Deep Learning for Process Monitoring and Defect Detection of Laser-Based Powder Bed Fusion of Polymers

Additive Manufacturing (AM) is increasingly leveraging Deep Learning (DL) to enhance process monitoring, defect detection, and predictive simulation. This paper synthesizes our results in applying DL to laser-based powder bed fusion of polymers (PBF-LB/P), fo-cusing on four key architectures: convolutional neural networks (CNNs) for real-time spa-tial anomaly detection, recurrent neural networks (RNNs/LSTMs) for capturing temporal dynamics, generative models (GANs and autoencoders) for unsupervised anomaly detec-tion and data augmentation, and physics-informed neural networks (PINNs) for embed-ding governing equations into predictive models. Each approach demonstrates distinct advantages: CNNs deliver high accuracy, LSTMs capture evolving defects, GANs mitigate data scarcity, and PINNs improve generalizability, yet critical limitations persist, includ-ing heavy reliance on labelled datasets, instability of generative models, limited interpret-ability, and lack of scalability for real-time industrial deployment. This paper delineates a roadmap for advancing DL-driven monitoring of the PBF-LB/P process from academic feasibility to robust industrial practice, with particular emphasis on resource efficiency and the promotion of both ecological and economic sustainability.