Authors
E. Ross, D. Hambleton
Abstract
The promise of computer aided manufacturing (CAM) is to make materializable structures that could not be fabricated using traditional methods. An example is 3D lattices, which may be arbitrarily complex. Variation in the lattice geometry and print media can define a vast spectrum of resulting material behaviour, ranging from fully flexible forms to completely stiff examples with high strength. Panetta et al. (2015) outline a methodology to generate lattice geometries with specified material properties. However, their method relies heavily on finite element analysis of beam models to determine the material properties of a discretely sampled space of lattices. In the present study, we use machine learning to perform a stiffness analysis on highly symmetric lattice geometries with periodic boundary conditions. We train a graph convolutional network on a dataset of lattices sampled continuously from the space of all lattices, then use the trained model to predict deflections for previously unseen lattices. With this approach we are able to approximate the material behaviour of the vast space of all lattice geometries, which offers potential for real-time material feedback at the design stage. It also offers a method to explore a space of building components that are materially sparse yet offer high strength and stiffness. The symmetry of the lattice geometry together with the stiffness analysis creates a homogenized material model, which can be applied to different designs to obtain similar material performance. We illustrate the approach with several examples across different scales.
