Research map / Image generation / Autoregressive and latent generation
Scalable Diffusion for Materials 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 UniMat, a periodic-table-based 4D tensor representation of crystals that lets a 3D U-Net diffusion model generate materials, evaluated with proxy metrics and DFT-based formation and decomposition energies.
Problem
Generating stable novel crystal materials with generative models scales poorly with atom count when inter-atomic relationships are modeled explicitly, and learned proxy metrics are unreliable for judging stability.
Approach
Represent any crystal as a 4D tensor indexed by periodic-table period, group and per-element atom index, storing xyz coordinates (null location for absent atoms), and train a diffusion model with interleaved 3D convolution and attention over this tensor; evaluate with DFT per-composition formation energy and convex hull decomposition energy.
Key results
- On MP-20 proxy metrics UniMat achieves best property statistics, coverage and composition validity, though structure validity (97.2) is below CDVAE (100).
- Against CDVAE, UniMat materials have average -0.216 eV/atom lower formation energy on overlapping compositions, with 86.3% of pairs favoring UniMat.
- Against AIRSS on unseen compositions, conditional UniMat reaches 0.81 convergence rate versus 0.55 for AIRSS (estimated overall 0.817 in appendix).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UniMat instead of CDVAE for unconditional generation | Unconditional material generation · MP20 per-composition pairs | Formation energy difference (eV/atom): -0.216 eV/atom | Helped |
| Use UniMat instead of CDVAE for unconditional generation | Unconditional material generation · MP20 per-composition pairs | Formation energy reduction rate: 0.863 | Helped |
| Use UniMat instead of CDVAE for unconditional generation | Unconditional material generation · MP 2021 convex hull | Stable materials: 56 -> 414 | Helped |
| Use UniMat instead of CDVAE for unconditional generation | Unconditional material generation · MP 2021 convex hull | Metastable materials: 90 -> 2157 | Helped |
| Use UniMat instead of CDVAE for unconditional generation | Unconditional material generation · GNoME convex hull | Stable materials: 1 -> 32 | Helped |
| Condition generation on composition using UniMat instead of AIRSS initialization | Composition-conditioned structure generation · Zero-shot composition-conditioned DFT evaluation | Convergence rate: 0.55 -> 0.81 | Helped |
| Condition generation on composition using UniMat instead of AIRSS initialization | Composition-conditioned structure generation · Zero-shot composition-conditioned DFT evaluation | Formation energy difference (eV/atom): -0.68 eV/atom | Helped |
| Increase UniMat feature dimension (ablation) | UniMat unconditional generation · MP20 | Validity and coverage | Helped |
| Use UniMat instead of CDVAE | Unconditional material generation · MP20 | Structure validity: 100 -> 97.2 | Hurt |
Surprising findings
- Learned formation-energy predictors can fail to reproduce DFT-calculated relative stabilities, so proxy metrics may be misleading for materials discovery.
Limitations
- Representation is sparse for small chemical systems, incurring computational cost.
- Proxy metrics are limited; DFT verification used only for a subset of generated structures.
- AIRSS and UniMat conditional settings could be further optimized for convergence rate; results are initial signals.
Experimental setup
Datasets: Perov-5 Carbon-24 MP-20 GNoME Materials Project 2021
Benchmarks: Perov-5 Carbon-24 MP-20
Metrics: structure validity composition validity COV recall COV precision density energy number of elements per-composition formation energy difference formation energy reduction rate number of stable materials via decomposition energy
Baselines: CDVAE language model crystal generation (LM) AIRSS
Builds on: denoising diffusion probabilistic models CDVAE 3D U-Net classifier-free guidance ab initio random structure search
Closest papers by meaning
- LLM Meets Diffusion: A Hybrid Framework for Crystal Material Generation (NeurIPS 2025 · similarity 0.92)
- Vector Field Oriented Diffusion Model for Crystal Material Generation (AAAI 2024 · similarity 0.92)
- Periodic Materials Generation using Text-Guided Joint Diffusion Model (ICLR 2025 · similarity 0.91)
- Generative Hierarchical Materials Search (NeurIPS 2024 · similarity 0.91)
- All that structure matches does not glitter (NeurIPS 2025 · similarity 0.91)
- WyckoffDiff -- A Generative Diffusion Model for Crystal Symmetry (ICML 2025 · 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.