Accelerating metamaterial topology optimization using deep super-resolution networks

TL;DR

This paper introduces a deep learning framework using super-resolution networks to efficiently generate high-resolution metamaterial topologies from low-resolution designs, significantly reducing computational costs in topology optimization.

Contribution

The novel use of an enhanced deep super-resolution network for metamaterial topology optimization enables accurate high-resolution design prediction from low-resolution inputs, reducing computational effort.

Findings

Predicts high-resolution topologies with 5-7% of traditional computational cost

Achieves accurate topology reconstruction validated by multiple metrics

Enables smoother, printable designs for 3D applications

Abstract

Designing metamaterials for extreme mechanical behavior involves the optimal selection of design parameters. However, identifying these optimal parameters through topology optimization (TO) across a large parametric space requires extensive computational resources. To address this challenge, we propose a novel deep learning framework for metamaterial topology optimization using an enhanced deep super-resolution (EDSR) approach. Generating low-resolution topologies significantly reduces computational cost compared to high-resolution designs. Therefore, an EDSR network is trained to learn the mapping between low- and high-resolution metamaterial topologies. The training dataset is generated using solid isotropic material with penalization (SIMP)-based TO. We demonstrate the proposed approach for the design of mechanical metamaterials targeting objectives such as maximization of bulk modulus, shear modulus, and elastic modulus, and minimization of Poisson's ratio. Quantitative assessments -including (i) pixel value error, (ii) objective function error, (iii) intersection over union, and (iv) volume fraction error -validate the accuracy of the EDSR-based TO. Our framework predicts high-resolution topologies of size $192 \times 192$ from optimized low-resolution topologies of size $48 \times 48$. Once trained, the proposed network predicts these high-resolution topologies with only $5-7\%$ of the computational cost required by conventional SIMP-based TO at the same resolution. Moreover, by adding upscale blocks, the framework can generate smoother, higher-resolution topologies suitable for 3D printing. This approach offers a scalable and efficient solution with strong potential for multidisciplinary metamaterial design applications.