Machine Learning-Based Diagnosis of Defects in 2.5-D and 3-D Interconnects

Abstract

Diagnosing interconnect line defects becomes increasingly challenging in advanced chiplet integration due to the immaturity of fabrication processes, reduced interconnect spacing, and increased density. In this article, a nondestructive interconnect defect diagnosis method is proposed. First, redistribution layers (RDLs) on the 2.5-D silicon interposer and 3-D through silicon via-RDL (TSV-RDL) interconnect channels are simulated by ANSYS HFSS, with open and short defects injected at various positions to analyze signal integrity; utilizing S parameters and group delay time (GDT) as feature vectors, machine learning algorithms are employed specifically for the classification and identification of defects. For defect localization, a novel high-precision scheme driven by adaptive regression and clustering is proposed, achieving highly accurate and reliable localization of defects. The results show that the algorithm can accurately identify open and short defects. In defect localization, the mean relative error (MRE) of the proposed method is less than 8%, and the maximum relative error (MaxRE) does not exceed 13%. Compared with the related algorithms, the localization accuracy is significantly improved, providing a novel perspective for the identification and localization of defects within package interconnect lines.