Who needs scientists anyway? A global consortium of six automated laboratories, overseen by artificial intelligence (AI), set out to produce new laser materials, dividing the labor from synthesis to testing. The effort yielded a compound that emits laser light with record-setting efficiency, researchers report today in Science. Along with other recent results, the feat suggests that, in some areas, self-driving labs can surpass the best scientists, making discoveries missed by humans.
“Automated labs are going beyond proof-of-concept demonstrations,” says Milad Abolhasani, a chemical engineer at North Carolina State University who developed a self-driving lab unaffiliated with the new work. “They have started to push the edge of science to the next level.”
The allure of AI-driven labs for developing new drugs, industrial catalysts, and energy and emission-reduction technologies is clear. Creating new molecules and materials is normally slow and tedious. Researchers must explore not only myriad recipes for making molecules, but also different reaction conditions. They have to test new compounds at every step and evaluate schemes for scaling up production and assembling materials into devices.
Over the past decade, robots have begun to automate many of these repetitive steps. In 2015, for example, Martin Burke, a chemist at the University of Illinois Urbana-Champaign, unveiled an automated system for synthesizing small molecules. Later, by incorporating AI, researchers added feedback loops, so data from newly characterized compounds could guide decisions on what to synthesize next. But discovering new materials and assembling them into devices requires robots to work in concert across even more steps, Burke says. “Nobody has all those tools and perspectives in one lab.”
Burke and Alán Aspuru-Guzik, a theoretical chemist at the University of Toronto, thought they could unite these disparate functions, hosted in different labs. “We thought, let’s make a self-driving lab made of self-driving labs,” Aspuru-Guzik says.
So, the duo teamed up with labs at the Institute for Basic Science in South Korea, the University of Glasgow, the University of British Columbia (UBC), and Kyushu University to focus on a specific goal: discovering organic compounds that can emit highly pure laser light. Such materials could power advanced displays and telecommunications devices because they can be made into thin, flexible, light-emitting films. But despite more than a decade of work in the field, only about a dozen candidate organic laser emitters have been discovered.
To start, the Glasgow and UBC labs made sugar cube–size quantities of building blocks for the materials. These colored powders were packaged up and sent to Burke’s and Aspuru-Guzik’s groups, where robots knitted them in different combinations into candidate emitters. All of those were passed to Toronto, where other robots characterized their light-emitting properties in solution. For the best ones, the UBC lab determined how to synthesize and purify the larger quantities needed for making devices. In batches of a few grams, the materials were then shipped to Japan, where the Kyushu lab incorporated them into working lasers and tested their properties.
The whole operation was overseen by a cloud-based AI platform designed primarily by the teams in Toronto and South Korea to learn from each experiment and incorporate feedback into subsequent iterations. “It was almost like a symphony,” says Lee Cronin, who leads the lab in Glasgow. The main hurdle became shipping compounds around the world in time. “FedEx became the bottleneck,” Burke says.
The collaboration paid off. The effort produced 621 new compounds, including 21 that rivaled state-of-the-art laser emitters and one that emits blue laser light more efficiently than any other organic material. “It’s really impressive to make all of these different components work together,” says Philippe Schwaller, an expert in self-driving labs at the Swiss Federal Institute of Technology. And the pace of discoveries was “fantastic,” says Donna Blackmond, a chemical engineer at Scripps Research. “Their methods got them to the good candidates much faster than usual,” she says.
It’s not the only recent success. Last year, for example, Abolhasani’s lab reported[https://onlinelibrary.wiley.com/doi/full/10.1002/aenm.202302303] creating nanoparticles of so-called perovskite minerals that showed record-setting photoluminescence, a property that can identify materials likely to work well in solar cells. And in a preprint posted last year on ChemRxiv, Burke’s team reported an AI setup that not only synthesized a bevy of new light-harvesting compounds, but also revealed what made them stable rather than prone to rapid breakdown, offering a rare glimpse of how an AI—normally a black box—made its decisions.
Burke hopes that advances in automation and AI will allow more and more labs to join forces. “That’s something we desperately need,” he says. It might ultimately allow scientists to stop pursuing robotic tasks and become robot overlords.
