[Paper Review] Guided Diffusion for Fast Inverse Design of Density-based Mechanical Metamaterials

Yanyan Yang, Lili Wang+6 more|arXiv (Cornell University)|Jan 24, 2024

Acoustic Wave Phenomena Research4 citations

TL;DR

This paper proposes a guided diffusion model that generates high-resolution (128³) voxel-based mechanical metamaterials in just 3 seconds, achieving target homogenized elastic properties via a self-conditioned diffusion process. The method enables fast inverse design, extreme property exploration, and diverse microstructure generation for multi-scale applications.

ABSTRACT

Mechanical metamaterial is a synthetic material that can possess extraordinary physical characteristics, such as abnormal elasticity, stiffness, and stability, by carefully designing its internal structure. To make metamaterials contain delicate local structures with unique mechanical properties, it is a potential method to represent them through high-resolution voxels. However, it brings a substantial computational burden. To this end, this paper proposes a fast inverse design method, whose core is an advanced deep generative AI algorithm, to generate voxel-based mechanical metamaterials. Specifically, we use the self-conditioned diffusion model, capable of generating a microstructure with a resolution of $128^3$ to approach the specified homogenized tensor matrix in just 3 seconds. Accordingly, this rapid reverse design tool facilitates the exploration of extreme metamaterials, the sequence interpolation in metamaterials, and the generation of diverse microstructures for multi-scale design. This flexible and adaptive generative tool is of great value in structural engineering or other mechanical systems and can stimulate more subsequent research.

Motivation & Objective

• To address the high computational cost and limited diversity in traditional topology optimization for high-resolution density-based mechanical metamaterials.
• To overcome the dependence on initialization and non-convexity in inverse homogenization by leveraging deep generative AI.
• To construct a comprehensive, high-resolution (128³) voxel-based metamaterial dataset with broad coverage of elastic moduli and Poisson’s ratio.
• To enable rapid, diverse, and geometrically connected microstructure generation for multi-scale mechanical system design.
• To facilitate exploration of extreme mechanical properties and sequence interpolation in metamaterials.

Proposed method

• A self-conditioned 3D U-Net-based diffusion model is trained on a large-scale, high-resolution (128³) voxel-based mechanical metamaterial dataset to generate microstructures with specified homogenized elastic tensors.
• The dataset is constructed using the LIVE3D framework for topology optimization, covering bulk modulus, shear modulus, and Poisson’s ratio across volume fractions from 0.2 to 0.9.
• An iterative perturbation algorithm enhances dataset diversity by targeting underrepresented Poisson’s ratio ranges with Gaussian noise-based local elastic deformations.
• The diffusion model uses a U-Net backbone with five levels (64³ to 4³), ResNet blocks with 3×3 convolutions, and a final convolution layer to predict surface-occupancy values from latent features.
• The model is conditioned on target elastic tensor matrices, enabling fast inference with only 3 seconds of generation time per sample.
• Symmetric operations on the 1/8 unit cell lattice ensure cubic symmetry and reduce design space complexity.

Experimental results

Research questions

• RQ1 Can a deep generative diffusion model achieve fast and accurate inverse design of 128³ voxel-based mechanical metamaterials with specified homogenized elastic properties?
• RQ2 How can iterative perturbation strategies improve the diversity and coverage of elastic properties in a voxel-based metamaterial dataset?
• RQ3 To what extent can a self-conditioned diffusion model generate geometrically connected and mechanically functional microstructures across extreme property ranges?
• RQ4 Can the proposed method enable efficient sequence interpolation and multi-scale design of metamaterials?
• RQ5 What is the trade-off between generation speed, resolution, and accuracy in inverse design of density-based metamaterials?

Key findings

• The proposed guided diffusion model generates 128³ voxel mechanical metamaterials with target homogenized elastic tensors in just 3 seconds, achieving high fidelity and diversity.
• The dataset comprises 74,717 microstructures after removing disconnected components, with 10,000 uniformly sampled for training to ensure efficiency and coverage.
• The iterative perturbation method successfully expanded the dataset’s coverage of Poisson’s ratio, especially in underrepresented ranges, enhancing overall property diversity.
• The method enables rapid exploration of extreme mechanical properties, such as high bulk and shear moduli, and facilitates sequence interpolation across property spaces.
• The self-conditioned diffusion model demonstrates strong generalization, generating diverse, geometrically connected microstructures suitable for multi-scale design.
• The framework supports fast, flexible, and adaptive inverse design, significantly reducing computational burden compared to traditional topology optimization.