Many familiar alloys are simple in structure: one or two main metals plus small additives. But there are also so-called high-entropy alloys - materials in which four or more metals are mixed in roughly equal proportions. They have promising mechanical, thermal and electromagnetic properties useful for structures, high-temperature technology, coatings, catalysis and electronics. However, selecting such compositions is difficult: the number of possible combinations is enormous, and traditional experiments quickly hit a limit.
Computational materials science comes to the rescue - a field where the properties of materials are studied using calculations and models. But to use such tools correctly, serious training is needed. "Experimenters do not have time to learn our programs," explains researcher Vladislav Turlo from the Empa laboratory of computational engineering.
The solution was the OptiMat Chat platform, created by Turlo together with colleague Yan Hu. It allows users to download an "AI materials scientist" onto their computer. The first such virtual specialist, OptiMat Alloys, selects the necessary tools itself, performs specified chains of calculations and formats the results as tables, graphs and texts resembling publication-quality output.
The user does not need to program or configure complex parameters - it is enough to formulate a request in ordinary language. The similarity to ChatGPT is no accident: the interface uses a large language model (LLM) - a system that understands natural speech and generates responses. But the main thing in OptiMat Alloys is the so-called AI agent, a digital assistant capable of performing complex tasks and interacting with specialized tools.
"We made it so that the agent applies our programs for modeling high-entropy alloys according to validated protocols," Turlo says. The results are saved in a database - for traceability, reuse and future comparisons. This distinguishes the development from similar AI tools.
The authors emphasize: the agent reduces the risk of "hallucinations" - invented answers - through structured tool calls, predefined computational scenarios and calculation records. Scientific data are taken from connected programs and databases, not from the language model alone. In the future, they plan to add agents for ceramics, metal oxides and amorphous materials to OptiMat Chat.
The work was published in the journal npj Computational Materials. The researchers are not afraid that AI will replace them: "Agents take on routine calculations, and we have more time left for interesting scientific questions," Hu says. Turlo warns: the agent's data are a starting point for expert interpretation and experiment planning, not ready-made publications. According to him, AI can predict which material is better, but explaining and verifying why is a human task.