Anomaly Detection for Nondestructive Defect Detection and Localization in Acoustic Signals

Abstract Scanning acoustic microscopy (SAM) is a critical nondestructive technique for detecting buried defects in semiconductor devices [1]. However, interpreting the complex acoustic data requires expert knowledge, especially as microelectronic structures grow increasingly intricate. This study presents a novel, unsupervised approach for defect detection in SAM data using an autoencoder neural network trained exclusively on intact reference samples. The method exploits reconstruction error as an indicator of signal anomalies, identifying potential defects without prior knowledge of their characteristics or the need for defective training samples. The resulting similarity parameter facilitates intuitive visualization of outlier regions through overlay on high-resolution SAM images. Furthermore, by analyzing the spatial and temporal deviation of the reconstructed signals, the method enables localization of the signal interval from which the deviation originates. This provides a rough estimation of the defect depth, adding an additional layer of interpretability to the detected anomalies. This enhances defect detection sensitivity and operator support, making SAM more accessible for less experienced users and highly promising for automation in production environments. While the method effectively highlights signal deviations, it may also flag non-defect-related anomalies such as sample tilting. Nonetheless, it significantly advances automated and operator-independent semiconductor failure analysis by enabling reliable detection of structural anomalies through learned representations of intact signal patterns.