TCSP3.0 fixes composition errors and lifts crystal prediction match rates
Original title: TCSP3.0: Template-Based Crystal Structure Prediction with Coefficient-Preserving Element Substitution and Structural Diversity
On T180, Top-1 rises from 28.3% to 72.2%, beating CrystalFlow and MaskGXT even before relaxation.
Auditing TCSP2.0 on the T180 benchmark found 469 of 2,347 selected structures, 20%, had the wrong composition because mappings ignored stoichiometric coefficients and sequential text substitution corrupted symbols. TCSP3.0 enforces coefficient-preserving mappings, substitutes species simultaneously, and validates composition and cell. It adds periodic-number retrieval, oxidation-state and radius scoring, deduplication, and ORB-v3 relaxation. Top-20 reaches 90.6% on T180 and 74.0% on a new 800-target benchmark.
- Without relaxation, Top-1 is 67.8% in a median of 28 seconds on two CPU cores.
- That unrelaxed run is about ten times faster than CrystalFlow.
- Ablations attribute most gains to the correctness fixes.
- Deduplication improves Top-5, while embedding retrieval is required for coverage.
- AI models
- ORB-v3CrystalFlowMaskGXT
From the original article
1 min read at arxiv.orgTemplate-based crystal structure prediction (CSP) transfers a known crystal structure to a new composition by element substitution, so its accuracy depends on retrieving suitable templates, mapping elements and ranking the resulting candidates correctly. Auditing the original TCSP2.
By Ying Feng · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.