A Northwestern University study suggests AI-driven materials discovery can move beyond hunting for promising compounds and begin designing materials with specific properties on demand. Published in Science Advances on May 22, the U.S. Department of Energy-funded research used a high-speed "megalibrary" platform to generate and test millions of material combinations at once.
Designing a material to order
The team, led by Northwestern's Chad Mirkin with collaborators including materials-discovery startup Mattiq, challenged the system to identify and then intentionally design a new piezoelectric material. Piezoelectric materials generate electricity when compressed, bent or stressed and are used in sensors, ultrasound systems and energy-harvesting devices. The platform first searched thousands of chemical combinations to find a promising candidate, then engineered one built to keep performing at a target operating temperature of up to 80 degrees Celsius.
The process took hours rather than the months or years typical of conventional trial-and-error development. The study, titled "High entropy 1D halide perovskite piezoelectrics discovered by megalibrary synthesis and rapid nonlinear optical screening," condenses large-scale experimentation onto a single chip, letting researchers examine how subtle shifts in chemistry change a material's behaviour.
Reading a million samples in 30 minutes
"We have developed a screening capability based on a technique called second harmonic generation (SHG) microscopy that allows researchers to review more than a million different material samples in less than 30 minutes," Mirkin said. "In this study, we show we can not only build a library of a million different materials, but we also can interrogate them at the individual particle level. We're about to witness the meteoric rise of materials discovery, and this is just the start."
According to the researchers, that throughput tackles a chokepoint in AI-assisted science: while robotics and automation increasingly let labs produce samples at scale, collecting detailed performance data lags behind. The megalibrary generates large experimental datasets linking chemistry to measurable outcomes, exactly the fuel needed to train machine-learning models.
A faster path than self-driving labs
Unlike newer "self-driving labs" that test one experiment after another in an iterative loop, the megalibrary operates in parallel, generating and screening enormous numbers of candidates simultaneously. The researchers argue that approach could accelerate both discovery and the data generation that powers it, with potential payoffs from energy systems to next-generation batteries.
"We've developed a screening capability that allows researchers to look at literally a million different materials, generating a million data points," said Jun Li, a former Northwestern postdoctoral fellow now an assistant professor of mechanical engineering at the University of Colorado Boulder and a co-first author. "We can use that data to train algorithms," he added, drawing a parallel to how large datasets have driven recent AI breakthroughs in the sciences. The researchers noted that Mirkin and Northwestern have financial relationships with Mattiq.
Reporting based on coverage from AI Insider, Northwestern University and Science Advances.