gas phases (e.g., blueshifted upstream, redshifted wake, dense absorbing clumps) from integral-field spectroscopy, using priors from BHL accretion theory. It would output density, temperature, velocity, and column density for each component simultaneously, even in low-S/N data. This would enable rapid detection of similar accreting IMBHs in large surveys (e.g., SDSS-V, 4MOST).

Improvement 2: Time-Domain Variability Classifier for Changing-Look AGN

• What to improve: Existing variability classifiers (e.g., for quasars) do not distinguish between stochastic AGN flickering and discrete state transitions (line disappearance/reappearance).

• Improved AI capability: A recurrent neural network trained on multi-epoch spectra that predicts the probability of a changing-look event (line vanishing/appearing) and links it to physical parameters (accretion rate, column density of intervening clumps). It would flag objects like UGCA320-IMBH for follow-up, and estimate the timescale of future reappearances based on Keplerian orbital dynamics of absorbing clumps.

Improvement 3: SED Fitting with Optically Thick Partial Covering

• What to improve: Standard SED fitting codes (e.g., CIGALE, Prospector) assume uniform dust attenuation and stellar-only or simple power-law AGN models.

• Improved AI capability: A Bayesian neural network that fits UV-to-IR SEDs with a physically motivated AGN continuum plus partial covering by optically thick gas (τ b ≫ 1 at the Balmer edge) and dust. It would automatically reject stellar-only solutions and recover black hole mass, Eddington ratio, and covering fraction. This would allow AI to identify off-nucleus IMBH candidates from photometric surveys (e.g., Rubin/LSST) without requiring spectra.

Improvement 4: Gravitational Wake Accretion Detection from Kinematic Maps

• What to improve: Current AI tools for galaxy kinematics (e.g., 3D barolo) do not search for asymmetric wake signatures behind moving massive objects.

• Improved AI capability: A convolutional neural network applied to velocity and density maps from ALMA or MUSE that detects the characteristic redshifted dense wake and blueshifted upstream flow around a point source. It would output a likelihood score for BHL accretion, plus the inferred black hole mass and velocity relative to the ISM. This enables blind searches for wandering IMBHs in nearby galaxies.

Improvement 5: Self-Supervised Learning for Rare Object Discovery

• What to improve: Supervised classifiers fail on rare classes (e.g., off-nucleus IMBHs) due to lack of labeled examples.

• Improved AI capability: A self-supervised contrastive model pre-trained on millions of galaxy spectra and images, then fine-tuned with a few dozen synthetic BHL-accreting IMBH spectra (generated from the paper’s parameters). The AI would cluster unseen objects by physical similarity, surfacing new candidates with broad Balmer lines, off-nucleus positions, and AGN-like variability—even without prior labels.

Improved AI system summary:

The enhanced AI can (a) decompose multi-phase gas kinematics in IFS data, (b) predict changing-look behavior from time-series spectra, (c) fit SEDs with partial covering to reject stellar mimics, (d) detect gravitational wake signatures in kinematic maps, and (e) discover new wandering IMBHs via self-supervised anomaly detection. This would accelerate the census of IMBH growth channels and inform seed black hole formation models.