Technology
Machine learning
At the University of Toronto, Ted Sargent runs a test kitchen of sorts. His team, composed of researchers and students, develops recipes, measures and mixes ingredients carefully, and then evaluates the aftermath. The concoctions mostly—if not always—turn out to be inedible.
Fortunately, though, flavor is not the point. They’re not making food. Sargent’s team cooks with carbon dioxide. Their goal is to invent recipes to “upgrade” the greenhouse gas into useful materials, says Sargent, an electrical engineer. Instead of releasing the pollutant into the air, or capturing and sequestering it underground, factories and power plants of the future could use renewable energy to convert the carbon dioxide into raw materials they could sell.
One promising class of recipes involves electrically zapping carbon dioxide with other reactants to transform it into the six-atom molecule ethylene, composed of two carbon atoms and four hydrogens. Ethylene is a raw material used to make common plastics, including those found in supermarket and Ziploc bags. “It’s about a $60 billion market,” says Sargent. “It’s a pretty valuable commodity chemical.”
But the real significance of Sargent’s work isn’t just the value of his recipes. It’s that he’s cooking with artificial intelligence. Sargent’s team discovered new ingredients for making ethylene by using new AI and supercomputer-driven techniques that have been gaining popularity among materials scientists in the past decade.
Sargent teamed up with Zachary Ulissi of Carnegie Mellon University, who specializes in using algorithms to invent new materials. Ulissi simulated 12,229 microscopic close-ups of 244 different crystals to zero in on the most promising candidates for making ethylene. In particular, they wanted to find materials that carbon monoxide molecules—which are produced from the breakdown of carbon dioxide—would stick to easily. Ulissi used a supercomputer to run a fraction of the simulations, a time-intensive task that would be impractical to complete for all 12,229 close-ups. Then, he trained a machine-learning algorithm with those supercomputing results, and that algorithm learned to make the remaining simulations quickly.
.jpg)
These computer-based methods provide researchers with a faster and more comprehensive strategy for discovering new materials. It can take two decades for scientists to discover a material and fine-tune it well enough for commercialization. Conventionally, says Ulissi, “it has been really difficult to search wide swaths of materials.” As the story goes, in the 1870’s, Thomas Edison tested more than 3,000 different materials to find the right filament for the first affordable incandescent light bulb. Ultimately, the filament that reigned in the next century was made of tungsten—a material Edison never even tried.
Likewise, the Toronto/Carnegie Mellon team may not have found their winning ingredients yet. Their recipes require a lot of electricity to make, which means that right now, producing ethylene from carbon dioxide is not profitable. Sargent and his colleagues are working to design more economically viable recipes. In new work published last week in Nature, they report the discovery of multiple new materials, so-called catalysts, that enable faster, more energy-efficient recipes for converting carbon dioxide to ethylene. These catalysts could be the secret ingredient that finally makes this technology scalable. “We need to decrease our carbon footprint, but we don’t want to do so at the expense of the increasing prosperity of people around the world,” says Sargent.
With computers, material recipes are no longer strictly limited to an individual scientist’s expertise. To find their catalysts, Ulissi and Sargent’s team used a public database called the Materials Project, which aims to serve materials scientists as a Google-like search engine. It contains data about more than 120,000 different inorganic chemical compounds. Anyone can log onto its website, specify which atomic elements they’d like to investigate and properties they’d like to pursue, and quickly find many candidate materials.
.jpg)
Sargent and Ulissi, for example, knew from previous experience that copper-containing materials make good catalysts, so they searched the Materials Project specifically for nonreactive alloys made of copper. The site suggested 244 crystals as a starting point. From that list, the team’s algorithms pointed to aluminum-containing copper alloys that might be the best fit. As the algorithms made predictions for the best aluminum-copper ratios and how evenly the two metals should be mixed, scientists in the lab synthesized materials based on those predictions and fed their results back into the algorithm. This iterative process between computer and lab scientists eventually led them to discover and produce 17 efficient catalysts in the lab. (All of this work was completed before their lab shut down two months ago due to the Covid-19 pandemic.)
More scientists are now relying on computing tools to invent new materials. “There’s really been a paradigm change in the last 20 years,” says Kristin Persson, a physicist at Lawrence Berkeley National Laboratory who began the Materials Project database in 2011. Computational techniques “have gone from niche applications to driving innovation,” she says.
In 2017, Boeing-affiliated researchers reported using AI to invent a powder alloy for 3D-printing airplane parts. That same year, researchers at Los Alamos National Laboratory used an AI to design alloys that could be repeatedly heated and cooled without weakening. Last July, Duracell launched a new battery called Optimum that contains a new material that Persson first discovered in 2004 using computer simulations. Oil giant BP has recently partnered with Massachusetts-based startup Kebotix, which has created an AI-driven tool for designing recipes for more ecologically friendly plastics.
Still, even with the AI and supercomputers, it took Sargent and Ulissi’s team about three years to identify these new catalysts, test them, and publish the results. The current materials discovery bottlenecks are in the lab—mixing chemicals and testing them, says Persson. Algorithms can only help so much, she says, because you still need to test all the ideas at the workbench.
Persson thinks that incorporating robots into the workflow would speed up materials discovery. "The only answer there is robotics,” she says. “We can’t hire every single grad student in our school to stand on an assembly line, trying all the different possibilities that the computations are spitting out at them."
In fact, Kebotix has already begun using robots to discover chemicals. The company is moving toward fully autonomous materials design, a setup that CEO Jill Becker refers to as a “self-driving laboratory,” where computer simulations suggest recipes for new materials and robots test those recipes. Kebotix’s customers may choose to use these capabilities independently, with one National Institutes of Health laboratory recently using the company’s AI software to more efficiently conduct an experiment for drug development.
Still, with robots, materials discovery will require human oversight. Algorithms are not super precise, and synthesizing new materials still requires “a fair amount of artisanship,” says Sargent. “It’s not like the experimentalists never surprise the theorists,” he adds. Even with stand mixers, Instapots, and bread machines, the kitchen still needs a cook.
- How a Chinese AI giant made chatting—and surveillance—easy
- The confessions of Marcus Hutchins, the hacker who saved the internet
- We’ll learn to sing together when we’re far apart
- 27 days in Tokyo Bay: What happened on the Diamond Princess
- Tips and tools for cutting your hair at home
- 👁 AI uncovers a potential Covid-19 treatment. Plus: Get the latest AI news
- 🏃🏽♀️ Want the best tools to get healthy? Check out our Gear team’s picks for the best fitness trackers, running gear (including shoes and socks), and best headphones