Machine Learning-Based Defect Classification and Anomaly Detection in Advanced IC Package 3D Interconnects Using Scanning Acoustic Microscopy

Title: Machine Learning-Based Defect Classification and Anomaly Detection in Advanced IC Package 3D Interconnects Using Scanning Acoustic Microscopy

Technical Description:

As transistor scaling approaches its physical limits, advanced packaging has become essential to drive semiconductor performance, system integration, and reliability beyond traditional Moore's Law. In such packaging architectures, manufacturing defects such as micro-cracks, bad wetting, and underfill voids can jeopardize device reliability. Scanning Acoustic Microscopy (SAM) can non-destructively detect these defects, but current inspection relies heavily on slow, subjective manual review. Automating this process with machine learning would greatly speed up industrial failure analysis, yet two main bottlenecks remain: defects occur too rarely to build large training datasets, and tiny micro-cracks look almost identical to healthy joints in acoustic scans.

Preliminary work has demonstrated that depth-resolved C-SAM gating combined with anomaly detection substantially improves defect screening over conventional supervised models. However, 2D C-SAM images discard vital phase, frequency, and time-of-flight information embedded in the raw acoustic waveform.

Objectives:

The primary goal of this project is to build an automated, highly accurate failure analysis pipeline for advanced micro-interconnects by overcoming the severe data scarcity inherent to packaging failure analysis. This project will investigate novel machine learning frameworks capable of achieving high detection accuracy despite extreme class imbalance and limited defective labels. The proposed methodology should aim to maximize model generalization and accuracy across varied interconnect architectures. In summary, some potential key objectives include:

• Expand inspection beyond C-SAM images by processing raw A-scan waveforms, extracting phase shifts, time-of-flight variations, and defect signatures.

• Benchmark modern data-efficient architectures (e.g., few-shot prototypical networks, metric learning, etc.) to maximize sensitivity toward rare defect classes.
• Leverage forward acoustic simulation tools to synthesize realistic, defect-injected waveforms and scans, directly mitigating class imbalance and preventing model overfitting.
• Test model robustness across diverse package types (e.g., FCBGA with exposed die, over molded, etc.), validating predictions against destructive cross-sectional inspections.

References:

[1] M. Gomes Abrantes, "Defect Detection in Advanced IC Interconnects," Master's Thesis, Faculty of Engineering Sciences, KU Leuven & imec, 2026.

[2] J. Snell, K. Swersky, and R. Zemel. Prototypical networks for few-shot learning. In Advances in Neural Information Processing Systems (NeurIPS), 2017.

[3] S. Brand, P. Czurratis, P. Hoffrogge, and M. Petzold. Automated inspection and classification of flip-chip-contacts using scanning acoustic microscopy. Microelectronics Reliability, 50(9–11):1496–1501, 2010.

[4] C. S. Lee, G.-M. Zhang, D. M. Harvey, and A. Qi. Characterization of microcrack propagation through analysis of edge effect in acoustic microimaging of microelectronic packages. NDT & E International, 79:19–27, 2016.

[5] M. Wolf, A. Sukumaran Nair, P. Hoffrogge, E. Kühnicke, and P. Czurratis. Improved failure analysis in scanning acoustic microscopy via advanced signal processing techniques. Microelectronics and Reliability, 138:114618, 2022.

Desired Qualifications:

• Strong programming skills in Python (or MATLAB) and machine learning frameworks (e.g., PyTorch, scikit-learn).
• Knowledge of signal processing, computer vision, or acoustic wave propagation.
• Interest in semiconductor packaging, failure analysis, and hands-on lab metrology (SAM).

Contacts: Jeffrey.Gutierrez@imec.be