HMC Method for Defect Classification and Location of Postbond TSVs

In this paper, based on S-parameter testing, a defect detection flow using a hierarchical-level neural network, can successfully achieve the classification and location of five common defects. This work categorizes TSV defects into A- and B-types, framing the defect issue as a hierarchical multi-label classification (HMC) problem. First, a defect modeling approach based on measurement benchmarks is used to construct datasets. Subsequently, based on the proposed training strategy, a neural network model with multiple local models for three hierarchical-levels has been trained, named hierarchical multi-label classification with local deep neural networks (HMC-LDNN) model. Then, a complete flow for TSV defect detection, classification, and location is proposed. This flow covers various aspects such as EM modeling strategy, neural network training strategy, and measuring strategy. After testcases, confusion matrices are verified to have good classification indices. The proposed classification method based on the HMC-LDNN model demonstrates superior defect classification and localization performance.