SAN FRANCISCO—Imagine a cookbook with 150,000 tempting dishes—but few recipes for making them. That’s the challenge facing an effort at the Lawrence Berkeley National Laboratory (LBNL) known as the Materials Project. It has used computers to predict some 150,000 new materials that could improve devices such as battery electrodes and catalysts. But the database’s users around the globe have managed to make just a fraction of these for testing, leaving thousands untried. “Synthesis has become the bottleneck,” says Gerbrand Ceder, a materials scientist at LBNL.

Now, Ceder and his colleagues have married artificial intelligence (AI) and robotics to eliminate that bottleneck. The AI system makes a best guess at a recipe for a desired material and then iterates the reaction conditions as robots try to create physical samples. The new setup, known as the A-Lab, is already synthesizing about 100 times more new materials per day than humans in the lab can manage. “This is the way to go,” says Ali Coskun, a chemist at the University of Freiburg who isn’t involved with the A-Lab, but attended the Materials Research Society meeting here last week, where the new AI approach was announced.

AI-driven robotics labs are becoming commonplace among pharmaceutical companies searching for new drugs and even some academic materials labs. But those efforts primarily use liquid precursor compounds that are relatively straightforward to mix and process. “It’s a lot more difficult to do this with solid materials,” Coskun says. Synthesizing these materials typically requires mixing solid powders together and then adding different combinations of solvents, and experimenting with heat, drying time, and other inputs to try to get them to crystallize into the predicted material.

The number of recipes is essentially infinite, Ceder says. Although computers can predict which final compounds should lead to better devices, “there is no theory for synthesis that tells us what can and cannot be made,” says Kristin Persson, who heads LBNL’s Materials Project and announced the new A-Lab.

Previous automation efforts randomly mixed compounds in search of new materials, Ceder says, but the new AI-driven approach is more akin to the way traditional chemists do their jobs. The AI starts by coming up with a plausible way to synthesize a material, using its understanding of chemistry. It guides robotic arms to select among nearly 200 different powdery starting materials, containing elements such as lithium, nickel, copper, iron, and manganese. After mixing the precursors, another robot parcels out the mix into a set of crucibles, which are loaded into furnaces where they can be mixed with gases such as nitrogen, oxygen, and hydrogen. The AI then determines how long to bake the different mixes, the temperatures, drying times, and so on.

After the baking, a gumball-like dispenser adds a ball bearing to each crucible and shakes it to grind the new substance into a fine powder that’s loaded onto a slide. A robot arm then grabs each sample and slides it into an x-ray machine or other equipment for analysis. Results are fed back into the Materials Project database of materials structures and properties, and if the outcome isn’t what was predicted, the AI setup iterates the reaction conditions and starts anew.

LBNL researchers have spent the past several months working out the kinks in their system and testing it. In the process, the A-Lab has produced more than 40 target materials—about 70% of the compounds it has set out to produce. “I have made more new compounds in the last 6 weeks than my whole career,” Ceder says.

LBNL’s AI materials lab may not be alone for long. In a 3 April preprint, researchers from the Samsung Advanced Institute of Technology reported that they, too, have set up a computer-driven robotics lab to search for new electronic materials. Results from that report show their setup performed more than 200 reactions to make 35 inorganic compounds, including certain oxides commonly used in battery electrodes, solid oxide fuel cells, and superconductors. In each stage of their robotic experiments “AI is used to some degree,” says Samsung’s Jeong-Ju Cho.

Ceder notes that despite the move to fully automated synthesis and analysis, researchers are just as likely as ever to make unexpected discoveries. “That’s no different with the A-Lab.” Except now, the hits and the surprises will likely come faster.