DiffCSP

By Tsinghua University, Renmin University of China, 2023. Materials. Oracle-trained generation.

Jointly diffuses unit-cell lattice matrices and fractional atomic coordinates under periodic E(3)-equivariant symmetry constraints.

The loop

Proposes
A periodic E(3)-equivariant denoising network jointly updates lattice vectors and fractional coordinates from a target composition.
Judged by
Periodic crystal geometry validators, RMSD matchers, and DFT energy evaluations score the generated polymorphs.
Steers
Space-group symmetry constraints and multi-sample energy ranking select the lowest-energy valid crystal structure.

Outperformed USPEX evolutionary crystal search and CDVAE on MP-20 and Perov-5 crystal structure prediction at a fraction of the compute time.

Requires specifying the target unit-cell stoichiometry upfront unless paired with a composition generator.

Uses a periodic E(3)-equivariant diffusion model over fractional coordinates and lattice matrices, ranked by physical energy.

Classification of DiffCSP
What it proposes	Crystal structure
Generative model	Periodic E(3)-equivariant joint diffusion network
Algorithmic authority	Space-group symmetry validator and DFT energy evaluator
Feedback	Score
What gets updated	Population of candidates
When it runs	At inference time
Validity	Approximate physics (simulation, not lab results)
Evidence	Code released
Lineage	Original system