A virtual materials scientist: Making complex alloy design more accessible through AI
Thun, 17.11.2026 — Computational materials science develops models and simulations that are indispensable for the development of complex materials, such as high-entropy alloys. Researchers at Empa in Thun have developed an AI framework that allows researchers to operate these complex tools via a simple text chat – a kind of ChatGPT for materials but grounded in scientific knowledge and calculated on-demand on the user's own computer.

Take copper and add a little tin: You have bronze, one of the oldest known alloys. Many other alloys work in a similar way: one or two metals as the main components and a few additives to improve the material’s properties. The situation is quite different for so-called high-entropy alloys. These novel materials consist of four or more metals mixed together in roughly equal proportions.
High-entropy alloys have promising mechanical, thermal, and electromagnetic properties that make them attractive for applications ranging from structural and high-temperature materials to functional coatings, catalysis, and electronic or magnetic components. However, their development is complex. “There are an extremely large number of possible combinations,” explains Empa researcher Yang Hu from Empa’s Advanced Metallurgy & Manufacturing of Metals laboratory in Thun. “Traditional methods of materials development, in which one produces and examines a few varieties of the material, quickly reach their limits here.”
A “virtual” scientist
The development of high-entropy alloys and other complex materials therefore increasingly relies on an emerging branch of materials science: computational materials science. Using calculations, simulations, and models, researchers can understand physical relationships in complex materials and identify which experiments are most likely to succeed. However, applying computational tools correctly requires a great deal of expertise. “Experimental materials scientists do not have the time to learn how to use our tools,” explains computational materials scientist Vladyslav Turlo, who works in the Computational Engineering lab at Empa.
To make these computational tools more easily accessible to their colleagues in the lab, Turlo and Hu jointly developed the OptiMat Chat platform, which allows scientists to download an “AI materials scientist” onto their computer. The first such virtual scientist, called OptiMat Alloys, selects appropriate domain-specific tools, executes predefined workflows, and summarizes the results in publication-style figures, tables, and texts for experts to review. To use OptiMat Alloys, researchers need neither to code nor to configure complicated settings. They can formulate their query in plain language – just as they would with a “real” computational materials scientist, or with ChatGPT.
The comparison with ChatGPT is apt: OptiMat Alloys uses a similar large language model (LLM) interface that can understand natural language and generate responses. But that is by no means the only foundation of this innovative research instrument. At the heart of OptiMat Alloys is a so-called AI agent. These digital assistants are capable of handling complex tasks and processing various inputs and outputs. Their key capability is interacting with different domain-specific tools – such as those used in computational materials science.
Rigorous, but not a replacement for human experts
“We have designed our agent to use our simulation tools for high-entropy alloys, and to do so according to the best-practice protocols,” explains Turlo. The OptiMat Alloys agent can be accessed via the OptiMat Chat platform and tasked with simulating specific types of alloys, thereby providing a theoretical foundation for future experiments. Unlike comparable AI tools, the results are stored in a database for traceability, reuse, and future benchmarking or refinement. In future, the researchers plan to add further agents to OptiMat Chat, such as those for ceramics and metal oxides or for amorphous materials. This will make the powerful tools of computational materials science more accessible in these areas as well.
Vladyslav Turlo and Yang Hu do not worry about this “artificial computational materials scientist” they have created taking their jobs. “Our agents handle routine calculations. This gives us more time to focus on the truly interesting research questions,” says Hu. Turlo adds: “Computational materials science is a relatively young field, and there is still an extraordinary amount to explore.”
In developing the platform and OptiMat Alloys, the researchers focused on reliability, traceability, and human oversight. The agent is designed to reduce hallucinations by using structured tool calls, predefined computational workflows, and stored calculation records. The scientific outputs come from connected computational tools, databases, and validated procedures rather than from the LLM alone. Its outputs are formatted as publication-style figures, tables, and summaries. “However, we by no means recommend creating scientific publications based solely on this data,” Turlo cautions. “The data are intended as starting points for expert interpretation and experimental planning rather than as standalone publishable scientific results.”
This is where the researchers also see the future of AI-supported research: AI can assist with literature reviews, design experiments, and generate data. However, human experts will still be needed to interpret and validate this data. “AI may be able to predict whether material A or material B is better for property C, but only human experts can dig deeper, validate and explain why that is the case,” summarizes Turlo.
Literature
Y. Hu, V. Turlo: OptiMat Alloys: a FAIR-aligned living database of multi-principal element alloys enabled by a conversational agent; npj Computational Materials (2026); doi: 10.1038/s41524-026-02377-3
Further information
Dr. Vladyslav Turlo
Computational Engineering
Phone +41 58 765 63 18
vladyslav.turlo@empa.ch
Dr. Yang Hu
Advanced Metallurgy & Manufacturing of Metals
Phone +41 58 765 63 68
yang.hu@empa.ch