Research map / Image generation / Autoregressive and latent generation
Generative Hierarchical Materials Search
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 GenMS, a hierarchical LLM-plus-diffusion-plus-GNN search that turns natural-language instructions into candidate crystal structures (perovskites, pyrochlores, spinels) and verifies them by energy prediction and DFT.
Problem
Generating crystal structures directly from high-level natural-language specifications is hard because no paired language-to-structure data exists and LLMs lack low-level atomic geometry knowledge.
Approach
Factorize the language-to-structure generator into a high-level LLM that proposes chemical formulae (with retrieval augmentation and deduplication) and a low-level diffusion model that generates structures conditioned on formulae, combined with a GNN energy heuristic and best-first multi-objective tree search; uses a compact A x 4 crystal representation for efficient diffusion.
Key results
- GenMS achieves 0.93 match rate for perovskites, 0.86 for pyrochlores, and 0.89 for spinels versus at most 0.37 for prompting baselines.
- On holdout compositions, GenMS's diffusion model reaches 0.93 DFT convergence versus 0.62 for UniMat, with mean E_f -0.49 vs -0.40.
- Best-of-N (N=10) with GNN selection improves formation energy over a single sample for over 80% of 1,000 compositions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GenMS instead of prompting an LLM to generate CIF files | Crystal-family generation · Perovskites | Structural validity: 0.66 -> 1.00 | Helped |
| Use GenMS instead of prompting an LLM to generate CIF files | Crystal-family generation · Pyrochlore | Match rate: 0.04 -> 0.86 | Helped |
| Use GenMS instead of prompting an LLM to generate CIF files | Crystal-family generation · Spinel | Uniqueness: 0.51 -> 0.44 | Hurt |
| Increase CIF examples in context from 1 shot to 25 shot | Prompting CIF baseline · Perovskites | Structural validity: 0.04 -> 0.66 | Helped |
| Increase CIF examples in context from 25 shot to Max | Prompting CIF baseline · Perovskites | Structural validity: 0.66 -> 0.22 | Hurt |
| Increase language branching from 1 to 25 | GenMS · Perovskite | DFT formation energy: -0.55 -> -2.76 | Helped |
| Increase structure branching from 1 to 100 | GenMS · Pyrochlore | DFT formation energy: -2.40 -> -2.51 | Helped |
| Add retrieval-augmented generation (ablation) | GenMS formula generation · Three crystal families | Valid formula rate: 0.97 -> 1.00 | Helped |
| Add retrieval-augmented generation (ablation) | GenMS formula generation · Three crystal families | Match rate: 0.72 -> 0.89 | Helped |
| Add compact crystal representation (ablation) | GenMS diffusion model · Holdout structures | DFT convergence rate: 0.62 -> 0.93 | Helped |
| Add compact crystal representation (ablation) | GenMS diffusion model · Holdout structures | Mean formation energy: -0.40 ± 0.06 -> -0.49 ± 0.03 | Helped |
| Use best-of-10 sampling instead of a single sample (ablation) | GenMS structure generation · 1,000 holdout compositions | Formation energy | Helped |
Surprising findings
- Few-shot prompting of Gemini to generate CIF files yields near-zero match rates for perovskites and pyrochlores at 1 shot, and adding too many CIF files in context hurts performance.
Limitations
- Less effective on complex structures such as MXenes and Kagome lattices.
- Uniqueness is low for pyrochlore and spinel (less than half of formulae unique), and synthesizability is not assessed.
- Structure matching may produce false positives since not all A2B2O7 compositions are pyrochlores.
Experimental setup
Datasets: Materials Project ICSD OQMD NOMAD Wikipedia
Benchmarks: Materials Project-based perovskite, pyrochlore, spinel evaluation holdout compositions following UniMat CrystaLLM unseen-composition test
Metrics: CIF validity structural validity composition validity formation energy (GNN and DFT) uniqueness match rate DFT convergence rate
Baselines: Prompting CIF with Gemini (1, 5, 25, max shot) UniMat CrystaLLM (small LLM) CrystaLLM (large LLM)
Builds on: latent diffusion models cascaded diffusion models UniMat CrystaLLM
Closest papers by meaning
- Scalable Diffusion for Materials Generation (ICLR 2024 · similarity 0.91)
- Fine-Tuned Language Models Generate Stable Inorganic Materials as Text (ICLR 2024 · similarity 0.91)
- FlowLLM: Flow Matching for Material Generation with Large Language Models as Base Distributions (NeurIPS 2024 · similarity 0.89)
- Periodic Materials Generation using Text-Guided Joint Diffusion Model (ICLR 2025 · similarity 0.89)
- All that structure matches does not glitter (NeurIPS 2025 · similarity 0.89)
- LLM Meets Diffusion: A Hybrid Framework for Crystal Material Generation (NeurIPS 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.