Mitra Taheri, Johns Hopkins University
Teaching Machines to Solve Global Challenges, One Atom at a Time
Written by Sophia Chen
In 2011, the US federal government launched its Materials Genome Initiative, a program whose goal was to accelerate materials discovery. To that end, the community developed so-called “high-throughput” methods that used data-heavy computational simulations to inform human-led experiments. Researchers have also worked to make various steps of the materials discovery process autonomous, with the aid of machine learning tools.
However, focusing solely on the speed of discovery can exacerbate or ignore other global problems, such as climate change, according to Mitra Taheri of Johns Hopkins University. It’s time for the field to consider a more holistic approach to materials discovery, as Taheri discussed in Thursday’s Symposium X, titled, “Teaching Machines to Solve Global Challenges, One Atom at a Time.” “We need to start making better decisions about how we set up autonomous labs,” she said.
Taheri highlighted the field’s problematic reliance on rare earth metals. Her group focused on cobalt, which is used in many types of so-called Heusler alloys, a class of materials with desirable magnetic properties. Researchers are investigating cobalt-containing Heusler alloys for spintronics, an emerging information technology that could be more energy-efficient than current computers.
However, most cobalt originally comes from mines in the Democratic Republic of the Congo, where they employ workers, including children, to perform dangerous work in an environment full of toxins for the equivalent of dollars per day. Taheri thinks that materials scientists need to consider this human cost in their research.
To this end, Taheri’s group used high-throughput methods to discover Heusler alloys with the desired magnetic properties, but with little or no cobalt. Using characterization methods such as the Magneto-Optical Kerr Effect and Extended X-ray Absorption Fine Structure, they found several promising candidate materials. In addition, the high-throughput methods allowed them to “reduce the time to […] less than a day versus months,” said Taheri.
In addition, the field may find it fruitful to take more “good enough” attitude toward materials discovery. “How many of you […] have come home from vacation and eaten something that you really didn't want to eat just because you were hungry and you didn't go to the store?” she asked the audience. Spaghetti sauce without herbs may not be delicious, but it still feeds you.
Taheri’s group took this attitude to a project to discover high-entropy alloys. Instead of using the more precise density functional theory to simulate candidate materials, they looked at using graph neural network, a machine learning method that is less precise but quicker than DFT. “We were able to discover a number of alloys that actually matched the prediction,” she says.
Using faster but less precise methods could help researchers move toward autonomous laboratories that can “learn on the fly” and make decisions in real time, she says.
Symposium X—MRS/The Kavli Foundation Frontiers of Materials features lectures aimed at a broad audience to provide meeting attendees with an overview of leading-edge topics.