Research map / Large language models / Language, safety and interpretability
MatExpert: Decomposing Materials Discovery By Mimicking Human Experts
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 MatExpert, a three-stage (retrieval, transition, generation) LLM framework that mimics human materials experts, using contrastive T5 retrieval and LoRA-finetuned Llama models to generate crystal structures as CIFs.
Problem
Generating new solid-state crystal structures that meet user-specified property constraints; existing LLM and diffusion methods generate in a single step and struggle to integrate textual property descriptions with structural information.
Approach
Decompose material design into retrieval of a similar existing material via contrastive learning between property and RoboCrystallographer text embeddings (two T5 encoders), a GPT-4-distilled transition-pathway step (chain-of-thought), and LLM generation of lattice and atomic coordinates converted to CIF, all within one conversation.
Key results
- On MP-20 unconditional generation, MatExpert with Llama-3 70B (tau=0.7) reaches 0.998 structure validity, 51.0% metastable and 12.0% stable, beating Crystal-LLM 70B (49.8%, 10.6%).
- On MP-20, MatExpert's density Wasserstein distance drops to 0.18 versus 0.81 for Crystal-LLM 70B (tau=0.7).
- On NOMAD conditional generation, MatExpert with Llama-3 70B (tau=0.7) achieves 0.999 structure validity and 0.986 composition validity versus 0.997 and 0.971 for Crystal-LLM 70B.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MatExpert instead of Crystal-LLM with LLaMA-2 70B (τ = 0.7) | Unconditional generation on MP-20 · MP-20 | wdist(ρ): 0.81 -> 0.18 | Helped |
| Use MatExpert instead of Crystal-LLM with LLaMA-2 70B (τ = 0.7) | Unconditional generation on MP-20 · MP-20 | stable rate (DFT): 10.6% -> 12.0% | Helped |
| Use MatExpert instead of Crystal-LLM with LLaMA-2 70B (τ = 0.7) | Conditional generation · NOMAD | wdist(ρ): 0.77 -> 0.14 | Helped |
| Use MatExpert instead of Crystal-LLM with LLaMA-2 70B (τ = 0.7) | Conditional generation · NOMAD | coverage precision: 0.988 -> 0.985 | Mixed |
| Use T5 rather than BERT as the retrieval encoder (ablation) | MatExpert retrieval stage · Material retrieval | top-1 accuracy: 59.96 -> 71.35 | Helped |
| Use T5 rather than BERT as the retrieval encoder (ablation) | MatExpert retrieval stage · Material retrieval | average rank: 2.36 -> 1.59 | Helped |
| Add the transition (CoT) stage (ablation) | MatExpert conditional generation · NOMAD condition satisfaction | condition satisfaction rate | Helped |
| Add the retrieval stage with contrastive learning (ablation) | MatExpert conditional generation · NOMAD condition satisfaction | condition satisfaction rate | Helped |
Surprising findings
- Retrieval-based transition from existing materials gives much better distribution metrics than Crystal-LLM, which the authors attribute to fewer LLM hallucinations.
- Crystal-LLM's novelty decreases as model size grows, while MatExpert maintains high novelty regardless of size.
Limitations
- Paper's NOMAD dataset and code are promised but not released; the evaluation relies on M3GNet estimates for stability, with DFT only on metastable subsets.
- Coverage recall is lower than baselines (e.g., CDVAE and LM-AC are best on coverage), and the paper concedes it does not win all metrics.
- Inconsistent model naming (LLaMA-2 vs Llama-3 8B) and dataset size (2,886,120 vs 2,928,355) across sections.
Experimental setup
Datasets: NOMAD MP-20 Materials Project
Benchmarks: MP-20 NOMAD
Metrics: structure validity composition validity COV-Recall COV-Precision Wasserstein distance density Wasserstein distance number of elements metastable rate (M3GNet) stable rate (DFT) diversity novelty
Baselines: CDVAE Crystal-LLM LM-CH LM-AC
Builds on: Crystal-LLM CDVAE RoboCrystallographer LoRA NT-Xent
Closest papers by meaning
- FlowLLM: Flow Matching for Material Generation with Large Language Models as Base Distributions (NeurIPS 2024 · similarity 0.89)
- Generative Hierarchical Materials Search (NeurIPS 2024 · similarity 0.88)
- Offline Materials Optimization with CliqueFlowmer (NeurIPS 2026 · similarity 0.87)
- MADE: Benchmark Environments for Closed-Loop Materials Discovery (ICML 2026 · similarity 0.87)
- Periodic Materials Generation using Text-Guided Joint Diffusion Model (ICLR 2025 · similarity 0.87)
- Fine-Tuned Language Models Generate Stable Inorganic Materials as Text (ICLR 2024 · similarity 0.86)
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.