Adapted from this Berkeley Lab press release
A research team led by the Department of Energy’s Lawrence Berkeley National Laboratory (Berkeley Lab) has built and successfully demonstrated an automated experimentation platform to optimize the fabrication of advanced materials. The platform, called AutoBot, uses machine learning algorithms to direct robotic devices to rapidly synthesize and characterize materials. The algorithms automatically refine the experiments based on analysis of the characterization results.
The researchers tested the platform on an emerging class of materials called metal halide perovskites that show promise for applications such as light-emitting diodes (LEDs), lasers, and photodetectors. It took AutoBot just a few weeks to explore numerous combinations of fabrication parameters to find the combinations that yield the highest quality materials.
Informed by machine learning algorithms with a super-fast learning rate, AutoBot needed to experimentally sample just 1% of the 5,000 combinations to find this ‘sweet spot.’ This process would have taken up to a year with the traditional trial-and-error approach, where researchers manually test one set of parameters at a time, guided by previous experience and intuition.
“AutoBot represents a paradigm shift for material exploration and optimization,” said Carolin Sutter-Fella, a Berkeley Lab scientist and one of the study’s corresponding authors. “By integrating synthesis, characterization, robotics, and machine learning capabilities in a single platform, AutoBot dramatically accelerates the process of screening synthesis recipes. Its rapid learning approach is a significant step toward establishing autonomous optimization laboratories and can be expanded to a wide range of materials and devices.”
Scientists at the Molecular Foundry—a Department of Energy Office of Science User Facility located at Berkeley Lab—conceived the idea for AutoBot, expanded on a commercial robotics platform, and implemented solutions for data processing, analysis, and machine learning infrastructure. The multidisciplinary team included researchers from the University of Washington, University of Nevada, University of California-Davis, University of California-Berkeley, and Friedrich-Alexander-Universität Erlangen-Nürnberg. The scientists report their work in the journal Advanced Energy Materials.
AutoBot found that high-quality films can be synthesized at relative humidity levels between 5 and 25% by carefully tuning the other three synthesis parameters.
“This humidity range does not require stringent environmental controls,” said Ansuman Halder, a Berkeley Lab postdoctoral researcher and co-first author of the research paper. “The finding lays important groundwork for the development of commercial manufacturing facilities.”
Another insight was that humidity levels above 25% destabilized the material during the deposition process, resulting in poor film quality. The team explained and validated this finding by manually performing photoluminescence spectroscopy during film synthesis.
AutoBot’s performance was impressive. By identifying the most informative experiments, the algorithms rapidly learned how the synthesis parameters influence film quality.
“This strong performance was demonstrated by a dramatic decline in the algorithms’ learning rate after AutoBot sampled less than 1% of the 5000-plus parameter combinations,” said Maher Alghalayini, a Berkeley Lab postdoctoral scholar and co-first author. “Because new experiments were not changing the algorithms’ material quality predictions at this point, we decided to stop performing experiments.”
An innovative aspect of the study was “multimodal data fusion.” This involved using various data science and mathematical tools to integrate the disparate datasets and images from the three characterization techniques into a single metric for material quality. The idea was to quantify the results so that they were usable by the machine learning algorithms. For example, collaborators at the University of Washington designed an approach to convert the photoluminescence images into a single number based on how the light intensity varied across the images.
Read the full press release
