Increasing the performance of solar cells, transistors, LEDs and batteries requires better electronic materials made of new, yet undiscovered compositions.
To speed up the search for advanced functional materials, scientists are using AI tools to identify promising materials from hundreds of millions of chemical agents. Alongside this, engineers are building machines that can print hundreds of material samples at a time based on chemical composition tagged with AI search algorithms.
However, to date there is no similarly quick way to verify that these prints actually perform as expected. The final step of materials characterization has been a major bottleneck in advanced materials screening pipelines.
Now, a new computer vision technique developed by MIT engineers has significantly improved the speed of characterization of newly synthesized electronic materials. This technology automatically analyzes images of printed semiconductor samples and quickly estimates two key electronic properties of each sample: band gap (a measure of electron activation energy) and stability (a measure of lifetime).
New technology accurately characterizes electronic materials 85 times faster than standard benchmark approaches.
The researchers plan to use this technique to speed up the search for promising solar cell materials. They also plan to integrate this technology into a fully automated materials screening system.
“Ultimately, we envision applying this technology to autonomous laboratories of the future,” said MIT graduate student Eunice Aissi. She said, “The entire system allows you to give a computer a materials problem, predict potential compounds, and then build and characterize the predicted materials 24/7 until the desired solution is reached.”
“Applications for these technologies range from solar energy improvements to transparent electronics and transistors,” adds MIT graduate student Alexander (Aleks) Siemenn. “It spans all areas where semiconductor materials can benefit society.”
Aissi and Siemenn detail the new technology in research published today. Nature Communications. Co-authors from MIT include graduate student Fang Sheng, postdoctoral fellow Basita Das, mechanical engineering professor Tonio Buonassisi, former visiting professor at Cukurova University Hamide Kavak, and Aalto University visiting postdoctoral fellow Armi Tiihonen.
power of optics
Once a new electronic material is synthesized, characterization of its properties is typically handled by “domain experts” who examine one sample at a time using a benchtop tool called UV-Vis. The semiconductor begins to absorb more strongly. This manual process is accurate but time consuming. Domain experts typically characterize approximately 20 material samples per hour. This is a snail’s pace compared to some printing tools that can produce 10,000 different material combinations per hour.
“The manual characterization process is very slow,” says Buonassisi. “It provides high confidence in the measurements, but nowhere near the speed at which materials can be deposited onto the substrate today.”
To speed up the characterization process and address one of the biggest bottlenecks in materials screening, Buonassisi and his colleagues looked to computer vision, a field that applies computer algorithms to quickly and automatically analyze optical features in images.
“There is great power in optical characterization methods,” says Buonassisi. “You can get information very quickly. “There is an abundance of images with many pixels and wavelengths that humans cannot process, but computer machine learning programs can.”
The team realized that certain electronic properties, such as band gap and stability, could be estimated from visual information alone, if the information was captured in sufficient detail and interpreted correctly.
With these goals in mind, researchers developed two new computer vision algorithms to automatically interpret images of electronic materials. One estimates the band gap and the other determines the stability.
The first algorithm is designed to process visual data from highly detailed hyperspectral images.
“Instead of standard camera images with three channels: red, green and blue (RBG), hyperspectral images have 300 channels,” explains Siemenn. “The algorithm takes that data, transforms it, and calculates the band gap. We run that process very quickly.”
The second algorithm analyzes standard RGB images and evaluates the stability of the material based on the visual change in material color over time.
“We found that color change can be a good indicator of the rate of degradation of the material systems we are studying,” says Aissi.
Material Composition
The team applied two new algorithms to characterize the band gap and stability of about 70 printed semiconductor samples. They used a robotic printer to place the samples on a single slide, like cookies on a baking sheet. Each deposit was made from a slightly different combination of semiconductor materials. In this case, the team printed varying proportions of perovskites, a type of material expected to be a promising solar cell candidate but known to decompose quickly.
“People are trying to change the composition of (perovskites) to make them more stable and high-performance, adding little bits of this and that,” Buonassisi said.
After printing 70 different perovskite sample configurations on a single slide, the team scanned the slide with a hyperspectral camera. We then applied an algorithm to visually “segment” the image, automatically separating samples from the background. They ran the new band gap algorithm on the isolated samples and automatically calculated the band gap for all samples. The entire bandgap extraction process took approximately 6 minutes.
“Typically, it would take a domain expert several days to characterize the same number of samples manually,” says Siemenn.
To test stability, the team placed identical slides in a chamber that varied environmental conditions such as humidity, temperature, and light exposure. They used a standard RGB camera to take sample images every 30 seconds for two hours. They then applied a second algorithm to images of each sample over time to estimate the extent to which each droplet changed color or degraded under different environmental conditions. Ultimately, the algorithm produced a “stability index,” a measure of the durability of each sample.
For confirmation, the team compared their results with manual measurements of the same water droplets performed by domain experts. Compared to expert benchmark estimates, the team’s bandgap and stability results were 98.5% and 96.9% accurate and 85x faster, respectively.
“We are constantly amazed by how these algorithms can not only speed up characterization but also achieve accurate results,” says Siemenn. “We envision adapting this slot to the current automated materials pipeline we are developing in our lab. So we can use machine learning to discover these new materials and run them in a completely automated way that guides us where we want to print them. We actually characterize them with very fast turnaround.”
This work was supported in part by First Solar.
