Machine Learning Aided Fracture Surface Analysis of Additively Manufactured Lattice Structures Subjected to Uniaxial Tensile Loading

Abstract

Additive manufacturing (AM) is a continuously developing and growing manufacturing process that provides a means for more efficient and rapid production of 3D parts. A major application of AM includes production of lattice structures which are highly complex, periodic parts that are difficult to fabricate with conventional manufacturing methods. These lattice structures can be used in a multitude of industries because of their desirable mechanical properties and high strength-to-weight ratio and are particularly attractive for load-bearing biomedical implants for bone regeneration. The need for these structures as implants requires an understanding of their mechanical properties as well as their fracture behavior in the event of unexpected failure. This study examined and characterized the mechanical properties and fracture mechanics of cylindrical AM lattice structures subjected to uniaxial tension. These lattice structures were designed using two different unit cells, four porosity values, ten materials and six printing technologies. The structures were imaged using micro-computed tomography (micro-CT) and scanning electron microscopy (SEM) to observe the fracture surfaces. These fracture surfaces were then converted as inputs for predictive machine learning models that worked to determine the fourth order plane of best fit to the broken lattice structure. The results of the study showed that lattice structures fabricated using Multi Material Jetting (MMJ) had the best mechanical performance in terms of normalized lattice strength and normalized strain at break. Gyroid structures had superior mechanical properties to the octet unit cell structures. It was also observed that a power function can describe the relationship between material ultimate strength and mechanical ultimate strength. The machine learning models used were Kernel Ridge Regression, Polynomial Regression, and Random Forest, with the Random Forest demonstrating the lowest root mean squared error for test data at a value of 0.333.