To address the challenge of characterizing bonding strength in hybrid bonding, this paper proposes a method based on post-bonding crack-induced di-cantilever bending (PBC-DCB). The commonly used industrial blade insertion (BI) method only measures the edge strength of wafer-to-wafer samples via bevel geometry, failing to characterize the wafer center or die-to-wafer samples (as they lack a bevel) and exhibiting limited accuracy due to irregular crack edges. This study develops a di-cantilever beam test initiated by a specially prepared post-bonding notch. By deriving the relationship between load–displacement and bonding strength based on fracture mechanics and optimizing sample preparation and testing parameters, quantitative measurement of interface fracture strength is achieved. Experimental results show the method has a deviation < 5% from BI with better repeatability, verifying its reliability. It successfully distinguishes bonding strength between wafer center and edge and reveals that SiCN dielectric layers have 17.34% higher strength and better uniformity than SiO2. This provides a reliable testing tool for hybrid bonding quality and process optimization, with broad prospects in industrial production and scientific research.
Inorganic gap filling technology is an effective method to improve reliability and heterogeneous integration density in 2.5D and 3D integration. It uses plasma-enhanced chemical vapor deposition (PECVD) to deposit silicon dioxide (SiO2) filler layers in gaps between chiplets. This technology is used to replace the Epoxy Mold Compound (EMC) commonly used in traditional packaging. However, as an inorganic filling material, SiO2 poses reliability challenges such as cracking and peeling during or after deposition. Furthermore, there lacks quantitative characterization and modeling of the microscale mechanical properties, thermal stress distribution, and fracture failure risk in the filler layer. By combining nanoindentation technology with three-point bending tests, this study reports a comprehensive characterization route for quantitative characterization of mechanical behavior of the filler. A finite element method (FEM) model was also established to predict the thermomechanical reliability of the gap filling process. Raman spectroscopy measured data confirm the model’s reliable predictive ability. The results reveal the impact of filler thickness on the stress. The microscale SiO2 mechanical characterization method and the thermal stress and fracture risk FEM prediction model in this study not only address the limitations of traditional testing and simulation but also provide support for process optimization and structural design of gap filling in high-density 2.5D/3D packaging. This work promotes the understanding of inorganic filling process reliability in chiplet integration.
The quantitative measurement of bonding strength in hybrid bonding (HB) is an indispensable metrology for process and reliability evaluation. Methods currently used such as blade insertion (BI) and double cantilever bending (DCB) have suffered from a wafer edge-only capability and limited repeatability due to a lack of precise interface crack initiation. This study reports an improved DCB method by introducing a post-bonding crack (PBC) to avoid undesired cracking in a wafer substrate during DCB propagation. This method is firstly applied to measure the bonding strength of SiCN-SiCN under O2 and N2 activation. The test data shows that the bonding strengths are 3.53 J/m2 and 2.93 J/m2 with the deviations less than 3.84% and 1.84%. Based on the experimental data, finite element analysis (FEA) methods are used to simulate the crack propagation process of the PBC-DCB method. The results show that the bonding interface crack propagation can be accurately described by an optimized viscoelastic exponential model. The accuracy of simulation increased from 16.06% to 1.77%. Finally, it was found that the PBC-DCB method can solve the issue in conventional DCB where the crack may be initiated away from the target interface, therefore measuring the wrong interface. This advantage is further validated by simulations considering the offset of the PBC away from the bonding interface.
Polymer insulation layers such as polyimide (PI) have gradually replaced inorganic dielectric layers (SiO2, SiCN) in the integrated packaging process of hybrid bonding (HB). PI can fill the gaps in the thermal compression bonding process and help to obtain a good Cu/Polymer bonding interface. At present, the existing post-crack double cantilever beam tensile test (PBC-DCB) has been successfully applied to the quantitative measurement of bonding strength of hybrid bonding with inorganic materials, but this method only considers elastic behavior. Since PI exhibits viscidity, elasticity and plasticity, knowing how to correlate these properties to the bonding process is challenging. Whether PBC-DCB is suitable for the characterization of PI bonding is unclear. This paper presents a comprehensive experimental and finite element analysis (FEA) study on the PI–PI bonding interface. Firstly, nanoindentation experiments and simulations are performed on the prepared PI interface to obtain key elasticity and plasticity parameters. Then, the bonding strength is characterized by the PBC-DCB test. Theoretical and experimental results show that the plasticity of PI causes energy dissipation during stretching, resulting in a deviation of approximately 2.51% compared with pure elasticity. Based on experimental data, the Cohesive Zone Model (CZM) FEA method is used to simulate the crack propagation. The results indicate that the Embedded Process Zone (EPZ) model can accurately describe crack initiation and delamination behavior, with a margin of error of about 3.61%. Finally, based on the EPZ CZM, defects such as bonding void and wafer warpage are further discussed in relation to bonding strength measurement.
Reliable wafer map defect classification is important for semiconductor yield analysis, but it remains challenging because labeled wafer maps are often limited, class distributions are highly imbalanced, and many deep models provide limited interpretability. This paper presents a compact data-centric framework that integrates conditional generative augmentation, dual-branch deep feature extraction, and hand-crafted geometric descriptors for imbalanced wafer map defect classification. A class- and mask-conditioned StyleGAN2-ADA generator is used offline to synthesize minority-class wafer maps, and the resulting samples are filtered before being merged with the training set. For classification, a global-local encoder models whole-wafer context and local defect patterns, while a branch-relative positional cross-attention module enables local patch features to selectively attend to wafer-level contextual information. The positional encodings preserve token structure within each branch but do not impose a one-to-one absolute-coordinate alignment between a crop and the full wafer. To complement the deep representation, five interpretable hand-crafted feature groups are extracted, including radial profile, angular profile, connected component count, eccentricity, and edge concentration. The global and local deep encoders are shallow MobileNetV2-derived branches, while their cross-attention-fused representation is combined with a projected hand-crafted feature vector and classified by a compact classification head trained with Class-Balanced Focal Loss. Experiments on the public WM-811K dataset show that the proposed framework achieves 98.89% accuracy and 94.41% macro F1-score under the reported split. Additional ablation studies indicate that mask-conditioned generation, positional cross-attention, hand-crafted features, multi-scale local patches, and class-balanced focal learning each contribute to the final performance. The deployed classifier contains 64,967 trainable parameters, excluding the offline generator and non-trainable hand-crafted feature computation.