University of Toronto researchers have developed a self-driving laboratory that uses artificial intelligence, automated metal printing, and experimental testing to accelerate the discovery of alloys for extreme environments.
The system identified six printable alloys made from nickel, cobalt, and chromium, including candidates that retained hardness or resisted oxidation better than established materials at high temperatures. The researchers designed the alloys for additive manufacturing applications in aerospace, power generation, and other industries where components face intense heat, pressure, and rapid temperature changes.
The study, published in npj Advanced Manufacturing, demonstrates how closed-loop laboratory systems can explore large materials-design spaces while accounting for whether a proposed material can be manufactured successfully.
A closed loop for data-lean discovery
Machine-learning models typically require large datasets, but researchers often lack sufficient data when investigating new material compositions.
“Most machine learning models require lots of data about material properties to learn from,” said lead author Ajay Talbot, a PhD student in the department of materials science and engineering at the University of Toronto.
To work with limited data, the team developed ALPHA-AM, an active-learning framework that combines physics-informed computer models, high-throughput experiments, process optimization, and laser-directed energy deposition.
The researchers began with experimental data from 27 alloys and used 21 descriptors related to thermodynamics, phase composition, and material properties. The model then selected promising compositions for fabrication and testing. Results from each experiment—including failed, crack-prone builds—returned to the model to refine its predictions and guide the next selection.
Across six active-learning rounds, the system moved between lower-risk searches focused on printability and higher-risk searches that pursued greater hardness. When one candidate developed hot cracks, the model incorporated the failure data and redirected its search toward a printable region of the design space.
Designing for performance and printability
The team also optimized the manufacturing process before comparing alloy performance. Researchers evaluated 90 combinations of laser power, scanning speed, and powder feed rate to identify a processing window that produced stable, dense deposits.
The system ultimately identified six printable alloys that were approximately 40 percent harder at room temperature than an alloy containing equal proportions of nickel, cobalt, and chromium.
Two compositions stood out during higher-temperature testing. An alloy containing 12 percent nickel, 62 percent cobalt, and 26 percent chromium maintained 46.2 percent greater hardness than the equal-proportion alloy at 600 °C. It also retained a 4.4 percent hardness advantage over Inconel 625, a nickel-based superalloy used in demanding industrial applications.
A second alloy, containing 36 percent nickel, 14 percent cobalt, and 50 percent chromium, showed an 84.6 percent reduction in oxidation-related mass gain compared with Inconel 625 during testing at 1,000 °C.
Operational implications for AI-enabled labs
For laboratories developing autonomous or AI-assisted workflows, the project highlights the importance of connecting modeling, fabrication, characterization, and data management within one iterative process. The system did not treat unsuccessful experiments as wasted work; it captured those outcomes as structured data that improved later decisions.
The researchers noted that the platform has so far been demonstrated only with a three-element alloy system. The study also used hardness as an initial screening measure and did not evaluate more complex properties such as ductility or fatigue life. Expanding the system to alloys with additional elements and validating candidates for production components will require further research.
This article was created with the assistance of Generative AI and has undergone editorial review before publishing.
