Research map / Image generation / Autoregressive and latent generation
LLM Meets Diffusion: A Hybrid Framework 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 CrysLLMGen, a hybrid that samples atom types and coarse coordinates/lattice from a LoRA-fine-tuned LLaMA-2-7B, retains the atom types, and refines coordinates and lattice with a DiffCSP-style CSPNet diffusion model for crystal generation.
Problem
Generating periodic crystal structures requires jointly modeling discrete atom types and continuous coordinates and lattice parameters; LLMs get composition right but structure poorly, while equivariant diffusion models get structure right but composition poorly.
Approach
Use an LLM to produce atom types, fractional coordinates and lattice as an intermediate representation, keep the atom types, and inject the continuous variables into a pre-trained equivariant diffusion model at an intermediate timestep tau to refine them, with the two modules trained independently and in parallel.
Key results
- On MP-20 de novo generation, CrysLLMGen (7B) reaches 99.94 structural and 93.55 compositional validity, versus 93.55 for LLaMA-2 and 89.40 for UniMat among baselines.
- On the S.U.N. evaluation with 1,000 generated structures on MP-20, CrysLLMGen reports 16.79% stable and 9.21% S.U.N., above LLaMA-2 (12.67%, 4.84%) and DiffCSP++ (8.58%, 6.55%).
- In CSP on MP-20, CrysLLMGen achieves match rate 74.24 versus 73.22 for LLaMA-2 (7B) and 61.39 for FlowMM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine LLaMA-2-7B generation with diffusion refinement of coordinates and lattice | CrysLLMGen · MP-20 Gen | COV-Precision: 99.32 -> 99.84 | Helped |
| Combine LLaMA-2-7B generation with diffusion refinement of coordinates and lattice | CrysLLMGen · MP-20 Gen | Compositional validity: 93.55 -> 93.55 | No clear effect |
| Combine LLaMA-2-7B generation with diffusion refinement of coordinates and lattice | CrysLLMGen · MP-20 S.U.N. | % Stable: 12.67 -> 16.79 | Helped |
| Combine LLaMA-2-7B generation with diffusion refinement of coordinates and lattice | CrysLLMGen · MP-20 S.U.N. | % S.U.N.: 6.76 -> 9.21 | Helped |
| Combine LLaMA-2-7B generation with diffusion refinement of coordinates and lattice | CrysLLMGen · Perov-5 CSP | Match Rate: 97.60 -> 98.30 | Helped |
| Combine LLaMA-2-7B generation with diffusion refinement of coordinates and lattice | CrysLLMGen · Carbon-24 CSP | Match Rate: 97.21 -> 98.81 | Helped |
| Combine LLaMA-2-7B generation with diffusion refinement of coordinates and lattice | CrysLLMGen · MP-20 CSP | Match Rate: 61.39 -> 74.24 | Helped |
| Combine LLaMA-2-7B generation with diffusion refinement of coordinates and lattice | CrysLLMGen · MP-20 CSP | RMSE: 0.0566 -> 0.1276 | Hurt |
| Combine LLaMA-2-7B generation with diffusion refinement of coordinates and lattice | CrysLLMGen · MPTS-52 CSP | Match Rate: 39.96 -> 42.16 | Helped |
| Use diffusion refinement to generate structures from prompted atomic composition | CrysLLMGen · MP-20 conditional generation | Space group match rate: +42% | Helped |
Surprising findings
- Its MP-20 density EMD (0.972) is worse than DiffCSP (0.352) and UniMat (0.088), so the hybrid does not win on all property statistics.
Limitations
- No interaction or feedback between the LLM and diffusion modules; they are trained independently.
- Only LLaMA-2-7B and a vanilla DiffCSP-style diffusion model were explored; larger LLMs were left to future work.
- No error bars or statistical significance reported for results.
Experimental setup
Datasets: MP-20 Perov-5 Carbon-24 MPTS-52
Benchmarks: MP-20 Perov-5 Carbon-24 MPTS-52
Metrics: structural validity compositional validity COV-Precision COV-Recall density EMD element number EMD match rate RMSE % metastable % stable
Baselines: CDVAE DiffCSP DiffCSP++ MatterGen UniMat SymmCD FlowMM FlowLLM CrysBFN LLaMA-2 (7B)
Builds on: DiffCSP LLaMA-2 CSPNet FlowLLM LoRA
Closest papers by meaning
- Scalable Diffusion for Materials Generation (ICLR 2024 · similarity 0.92)
- Periodic Materials Generation using Text-Guided Joint Diffusion Model (ICLR 2025 · similarity 0.92)
- Fine-Tuned Language Models Generate Stable Inorganic Materials as Text (ICLR 2024 · similarity 0.91)
- WyckoffDiff -- A Generative Diffusion Model for Crystal Symmetry (ICML 2025 · similarity 0.91)
- SymmCD: Symmetry-Preserving Crystal Generation with Diffusion Models (ICLR 2025 · similarity 0.90)
- FlowLLM: Flow Matching for Material Generation with Large Language Models as Base Distributions (NeurIPS 2024 · similarity 0.90)
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.