Multimodal machine learning model detects keyhole pores at high speed

Universities collaborate with Argonne National Laboratory to enhance laser powder bed fusion.

Engineers from Northwestern University, the University of Virginia, Carnegie Mellon and Argonne National Laboratory have made a breakthrough in defect detection for laser powder bed fusion (LPBF).

Using a set of widely available sensors in combination with machine learning, the team has achieved over 90% accuracy in detecting keyhole pore formation with a temporal resolution of 0.1 milliseconds. This could be the key to developing true closed-loop control systems for LPBF machines and a faster, more reliable qualification and certification process for parts made with metal additive manufacturing (AM).

Keyhole pores – microscopic defects which form as the result of trapped gases within the melt pool – can significantly weaken metal AM parts, reducing their service life and making them unsuitable for demand applications. The speed and complexity of the LPBF process has made it difficult to detect keyhole pores accurately and in real time.


Image: Zhongshu Ren, Jiayun Shao, Tao Sun at Northwestern University.

To address this challenge, the research team used a combination of microphones and photodiodes to monitor the LPBF process and, with some help from machine learning, detect when and where keyhole pores form. The core of this approach lies in measuring the oscillations of the keyhole – a vapor depression formed in the melt pool during printing.

Using high-speed synchrotron x-ray imaging to establish a precise “ground truth” the researchers trained a machine learning model to recognize conditions that lead to pore formation. The potential impact of this research is considerable, especially given the relatively low cost and wide availability of their equipment.

Future work will aim to improve the accuracy of the method by integrating additional sensors. Additionally, while the study achieved high accuracy in single-track laser melting experiments, future efforts will need to extend this approach to whole additive parts, including analyzing pore movements within the melt pool and assessing pore removal during repeated melting cycles.

The research is published in the journal Materials Futures.

Written by

Ian Wright

Ian is a senior editor at engineering.com, covering additive manufacturing and 3D printing, artificial intelligence, and advanced manufacturing. Ian holds bachelors and masters degrees in philosophy from McMaster University and spent six years pursuing a doctoral degree at York University before withdrawing in good standing.