Benchmark / XRD phase identification
SciPhys runs a public XRD benchmark across clean powders, textured thin films, multiphase mixtures and substrate-containing scans. Each case exposes the expected phases, recovered phases and failure reasons.
8/8
Cases passed
100.0%
Top-1 accuracy
100.0%
Mean phase recall
100.0%
Strong peak coverage
0.0%
False-positive rate
PASS
Suite status
A match must explain peaks. The benchmark checks recovered phases and how much of the strong observed signal is covered.
Cases include anatase/rutile, iron oxide mixtures and ZnO on Si so the engine is tested beyond single-phase examples.
The same suite runs from the CLI, so XRD algorithm changes can be checked before shipping.
Canonical ZnO peaks with small zero-shift and intensity jitter.
Top phase hit
Required phases recovered with sufficient strong-peak coverage.
A c-axis textured ZnO pattern dominated by (002), similar to lab thin-film scans.
Top phase hit
Required phases recovered with sufficient strong-peak coverage.
Anatase polymorph should not be mislabeled as rutile.
Top phase hit
Required phases recovered with sufficient strong-peak coverage.
Rutile polymorph should not collapse into anatase when the main 27.45 degree peak is present.
Top phase hit
Required phases recovered with sufficient strong-peak coverage.
Two TiO2 polymorphs must both survive the multiphase ranking.
Top phase hit
Required phases recovered with sufficient strong-peak coverage.
Primary ZnO with a strong substrate contribution should keep Si as a secondary phase.
Top phase hit
Required phases recovered with sufficient strong-peak coverage.
Common iron-oxide ambiguity case with overlapping peaks.
Top phase hit
Required phases recovered with sufficient strong-peak coverage.
Fluorite oxide control case with well-separated principal peaks.
Top phase hit
Required phases recovered with sufficient strong-peak coverage.
What comes next
The public suite is intentionally small and readable. Internally, SciPhys can extend it with real lab cases, expert-reviewed corrections, low-signal scans and instrument-specific edge cases.
The goal is not to hide uncertainty. The goal is to show which claims are backed by measured peaks, which require expert review and which should be rejected before they enter lab memory.