Porous zeolite minerals work like tiny sponges with holes of a strictly defined size. Through such holes, molecules of one kind pass while others get stuck: this is the basis for water purification, gas separation, and catalysis in petrochemistry. But inventing a new zeolite on paper is much easier than growing it in the lab: dozens of delicate conditions must align for the crystal to form.

The new model neural network looks at the three-dimensional map of atoms in the crystal the way a doctor looks at an X-ray. It compares the pattern with those already seen in real zeolites and delivers a verdict: synthesizable or not. Out of more than three hundred thirty thousand fictional structures, the model mistakenly recognized only about twelve hundred as real: this is dozens of times more accurate than previous methods.

An error of three percent is not a miss but a hint. The very structures that the model took for real, although no one has synthesized them, might actually be feasible. Chemists just need to take this short list and try growing the crystals in an autoclave.