Research map / Image generation / Autoregressive and latent generation
Vector Field Oriented Diffusion Model for Crystal Material Generation
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 GemsDiff, a diffusion model over full crystal geometry (atomic positions on a torus and lattice cells via a Lie-algebra vector space) using an equivariant GemsNet-based denoiser, plus a Fréchet ALIGNN Distance metric for crystal generation.
Problem
Generating crystal structures conditioned on chemical composition requires modeling both atomic positions and lattice cells, which prior generative models handle only partially (e.g., positions only, or cubic cells), and existing metrics (validity, density, energy) are limited for evaluating generat
Approach
Diffusion over the whole crystal geometry: atomic positions diffused on a torus (periodic space, shortest-path loss over lattice translations k*) and lattices diffused in a 9-dimensional Lie-algebra vector space via a log/exp map, denoised by an SL3(Z)- and E(3)-equivariant GemsNet-based autoencoder; plus a Fréchet distance computed on ALIGNN latent features.
Key results
- GemsDiff achieves 99.05% validity, EMD density 0.3331 g/cm^3 and EMD energy 0.1984 eV/atom, outperforming baselines on density and energy (CDVAE validity 100.0%).
- GemsDiff reaches FAD 1.509±0.115 versus 1.777 for CDVAE, 16.562 for FTCP and 30.692 for PGCGM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add joint diffusion over lattice cells and atomic positions | GemsDiff · Materials Project test set | density EMD: 0.6875 -> 0.3331$\pm$0.0327 | Helped |
| add joint diffusion over lattice cells and atomic positions | GemsDiff · Materials Project test set | validity: 100.0 -> 99.05$\pm$0.20 | Hurt |
| decode the lattice from the noisy cell instead of a cubic cell (ablation) | GemsDiff · Materials Project test set | validity: 99.17\% -> 15.23\% | Hurt |
| remove temporal embedding (ablation) | GemsDiff · Materials Project test set | validity: 99.17\% -> 98.78\% | Hurt |
| remove the GNN encoder and use the atom embedding directly (ablation) | GemsDiff · Materials Project test set | validity: 99.17\% -> 11.22\% | Hurt |
| diffuse directly on lattice matrices instead of the lattice vector space (ablation) | GemsDiff · Materials Project test set | EMD density: 0.289 -> 0.392 | Hurt |
| remove the shortest-path periodic constraint on atomic positions (ablation) | GemsDiff · Materials Project test set | FAD: 1.602 -> 1.691 | Hurt |
Surprising findings
- Validity is lower for GemsDiff than CDVAE (99.05% vs 100.0%), which the authors attribute to the limitations of the validity metric.
- Decoding the lattice from the noisy cell instead of a cubic cell, or removing the encoder, causes severe instability (validity 15.23% and 11.22%, density EMD in the hundreds or tens of thousands).
Limitations
- Validity is lower than CDVAE's, and the authors concede validity is a limited metric.
- The model is not designed for composition generation; it conditions on test-set compositions, so it is less suited to composition-free generation.
- FAD depends on ALIGNN features, which would change if ALIGNN were retrained.
Experimental setup
Datasets: Materials Project
Benchmarks: Materials Project subset (test split)
Metrics: validity EMD density EMD energy Fréchet ALIGNN Distance (FAD)
Baselines: CDVAE FTCP PGCGM G-SchNet P-G-SchNet
Builds on: denoising diffusion probabilistic models GemsNet CDVAE ALIGNN
Closest papers by meaning
- Periodic Materials Generation using Text-Guided Joint Diffusion Model (ICLR 2025 · similarity 0.92)
- Scalable Diffusion for Materials Generation (ICLR 2024 · similarity 0.92)
- Kinetic Langevin Diffusion for Crystalline Materials Generation (ICML 2025 · similarity 0.91)
- Equivariant Diffusion for Crystal Structure Prediction (ICML 2024 · similarity 0.90)
- WyckoffDiff -- A Generative Diffusion Model for Crystal Symmetry (ICML 2025 · similarity 0.90)
- Open Materials Generation with Stochastic Interpolants (ICML 2025 · similarity 0.89)
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.