AI-driven robots start hunting for novel materials without help from humans Science.

SAN FRANCISCO, CA—Consider a cookbook with 150,000 delicious recipes but little instructions for making them. That is the dilemma confronting the Materials Project at the Lawrence Berkeley National Laboratory (LBNL). It employed computers to forecast 150,000 new compounds that may improve gadgets like battery electrodes and catalysts. However, database users all across the world have only managed to create a portion of these for testing, leaving thousands untested. "Synthesis has become the bottleneck," explains LBNL materials scientist Gerbrand Ceder.

Ceder and his colleagues have now combined artificial intelligence (AI) and robots to remove the bottleneck. As robots attempt to manufacture actual samples, the AI system makes a best estimate at a formula for a desired substance and then iterates the reaction conditions. The new setup, known as the A-Lab, is already capable of producing 100 times more novel materials each day than humans in the lab can. "This is the way to go," says Ali Coskun, a chemist at the University of Freiburg who isn't associated with the A-Lab but was in town last week for the Materials Research Society meeting, when the new AI technique was revealed.

AI-powered robotics labs are becoming more widespread in pharmaceutical businesses looking for new medications, as well as some academic materials labs. However, those attempts mostly make use of liquid precursor chemicals that are relatively simple to combine and process. "It's a lot more difficult to do this with solid materials," explains Coskun. To synthesize these materials, solid powders are often mixed together and then different solvent combinations are added, as well as experimenting with heat, drying time, and other inputs to try to get them to crystallize into the predicted material.

Ceder claims that the number of recipes is essentially unlimited. Although computers can forecast which end chemicals will result in better gadgets, "there is no theory for synthesis that tells us what can and cannot be made," says Kristin Persson, director of LBNL's Materials Project and the person who introduced the new A-Lab.

Previous automation initiatives, according to Ceder, randomly combined substances in search of novel materials, but the new AI-driven technique is more akin to how traditional chemists do their work. Using its understanding of chemistry, the AI begins by devising a plausible method of synthesizing a substance. It directs robotic arms to choose from almost 200 distinct powdery starting ingredients containing lithium, nickel, copper, iron, and manganese.

After combining the precursors, another robot divides the mixture into crucibles, which are then loaded into furnaces and combined with gases such as nitrogen, oxygen, and hydrogen. The AI then determines how long to bake the various ingredients, as well as the temperatures and drying timeframes.

Following baking, a gumball-like dispenser inserts a ball bearing into each crucible and shakes it to ground the new substance into a fine powder that is put onto a slide. A robot arm then collects each sample and slips it into an x-ray machine or other analysis equipment. If the outcome is not what was projected, the AI setup iterates the reaction conditions and starts over.

Researchers at LBNL have spent months ironing out the flaws in their technology and testing it. During the procedure, the A-Lab created over 40 target materials—roughly 70% of the compounds it set out to produce. "I've created more new compounds in the last six weeks than I have in my entire career," Ceder says.

The AI materials lab at LBNL may not be alone for long. Researchers at the Samsung Advanced Institute of Technology claimed in an April 3 preprint that they, too, have set up a computer-driven robotics lab to explore for novel electronic materials. According to the findings of that report, their system accomplished over 200 reactions to produce 35 inorganic compounds, including particular oxides utilized in battery electrodes, solid oxide fuel cells, and superconductors. "AI is used to some extent" in each stage of their robotic studies, according to Samsung's Jeong-Ju Cho.
Ceder observes that, despite the transition to fully automated synthesis and analysis, researchers are still as likely as ever to produce surprising findings. "That's no different with the A-Lab." Except that the hits and surprises are likely to come faster now.

‍