Increasing the Throughput of Hybrid Bonding Inspection Using Novel Methods and Machine Learning

Hybrid bonding is enabling denser and finer pitch interconnections, which is key to meeting demands for next generation applications like artificial intelligence. But careful control of surface topography and defectivity is required to realize the benefits of hybrid bonding.

At the 2024 IEEE Electronic Components and Technology Conference (ECTC), my colleague, Dr. Bongsub Lee, and I presented two papers that explain how to efficiently monitor both the surface topography and defectivity before bonding, so any bad material can be screened out. The first paper discusses a method to reliably measure the surface topography 1,000 times faster than existing methods. The second paper details how we streamlined the inspection process for hybrid bonding defects using convolutional neural networks (CNNs).

2 Key Factors for Pre-Bond Inspection

During wafer inspection, there are two main factors that need to be defined and controlled: metal recess and defectivity.

• Metal recess is typically defined as the depth of a metal interconnect (usually copper), compared to the dielectric surface.
• Defectivity is defined as the measure of the surface flaws in the material, such as foreign particles which may come from dust in the air, metal corrosion, or lithography pattern defects.

In the hybrid bonding process flow, the inspection process to measure both of these factors takes place after the silicon wafer is cleaned and before the wafer is bonded. Inspection is necessary to determine which areas of a wafer have an anomalous metal recess and which areas are defective, so that it is possible to avoid bonding dies to the sites that include those locations.

Efficient Metal Recess Measurements with Phase Shift Interferometry

Metal recess is conventionally characterized by a technique called atomic force microscopy (AFM). AFM is very accurate and straightforward, but the throughput is very limited. We developed a protocol to analyze topography by using an optical technique called phase shift interferometry (PSI) with about 1,000 times improvement in throughput over AFM. We calibrated this PSI technique for hybrid bonding samples to interpret the data properly. We found that this technique could easily analyze hundreds of thousands of metal pads within hours, using our automated method to take and analyze the data. This optical technique with automation can be useful in high volume settings, where a large number of samples must be analyzed and compared with the previous samples. For more information, please see our 2024 ECTC paper "High-Throughput Characterization of Nanoscale Topography for Hybrid Bonding by Optical Interferometry."

Improving Defect Detection with CNNs

At Adeia, we developed, implemented, and trained convolutional neural networks to improve the accuracy and speed of the defect inspection process. Our automated method can inspect wafers at a rate of over 100 minutes per wafer faster compared to a manual inspection method, which will usually take on the order of a couple hours per wafer, allowing us to collect larger datasets and improve our understanding of process-materials and parameter-performance relationships.

Combined with improved accuracy, these changes also increase throughput in hybrid bonding manufacturing cycles and reduce time to market. To learn more, please read our 2024 ECTC paper, "Deep Convolution Neural Networks for Automatic Detection of Defects which Impact Hybrid Bonding Yield."

Next