Gravitational lenses are natural telescopes that can boost the resolution of X-ray observatories hundreds of times. The effect, a consequence of Einstein field equations, was proposed nearly a century ago by Fritz Zwicky for studying distant objects. Today, lenses are a key tool for probing active galactic nuclei at cosmological distances. But to harness that magnification, you need an accurate map of the lens's invisible mass distribution. Until now, constructing such maps has been like painstakingly sculpting in the dark: astrophysicists would tweak parameters by hand, spending days on a single object — an unworkable approach for the thousands of new systems expected from the Euclid and Rubin surveys.
Rather than guessing, the researchers built a digital “shadow sculptor” based on Machine Learning — two fully connected neural networks trained on over 6 million numerical simulations of lensed sources. Modeling gravitational potentials as singular isothermal ellipsoids (SIE), they varied the mass parameter (from 0,1 to 5,5 arcseconds) and ellipticity (0,05–0,60). The only inputs were the coordinates of the four images and the ratios of distances between them, computed relative to the intersection points of opposite pairs. Flux data, which are contaminated by microlensing, were deliberately excluded to avoid biases. To account for the rotation of real systems, the orientation angle of the ellipticity was estimated geometrically using the Cassiola & Kovner formula (average error 2,73 degrees).
On a test set of 200 simulated systems not used in training, 151 (75%) achieved the target positional accuracy of 0,5 milliarcseconds — matching Gaia's astrometric precision. The average convergence time was 126 seconds, and in 133 of the 151 successful cases it stayed under three minutes. The RMS error in predicted lens mass was 0,15 arcseconds, and for ellipticity it was 0,18 (for systems with random orientation). The networks showed clear systematics: for low-mass lenses (<1,5″) the mass was overestimated, for high-mass lenses (>2,0″) it was underestimated, but the discrepancy never exceeded 0,3″. Real quasars, including HE0435-1223 and PG1115+080, were optimized to a χ²<8 criterion in 22–319 seconds, demonstrating that the method works even without flux data.
This automated “lens labeler” turns tedious manual tuning into a rapid procedure, opening the door to systematic population studies of active galactic nuclei at high redshift. It enables real-time processing of data streams from future surveys and the hunt for rare objects: binary black holes after galaxy mergers, jets, outflows, and offset nuclei. All of this helps probe feedback mechanisms and the hierarchical growth of structure. The method requires no expensive GPU computing and slots easily into existing software pipelines.
Next, the team plans to train the network on systems with external shear, which will break the degeneracy between ellipticity and shear. Adding an uncertainty estimation module (e.g., via a Bayesian approach) will provide confidence intervals for parameters. They also plan to incorporate direct measurements of the lensing galaxy's center of mass from optical or radio images, which will dramatically improve the accuracy of initial guesses.
The method will directly accelerate X-ray astrometry, cosmography, and studies of galaxy evolution, revealing sub-parsec scales in the early Universe.
The immediate priority is to apply the pipeline to all known quadruply imaged quasars with archival Chandra and Gaia observations, and then to new candidates from the Euclid survey.
This work brings us closer to solving the inverse modeling problem — a core computational challenge in astrophysics. Like a sculptor who has learned to discern the faintest folds in a shadow, the neural network opens the way to systematically probing dark matter through line-of-sight substructure and the physics of gravitational waves from binary black holes.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The paper proposes a method for rapid estimation of gravitational lens parameters (mass, ellipticity) from the coordinates of four quasar images using two fully connected neural networks trained on millions of SIE lens simulations. The model achieves astrometric accuracy for 75% of test systems within minutes and has been successfully tested on real quasars without using flux information.
Where its strength is
The method turns laborious manual fitting into an automatic procedure that relies solely on positional information, making it robust to microlensing effects and suitable for large-scale application in future surveys.
Where it could go next
Worth trying
Why this may matter
Fast and reliable lens parameter estimation without flux data opens the way to statistical analysis of thousands of systems from the Euclid and Rubin surveys, potentially leading to new constraints on dark matter distribution and the population of compact objects.
Nearby works in our archive
The closest work in the archive sits at 0.63 by meaning — the usual level for our corpus.
Neither network ever saw a real image during training: they were taught on synthetic data, yet they performed on actual quasars as reliably as if they had learned from real snapshots.