Research map / Image generation / Autoregressive and latent generation
UniMate: A Unified Model for Mechanical Metamaterial Generation, Property Prediction, and Condition Confirmation
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 UniMate, a unified discrete-latent tokenizer plus partially frozen diffusion model that handles mechanical metamaterial topology generation, property prediction, and density condition confirmation, with a new benchmark.
Problem
Mechanical metamaterial design involves three modalities (3D topology, relative density, mechanical property), but existing ML models handle only two of them at a time, and no benchmark covers all tasks jointly.
Approach
Encode each modality into a shared discrete codebook latent space, align the modalities with a tripartite Sinkhorn optimal transport loss, and then run a score-based diffusion transformer over the latent tokens with known-token freezing so any subset of modalities can be given as context and the rest generated.
Key results
- Topology generation F_qua 2.74e-2 vs 16.94e-2 for best baseline SyMat (UniMate reports up to 80.2% improvement).
- Property prediction NRMSE_pp 2.44e-2 vs 2.57e-2 for MACE-ve (about 5.1% improvement).
- Condition confirmation NRMSE_cc 4.43e-2 vs 8.89e-2 for UniTruss (about 50.2% improvement).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UniMate instead of the strongest baseline for topology generation | Mechanical metamaterial topology generation · Mechanical metamaterial topology generation | F_qua: 16.94 -> 2.74 | Helped |
| Use UniMate instead of the strongest baseline for topology generation | Mechanical metamaterial topology generation · Mechanical metamaterial topology generation | F_cond: 32.71 -> 7.81 | Helped |
| Use UniMate instead of the strongest baseline for property prediction | Mechanical metamaterial property prediction · Mechanical metamaterial property prediction | NRMSE_pp: 2.57 -> 2.44 | Helped |
| Use UniMate instead of the strongest baseline for condition confirmation | Mechanical metamaterial condition confirmation · Mechanical metamaterial condition confirmation | NRMSE_cc: 8.89 -> 4.43 | Helped |
| Add latent-space unification rather than generating directly in raw MTR space (ablation) | UniMate topology generation · Mechanical metamaterial topology generation | F_qua: 0.0531 -> 0.0274 | Helped |
| Add latent-space unification rather than generating directly in raw MTR space (ablation) | UniMate topology generation · Mechanical metamaterial topology generation | F_cond: 0.1136 -> 0.0781 | Helped |
| Add TOT alignment (ablation) | UniMate latent-space generation · Mechanical metamaterial topology generation | F_qua: 0.0295 -> 0.0274 | Helped |
| Add TOT alignment (ablation) | UniMate latent-space generation · Mechanical metamaterial property prediction | NRMSE_pp: 0.0263 -> 0.0244 | Helped |
| Add partially frozen diffusion instead of vanilla diffusion (ablation) | UniMate generation · Mechanical metamaterial topology generation | F_qua: 0.0405 -> 0.0274 | Helped |
| Add partially frozen diffusion instead of vanilla diffusion (ablation) | UniMate generation · Mechanical metamaterial property prediction | NRMSE_pp: 0.0322 -> 0.0244 | Helped |
Surprising findings
- Partially frozen diffusion gives a 13.9% average gain over vanilla diffusion in ablation.
Limitations
- Dataset limited to 500 topologies with at most 20 vertices and three densities per topology.
- Baselines for condition confirmation were revised by forcing prediction-oriented models to output density, which is not their native design.
- Time efficiency is only medium-level according to the authors' own slope analysis.
Experimental setup
Datasets: dataset from prior work with 17087 topologies (500 selected)
Benchmarks: UniMate metamaterial benchmark
Metrics: F_qua F_cond NRMSE_pp NRMSE_cc
Baselines: CDVAE SyMat Equiformer ViSNet UniTruss MACE-ve
Builds on: VQ-VAE Sinkhorn algorithm score-based diffusion CDVAE uniTruss
Closest papers by meaning
- Energy-conserving equivariant GNN for elasticity of lattice architected metamaterials (ICLR 2024 · similarity 0.87)
- Designing Mechanical Meta-Materials by Learning Equivariant Flows (ICLR 2025 · similarity 0.84)
- LEIA: Learned Environment for Interactive Architected Materials (NeurIPS 2026 · similarity 0.82)
- OmniPhysGS: 3D Constitutive Gaussians for General Physics-Based Dynamics Generation (ICLR 2025 · similarity 0.82)
- MatExpert: Decomposing Materials Discovery By Mimicking Human Experts (ICLR 2025 · similarity 0.82)
- Unveiling the Potential of AI for Nanomaterial Morphology Prediction (ICML 2024 · similarity 0.82)
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.