By Gregor Dairaghi
Helix
From drafting and proofreading your emails to organizing your calendar, tech companies market artificial intelligence as a tool to unburden us from the tedium of our jobs. Meanwhile, social media bursts at the seams with ever more realistic AI-generated video content. With so much hype, it can be hard to predict which AI tools will permanently change the landscapes of productivity and entertainment and which will fade into distant memory.
Some scientists have a brighter vision: using AI to guide materials discovery. The discovery of new materials drives improved solar panel efficiency, longer lasting rechargeable batteries, durable spacecraft materials and more. Researchers are increasingly looking to AI algorithms to screen candidate materials for experimental testing. But experts disagree on which tasks are better suited for humans or computers and to what extent we can trust what computers tell us. So far, the rollout of automated materials generation hasn’t been entirely smooth.
The traditional approach to discover new materials, sans AI, often amounts to a shot in the dark. To synthesize a new crystal — a periodically repeating arrangement of atoms — you first choose a subset of the 118 elements on the periodic table. For each set of elements, there are multiple possible structures, each with unique physical properties. Think of the difference between pencil lead and diamond — both are crystals of pure carbon but the layout of atoms differs. For the AI-generated arrangements, most of these hypothetical crystals are duds; either their chemical bonds are unstable and the crystals will never form or else the crystal lacks any desirable properties. A synthetic chemist must traverse the golden path through this forest of possibilities that leads to a useful material. And the process is slow. “A chemist can produce one material, two materials in five years. It’s very challenging” said Basita Das, a materials scientist at MIT. With so many options to explore at a glacial pace, whittling down the pool to the most promising candidates is crucial. For that, researchers have turned to AI.
Rapid screening, careful validation
Researchers have long used computer simulations to predict the properties of materials prior to synthesis. The problem with these simulations is time, said Lauren Walters, a materials scientist at Lawrence Berkeley National Lab. A simulation for just one crystal frequently takes days to run. AI presents a faster alternative.
Researchers have developed a machine learning algorithm to fast-track the screening process. Focusing on a class of three-element crystals called ternary oxides as a case study, the team fed the compiled results of computer simulations to an AI algorithm and set it loose on a long list of possible three-element crystal structures. Unlike the computational models, which solve physics equations to build knowledge from the ground up, the machine learning approach takes a bird’s-eye view to outline the landscape of chemical possibilities. It was not only fast, but cheap as well. When Joseph Montoya, a materials scientist at the Toyota Research Institute who led the research, checked the electricity bill, the screening process cost a mere $2.5 per compound identified. Admittedly, this price doesn’t account for the initial upfront investment in the computing infrastructure.
This work, from 2024, was not the first time AI has predicted new crystal structures. A machine learning algorithm from Google identified an impressive 2.2 million new crystals, according to a 2023 study. The scale of the finding reflects the wealth of computational resources at Google’s command. However, identification is not the same as discovery, which requires experimental verification. “Discovery to me, it means I conceptualized something, I figured out how to make it, and I made it, and I validated that,” Das said.
Montoya and his co-authors followed through with the complete discovery process. The team sifted through the AI’s predictions to remove compounds containing rare or toxic elements and selected six finalists to target. While the report only identifies one new crystal, a compound of calcium, ruthenium and oxygen, the method demonstrates an expedited process for future materials discovery. “Now that we’ve done it once, doing it the next time, we could do it faster,” said Ben Savitzky, a physicist at Lawrence Berkeley National Lab who was part of the characterization effort.
Doing it faster may be necessary to stay ahead of the field. Just a year later, in 2025, an international collaboration headed by the Nanyang University of Technology in Singapore turned their machine learning models to the same class of ternary oxides and came away with two new crystals. As with Montoya’s research, the group sourced training data from The Materials Project, one of several open access databases of verified and hypothesized crystals. The successfully synthesized crystals get updated back into the database, which provides more data to train the machine learning programs to better predict more crystals.
The round-the-clock mechanical chemist
Other groups have ventured further into full robot autonomy. Scientists at Lawrence Berkeley National Lab have created what’s known as the A-Lab, a human-free materials synthesis and characterization laboratory. In A-Lab, not only does artificial intelligence select candidate materials and recipes, but a robot arm measures, mixes and heats the chemicals all by itself. This self-driving laboratory even characterizes the compounds independently using x-ray diffraction, a common technique used to quickly identify a crystal structure by shooting x-rays at it and observing how they deflect from the atoms at specific angles.
Originally, in 2023, the researchers reported the synthesis of 41 new compounds within the first 17 days of continuous operation. This claim swiftly incited pushback from fellow chemists on social media. Researchers disagreed about whether the compounds could even be called new, with some of them having been previously synthesized by chemists. The authors offered a semantic retort: “novelty” referred to compounds that were previously missing from their databases, and their databases were not exhaustive.
Another line of criticism targeted the accuracy of their identification strategy. X-ray diffraction is a standard verification technique due to its speed and versatility, but sometimes its results aren’t so clear. A unique X-ray diffraction signal might be indicative of a novel material, or it could simply be two different crystal structures mixed together. A speedy interpretation might also be part of the problem. “Faster really isn’t always better, and the push to do things faster is what’s causing a lot of errors,” Savitzky said. For the researchers behind A-Lab, manual inspection of the X-ray diffraction spectra revealed a 12 percent rate of misattribution. The authors have since softened their claim to just 36 compounds in an official correction published in January of 2026.
Respectful criticism can also inspire improvement. The group behind A-Lab recently published an updated framework for automated X-ray diffraction analysis that explains how ambiguous results from the automated process are flagged for review by a human expert.
Expanding the researcher’s toolbox
Even as the technology and its implementation improves, it’s not clear whether AI can capture the complexity of real life or instincts honed from decades of lab work. Just like a well-written recipe does not make you a Michelin-star chef, chemists have tricks and habits that affect the success of a synthesis that can’t easily be programmed into a computer. “There is a lot of human intuition that goes into the process that is not captured,” Das said.
But some researchers are optimistic. While robots may lack human intuition, they have the advantage of being much more consistent over time. Additionally, automated synthesis allows researchers to easily track not just successes but also failures, which usually are not published. Robot labs working around the clock can systematically test large numbers of recipes and store all the results, good and bad, into a massive database. “Those data sets are going to absolutely revolutionize the field,” Walters said.
While still in its early stages, AI-guided materials discovery carries a lot of hopes with it, and more than a few doubts as well. With greater responsibility allotted to computers, scientists must ponder how to verify AI’s predictions and how to steer it towards materials of interest. The future of AI-guided materials discovery, like the incorporation of AI in a broad range of scientific disciplines, lies in smoothly incorporating computer and human decision into a single integrated workflow. “There’s a lot of promise in using these AI assisted approaches,” said Carolyn Grimley from the materials science consulting firm Lucideon. “But they’re just tools, right? They’re not going to replace people.”
Gregor Dairaghi is a 5th year PhD candidate in Applied Physics at Northwestern University.
Editor’s note: This story was written in 2024. Since its writing, AI tools have become more common in chemistry and the broader world. The story has been updated to reflect new developments related to the original research covered. Statements from sources are from 2024 and do not necessarily reflect the state of AI-guided materials development in 2026.
