Real-Time Porosity Prediction for Metal Additive Manufacturing using Convolutional Neural Networks

W. Young, S. Ho, S. Al Jufout, M. Mozumdar, M. Buchholz, W. Zhang, K. Dajani
California State University, Long Beach,
United States

Keywords: anomaly detection, convolutional neural network, deep learning, metal additive manufacturing, porosity prediction, recurrent neural network, thermal images analysis


The development of 3D printing technologies has had a significant impact across the automotive, biomedical, construction and aerospace industries. While additive manufacturing (AM) can both lower manufacturing costs and increase efficiency, the structural properties of additively-manufactured parts are not as reliable as for traditional manufacturing methods. To this end, understanding and modelling the thermal profiles of the melt pool is essential to predicting AM process anomalies. Recent progress in Machine Learning (ML) algorithms and pattern recognition can provide a data-driven framework for improving the quality of metal printers through in-situ anomaly detection and improved control algorithms. The aim of this research is to propose an end-to-end deep learning pipeline which can accurately predict porosity within Ti-6AL-4V, 3D printed parts based upon thermal images. (I did not receive a confirmation e-mail about my first submission, so this is second attempt in case the first did not go through.)