Four points in the sky — they are the same quasar. Its light was bent by a gravitational lens — a massive galaxy along the line of sight. From the shape of these images, like a sculptor from shadows, one can reconstruct how the invisible mass of the lens is distributed. But previously this took days of manual calculations.
A neural network, trained on millions of examples, now does it in minutes. It doesn't need images — only the coordinates of the four images. It predicts how flattened the lens is and its total mass. Like a sculptor guessing the curvature of glass from a few shadows.
The method has already been tested on real quasars. Now astronomers can quickly process thousands of new systems and search for unique objects — binary black holes, streams of matter, clues about dark matter and gravitational waves.
Both neural networks never saw real images during training: they were trained on synthetic data, but on real quasars they worked just as reliably as if they had learned from real images.