Scientists at Ames National Laboratory have developed a new artificial intelligence (AI) tool called DuctGPT to speed up the search for materials needed for fusion energy systems. Fusion reactors generate power by joining atoms together at very high temperatures, similar to the process inside the sun. The reactors create extreme conditions of intense heat, radiation, and mechanical stress that ordinary materials cannot withstand.
DuctGPT combines advanced artificial intelligence with physics-based modeling. The tool helps identify suitable alloys, which are mixtures of different metals, that balance high-temperature strength with ductility. Ductility is the ability of a material to be stretched or formed into shapes without breaking.
The tool started from an existing model called AtomGPT developed by the National Institute of Standards and Technology. Scientists modified it using existing materials science data. DuctGPT can examine thousands of possible element combinations in seconds. Researchers can ask questions in ordinary conversational language. For instance, a user can request combinations of elements that meet specific properties required for fusion reactors.
Special attention given to tungsten alloys
Tungsten is a metal of particular interest for fusion reactors. It withstands very high temperatures well and becomes only mildly radioactive after exposure. Its main drawback is poor ductility at lower temperatures, which makes it hard to shape into complex parts. DuctGPT can suggest additions of other elements, such as titanium, zirconium, or hafnium, to tungsten. These new alloys keep the high strength and melting point while improving flexibility.
A major practical advantage is that DuctGPT runs on a normal desktop computer rather than requiring expensive supercomputers. This change reduces discovery time from months to days or hours. Scientists can then synthesize and test the predicted materials in the laboratory to confirm they perform as expected. The developers are expanding the platform with new data and models to better predict real-world behavior inside operating reactors. The work supports wider efforts to accelerate advanced materials for clean energy technologies.
This research is published in Acta Materialia.