Using artificial intelligence to engineer materials' properties

[The SC]

Applying just a bit of strain to a piece of semiconductor or other crystalline material can deform the orderly arrangement of atoms in its structure enough to cause dramatic changes in its properties, such as the way it conducts electricity, transmits light, or conducts heat.

Now, a team of researchers at MIT and in Russia and Singapore have found ways to use artificial intelligence to help predict and control these changes, potentially opening up new avenues of research on advanced materials for future high-tech devices.

The findings appear this week in the Proceedings of the National Academy of Sciences ("Deep elastic strain engineering of bandgap through machine learning"), in a paper authored by MIT professor of nuclear science and engineering and of materials science and engineering Ju Li, MIT Principal Research Scientist Ming Dao, and MIT graduate student Zhe Shi, with Evgeni Tsymbalov and Alexander Shapeev at the Skolkovo Institute of Science and Technology in Russia, and Subra Suresh, the Vannevar Bush Professor Emeritus and former dean of engineering at MIT and current president of Nanyang Technological University in Singapore.

Introducing a small amount of strain into crystalline materials, such as diamond or silicon, can produce significant changes in their properties, researchers have found. The mechanical strain is represented here as a deformation in the diamond's shape. (Image: Chelsea Turner, MIT)
​

Already, based on earlier work at MIT, some degree of elastic strain has been incorporated in some silicon processor chips. Even a 1 percent change in the structure can in some cases improve the speed of the device by 50 percent, by allowing electrons to move through the material faster.

Recent research by Suresh, Dao, and Yang Lu, a former MIT postdoc now at City University of Hong Kong, showed that even diamond, the strongest and hardest material found in nature, can be elastically stretched by as much as 9 percent without failure when it is in the form of nanometer-sized needles. Li and Yang similarly demonstrated that nanoscale wires of silicon can be stretched purely elastically by more than 15 percent. These discoveries have opened up new avenues to explore how devices can be fabricated with even more dramatic changes in the materials’ properties.

Strain made to order
Unlike other ways of changing a material’s properties, such as chemical doping, which produce a permanent, static change, strain engineering allows properties to be changed on the fly. “Strain is something you can turn on and off dynamically,” Li says.

But the potential of strain-engineered materials has been hampered by the daunting range of possibilities. Strain can be applied in any of six different ways (in three different dimensions, each one of which can produce strain in-and-out or sideways), and with nearly infinite gradations of degree, so the full range of possibilities is impractical to explore simply by trial and error. “It quickly grows to 100 million calculations if we want to map out the entire elastic strain space,” Li says.

That’s where this team’s novel application of machine learning methods comes to the rescue, providing a systematic way of exploring the possibilities and homing in on the appropriate amount and direction of strain to achieve a given set of properties for a particular purpose. “Now we have this very high-accuracy method” that drastically reduces the complexity of the calculations needed, Li says.

“This work is an illustration of how recent advances in seemingly distant fields such as material physics, artificial intelligence, computing, and machine learning can be brought together to advance scientific knowledge that has strong implications for industry application,” Suresh says.

The new method, the researchers say, could open up possibilities for creating materials tuned precisely for electronic, optoelectronic, and photonic devices that could find uses for communications, information processing, and energy applications.

The team studied the effects of strain on the bandgap, a key electronic property of semiconductors, in both silicon and diamond. Using their neural network algorithm, they were able to predict with high accuracy how different amounts and orientations of strain would affect the bandgap.

“Tuning” of a bandgap can be a key tool for improving the efficiency of a device, such as a silicon solar cell, by getting it to match more precisely the kind of energy source that it is designed to harness. By fine-tuning its bandgap, for example, it may be possible to make a silicon solar cell that is just as effective at capturing sunlight as its counterparts but is only one-thousandth as thick. In theory, the material “can even change from a semiconductor to a metal, and that would have many applications, if that’s doable in a mass-produced product,” Li says.

While it’s possible in some cases to induce similar changes by other means, such as putting the material in a strong electric field or chemically altering it, those changes tend to have many side effects on the material’s behavior, whereas changing the strain has fewer such side effects. For example, Li explains, an electrostatic field often interferes with the operation of the device because it affects the way electricity flows through it. Changing the strain produces no such interference.

Diamond’s potential
Diamond has great potential as a semiconductor material, though it’s still in its infancy compared to silicon technology. “It’s an extreme material, with high carrier mobility,” Li says, referring to the way negative and positive carriers of electric current move freely through diamond. Because of that, diamond could be ideal for some kinds of high-frequency electronic devices and for power electronics.

By some measures, Li says, diamond could potentially perform 100,000 times better than silicon. But it has other limitations, including the fact that nobody has yet figured out a good and scalable way to put diamond layers on a large substrate. The material is also difficult to “dope,” or introduce other atoms into, a key part of semiconductor manufacturing.

By mounting the material in a frame that can be adjusted to change the amount and orientation of the strain, Dao says, “we can have considerable flexibility” in altering its dopant behavior.

Whereas this study focused specifically on the effects of strain on the materials’ bandgap, “the method is generalizable” to other aspects, which affect not only electronic properties but also other properties such as photonic and magnetic behavior, Li says. From the 1 percent strain now being used in commercial chips, many new applications open up now that this team has shown that strains of nearly 10 percent are possible without fracturing. “When you get to more than 7 percent strain, you really change a lot in the material,” he says.

“This new method could potentially lead to the design of unprecedented material properties,” Li says. “But much further work will be needed to figure out how to impose the strain and how to scale up the process to do it on 100 million transistors on a chip [and ensure that] none of them can fail.”

“This innovative new work demonstrates potential to significantly accelerate the engineering of exotic electronic properties in ordinary materials via large elastic strains,” says Evan Reed, an associate professor of materials science and engineering at Stanford University, who was not involved in this research. “It sheds light on the opportunities and limitations that nature exhibits for such strain engineering, and it will be of interest to a broad spectrum of researchers working on important technologies.”


Source: By David L. Chandler, MIT

 

[Bogeyman]

AI-driven robots are making new materials, improving solar cells and other technologies

In July 2018, Curtis Berlinguette, a materials scientist at the University of British Columbia in Vancouver, Canada, realized he was wasting his graduate student's time and talent. He had asked her to refine a key material in solar cells to boost its electrical conductivity. But the number of potential tweaks was overwhelming, from spiking the recipe with traces of metals and other additives to varying the heating and drying times. "There are so many things you can go change, you can quickly go through 10 million [designs] you can test," Berlinguette says.

So he and colleagues outsourced the effort to a single-armed robot overseen by an artificial intelligence (AI) algorithm. Dubbed Ada, the robot mixed different solutions, cast them in films, performed heat treatments and other processing steps, tested the films' conductivity, evaluated their microstructure, and logged the results. The AI interpreted each experiment and determined what to synthesize next. At a meeting of the Materials Research Society (MRS) here last week, Berlinguette reported that the system quickly homed in on a recipe and heating conditions that created defect-free films ideal for solar cells. "What used to take us 9 months now takes us 5 days," Berlinguette says.

Other material scientists also reported successes with such "closed loop" systems that combine the latest advances in automation with AI that directs how the experiments should proceed on the fly. Drug developers, geneticists, and investigators in other fields had already melded AIs and robots to design and do experiments, but materials scientists had lagged behind. DNA synthesizers can be programmed to assemble any combination of DNA letters, but there's no single way to synthesize, process, or characterize materials, making it exponentially more complicated to develop an automated system that can be guided by an AI. Materials scientists are finally bringing such systems online. "It's a superexciting area," says Benji Maruyama, a materials scientist with the U.S. Air Force Research Laboratory east of Dayton, Ohio. "The closed loop is what is really going to make progress in materials research go orders of magnitude faster."

With more than 100 elements in the periodic table and the ability to combine them in virtually limitless ways, the number of possible materials is daunting. "The good news is there are millions to billions of undiscovered materials out there," says Apurva Mehta, a materials physicist at the Stanford Synchrotron Radiation Lightsource in Menlo Park, California. The bad news, he says, is that most are unremarkable, making the challenge of finding gems a needle-in-the-haystack problem.

Robots have already helped. They are now commonly used to mix dozens of slightly different recipes for a material, deposit them on single wafers or other platforms, and then process and test them simultaneously. But simply plodding through recipe after recipe is a slow route to a breakthrough, Maruyama says. "High throughput is a way to do lots of experiments, but not a lot of innovation."

To speed the process, many teams have added in computer modeling to predict the formula of likely gems. "We're seeing an avalanche of exciting materials coming from prediction," says Kristin Persson of Lawrence Berkeley National Laboratory (LBNL) in California, who runs a large-scale prediction enterprise known as the Materials Project. But those systems still typically rely on graduate students or experienced scientists to evaluate the results of experiments and determine how to proceed. Yet, "People still need to do things like sleep and eat," says Keith Brown, a mechanical engineer at Boston University (BU).

So, like Berlinguette, Brown and his colleagues built an AI-driven robotics system. Their goal was to find the toughest possible 3D-printed structures. Toughness comes from a blend of high strength and ductility, and it varies depending on the details of a structure, even if the material itself doesn't change. Predicting which shape will be toughest isn't feasible, Brown says. "You have to do the experiment."

As a test case, the BU team set out to make salt shaker–size, barrel-shaped structures from a plastic. They varied the number of struts that make up the outer wall of the barrel and details of each strut's shape and orientation. Testing all possible combinations, about a half-million, wasn't realistic. So, they initially had their robots fabricate 600 structures that sampled the full array of options. A kind of vise then squeezed each one until it gave way.

The group then added an AI decision-making algorithm that calculated the most likely next best design after each test. The program spots trends in attributes that confer toughness, such as the thickness and radius of each strut, in order to predict even sturdier structures. "We basically turned on the machine and walked out the door," Brown says. After 24 hours and just over 60 designs, the AI-driven system had come up with a tougher barrel than any of the original designs.

Many more closed loop efforts were showcased at the MRS meeting. Researchers at the Massachusetts Institute of Technology in Cambridge and LBNL have independently developed autonomous systems to find better perovskite photovoltaics—cheap, lightweight materials that are poised to revolutionize solar energy. A team at Carnegie Mellon University in Pittsburgh, Pennsylvania, reported using another AI system to find safer charge-carrying electrolytes for lithium-ion batteries, which are now prone to catching fire. And researchers at the University of Liverpool in the United Kingdom have developed a suite of AI-driven robots to discover novel catalysts for generating hydrogen gas, a potential carbon-free fuel, from water.

Few of these projects have turned up blockbuster results, researchers acknowledge. However, Maruyama says, "It's still early days." One challenge is that materials scientists themselves often don't agree on how best to relate a material's conductivity or other testable properties to its structure, says John Gregoire, a physicist at the California Institute of Technology in Pasadena. "If we haven't figured out how to break that down in the community, it's hard to imagine how we will teach a computer to do it," he says.

Another issue is that each team must design its own robotics and software systems, as standards have yet to take shape. "Everyone is exploring different ways to do this," says Joshua Schrier, a computational chemist at Fordham University in New York City. Eventually, the materials community may coalesce around a handful of systems that can be used by a wide swath of researchers, Schrier says. "Over the next year or two I think we'll begin to see things converge."

https://www.sciencemag.org/news/201...-improving-solar-cells-and-other-technologies