Research map / Image generation / Autoregressive and latent generation
WyckoffDiff -- A Generative Diffusion Model for Crystal Symmetry
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes WyckoffDiff, a discrete diffusion (D3PM) model over Wyckoff-position-based protostructures (space group plus element occupancies) with a GNN denoiser, evaluated with a new Fréchet Wrenformer Distance and used to find low-energy candidate materials.
Problem
Most crystal generative models ignore symmetry and model atoms with continuous coordinates, producing low-symmetry, less realistic structures; symmetry-based (protostructure) generation is underexplored.
Approach
Represent a crystal as a space group plus discrete Wyckoff-position occupancies (constrained positions get one element, unconstrained get per-element counts), and denoise these discrete variables with D3PM using a space-group-agnostic GNN (WyckoffGNN) that operates on a complete graph of Wyckoff positions.
Key results
- WyckoffDiff-zeros achieves FWD 0.48±0.02 on WBM vs 0.83 for DiffCSP++ and 1.47 for SymmCD, and 119 novel materials/min vs 46 for DiffCSP++ and 115 for SymmCD.
- Among 10k novel protostructures, WyckoffDiff-unif. yields 1214±32 unique novel prototypes vs 527 for DiffCSP++ and 780 for SymmCD.
- Materials discovery pipeline: from 200 selected generated structures, DFT found CsSnF6 below the convex hull and NaNbO2 and Ca2PI on the hull, with seven other fluorides distinctly below the hull.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the uniform prior with the empirical marginal prior (ablation) | WyckoffDiff · WBM | FWD: 2.29±0.15 -> 1.65±0.07 | Helped |
| replace the uniform prior with an all-zero prior (ablation) | WyckoffDiff · WBM | FWD: 2.29±0.15 -> 0.48±0.02 | Helped |
| replace the uniform prior with the empirical marginal prior (ablation) | WyckoffDiff · WBM | novelty (%): 40.2±0.40 -> 55.7±1.95 | Mixed |
| replace the uniform prior with an all-zero prior (ablation) | WyckoffDiff · WBM | novelty (%): 40.2±0.40 -> 30.2±0.97 | Hurt |
| replace the uniform prior with the empirical marginal prior (ablation) | WyckoffDiff · Carbon24 | FWD: 0.78±0.14 -> 0.78±028 | No clear effect |
| replace the uniform prior with the empirical marginal prior (ablation) | WyckoffDiff · Carbon24 novel protostructures | unique and novel prototypes: 237±25 -> 290±48 | Helped |
Surprising findings
- Initializing the discrete diffusion from a uniform prior severely underperforms marginal or all-zero priors, even though the model is meant to denoise.
- Novelty of WyckoffDiff is lower than baselines, but it produces more novel materials per unit compute because it does not generate full geometries.
Limitations
- Generates only protostructures; full geometry requires additional steps such as Pyxtal realization and MACE relaxation.
- Novelty is lower than baselines, suggesting some memorization of the training distribution.
- Complete-graph message passing scales poorly with the number of Wyckoff positions and with max atoms per position P and max atomic number N_a.
Experimental setup
Datasets: WBM Carbon24 Materials Project
Benchmarks: WBM Carbon24
Metrics: Fréchet Wrenformer Distance novelty uniqueness novel prototypes novel materials per minute compositional validity energy above hull (DFT)
Baselines: CDVAE DiffCSP++ SymmCD
Builds on: D3PM DiGress CDVAE DiffCSP++ Wrenformer
Closest papers by meaning
- SymmCD: Symmetry-Preserving Crystal Generation with Diffusion Models (ICLR 2025 · similarity 0.94)
- Wyckoff Transformer: Generation of Symmetric Crystals (ICML 2025 · similarity 0.94)
- Generating Symmetric Materials using Latent Flow Matching (NeurIPS 2026 · similarity 0.91)
- LLM Meets Diffusion: A Hybrid Framework for Crystal Material Generation (NeurIPS 2025 · similarity 0.91)
- Kinetic Langevin Diffusion for Crystalline Materials Generation (ICML 2025 · similarity 0.91)
- Scalable Diffusion for Materials Generation (ICLR 2024 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.