ED-CSP: AI Predicts Crystal Structures from Electron Diffraction
A new machine learning framework called ED-CSP has been introduced to predict periodic 3D crystal structures using sparse and unindexed electron diffraction (ED) data. This innovative approach, detailed in an arXiv paper (2608.06448), addresses a challenging generative inverse problem in materials science. Instead of relying on crystallographic labels or indexed reflections, ED-CSP predicts crystal structures by considering chemical composition, atom counts, and various ED spot sets. It employs a relational set encoder and a periodic flow generator to estimate lattice parameters and atomic coordinates. The researchers also developed the ED-CS dataset with 4.85 million simulated multi-view ED structures, ensuring no overlap with CHILI-100K. Notably, training on 2,075 materials from CHILI-100K demonstrates its promise in improving crystal structure determination.
Key facts
- ED-CSP is a machine learning framework for crystal structure prediction from electron diffraction.
- It predicts structures from chemical composition, atom count, and multiple detector-plane ED spot sets.
- The framework uses a relational set encoder, permutation-invariant multi-view aggregation, and a periodic flow generator.
- Training dataset ED-CS contains 4.85 million simulated multi-view ED crystal structures.
- Dataset deduplicated across seven materials repositories and filtered to exclude CHILI-100K overlaps.
- Model tested on 2,075 held-out CHILI-100K materials.
- Paper available on arXiv with ID 2608.06448.
- Announcement type is cross.
Entities
Institutions
- arXiv