Utilizing Transformer Deep Learning Based Outlier Detection to Screen Out Reliability Weak ICs

This paper presents an innovative transformer deep learning-based outlier detection approach aimed at efficiently reducing Defective Parts Per Million (DPPM) while maintaining competitive testing costs. By leveraging deep learning techniques, key parameters are predicted. Outlier ICs are identified through a comparison of predicted and measured values. The comparative analysis performed on our proposed approach, along with traditional methodologies such as dynamic part average testing (D-PAT) and nearest neighborhood residual (NNR), and prior work based on machine learning, validates that our method delivers more reductions in DPPM. Specifically, our method yields a 62% reduction in DPPM compared to D-PAT, a 13% reduction compared to NNR, and an 8% reduction compared to machine learning methods after deploying a stricter stress test on the identified outliers. Physical failure analysis verifies one of the failed outlier ICs as having a gate-to-contact short, which highlights proposed method's sensitivity to latent defects. Overall, this approach provides a cost-effective solution for achieving automotive-grade quality.