AIDA: AI Detection of Anomalies for failure analysis at the single-transistor level
AIDA: AI Detection of Anomalies for failure analysis at the single-transistor level
Tuesday, November 18, 2025: 2:30 PM
1 (Pasadena Convention Center)
Summary:
Our group specializes in Transmission Electron Microscopy (TEM) data analysis, with a primary focus on failure analysis to identify defects that inform critical yield decisions. Given the large volume of TEM data we process, leveraging AI/ML is essential to accelerating our output, handling the high volume required by our customers, and making use of big data to improve accuracy and efficiency. Here, we demonstrate our completely unsupervised ML workflow for automated anomaly detection called AIDA. From a single TEM image containing multiple bitcells, we extract the repeating structures and train a PCA model solely on defect-free cells. The reconstructed image, composed of defect-free 'twins,' is subtracted from the original to reveal defects with a high precision. AIDA can be universally applied to any data that contains repeating structures making it compatible with any node technology out of the box. This enables real-time anomaly detection within just few seconds due to the speed and interpretability of PCA. AIDA achieves comprehensive analysis by automatically isolating even subtle defects in a large field of view, saving engineering resources. Furthermore, we can scale this approach with larger datasets to improve the robustness and adaptability of our offline computations, ensuring high-volume, high-accuracy analysis for our customers.
Our group specializes in Transmission Electron Microscopy (TEM) data analysis, with a primary focus on failure analysis to identify defects that inform critical yield decisions. Given the large volume of TEM data we process, leveraging AI/ML is essential to accelerating our output, handling the high volume required by our customers, and making use of big data to improve accuracy and efficiency. Here, we demonstrate our completely unsupervised ML workflow for automated anomaly detection called AIDA. From a single TEM image containing multiple bitcells, we extract the repeating structures and train a PCA model solely on defect-free cells. The reconstructed image, composed of defect-free 'twins,' is subtracted from the original to reveal defects with a high precision. AIDA can be universally applied to any data that contains repeating structures making it compatible with any node technology out of the box. This enables real-time anomaly detection within just few seconds due to the speed and interpretability of PCA. AIDA achieves comprehensive analysis by automatically isolating even subtle defects in a large field of view, saving engineering resources. Furthermore, we can scale this approach with larger datasets to improve the robustness and adaptability of our offline computations, ensuring high-volume, high-accuracy analysis for our customers.