머신러닝을 이용한 원하는 기계적 특성을 가지는 미세격자 구조물의 생성 및 설계
- Conference
- 한국정밀공학회 2023년도 추계학술대회
- Date
- 11월 2023
Machine learning algorithms are increasingly used for designing architectured materials with extreme properties. In this
study, we employ the variational autoencoders (VAE) to generate microlattice structures with desired mechanical
properties. We create unique datasets comprising 10,000 point cloud samples such as Kelvin foam and octet-truss by
manipulating design parameters such as voxel size and strut thickness. The stiffness matrices for the structures are
calculated using MATLAB. The datasets are split into training, validation, and test sets at a ratio of 8 : 1 : 1. During VAE
training, a contrastive loss term ensures that similar properties cluster closely in the latent space, with the regressor
predicting property. In the latent space, we generate new data by interpolating existing data with desired property. These
new data are subjected to surface reconstruction to produce 3D meshes, which are further validated through 3D printing
and mechanical testing. While our current framework relies on the Vanilla VAE and primarily focuses on the structural
property, future work might explore advanced models, potentially broadening the property domains beyond mechanical
properties to design multifunctional lattices tailored for varied applications.