CalPhaD Coffee Lecture – AI Agents in the CALPHAD Workflow: How Agentic Simulation Supports Alloy Design

Large language models can now call external software, read the output, and decide what to do next. These agentic workflows are starting to appear in computational materials science, and they raise practical questions for the CALPHAD community. Can an agent set up a meaningful equilibrium or Scheil calculation? How does it choose a database, and does it know when an alloy lies outside the composition range the database was assessed for? What happens when it is wrong?

In this lecture, I will show how to apply the agentic approach using alloy design as an example. Typically, alloy developers want to know which combinations of materials and processing methods will achieve a desired microstructure. To tackle this, the agent divides the larger problem into smaller, manageable parts. For instance, it looks at questions like phase stability, how precipitation occurs during heat treatment, and how the material behaves during solidification. This breakdown allows each specific question to be handed off to the most appropriate model for analysis. CALPHAD sits at the center of this chain, as many of the models that follow rely on the thermodynamic data it provides.

A central rule is that the language model never produces thermodynamic data itself. All numbers come from assessed databases or physical models, and every step is logged so that an expert can check how a result was obtained. I will close by discussing where agents currently fail, and how the CALPHAD community could help make them more reliable, for example, by providing machine-readable information on the validity range and uncertainty of thermodynamic descriptions.