Physics-Informed Ultrasonic C-scan Data Augmentation for Defect Detection in Metal Additive Manufacturing Parts

Metal additive manufacturing can introduce internal defects such as porosity, inclusions, and cracks, yet supervised ultrasonic C-scan detectors are hard to deploy at process introduction, where labeled images are scarce and rare defect classes are imbalanced. We propose a physics-informed C-scan augmentation method that explicitly models six ultrasonic effects, automatically synthesizing labeled images over ten classes. All evaluation is confined to the synthetic domain. Within that scope we verify that input physical parameters are quantitatively transferred to the rendered images: defect contrast increases monotonically with the theoretical reflection coefficient across six impedance levels, and the attenuation coefficient recovered from the images agrees with the input value. The best detector reached [email protected] of 0.705 at a fixed compute budget, and targeted synthetic augmentation of the most signal-sensitive class raised its [email protected] from 0.516 to 0.583, while a same-size neutral control gained only 0.009. Public measured laser-ultrasonic amplitude maps of an additively manufactured specimen are included as a quantitative external reference: under a common domain-distance measure the measured domain lies farther from the training distribution than any synthetic shift tested here, which bounds applicability and quantifies the adaptation scale required in future work.