A comparative study of principal component analysis and machine learning for semiconductor micro-defect detection using scanning acoustic microscopy

Accurate detection of microscopic defects in semiconductor structures is essential to ensure the reliability of next-generation electronic devices. This study presents a comparative evaluation of principal component analysis (PCA) and residual neural network (ResNet) methods for non-destructive defect detection using scanning acoustic microscopy (SAM). Artificial defects ranging from 10 μm to 500 μm were embedded in bonded silicon wafers, and ultrasonic A-scan signals were collected at multiple focal depths. Three types of input data (raw waveforms, frequency-domain signals, and merged multi-depth waveforms) were analyzed using C-mode imaging, PCA, and ResNet-based classification. PCA demonstrated stable performance across varying focal depths, especially for defects ≥20 μm, capturing dominant signal variations with minimal preprocessing. However, its sensitivity to sub-resolution defects (≤10 μm) was limited. In contrast, ResNet showed superior performance in detecting fine-scale defects under well-aligned focus conditions. However, the model performance tended to degrade under focal misalignment conditions.