Machine Learning and the Quest for Better Design
Designing an innovative new product may seem like a dream job to some engineers, but the reality is that it is an expensive and time-consuming challenge for most companies. To start, the company needs to hire a team of employees to conduct numerous experiments on many different materials before deciding which ones to use. The cost of the materials and time it takes to analyze those experiments can quickly add up, and months—even years—can pass before the product is commercially viable.
But what if there was a better way? How much money could companies save, and how many more products could hit the market sooner?
Tom Truskett, professor and chair of the McKetta Department of Chemical Engineering, believes there is, and it starts with machine learning. An emerging and growing field of study in which powerful computers are given access to data and then use that data to learn for themselves, machine learning can make actual decisions and provide real-time recommendations and predictions.
Machine learning gives us a way to use computers for what they’re really good at – analyzing data in an unbiased and systematic way to help solve problems.”
“If you have an idea for a new product and want to start designing it, you could hire 50 people to do experiments. Or, you could hire significantly fewer to think creatively about the goals and design challenges and then work with a computer algorithm that helps predict the most effective experiments to perform,” Truskett said. “If there’s a way to get to the ultimate answer with fewer resources, in a shorter amount of time and with a greater appreciation for the important human contribution, whether it’s anything from a new golf club to a new flu vaccine, we should leverage it. Machine learning gives us a way to use computers for what they’re really good at—analyzing data in an unbiased and systematic way to help solve problems.”
Three years ago, Truskett enlisted postdoctoral researchers Ryan Jadrich and Beth Lindquist to see if they could use these ideas to speed up discovery and transform the way research was conducted in the lab. This summer, their efforts paid off when they completed development of a new set of software tools that allows them to use machine learning to design new chemicals and materials. The tools provide an automated way to use data obtained from previous measurements or computer simulations to suggest the most promising directions for future research.
Whether we realize it or not, machine learning has become an inextricable part of our daily lives, enabling retailers to recommend online purchases and navigation apps to deliver the shortest route between two locations. However, given how research has traditionally been conducted, incorporating machine learning into fields like chemical engineering and materials science and engineering could be considered a paradigm shift, but it shouldn’t be perceived negatively, Truskett said.
“It would remove people from some tasks, but they would be the tasks that are less rewarding and that humans are frankly not as good at performing,” Truskett said. “In my mind, what machine learning is about is recognizing where computers can be effectively used to free up time and resources that are needed for creativity on the human side. We’re really at the beginning. No one knows what the future will bring, but it’s hard to imagine a world for certain companies where such methods won’t play a critical role in their research and development.”
The team’s machine learning tools are based on inverse design, an approach in which the end product is considered first.
“The idea is to start with the properties that you want the material to possess and then work backward to figure out how to make that possible,” Jadrich said.
They are using the technique to design structures — the arrangements of components that make up a material and give it its special properties, such as conductivity, optical properties or solubility, as well as performance, strength and resilience.
“Structures that we’ve targeted so far include materials with pores of specified size, shape and connectivity, as well of a variety of crystalline lattices,” Linquist said. “Once we have a set of targeted structures, we can use the machine learning technique to suggest what types of material systems are most likely to realize them. If all goes well, we can give researchers a map of how to best execute the design in their labs.”
Truskett said that Jadrich and Lindquist — who happen to be married — are the driving force behind the team’s new machine learning tools.
“Ryan and Beth are two of the best researchers I have had a chance to partner with over the past 20 years,” Truskett said. “Ryan’s deep interest and theoretical understanding of machine learning and statistical mechanics, coupled with his expertise in materials theory, allowed him to formulate the key design strategies and the theoretical framework. Beth, who has a talent for distilling the important physics of a problem and then planning an effective computational strategy for incorporating them, worked together with Ryan to build this strategy into a user-friendly simulation approach capable of addressing a wide variety of material design problems.”
After spending the last few years developing the algorithms, tailoring existing software and testing the strategy, Truskett, Jadrich and Lindquist said they are eager to apply the machine learning tools to real-world problems. They are currently working with chemical engineering professor Delia Milliron to accelerate the otherwise time-consuming synthetic discovery of nanomaterials for smart windows that separate light from heat and with professor Keith Johnston to find new formulations of therapeutic proteins appropriate for safe, at-home treatment of disease via injection (similar to insulin for diabetics). The team is also in discussions with companies about how such machine learning approaches can be applied.
The team’s machine learning tools aren’t necessarily the magic bullet for research, Truskett said, but they do bring us one step closer to automation of the research and development process.
