New Method Speeds Up Discovery Of Materials

Chemical engineers from Purdue University are using their method to discover new types of catalysts, compounds used extensively in numerous manufacturing processes. The same method also could be used for many types of materials research, said Jochen Lauterbach, an associate professor of chemical engineering.

Catalysts in American industry account for billions of dollars in annual business revenues. That means even small improvements in catalyst performance can result in significant increases in profits.

"If you have a reaction that runs a million pounds a year, a catalyst that is only 5 percent better makes a lot more money," Lauterbach said. "This approach will have a significant economic impact."

He will publicly discuss the research for the first time and unveil a commercial version of the new method Friday (1/25) during the 4th Annual International Symposium on Combinatorial Approaches for New Materials Discovery in San Diego, Calif. The conference is sponsored by the Knowledge Foundation Inc., an organization that provides an unbiased forum to showcase promising technologies through conferences and publications.

Lauterbach has developed an automated system that uses combinatorial chemistry, in which equipment systematically creates and tests thousands of chemical samples at the same time, or "in parallel." Thousands of tiny plastic beads about the width of a human hair are coated with different catalysts. All of the beads, each bearing its own individual catalyst, are tested simultaneously in the same experiment. Then the system uses a recently declassified infrared sensor technology to quickly screen each sample to evaluate its performance.

"We have the capability of making a hundred times more catalysts and screening them in the same amount of time that researchers previously needed to study one catalyst," Lauterbach said.

A small percentage of the catalysts created are effective. Information is collected from both the best catalysts and the failed catalysts and fed into software that mimics the logical and intuitive thought processes of chemists. Even though the majority of the catalysts created are not effective, the software uses the wealth of information gained from those failures to come up with entirely new catalysts.

"If a mixture doesn't work, the information about why it does not work is just as valuable as the information about why it does work," Lauterbach said. "We feed that information back into the software, and at some point we tell the program that we want a catalyst that does this and that. The software does its thing and it spits out a material combination, a range of completely new catalysts that nobody has ever thought of before, or had dared to even propose or synthesize because everybody would say, 'You've got to be crazy. This is never going to work.'

"It's something that is totally out of the box thinking for typical catalyst development."

Central to the method are two types of artificial intelligence software: hybrid neural networks and genetic algorithms. The software mimics the thought processes of chemists who create new formulas for everything from rubber compounds to rocket fuels, and plastic materials to snack foods, said James M. Caruthers, a professor of chemical engineering.

"There is this crazy-haired scientist, called a formulation chemist, who actually mixes dabs of this with dabs of that and stirs it up in a pot — cooks it, more or less — and makes new materials," Caruthers said. "You could say, 'Gee, these guys make this good stuff, and they are lucky.' Except that the people who are very good at it are lucky again and again and again, and they are actually some of the most valuable folks in an organization because they make the new materials.

"It's part science, it's part intuition."

The different types of software work together in a repeating two-phase cycle of discovery.

First, hybrid neural networks analyze the formulas of the numerous catalysts, or other materials, created by the parallel technique. The neural networks determine the properties of the materials, based on their chemical structures. In the second phase, genetic algorithms cull the best materials and eliminate the poor performers, just like survival of the fittest. The algorithms also generate "mutations" of the best materials to create even better versions, and the software determines the chemical structures of those mutations.

The resulting formulas are returned to the neural network software, and the cycle starts over again, progressively creating better and better materials, said Venkat Venkatasubramanian, a professor of chemical engineering who has been working with Caruthers to develop the software for more than a decade.

Caruthers said he observes how formulation chemists come up with new ideas. Then he models their trains of thought in software programs.

"Most experts don't think in terms of equations and mathematics," he said. "They think in terms of pictures. They have a picture of what goes in, and they have a picture of what comes out. What we should really be trying to do is model this sort of picture-to-picture reasoning that goes on.

"I look at eyes. I try to see when the eyes are excited, or a little confused or upset, and try to figure out what is the reasoning behind all of that. We want to learn why experts make the inferences that they do — why they jump from here to here."

The software isn't quite as smart as the human formulation chemists. While software programs can't match the creative brain power of people, they can mimic human thinking while simultaneously computing thousands of formulations, compared to about half a dozen for a human chemist.

"No human mind could keep all these balls in the air at the same time," Caruthers said. "Our idea is to reproduce 60, 70, 80 percent of the way these formulation chemists think, but now the computer can balance all of these balls in the air at the same time."

Neural networks are designed to think more like the human brain than a conventional computer program. The Purdue approach differs from previous methods that have used neural networks because it first takes a reaction's physics and chemistry into account, and then it lets the software take over, determining the properties of the materials. Because it combines known physics and chemistry with the software, it is called a hybrid neural network.

"Other methods assume that you know nothing about the physics and the chemistry of a process," Venkatasubramanian said. "However, in most cases you know something about the physics and chemistry governing a reaction, but that knowledge is not complete enough.

"We see how far our fundamental understanding will take us, and then we use neural networks and statistics. That way, your model is built on certain fundamentals of physics and chemistry, so it's more robust and its ability to extrapolate will be more reliable than otherwise."

Eventually, findings that identify the best new materials are handed over to human chemists who conduct experiments to validate that the formulations function well in real-life situations.

Lauterbach, Caruthers and Venkatasubramanian are working with W. Nicholas Delgass, a professor of chemical engineering, and Kendall Thomson, an assistant professor of chemical engineering.

Each component of the method — the software and the parallel screening technology — are equally important, Venkatasubramanian said.

"You generate this huge amount of data," Venkatasubramanian said. "But the data are not going to be very useful unless you can make sense of it all, and that's where the computation and the modeling software come in."

The Purdue researchers are working with several companies interested in the method. In research with one company, engineers using the method took about 30 minutes to find a new material that would have taken three months with conventional techniques, Caruthers said.

"We have a number of strategic relationships with companies," Caruthers said. "We customize the software for specific applications because there is no single commercialized package that fits all uses."

Researchers will present specific experimental findings in March during an American Physical Society meeting in Indianapolis.

The research is being conducted through a new Center for Integrated Materials-To-Product Design, headed by Caruthers, which works with industry to speed the process of making products from newly discovered materials.

"As far as we can tell, this is the first and only center of its kind in the United States," said Venkatasubramanian, associate director of the new center. "There is no other place in a university environment that is taking this kind of perspective, going all the way from molecular level modeling to final property design and then having a center dedicated to this."

The center has received $1.4 million in seed funding from the 21st Century Research and Technology Fund, administered by the state of Indiana. The research also is funded by the National Science Foundation and private industry.