Speaker
Description
The advent of novel technologies capable of producing small samples with complex shapes, such as additive manufacturing and 3D printing, has given rise to new challenges for defect detection and visualization techniques, including laser ultrasound imaging. When the ultrasound wavelength is comparable to the sample dimensions, multiple reflections can dominate the wave field, resulting in severe artifacts in conventional physics-based reconstructions. In this study, a simplified numerical test case is used to compare a physics-based reconstruction approach with a data-driven convolutional neural network for defect localization in samples exhibiting dominant higher-order scattering. The results indicate that, within the investigated setup, the data-driven approach can identify defect locations in cases where strong multiple reflections limit the applicability of physics-based reconstructions. For the considered synthetic dataset, accuracies of up to 95% were obtained on unseen test samples, and preliminary tests suggest that the trained model can generalize to selected cases outside the exact training grid. Overall, the study provides a first proof of concept that data-driven methods are a promising option for defect localization in laser-ultrasound scenarios dominated by higher-order scattering.