Smooth 3D transition cell generation based on latent space arithmetic

Lattice structures with multiple unit cell types diversify the property space by offering more design freedom, encouraging adaptation of metamaterials in engineering applications. It is essential to ensure structural connectivity and smooth transition among different cell types to avoid pre-mature failure. In this work, we propose a framework based on latent space operations to generate smoothly morphing and fully connected transition cells, addressing the current research gap in realising lattice designs of dissimilar unit cells. Latent embedding – a low-dimensional representation of the original microstructure – is obtained through a variational autoencoder. Different types of triply periodic minimal surface (TPMS) lattice were chosen as the targets to demonstrate the capability of the algorithm in handling complex 3D geometries within a physically restricted transition region. Both qualitative and quantitative evaluations are provided to illustrate the connectivity and geometric similarity of the generated transition. Benchmark comparisons against both analytical and existing machine learning (ML) based solutions indicate the superior efficacy and generality of the proposed framework.