Recognition and classification of single melt tracks using deep neural network: A fast and effective method to determine process windows in selective laser melting

Abstract

The identification of process windows in additive manufacturing (AM) out of a vast parameter space is a daunting task, especially for a large variety of feedstock powders used in selective laser melting (SLM). Despite many numerical simulations of the SLM process, the process window for each type of metal powder is typically determined through a sequential and time-consuming trial-and-error approach. Here, we present a fast and effective strategy that single melt tracks in SLM are produced and assessed in a high-throughput fashion using a computer vision approach. Using this strategy, we investigate the possibility of using deep neural network (DNN) models, as an image processing method, to conduct automatic assessments of SLM processing parameters. We identified the optimal laser power and scanning speed in the SLM process for 316L stainless steel and pure copper powders. Our trained models have achieved the highest mean average precision of 0.70 in a time-efficient manner. This method provides a versatile toolbox for accelerating the quality assessments of melt tracks in SLM with minimum human input, benefiting to the automatic AM parametric analysis and optimization.