Research map / Graph learning / Graph self-supervised learning
Code-Generated Graph Representations Using Multiple LLM Agents for Material Properties Prediction
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 Rep-CodeGen, an evolutionary-style multi-agent LLM framework (Qwen agents) that generates code for crystal graph representations satisfying six symmetry/continuity constraints, and uses the resulting representation for material property prediction on JARVIS and Materials Project.
Problem
Hand-designed graph representations of crystalline materials satisfy only specific structural constraints (invariances, equivariances, Lipschitz continuity), and existing methods do not extend to new constraints without expert effort.
Approach
Treat graph-representation construction as code generation: three LLM-agent stages (crossover generation via plan and code agents, evaluation summary via a reasoning agent, and Boltzmann-weighted parent selection by constraint signatures and code length) iterate as an evolutionary algorithm, with no model weight updates.
Key results
- Rep-CodeGen's representation (six constraints) achieves best MAE across all listed tasks, e.g., JARVIS Energy Hull 253 meV vs 265 for PerCNet (~13.9% gain noted in text as Energy Hull improvement).
- Within 14,076 iterations, one code satisfying all six constraints was found from 56,304 generated codes (39.60% executed successfully).
- With Qwen plus Rep-CodeGen over 1,000 generations, 66.53% of executable codes improved over the initial code, versus 13.51% for Qwen alone on the five-constraint task.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the six-constraint generated representation instead of ComFormer (5) | Material property prediction · JARVIS Energy Hull | MAE (meV): 47 -> 43.3 | Helped |
| Use the six-constraint generated representation instead of ComFormer (5) | Material property prediction · JARVIS Total Energy | MAE (meV/atom): 28.8 -> 28.3 | Helped |
| Use the six-constraint generated representation instead of ComFormer (5) | Material property prediction · JARVIS BandGap | MAE (meV): 260 -> 253 | Helped |
| Use the six-constraint generated representation instead of ComFormer (5) | Material property prediction · Materials Project Formation Energy | MAE (meV/atom): 27.2 -> 27.2 | No clear effect |
| Use the six-constraint generated representation instead of PerCNet (5) | Material property prediction · Materials Project Formation Energy | MAE (meV/atom): 18.1 -> 18.0 | Helped |
| Use the six-constraint generated representation instead of ComFormer (5) | Material property prediction · Materials Project BandGap | MAE (meV): 193 -> 191 | Helped |
| Add Rep-CodeGen to GPT 3.5 | GPT 3.5 code generation · Five-constraint code generation | Improve Rate: 0.00% -> 5.41% | Helped |
| Add Rep-CodeGen to DeepSeek | DeepSeek code generation · Five-constraint code generation | Improve Rate: 0.00% -> 26.27% | Helped |
| Add Rep-CodeGen to QWen | QWen code generation · Five-constraint code generation | Improve Rate: 13.51% -> 66.53% | Helped |
Surprising findings
- GPT-3.5 alone degraded to satisfying fewer constraints, while DeepSeek alone produced no improvement over the initial code.
Limitations
- Only one code out of 56,304 satisfied all six constraints, so the search is costly and low-yield.
- Constraint tests are built on MP-20 and the new representation's advantage relies on a specific design (lattice-based neighbors, dihedral angles).
Experimental setup
Datasets: MP-20 JARVIS-DFT-2021.8.18 3D Materials Project-2018.6
Benchmarks: JARVIS-DFT Materials Project
Metrics: MAE constraint satisfaction rate improve rate
Baselines: CGCNN ALIGNN MatFormer PotNet PerCNet ComFormer GPT-3.5 DeepSeek
Builds on: FunSearch evolution through large models ComFormer PerCNet CGCNN
Closest papers by meaning
- Generative Hierarchical Materials Search (NeurIPS 2024 · similarity 0.86)
- FlowLLM: Flow Matching for Material Generation with Large Language Models as Base Distributions (NeurIPS 2024 · similarity 0.86)
- Offline Materials Optimization with CliqueFlowmer (NeurIPS 2026 · similarity 0.86)
- MatExpert: Decomposing Materials Discovery By Mimicking Human Experts (ICLR 2025 · similarity 0.85)
- Complete and Efficient Graph Transformers for Crystal Material Property Prediction (ICLR 2024 · similarity 0.85)
- Invariant Tokenization of Crystalline Materials for Language Model Enabled Generation (NeurIPS 2024 · similarity 0.84)
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.