Key points are not available for this paper at this time.
Quality assessment in additive manufacturing (AM) relies on defect detection models, whose performance is limited by the lack of defect data and the high imbalance between classes. To address this, this paper introduces a physics-informed multimodal generative adversarial network (GAN) framework for synthetic defect data augmentation. It combines three innovations including physics-constrained models with thermodynamic consistency between visual and thermal representations, cross-modal attention mechanisms with realistic correlations between data features and adaptive augmentation strategies with dynamically computed synthetic-to-real ratios depending on model uncertainty and validation accuracy. The framework, when evaluated on a real-world AM process monitoring dataset of 1500 high-resolution layer images, achieved 89.2% defect detection accuracy, a 12.8% improvement over no augmentation and a 6.7% improvement over state-of-the-art baseline models. On average, recall for rare defect detection increased by 31.6%, particularly for the most severe cases, where the training prevalence was less than 5%. The documentation openly describes an initial training failure, with the discriminator achieving a fake accuracy of 1.00. The paper then details the systematic corrections applied: reducing the discriminator learning rate by 5x, automatically optimising the batch size and adding a regularisation method. This transparent account of failure and resolution makes the documentation highly valuable to practitioners.
Dhote et al. (Tue,) studied this question.