Research map / Probabilistic and scientific modeling / Molecules and proteins

Wyckoff Transformer: Generation of Symmetric Crystals

ICML 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Generative latent models for single-cell data

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.

Introduces WyFormer, a permutation-invariant Transformer encoder that autoregressively generates Wyckoff representations conditioned on space group to produce symmetric, stable crystal structures, and also predicts material properties without atomic coordinates.

Problem

Existing crystal generative models poorly handle internal crystal symmetry, making it hard to consistently generate stable, symmetrically valid crystal structures; Wyckoff-position-based representations are also fragmented by space-group-dependent letters.

Approach

Represent each crystal as an unordered set of tokens fusing chemical element, site symmetry (universal across space groups, encoded with spherical-harmonic descriptors and learned embeddings) and enumeration, with the space group as the initial token; train a position-free Transformer encoder with masked cascade prediction and shuffle augmentation so that it is permutation-invariant and autoregres

Key results

• WyFormer achieves 24% higher S.U.N. than the best baseline on a 10,000-structure DFT-evaluated sample (3.83% vs. baseline values reported by prior work, per the paper's comparison).
• WyFormer generates 186 novel unique templates per 1000 samples (WyFormerDiffCSP++) versus 10 for DiffCSP++ and 165 for WyCryst, with the lowest P1 fraction among top methods (1.46%).
• On MP-20 energy prediction WyFormer reaches 25 meV MAE and band gap 234 meV MAE without atomic coordinates; on AFLOW it achieves best thermal conductivity MAE (2.20).

Measured effects reported in the paper

Surprising findings

• Structures with P1 symmetry are found stable at a much higher rate by DiffCSP and FlowMM than they occur in nature, which the authors attribute more likely to flaws in stability estimation than to new asymmetric materials.
• Coverage and Property EMD protocol metrics are negatively correlated with structure novelty, so the authors argue the legacy protocol is misleading.

Limitations

• Stability estimated via formation-energy E_hull lacks configurational and vibrational entropy and depends on known structures.
• Novelty and diversity evaluation remains an open question; duplicate counting does not measure substantive diversity.
• Hard to ensure generated structures conform to a given stoichiometry; crystal structure prediction is future work.

Experimental setup

Datasets: MP-20 MPTS-52 AFLOW Materials Project

Benchmarks: MP-20 MPTS-52 AFLOW

Metrics: S.U.N. S.S.U.N. P1 percentage space group chi^2 novel unique templates novelty coverage property EMD validity MAE

Baselines: DiffCSP DiffCSP++ CrystalFormer SymmCD WyCryst FlowMM

Builds on: WyCryst DiffCSP++ pyXtal CHGNet

Closest papers by meaning

• WyckoffDiff -- A Generative Diffusion Model for Crystal Symmetry (ICML 2025 · similarity 0.94)
• SymmCD: Symmetry-Preserving Crystal Generation with Diffusion Models (ICLR 2025 · similarity 0.92)
• Generating Symmetric Materials using Latent Flow Matching (NeurIPS 2026 · similarity 0.91)
• Space Group Conditional Flow Matching (NeurIPS 2026 · similarity 0.90)
• Conformal Crystal Graph Transformer with Robust Encoding of Periodic Invariance (AAAI 2024 · similarity 0.89)
• Complete and Efficient Graph Transformers for Crystal Material Property Prediction (ICLR 2024 · similarity 0.89)