Deep Convolution Neural Networks for Automatic Detection of Defects which Impact Hybrid Bonding Yield

Fast and accurate defect detection is critical for hybrid bonding because surface defects directly impact yield and processing costs. Optical microscopy provides a low-cost and high-throughput means to assess the quality of hybrid bonded wafers, but manual inspection and defect categorization requires significant time and effort. This paper presents an efficient and accurate optical inspection scheme where an optical microscope with automated stage and customized software collects full wafer images for designs >10 μm pad diameter. All the images are fed sequentially into two Convolutional Neural Network (CNN) models: a defect identification model (Model #1) and a defect categorization model (Model #2). With careful tuning of model parameters and category selection, Model #1 and Model #2 achieve 97% and 96% accuracy on the test set, respectively. Model #1 is then further validated by comparing the accuracy and time saved to the performance of a fully manual inspection by a highly trained human operator. The CNN model performs with higher recall and saves over 1.5 hours of inspection time per wafer. Finally, limitations of the current models and methods for model refinement are presented with additional discussion on future methods to streamline the inspection process of hybrid bonding.