New AI Predicts Material Properties Even Without Knowing Their Structure

Finding a new material that works well in a battery, solar panel, or medical device is slow and expensive. Scientists often start with just the chemical recipe — the elements involved — but not the exact way atoms are arranged. Many AI tools need that full structural picture to make useful predictions, so they're stuck before the real research even begins.

DISTAL solves this by using a clever learning trick. First, it studies a huge library of virtual materials, learning patterns from their chemical compositions alone. Then, a structure-savvy AI — called a 'teacher' — passes along its knowledge to a simpler 'student' model. The student learns to connect structures to properties during training, but at the moment of real-world use, it doesn't need structure data at all. It can predict how a material will behave based only on its ingredient list.

The result is a unified system that combines three types of information: traditional chemical descriptors, patterns learned from pretraining, and structural hints from the teacher. Across 39 different prediction tests, this combined approach outperformed the reference model on 37 of them — a strong showing in an area where reliable predictions are hard to come by.

What this means in practice is speed. Early-stage scientists don't need to wait for expensive experiments or fully solved crystal structures to narrow down promising new materials. They can quickly screen thousands of possibilities on a laptop, focusing lab work only on the most promising candidates. That could compress years of material development timeline into months, helping bring everything from longer-lasting phone batteries to more efficient solar panels to consumers sooner.