Answers are grounded in this paper and cite their sources.
Scanning Acoustic Microscopy (SAM) is widely used for non-destructive evaluation of semiconductor devices. The current approach for defect detection mainly for the void detection in die attach layer of the power device has been performed using signal-based machine learning method, which operate directly on acoustic signals. However, in many practical scenarios, signal data may not be readily accessible, inspiring the development of image-based approaches. This study presents the development of an image-based deep learning method for the void segmentation in die attach layer of the power device in SAM C-scan images. Additionally, it provides a systematic comparison with an existing signal-based machine learning approach. Multiple UNet architectures, including transformer-based variants such as TransUNet and Swin-UNet, were applied and evaluated. UNet with an EfficientNetB3 backbone achieved the best performance amongst all, achieving Intersection over Union (IoU) score of 83%. Furthermore, experiments with reduced training data demonstrated that the image-based model can achieve comparable performance with approximately 30 annotated images, with a minor reduction in IoU score. Comparative analysis shows that the image-based approach is data intensive comparatively, while the signal-based method requires significantly fewer labeled samples and lower data preparation effort. Although both methods demonstrate comparable qualitative performance, their suitability depends on application constraints, including data availability, computational resources, and deployment requirements. The findings provide practical insight into the trade-offs between signal-based and image-based machine learning methods for SAM-based defect detection in power device.
No takes yet. Share an insight, caveat, or question.
Rathod et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: