A Compact Review on AI-Driven Defect Detection in Polymer Additive Manufacturing

The rapid expansion of polymer additive manufacturing (AM) is severely bottlenecked by the formation of process-induced macro-scale (warping, delamination) and micro-scale (porosity, micro-cracks) defects. Traditional ex-situ inspections are inefficient, leading to structural failures and significant material waste. The primary objective of this review is to systematically evaluate the transition from post-production quality control to real-time, AI-driven in-situ monitoring, and to propose a physics-grounded roadmap for future reliability. Based on the assumption that purely data-driven models lack cross-platform generalizability due to their "black-box" nature, this paper analyzes state-of-the-art computer vision architectures and micromechanics-coupled machine learning. The comparative synthesis demonstrates that advanced single-stage detectors, particularly YOLOv۱۱ variants and Chroma-YOLO, achieve exceptional precision (mAP > ۸۳%) for real-time surface anomaly detection at high frame rates. Conversely, predicting internal structural integrity relies on X-ray Computed Tomography (XCT) fused with algorithms like Random Forests, achieving over ۹۶% accuracy in fatigue life prediction. To transcend the limitations of purely empirical data, the study concludes that the future of AM reliability depends on integrating Physics-Informed Neural Networks (PINNs) to embed thermal partial differential equations directly into AI loss functions. Furthermore, coupling these networks with non-local mechanics, specifically Peridynamics, provides a robust, mesh-free methodology to predict micro-crack propagation, paving the way for autonomous, zero-defect manufacturing.

کلیدواژه ها:

Additive Manufacturing ، Defect Detection ، Computer Vision ، Physics-Informed Neural Networks