Four faint streaks on an image — that's all an astronomer sees. Like a sculptor reconstructing the shape of invisible glass from four shadows. About a hundred years ago, Albert Einstein predicted, and Fritz Zwicky proposed using gravitational lenses — natural telescopes made of dark matter, amplifying the light of distant active galactic nuclei by hundreds of times. But to apply such magnification, you need a precise map of the invisible mass.

Previously, astrophysicists manually tweaked parameters, spending days on a single system. The process was like sculpting by touch: each possible mass distribution was checked against the four image points. With upcoming sky surveys — Euclid, Rubin — the number of systems will reach thousands. Manual work became a bottleneck.

The solution came from machine learning. Researchers created a digital sculptor from shadows — two fully connected neural networks trained on six million computer simulations. The networks never saw sky images. They were fed only the coordinates of four images, and in return they predicted the mass distribution and ellipticity of the lens — like a sculptor deducing the curvature of glass from the arrangement of shadows.

On a validation set of 200 simulated systems, 75% of models met this benchmark in an average time of 126 seconds. The method was also tested on real quasars — HE0435-1223, PG1115+080. Optimization took between 22 and 319 seconds, never relying on brightness data, which is distorted by microlensing from stars inside the lens.

Now the 'lens labeler' turns tedious tuning into a fast procedure. It opens systematic analysis of active galactic nuclei at high redshifts, searching for binary black holes after galaxy mergers, studying dark matter and the physics of gravitational waves. The sculptor, who once worked for days, now finishes in minutes — and is ready to process thousands of new shadows.