Gan Jiang: self-learning XRD agent, 96.3% phase ID, no retraining

Gan Jiang, a new scientific agent for powder X-ray diffraction, reports single-phase top-1 identification accuracy of 96.30% on the MP500 dataset without being told the sample's composition. The strongest comparator in the paper scored 58.00%. The figures come from a preprint posted to arXiv on October 6 by Bin Cao, Huichi Zhou and eight co-authors.

The agent sits on top of four analysis engines the group built in-house, named "XMatcher, XQueryer, XDecomposer and WPEM," which the authors say "span phase identification, multiphase decomposition and physics-constrained whole-pattern modelling." Gan Jiang does not refit the diffraction physics or retrain its language model. Instead, the paper writes, it "converts analytical experience into executable skills by diagnosing failures, revising skill instructions and code, and validating revisions before reuse."

On the authors' own DeltaXRDbench, Gan Jiang hit 96.30%, 81.78% and 40.83% on MP500, RRUFF and opXRD, "compared with 58.00%, 58.47% and 26.45% for the strongest comparator." The opXRD score covers real experimental patterns, and the gap between it and MP500 is the headline reminder that simulated data is still much easier than what comes off a lab instrument.

The paper also walks through specific case studies: resolving overlapping reflections, quantifying a five-phase Egyptian cosmetic sample, tracking battery lattice evolution, and analyzing disordered oxide catalysts.

DeltaXRDbench is introduced in the same preprint that reports the scores. No external group has yet published a reproduction.