Powder diffraction structure solving goes to AI agents: exact recovery on pXRD-Bench rises from 27% to 69%
Synopsis
Crystal structure determination from powder X-ray diffraction has long depended on manual trial and error, which is hard to plug into self-driving laboratories. The task is recast as a verifiable agentic search problem: candidate structures can be scored during inference by forward-simulated profile agreement and during evaluation by structural matching to a known reference. On the 200-problem simulated single-phase pXRD-Bench, an adapted agent raises selected-structure exact recovery from 27% to 69%, with further gains from repeated sampling and verifier-guided inference-time scaling.
Interpretation
Powder diffraction structure solving is recast as a verifiable agentic search task, so the agent can judge during inference whether a candidate holds up by forward-simulating its diffraction profile and comparing it with the measurement. Earlier automation routes included crystallographic software, database phase retrieval, physics-guided global search and generative models, yet reliable autonomous powder-diffraction structure determination remained elusive. On the 200-problem simulated single-phase pXRD-Bench, a tool-free single-call model reaches 27.0% overall exact recovery, while adding a coding-agent harness and a materials-science Python environment raises it to 52.0%.
Agent adaptation leaves model parameters fixed and changes only the surrounding tools, interface and reusable task knowledge, yet pushes exact recovery to 69.0%. The same frontier model scores 27.0% without tools, 52.0% with the materials-science coding environment, and 57.0% once GSAS-II and DICVOL14 are added. The final adapted-codex condition reaches 69.0% selected-structure exact matches on pXRD-Bench, with a capability package of one 139-word skill and a DICVOL adapter, trained on a 50-problem curriculum disjoint from the benchmark.
On a 20-problem representative subset, repeated sampling plus ranking candidates by diffraction-profile residual, without access to the hidden reference, turns the correct answer in the candidate pool into the top pick. Repeated sampling alone raises only oracle coverage, the probability that at least one trajectory produces the correct structure; realizing that gain needs an inference-time verifier that does not rely on the hidden reference. Across eight independent adapted-codex trajectories on the 20-problem panel, the candidate pool contained an exact structure in 18 cases, and a lowest-residual selector ranked an exact structure first in all 18.
Perspective
The demonstrated scope is the 200-problem simulated single-phase pXRD-Bench, the 20-problem representative subset used for inference-time scaling, and 15 measured divalent-metal tungstate multiphase patterns from the Precursor Genome; the measured part is a developmental evaluation because those same patterns informed earlier design decisions about the framework and synthetic generator.
In the measured comparison, conditions used a mutable model alias on different dates and consumed different amounts of inference-time compute; the additional Sr repeats were selected after inspecting initial outcomes and concern the same specimen. Direct framework access gives the lowest mean residual at 5.98%, better than the evolved v6 package at 6.71%, so the average fit improvement from evolution is limited.
