Detecting microscale surface imperfections in powder bed fusion through light scattering and machine learning – validation of inspection principles

Microscale surface imperfections in laser beam powder bed fusion (PBF-LB) additively manufactured parts, such as balling, spattering, and surface pores, can substantially reduce component quality but are difficult to detect with current real-time measurement and monitoring methods. This paper introduces a novel, rapid, and cost-effective method for detecting microscale surface imperfections in PBF-LB, utilising light scattering combined with machine learning (ML) algorithms. In the proposed method, a laser beam illuminates the measured surface, and the scattered light is captured and analysed to detect surface imperfections. The scattering patterns, which are associated with the illuminated surface and the configuration of the setup, are used to train unsupervised ML algorithms, including autoencoders and anomaly detection models, to classify surfaces as either uniform, without any imperfections or non-uniform, with imperfections. The ML models were trained on simulated scattering patterns of synthetic surfaces generated by a generative adversarial network (GAN) and validated on experimental datasets. The use of unsupervised models eliminates the need for data labelling, whilst the use of simulated and synthetically generated data reduces the time required for actual experiments and data collection. Experimental validation demonstrates that the most effective trained ML model achieved a classification accuracy of over 97 %, highlighting the potential of this technique for detecting microscale surface imperfections. This paper demonstrates the capability of our method to detect such imperfections on PBF-LB surfaces as an ex-situ process. Nonetheless, with further development, this approach has the potential to be adapted as on-machine and real-time defect detection method, by integrating the illumination source into a commercial PBF-LB machine and capturing scattered light information for real-time quality monitoring during the manufacturing process.