Research map / Probabilistic and scientific modeling / Variational and generative sampling
Designing Mechanical Meta-Materials by Learning Equivariant Flows
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.
Designs 2D cellular solids by deforming a space-group-symmetric reference shape with a symmetry-preserving, volume-preserving neural network flow, optimized through a differentiable IGA nonlinear mechanics solver.
Problem
Cellular solid (mechanical metamaterial) design has been limited to translational periodicity of pore unit cells; the goal is to expand the design space to all 2D crystallographic (wallpaper) symmetry groups while keeping volume fraction fixed to target nonlinear mechanical responses such as auxetic
Approach
Define shapes implicitly as a flow of a reference symmetric shape, where the flow's vector field is built from a neural network via symmetrization over non-compact space groups (a finite-coset decomposition) and a divergence-free curl-like operator, so every shape is symmetric and volume-preserving by construction.
Key results
- Under uniaxial tension, p2gg symmetry achieved linear stress-strain responses for all target betas except 1.5, while p1 reached only up to beta 1.1.
- Among eight groups targeting Poisson's ratio -0.5, p4 reached -0.45 versus -0.05 for p1 and -0.12 for pg.
- Fabricated pg and p4 prototypes measured Poisson's ratios of -0.14 and -0.49, in strong agreement with simulation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use p4 rather than p1 symmetry | Cellular-solid design framework · Tension, target νef=-0.5 | effective Poisson's ratio: -0.05 -> -0.45 | Helped |
| Use pg rather than p1 symmetry | Cellular-solid design framework · Tension, target νef=-0.5 | effective Poisson's ratio: -0.05 -> -0.12 | Helped |
| Use p2mg rather than p1 symmetry | Cellular-solid design framework · Tension, target νef=-0.5 | effective Poisson's ratio: -0.05 -> -0.40 | Helped |
| Use p4 symmetry for the fabricated design | Cellular-solid design framework · Tension, p4 design | experimental Poisson's ratio: -0.49 | Helped |
| Use pg symmetry for the fabricated design | Cellular-solid design framework · Tension, pg design | experimental Poisson's ratio: -0.14 | Helped |
| Use pg rather than p6 symmetry | Cellular-solid design framework · Compression, negative Poisson's ratio | effective Poisson's ratio: -0.15 -> -0.2 | Helped |
| Explore higher-order symmetries in addition to translational symmetry | Cellular-solid design framework · Tension, target νef=-0.5 | effective Poisson's ratio | Helped |
| Use p1 symmetry for the compression force-displacement design | Cellular-solid design framework · Uniaxial compression, target plateau stress-strain response | target-response achievement | Helped |
Surprising findings
- Under compression, the only symmetry group that enabled the designed plateau responses was p1, not higher-order symmetries.
- The p2gg group was best for linear tension responses, outperforming p1 in range of achievable betas.
Limitations
- Scaling to 3D cellular solids is limited by the expensive mechanics simulation at large degrees of freedom.
- Experiments are restricted to 2D cross-sections and a single 7x7 cellular solid size with 50% volume fraction.
- Compression designs were only explored for specific groups; some plateau targets were achievable only with p1.
Experimental setup
Metrics: effective Poisson's ratio nominal stress-strain curve error
Baselines: p1 translational symmetry designs
Builds on: neural ODE isogeometric analysis JAX
Closest papers by meaning
- Energy-conserving equivariant GNN for elasticity of lattice architected metamaterials (ICLR 2024 · similarity 0.89)
- LEIA: Learned Environment for Interactive Architected Materials (NeurIPS 2026 · similarity 0.85)
- A Space Group Symmetry Informed Network for O(3) Equivariant Crystal Tensor Prediction (ICML 2024 · similarity 0.85)
- Space Group Equivariant Crystal Diffusion (NeurIPS 2025 · similarity 0.84)
- UniMate: A Unified Model for Mechanical Metamaterial Generation, Property Prediction, and Condition Confirmation (ICML 2025 · similarity 0.84)
- Generating Symmetric Materials using Latent Flow Matching (NeurIPS 2026 · 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.