Determining the atomic structure of a crystalline material from its powder X-ray diffraction pattern is slow and often requires human expertise. This is a major bottleneck for self-driving laboratories that aim to discover new materials automatically. The problem is that existing software and AI models struggle to find the correct structure among many possibilities.
The researchers formulated structure determination as a search problem where candidate structures can be verified by comparing their simulated diffraction pattern to the measured one. They used large language model agents, AI systems that can reason and use software tools, to propose and refine candidate structures. The agents were adapted by learning from verifiable feedback on synthetic problems.
On a of 200 simulated patterns, the adapted agents recovered the correct structure for 69% of cases, up from 27% without adaptation. The approach also improved interpretation of real experimental patterns containing multiple phases, even recovering a phase whose structure was not in the reference library. This shows that AI agents can automate a key material characterization task.
