Generating 3D porous structures using machine learning and additive manufacturing

Abstract

Complex structures, often found in nature, may be difficult to replicate or integrate with human-made designs. Generative machine learning may be a useful tool in extracting and transferring complex structure features. A generative adversarial network (GAN) was trained using x-ray microtomography images of porous and lattice structures. Three types of cellular materials were used. Two-dimensional images were generated by the generative network at two resolutions. A bag of features approach was used to sequence the generated images of porous structures. The combination of 2D GAN method and similarity based stacking resulted in 3D structures. The approach aimed at economising on computational cost whilst ensuring a degree of continuity through the structure. The original and generated open cell porous structure images were binarized and 3D surfaces were created using imaging tools. The surfaces were transformed into solid geometries, using computer aided design tools and exported for 3D printing. The compressive behaviour of the specimens was compared. The method generated qualitatively similar structures of consistent relative densities. However the relative density and compressive response of the generated structures diverged in relation to the reduction in resolution. The method shows promise for biomimicking, or generating hybrid natural-artificial structures, based on training sets.