Self-learning agent Gan Jiang hits 96.3% top-1 on XRD phase ID, beating expert skills
Yangtze-ailab · hf · 2026-10-08
Yangtze-ailab released Gan Jiang, a self-learning agent for powder X-ray diffraction analysis that converts analytical experience into reusable, self-improving skills.
- Ecosystem: Built on the authors' diffraction toolchain — XMatcher, XQueryer, XDecomposer, and WPEM — covering phase identification, multiphase decomposition, and physics-constrained whole-pattern modeling.
- Self-learning loop: The agent diagnoses failures, revises skill instructions and code, and validates revisions before reuse, without retraining the LLM or altering the underlying physical models.
- Results: Frozen skills outperform original expert-designed ones on refinement scores across FullProf, GSAS-II, and PyWPEM; without supplied composition, single-phase top-1 accuracy reaches 96.30% (MP500), 81.78% (RRUFF), and 40.83% (opXRD), versus 58.00%, 58.47%, and 26.45% for the strongest comparator.
- Demonstrations: Resolving strongly overlapping reflections, quantifying a five-phase ancient Egyptian cosmetic, tracking lattice evolution in an operating battery, and comparing atomic configurations in a disordered oxide catalyst.
The work shows how an integrated scientific tool ecosystem enables agents to accumulate validated analytical expertise that transfers to new samples.
More from Research
- KAIST Proposes SQAM: Scalar Adjoint Speeds Up Flow Policy RL, +18–35pp Success on OGBench — kaist-ai · 2026-10-08
- Debate Erupts Over 'Mathematicians Did Not Ask for This' in Terence Tao Blog Post — littmath · 2026-10-08
- Three papers claiming Hodge conjecture progress withdrawn over a sign error — ziv_ravid · 2026-10-08
- IBM Integrates Spyre AI Accelerator as a Native PyTorch Device via Existing Abstractions — PyTorch · 2026-10-08
- ResearchTrails: Mining Git Histories Into Human Research Decision Trajectories to Train Better Research LLMs — ChenhaoTan · 2026-10-08
- Toby Ord's interpolation-extrapolation-hyperpolation trichotomy: why AI fails off the hyperplane — tobyordoxford · 2026-10-08