AI and automation are speeding up science and chemistry by helping scientists pick which experiments to conduct and home in on promising new materials.
Why it matters: There’s pressure on these fields to produce new materials faster and cheaper to support and power technologies that could transform industries and economies.
The big picture: New materials and molecules are needed for the batteries, drugs and semiconductors envisioned to underpin a green grid, precise medicine, and the next generation of computing and communications.
- “Ultimately, this new technology drives the next revolution, maybe the next major scientific revolution,” says Olexander Isayev, a professor of chemistry at Carnegie Mellon University.
- The U.S., China, EU and Japan all have ongoing initiatives to spur the development of materials by building libraries of compounds that can be tested and potentially developed into new materials.
- The U.S. led the world in publications in the field two decades ago, but China now holds the top position in materials science research by this measure.
What’s happening: It can take decades to get a new material to market in a process that involves an almost “artisanal science,” Isayev says.
- But as some laboratory tasks are automated and AI is integrated in the analysis of scientific data, materials scientists and chemists are using machine learning and other tools to perform computations and simulations that can point them to candidates for new catalysts, polymers and other materials with unique properties.
- They’re also using AI models to remove noise in the data generated in experiments or direct microscopes to areas of interest, cutting the time researchers can spend on them.
Zoom in: In a new study, researchers combined machine learning, theories and calculations of physical properties and experiments to identify new alloys.
- There are so many (10^50) possible combinations of the elements typically used in alloys, including nickel, iron, cobalt and copper, that it would be nearly impossible to wade through them all using trial and error.
- The researchers were interested in a particular type of alloy, called high-entropy alloys, that are made up of multiple elements in similar proportions. They were also searching for alloys called invars that don’t expand or contract when the temperature changes, making them ideal for transporting and storing natural gas.
How it works: The team put data about different alloys — some of it more than 100 years old — into an AI model that determines correlations between alloy properties and the elements in them and generates hundreds of thousands of candidate materials. A neural network then whittles that down to about 1,000 remaining candidates.
- Those are then assessed based on physics theories and computations about how the alloys should behave, and about 20 compositions are suggested.
- The top three compositions are selected by researchers and measured physically.
- That data gets fed back into the AI, which tries to improve on it having learned about the underlying physics, creating an active learning loop.
What they found: The researchers identified two new alloys in six times through the loop.
- It took two to three months, compared to years of experiments typically required to find a new alloy, says Ziyuan Rao, a postdoctoral researcher at the Max Planck Institute for Iron Research in Germany and a co-author of the paper, published Thursday in the journal Science.
Yes, but: Finding a material or chemical is one hurdle. Actually making it is another.
- It’s much more difficult to train AI models to predict how to synthesize a material, in part because there isn’t data on what can’t be synthesized, says Keith Butler, a senior lecturer at Queen Mary University of London.
- Researchers are starting to employ AI to try to optimize manufacturing of materials like perovskites, which are used for advanced solar cells.
- The National Science Foundation awarded $20 million over five years to the new Center for Computer-Assisted Synthesis at Notre Dame University, which aims to tackle the problem.
What they’re saying: “Despite the rate of advancement in this field, groundbreaking potential of these approaches is yet to be realized,” a description for a conference on the topic being held this month states.
- Materials science doesn’t have the big datasets that fuel AI-enabled advances in genomics and other fields.
- The high cost in time and money to conduct experiments means less data is available to train AI systems — and much of what is available is collected in different experiments or under various conditions, and is spread across institutions or locked up in proprietary databases.
- The alloy work is “impressive” because they were able to get the results with sparse data, says Isayev, who uses AI to identify new materials and predict the properties of chemicals for solar energy technology and drug design.
What to watch: Another AI model — large language learning models that can write text — could be coming to materials science.
- “I think that next year that’s what’s going to be very hot in material sciences and in physical sciences,” says Gabriel Gomes, a professor at CMU who uses machine learning to develop new chemical reactions and catalysts.