Non-Destructive Defect Localization Based on Anomaly Detection in Acoustic Signals using Autoencoders
Non-Destructive Defect Localization Based on Anomaly Detection in Acoustic Signals using Autoencoders
Tuesday, November 18, 2025: 1:10 PM
1 (Pasadena Convention Center)
Summary:
This paper introduces a novel, unsupervised approach for defect detection in semiconductor failure analysis using Scanning Acoustic Microscopy (SAM). The method employs an autoencoder neural network trained exclusively on acoustic signals from intact reference samples to identify outliers in new, unknown samples. By reconstructing the time-domain signal and analyzing reconstruction errors, the system can detect anomalies that may indicate structural defects such as delaminations or fractures—without requiring prior knowledge of defect types. This approach addresses key challenges in conventional supervised learning, such as limited defect-labeled data and the complexity of modern microelectronic structures. The autoencoder's reconstruction error serves as a similarity metric, highlighting suspicious regions in the acoustic image and allowing axial defect localization through signal segmentation. While the method significantly enhances automation, sensitivity, and accessibility of SAM-based analysis, it may also flag non-defective anomalies (e.g., sample tilting), requiring user interpretation. Overall, the proposed technique represents a major step toward automated, high-resolution, non-destructive failure analysis, offering robust defect detection even with limited or no defective training data.
This paper introduces a novel, unsupervised approach for defect detection in semiconductor failure analysis using Scanning Acoustic Microscopy (SAM). The method employs an autoencoder neural network trained exclusively on acoustic signals from intact reference samples to identify outliers in new, unknown samples. By reconstructing the time-domain signal and analyzing reconstruction errors, the system can detect anomalies that may indicate structural defects such as delaminations or fractures—without requiring prior knowledge of defect types. This approach addresses key challenges in conventional supervised learning, such as limited defect-labeled data and the complexity of modern microelectronic structures. The autoencoder's reconstruction error serves as a similarity metric, highlighting suspicious regions in the acoustic image and allowing axial defect localization through signal segmentation. While the method significantly enhances automation, sensitivity, and accessibility of SAM-based analysis, it may also flag non-defective anomalies (e.g., sample tilting), requiring user interpretation. Overall, the proposed technique represents a major step toward automated, high-resolution, non-destructive failure analysis, offering robust defect detection even with limited or no defective training data.