Abstract
Al-Sn-Al wafer bonding is a new semiconductor manufacturing technology that plays an important role in device manufacturing. Optimization of the bonding process and testing of the bonding strength remain key issues. However, using only physical experiments to study the above problems presents difficulties such as repeating many experiments, high costs, and low efficiency. Deep learning algorithms can quickly simulate complex physical correlations by training large amounts of data, which is a good solution to the difficulties in studying wafer bonding. Therefore, this paper proposes the use of deep learning models (2-layer CNN and 50-layer ResNet) to achieve autonomous recognition of bonding strengths corresponding to different bonding conditions, and the results from a comparative test set show that the ResNet model has an accuracy of 99.17%, outperforming the CNN model with an accuracy of 91.67%. Then, the identified images are analyzed using the Canny edge detector, which showed that the fracture surface morphology of the wafer is a hole-shaped structure, with the smaller the area of hole movement on the wafer surface, the higher the bonding strength. In addition, the effects of bonding time and bonding temperature on bonding strength are verified, showing that relatively short bonding times and relatively low bonding temperatures resulted in better wafer bonding strength. This research demonstrates the potential of using deep learning to accelerate wafer bonding strength identification and process condition optimization.