Machine learning assisted investigation of defect influence on the mechanical properties of additively manufactured architected materials

Abstract

Additive manufacturing techniques can introduce defects that worsen the mechanical properties of 3D printed parts. Current techniques for quantifying the detrimental effects of these defects can only provide detailed analysis for a small number of geometries. Here, we investigate the effect of each defect feature (surface roughness and void position, number density and size) on the mechanical properties of a large number of truss lattices belonging to the stretch-dominated and bending-dominated topology. This is done by reducing each truss lattice into a single-beam sub-unit cell and conducting finite element simulations on it. The generated data is subjected to machine learning algorithms to identify the most important defect and design features that determine the mechanical properties of the overall structure. Our results indicate that surface roughness, Rmax (i.e. peak-to-trough height), exceeding 10% of the beam diameter strongly reduces the specific modulus and strength of lattice structures, especially for bending-dominated geometries. Interior voids, on the other hand, adversely affect stretch-dominated geometries but improve the specific properties of bending-dominated structures by removing under-stressed material in the core of the beams and causing them to become more “tube-like”. These insights are supported by first-principles analytical modeling and experimental data of additively manufactured metal lattices in the literature.