A machine Learning approach for image classification for additively manufactured parts

Additive Manufacturing (AM) technology is one of the most promising processes for the production of complex shape parts. A number of issues still remain unsolved, among which the fundamental one is the definition of a model able to predict the mechanical behavior of additively manufactured components starting from the knowledge of the alloy microstructure and the process-induced defects. Ongoing research aims at optimizing the use of AM technologies in several industrial fields, like medical, aerospace and mechanical. This goal requires in-depth characterization of the alloy microstructure and of the metallurgical defects in terms of morphology and distribution, the determination of mechanical properties of AM-produced specimens (like fatigue of materials, fatigue damage, fracture toughness,...), and the development of a predictive model based on artificial intelligence (AI) algorithms. The first step requires the ability of classifying images obtained from specimens produced with different process parameters and, consequently, presenting various mechanical properties. In this paper, this goal is pursued by means of a totally automatized procedure based on advanced methods of machine learning (ML); the first results, obtained on real specimens fabricated using Electron Beam Powder Bed Fusion (EB-PBF), are promising, showing the classifier ability of obtaining satisfactory results after a training on limited number of images.