ED-CSP: Crystal Structure Prediction from Electron Diffraction

We study crystal structure prediction from known composition and atom count plus sparse, unindexed detector-plane electron diffraction (ED) spot sets. Most learned ED approaches instead predict crystallographic labels, reconstruct from indexed reflections, or search a finite structure library. ED-CSP combines a relational set encoder, permutation-invariant aggregation across views, and a periodic flow generator to predict the lattice and fractional atomic coordinates. We also introduce Electron Diffraction Crystal Structures (ED-CS), a 4.85-million-structure resource with simulated multi-view ED, deduplicated and filtered against CHILI-100K. On 2,075 held-out CHILI-100K materials, CHILI-only ED-CSP reaches 57.49% MR@5, compared with 52.92% for PXRD-conditioned PXRDGen at the same five-candidate budget. Warm-starting the complete model from a separate one-million-structure precursor raises MR@5 to 66.27%. Exact-formula lookup has no candidate for 1,024 queries, yet the warm-started model retains 53.52% MR@5 on this subset. Replacing the target ED with a non-isomorphic same-formula donor on 67 queries lowers mean MR@5 by 22.09 percentage points across five generation seeds, evidence that the prediction depends on query-specific diffraction. ED-CSP returns atomistic candidates for subsequent crystallographic refinement.