The gravitational lensing effect, predicted by Einstein and applied to search for invisible objects by Zwicky, is like raindrops tracing the outline of an invisible object on a windowpane. Light from a distant star, passing by a hidden planet or black hole, bends, and on telescope images a shifting brightness pattern appears. Before, to reconstruct the mass and position of the unseen object from this pattern, astronomers spent hours sifting through countless possibilities.

CausticFlow, a neural network that's undergone training on millions of examples, does it in a second. It was shown a vast number of synthetic light patterns from different objects, and now it instantly finds a matching template in its library. Even if the telescope captured just a few drops of that pattern, CausticFlow reconstructs the entire silhouette from them.

Remarkably, the neural network doesn't need the planet or other object to emit any light at all: it sees them only through the background distortions. On real observations, the program correctly determines parameters seven times out of ten and is already ready for large-scale searches in data from Hubble and Webb. This will open the door to studying stellar evolution, the birth of new stars, and even dark matter candidates.