CalPhaD Coffee Lecture – Agentic Workflows for Accessible and Reproducible CALPHAD Simulations

The CALPHAD community has developed high-quality thermodynamic databases and powerful simulation tools. However, using these resources often requires substantial expertise in software-specific interfaces, scripting, and the construction of complex calculation workflows. Recent advances in large language models (LLMs) and AI agents offer an opportunity to make CALPHAD methods more accessible while retaining the transparency and reproducibility required for scientific research.
In this lecture, the concept and utilization of agentic workflows will be introduced: computational processes in which AI agents select and connect scientific tools, evaluate intermediate results, and adapt subsequent steps to the task at hand. Using our pyiron aiflow plattform as an example, it will be shown how agents can assemble reusable calculation components into explicit, inspectable workflows rather than generating opaque, monolithic scripts. Such workflows can support tasks ranging from routine thermodynamic calculations and parameter studies to adaptive exploration of composition–temperature spaces and automated analysis of results.