Failure classification of porous additively manufactured parts using Deep Learning

Abstract

Microstructural features are one of the most important factors that determine performance and failure of engineering structures. In metal parts produced through Additive Manufacturing (AM), pores formed during deposition are one such feature that can dominate material response. These pores can be a result of keyhole-mode melting, gas porosity contained within the powder particles prior to deposition, or poor process parameters. Predicting the mechanical reliability of as-built AM structures containing such pores is often challenging, even with modern computational methods. Here we show that Deep Learning can be used to predict performance based on pore structure in AM components. We used a three-dimensional convolutional neural network to predict peak load in simulated tensile specimen gauge sections with different realizations of porosity sampled from an experimentally-measured statistical distribution. The algorithm was also tested in two additional simulation scenarios: a modified specimen geometry and a different nominal stress state. The algorithm remained predictive in these scenarios, which were outside the bounds of its training data. Our results show that Deep Learning can effectively learn microstructural features and patterns that have important implications in part performance.