Communications of the ACM
Machine-Learning Tools Eye Quicker, Cheaper Solar Cell Designs
Credit: Getty Images
Researchers at Iowa State University are developing machine learning tools to help flip the design script and accelerate the development of better solar cells.
Rather than follow a traditional, incremental approach of running experiments, collecting data and applying those answers to technology designs, they start with the answer — they want better solar cells — and use machine learning to scan existing data, learn patterns, make predictions, and help them quickly reach their design goal.
"Can I circumvent the need for expensive data collection by using knowledge that the domain already has?" asks Baskar Ganapathysubramanian, the Joseph C. and Elizabeth A. Anderlik Professor in Engineering and leader of a new effort to develop machine-learning for technology design.
The researcher and his colleagues at Iowa State and elsewhere believe the answer is yes.
A recent grant of up to $2 million over two years from the U.S. Department of Energy's Advanced Research Projects Agency – Energy (ARPA-E) will support the team's exploration of that idea. The Iowa State-led project is one of 23 supported by up to $15 million from the research agency's Differentiate Program dedicated to accelerating the search for energy innovations.
In addition to Ganapathysubramanian, the team includes Adarsh Krishnamurthy, an Iowa State assistant professor of mechanical engineering; Soumik Sarkar, an Iowa State associate professor of mechanical engineering; Chinmay Hegde, an assistant professor at New York University's Tandon School of Engineering; Zhenan Bao, chair of chemical engineering and K. K. Lee Professor at Stanford University; Ross Larsen, a senior scientist at the National Renewable Energy Laboratory (NREL); and Bryon Larson, a researcher at NREL.
Ganapathysubramanian says the project's immediate goal is to develop machine-learning theory and software tools that will allow rapid identification of organic thin film structures that enhance solar cell performance and are easy to manufacture. The broader goal is to demonstrate that machine learning can help rapidly design all kinds of technologies.
"We're looking at a non-traditional way of doing machine learning — we're doing science with machine learning," says Sarkar, who studies machine learning and artificial intelligence. "Machine learning has been used to make your next Netflix recommendation. The new frontier is trying to see if machine learning can help engineers or scientists do engineering or science better."
In this case, the basic idea is to use machine learning to help the researchers invert and expedite the design process, Ganapathysubramanian says.
For example, set the goal of designing a car that goes 60 miles per gallon of gas. Instead of using the sometimes slow and expensive process of testing one change at a time — start with better aerodynamics, then try lighter materials, then a new engine technology — Ganapathysubramanian says the new process would use machine learning to analyze existing data as well as new computer models to come up with an initial design capable of hitting 60 miles per gallon.
Krishnamurthy, who studies computer-aided design and 3-D modeling, says he's worked with Sarkar on similar projects that have used machine learning to improve manufacturing. Specifically, they looked at how machine learning can generalize design-for-manufacturing rules to identify difficult-to-manufacture features in a complex part. This has accelerated the design process and helped identify manufacturing bottlenecks at the design stage, he says.
Once the research team's new machine learning tool has come up with designs for better solar cells, manufacturability and device performance will be tested by Bao's research group at Stanford. Larsen and Larson at NREL will provide data on materials properties and also test materials performance.
While better solar cells are certainly a good thing, Ganapathysubramanian says that's not the best thing that could come from this project.
"The most important outcomes are going to be theory and software tools that allow us to design new technologies in a fast and agile manner," he says. "That's the key outcome that ARPA-E expects."
No entries found