Amil Merchant

Amil Merchant, 27, built a machine-learning model that unearthed atomic recipes for a vast trove of new materials, some of which could lead to breakthroughs in fields such as supercomputing and renewable energy.

As technical lead at Graph Networks for Materials Exploration, or GNoME, a program run by Google’s AI subsidiary DeepMind, Merchant faced a tricky proposition. While models like those behind today’s AI chatbots are good at making predictions based on information that they’ve seen, they’re less adept at making new discoveries.

Merchant and his colleagues believed materials science was due for a shake-up: Efforts to use machine learning to find new molecular structures had struggled to pinpoint ones that would be unlikely to decay or combust.

With Merchant leading the charge, GNoME set out to find these stable structures. The project team trained a machine-learning model on an open-access database of known materials to learn the patterns that set the stable ones apart and then let the model loose on the entire periodic table. The model identified millions of potentially stable combinations. Next, the team used existing techniques from quantum mechanics to appraise those possibilities and plugged the vetted data back into the model, a process repeated several times with increasing levels of accuracy. In the end, they reported finding 380,000 new stable structures—a nearly tenfold increase from the 40,000 previously known to humanity.

For now, most of these new materials exist only on paper. But the team’s findings are available to outside researchers, and they could eventually help build more powerful solar panels, batteries, or semiconductors.