Capability benchmark · Materials science · Chemistry · Computer science
Gan Jiang improves frozen diffraction-analysis workflows on held-out samples
A diffraction agent improves held-out refinement scores and phase identification across simulated and experimental powder patterns.
Summary
Gan Jiang connects phase retrieval, mixture decomposition and physics-constrained refinement. Frozen learned procedures raise mean FullProf refinement score from 46.90 to 72.38, with improvements in GSAS-II and PyWPEM. Without supplied composition, opXRD top-1 identification reaches 40.83% versus AutoXRD’s 26.45%. Case studies analyze measured samples, including a five-phase cosmetic and a battery series.
AI role
The agent revises executable skills and refinement procedures from failure traces, selects revisions on development data and freezes them before testing.
Narrative role
Extends research-agent evidence from executing fixed software workflows to learning analytical procedures that transfer to unseen samples.
Caveat
Developer-run preprint evaluation, not independent replication. Refinement scores are not correctness percentages. Adaptation differs by engine, benchmark mixtures are constructed, and longitudinal deployment improvement is not demonstrated.