ED-CSP AI predicts crystal structures from electron diffraction with 66% accuracy
New ML framework beats PXRDGen using sparse electron diffraction, not powder X-ray — with 4.85M training structures.
ED-CSP, developed by Germain Poloudenny, Yaël Frégier, and Arnaud Demortière, tackles the generative inverse problem of recovering periodic 3D crystal structures from sparse, unindexed electron diffraction (ED) observations. Unlike prior methods that predict crystallographic labels, reconstruct from indexed reflections, or retrieve from finite libraries, ED-CSP combines a relational set encoder, permutation-invariant multi-view aggregation, and a periodic flow generator to jointly predict lattice parameters and fractional atomic coordinates from multiple detector-plane ED spot sets.
Trained on ED-CS, a dataset of 4.85 million simulated multi-view ED crystal structures deduplicated across seven repositories, ED-CSP achieves a 57.49% MR@5 on 2,075 held-out CHILI-100K materials, outperforming PXRDGen (52.92%). Scaling the training data to a one-million-structure precursor boosts MR@5 to 66.27%, and the model still hits 53.52% on 1,024 compositions absent from its retrieval library, demonstrating genuine generative generalization. Ablation tests confirm that predictions depend on input diffraction patterns, not just composition: swapping ED observations with non-isomorphic structures drops MR@5 by 22.09 percentage points.
- ED-CSP outperforms PXRDGen on CHILI-100K materials: 57.49% vs 52.92% MR@5, scaling to 66.27% with more training data
- Trained on ED-CS dataset of 4.85 million simulated multi-view ED crystal structures, deduplicated across seven repositories
- Achieves 53.52% MR@5 on compositions absent from the training retrieval library, proving true generative capability beyond exact-formula retrieval
Why It Matters
Crystal structure prediction is vital for materials discovery; ED-CSP could accelerate battery and semiconductor R&D by making electron diffraction analysis AI-driven.