Research map / Image generation / Autoregressive and latent generation
Periodic Materials Generation using Text-Guided Joint Diffusion Model
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 TGDMat, a text-guided joint diffusion model that jointly generates atom types, fractional coordinates, and lattice for periodic crystals, conditioned on Robocrystallographer or short text prompts via MatSciBERT.
Problem
Existing equivariant diffusion models for crystal generation do not jointly learn atom types, fractional coordinates and lattice end-to-end, and cannot be conditioned on user-specified material descriptions.
Approach
Jointly diffuse lattice (DDPM), atom types (D3PM absorbing state) and fractional coordinates (wrapped normal score matching) with a periodic-E(3)-equivariant CSPNet backbone, injecting a MatSciBERT embedding of a text description into the initial atom features at every denoising step.
Key results
- On CSP with 1 sample on MP-20, TGDMat (Long) reaches 55.15% match rate vs 51.49% for DiffCSP.
- On Carbon-24 Gen, TGDMat (Long) achieves EMD density 0.043 vs 0.0805 for DiffCSP.
- On MP-20 Gen, TGDMat (Long) compositional validity 92.97% vs 83.25% for DiffCSP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TGDMat (Long) instead of DiffCSP | CSP · Perov-5, k=1 | Match Rate: 52.02 -> 90.46 | Helped |
| Use TGDMat (Long) instead of DiffCSP | CSP · Carbon-24, k=20 | Match Rate: 88.47 -> 95.27 | Helped |
| Use TGDMat (Long) instead of DiffCSP | CSP · Perov-5, k=20 | Match Rate: 98.60 -> 98.59 | No clear effect |
| Use TGDMat (Long) instead of SyMat+ | Random Material Generation · Carbon-24 | COV-P: 97.63 -> 92.43 | Hurt |
| Use TGDMat (Long) instead of DiffCSP+ | Random Material Generation · Carbon-24 | Property EMD (ρ): 0.0763 -> 0.043 | Helped |
| Use TGDMat (Long) instead of SyMat | Random Material Generation · MP-20 | COV-P: 99.97 -> 99.95 | Hurt |
| Use long detailed descriptions instead of short prompts (ablation) | TGDMat · CSP Perov-5, k=1 | Match Rate: 56.54 -> 90.46 | Helped |
| Use long detailed descriptions instead of short prompts (ablation) | TGDMat · Gen MP-20 | Compositional validity: 86.60 -> 92.97 | Helped |
| Reduce sampling steps from 1K to 500 | TGDMat (Long) · CSP Perov-5 | Match Rate | No clear effect |
Surprising findings
- Shorter text prompts degrade Gen performance relative to long Robocrystallographer descriptions, and TGDMat with 500 sampling steps can match DiffCSP with 1K steps.
Limitations
- Relies on text descriptions derived from existing CIF-based datasets rather than independent expert prompts.
- MatSciBERT is frozen; larger LLM encoders not explored, limiting prompt flexibility.
- Performance drops on MP-20 for property/formula matching of generated structures (e.g., 70.54% formula match).
Experimental setup
Datasets: Perov-5 Carbon-24 MP-20 Robocrystallographer descriptions (derived) pymatgen-derived short prompts
Benchmarks: Perov-5 Carbon-24 MP-20
Metrics: Match Rate RMSE compositional validity structural validity COV-R COV-P EMD property statistics (num elements, density, energy) property-matching percentage
Baselines: P-cG-SchNet CDVAE DiffCSP SyMat CDVAE+ SyMat+ DiffCSP+
Builds on: DiffCSP CSPNet D3PM DDPM CDVAE
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)
- Scalable Diffusion for Materials Generation (ICLR 2024 · similarity 0.91)
- Kinetic Langevin Diffusion for Crystalline Materials Generation (ICML 2025 · similarity 0.90)
- Open Materials Generation with Stochastic Interpolants (ICML 2025 · similarity 0.90)
- WyckoffDiff -- A Generative Diffusion Model for Crystal Symmetry (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.