Researchers Develop Autonomous Materials Discovery Engine Using AI

Researchers Develop Autonomous Materials Discovery Engine Using AI

Researchers studying artificial intelligence (AI) algorithms to automate materials exploration have developed an autonomous platform where theory and experiment are naturally coupled in a closed-loop manner. 

A research team led by Ichiro Takeuchi, professor and chair of the Department of Materials Science and Engineering (MSE), developed an AI-based program to accelerate the experimental discovery of advanced materials in a self-driving mode, a contribution that was published in the journal Science Advances today. 

In this work, called Autonomous MAterials Search Engine (AMASE), researchers demonstrated an autonomous workflow that is able to self-navigate experimental mapping of a phase diagram, which serves as a blueprint for discovering new materials. After each experimental iteration, materials phase information is automatically fed to a computational prediction of the phase diagram, which in turn drives and decides what experiment to perform next. 

“Every scientific endeavor is ideally a cooperation of experiment and theory, with constant feedback between the two, the way Aristotle envisioned the scientific method to be more than two thousand years ago. But in reality, this is hard to carry out for a number of practical reasons,” said Takeuchi. 

With a relatively simple experimental set up using a thin-film combinatorial library, which can house a large number of compositionally varying samples, they enlisted robot science to autonomously update feedback between materials exploration experiment and theory-based prediction of the phase diagram. 

Here’s how the AMASE workflow operates: the AI algorithm instructs a diffractometer—an instrument that analyzes materials crystal structure—to study a combinatorial library covering a range of composition, at a particular temperature. From the acquired experimental data, a machine learning code figures out the crystal phase distribution landscape in the composition range at that temperature. 

Then, this information is fed into CALculation of PHAse Diagrams (CALPHAD), a platformbased on Gibb’s theory of thermodynamics of materials, to perform a computational prediction of the entire phase diagram in the composition–temperature space, which is in turn used to determine which part of the phase diagram will be studied by the diffractometer next. The live theory-experiment cycle continues, with each iteration resulting in a more accurate phase diagram, and the AMASE operates autonomously with no human intervention, reducing overall experimentation time by six fold. The work is a collaboration with Ji-Cheng "J.C." Zhao, the former chair of MSE and the current Dean of Engineering of the University of Connecticut, a subject-matter expert in CALPHAD. 

Haotong Liang, the materials science and engineering doctoral student who carried out the work for this paper, recently won the IMPACT Award from the American Physical Society’s Topical Group on Data Science, an accolade for his work toward autonomous thin film synthesis techniques.  

“I am truly honored to have won this award. I have enjoyed developing various autonomous materials science platforms. I hope my work can contribute to improving AI-based workflow for materials science as well as for materials manufacturing,” said Liang.

July 2, 2025