Process modeling in laser powder bed fusion towards defect detection and quality control via machine learning: The state-of-the-art and research challenges

In recent years, machine learning (ML) techniques have been extensively investigated to strengthen the understanding of the complex process dynamics underlying metal additive manufacturing (AM) processes. This paper presents a comprehensive review and discussion on the latest successful applications of ML to one category of metal AM processes, i.e., laser powder bed fusion or LPBF. This paper will focus on three aspects of LPBF including process modeling, in-situ process monitoring, defect detection, off-line process optimization, and on-line process control. Due to the multi-physics mechanisms of LPBF and associated heterogeneous process sensing, different ML techniques naturally play a significant role in discovering the patterns underlying sensing data. The unsupervised component analysis helps to fuse features extracted from sensing data to facilitate the efficiency of data processing and modeling. Supervised regression techniques are applicable to advancing the causal reasoning of relationship among process parameters, thermal dynamics, structural formation and evolution, and achieved property of printed parts, which is also termed as the process-thermal dynamics-structure-property (PTSP) relationship. Supervised classification and unsupervised clustering techniques can be applied to classify in-situ sensing data to detect defect occurrence and identify defect type (e.g., balling) and severity (e.g., porosity level, crack density). The obtained PTSP relationship can then be used as a basis for off-line optimization of process parameters to achieve better printing quality, while real-time processing of in-situ sensing data through advanced ML techniques (e.g., reinforcement learning) allows online feedback control. Knowledge gaps and future research directions in the three aspects are also identified in this paper.