The new workflow allowed the team to quickly create thousands of compositions and determine which ones show the most promise for optoelectronic applications. Credit: Adam Malin/ORNL, U.S. Dept. of Energy
Oak Ridge National Laboratory Research Scientist Maxim Ziatdinov led a team that was honored at ORNL’s annual Awards Night for a paper published in Nature Machine Intelligence. This “Outstanding Scholarly Output” highlighted the researchers’ work to produce an autonomous microscope that can characterize the minute properties of materials.
During the development process, the team discovered that similar AI-enabled tools can be incorporated into chemical synthesis experiments that focus on versatile families of materials (e.g., metal halide perovskites, or MHPs), which can be synthesized into a nearly infinite number of compositions ideal for optoelectronic applications.
Although the intrinsic stability of each individual MHP composition must be measured to determine their potential, the time-consuming nature of these tests has historically limited the scope of experimentation.
To examine many MHPs simultaneously, a separate team that includes Ziatdinov and collaborators from the University of Tennessee and Hong Kong Baptist University developed a new workflow that combines the latest advances in automated chemical synthesis and machine-learning techniques to set the stage for a new era of more efficient and cost-effective materials science research.
The researchers published a paper in Joule in which they proposed various methods for machine learning–based materials synthesis and analysis. They then put some of those methods into practice and published the results in the Journal of the American Chemical Society.
The star of the show was a programmable pipetting robot that can precisely extract and release liquid samples. This automation allows the researchers to focus on making high-level decisions rather than manually performing routine, repetitive tasks. The massive dataset compiled by the robot, however, would not be useful without the accurate analyses quickly calculated by a cast of machine-learning methods.
“There is this famous expression in machine learning: garbage in, garbage out. An advantage of this robotic setup is that it reduces factors that can affect the accuracy of experimental results, such as human error. This improvement is a pretty significant step forward.”
With the help of their robotic lab assistant, the researchers applied two antisolvents – toluene and chloroform – to 15 combinatorial libraries, or groups of MHPs. In contrast to solvents, which dissolve materials they come into contact with, antisolvents can be added to liquid compounds to form microcrystal samples.
Each library contained 96 unique MHPs, and applying both antisolvents to each one resulted in a total of 2,880 new samples. The team completed the synthesis and characterization in about 1.5 weeks, which is a fraction of the time it would take to synthesize each sample by hand.
“If we wanted to make this many different compositions manually, one by one, it would take a year,” said Mahshid Ahmadi, an assistant professor at UT Knoxville. “Automated synthesis makes it much faster to explore these materials and optimize their properties.”
To make sense of this massive dataset, the researchers turned to machine-learning methods to identify key properties within each sample. They primarily focused on luminescence emissions, which refer to the color a material emits when exposed to a specific wavelength of light and indicate whether that material might be suitable for constructing electronic devices such as solar cells, which can be combined to form solar panels.
For a period of six hours, they collected data on the samples every 15 minutes to determine how these emission wavelengths changed over time. Compositions that change only slightly or do not change at all in this environment are the most likely to exhibit long-term stability in real-world scenarios, even when exposed to external factors such as humidity and other weather conditions.
“These are semiconducting materials, which means they can be used for many optoelectronic applications. The most common one is to make solar cells, which require fine-tuned materials capable of absorbing sunlight to create electronic charges, but you can also make LEDs – light-emitting diodes – and other types of detectors ranging from photodetectors to even high-energy devices like X-ray detectors.”
Instead of producing a single material with a specific application in mind, the team reversed the usual order of operations by producing many materials and then matching them to their potential applications. MHPs are suitable for this approach because their crystal structures contain multiple sites where various elements can be placed to create different combinations, and the ability to make so many materials from the same source also makes them a cost-effective resource for materials discovery.
The researchers used machine-learning techniques such as non-negative matrix factorization to examine the raw data and pinpoint where properties of interest were located. The team then used another form of machine learning known as a Gaussian process algorithm to transform discrete, or separate, data points collected by other methods into a continuous stream of observations to uncover promising compositions throughout the broader dataset.
“These compositional spaces can be huge, and it’s usually not practical to measure all of it, so instead we focused on regions where particular properties we had in mind for practical applications were maximized or minimized,” Ziatdinov said.
As a result of this research, scientists now have the option to further examine and optimize promising MHPs that could eventually be commercialized to compete with materials such as silicon, which is commonly used to make solar cells. The new workflow could be used to study other types of materials as well.
Going forward, the researchers plan to develop custom machine-learning models that can search compositional spaces and identify useful properties even faster. By further refining the workflow, they also hope to produce compositions in different forms.
“In the future, we want to be able to make a large batch of these samples, not as microcrystals, but as thin films needed to build solar cells and other devices,” Ziatdinov said.
Individuals and teams chosen by a highly selective committee are celebrated each year at the Awards Night event hosted by UT-Battelle, which manages ORNL for the U.S. Department of Energy.
This research was supported by the National Science Foundation and the Support for Affiliated Research Teams program started by the Science Alliance, a Tennessee Center of Excellence focused on UT Knoxville-ORNL collaboration. Additional support was provided by the Center for Nanophase Materials Sciences, a DOE Office of Science user facility located at ORNL; the Center for Materials Processing, a Center of Excellence at UT Knoxville funded by the Tennessee Higher Education Commission; Hong Kong Baptist University’s Faculty Niche Research Area and Interdisciplinary Matching Scheme; and the Hong Kong Research Grant Council’s Early Career Scheme.
UT-Battelle manages ORNL for DOE’s Office of Science, the single largest supporter of basic research in the physical sciences in the United States. DOE’s Office of Science is working to address some of the most pressing challenges of our time. For more information, visit https://energy.gov/science.
