EP-Flow: Disordered Crystal Structure Prediction without Site-Level Annotations
EP-Flow predicts disordered crystal structures from formulas without site-level disorder labels.
EP-Flow generates disordered crystal structures from chemical formulas without requiring site-level disorder annotations. It represents substitution, vacancies, and interstitial occupancy with an Occupancy Distribution Matrix constrained to a formula-dependent transportation polytope, then learns a marginal-preserving flow and recovers feasible occupancies with a Sinkhorn inverse map. Jointly generating occupancies, fractional coordinates, and lattice parameters, it reports state-of-the-art results on formula-conditioned benchmarks derived from COD and MPDS and outperforms adapted ordered-crystal generators. Analyses indicate it recovers sparse, chemically meaningful local disorder rather than only global composition statistics.
- EP-Flow predicts disordered crystals from formulas without site annotations.
- Occupancy Distribution Matrices encode mixing, vacancies, and interstitials.
- Flow matching uses a shared space and a Sinkhorn inverse map.
- State-of-the-art on COD- and MPDS-derived disordered CSP benchmarks.
- Tags
- crystal-structure-predictionflow-matchinggenerative-modelsmaterials-scienceep-flowdisordered-crystals
- AI models
- EP-Flow
From the original article
1 min read at arxiv.orgGenerative models have made rapid progress in ordered crystal structure prediction, yet many functional materials are intrinsically disordered, with substitutional mixing, vacancies, or interstitial species controlling their properties.
By Qiuliang Liu · 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.