Critical Dimension Scatterometry as a Scalable Solution for Hybrid Bonding Pad Recess Metrology

This paper demonstrates a successful implementation of Machine Learning assisted Spectroscopic Critical Dimension (SCD) Scatterometry as a scalable, high throughput, and precision metrology for 9 μm and 3 μm pitch hybrid bonding interconnect pad recess/protrusion. A novel larger spot size (˜60 μm) was implemented to allow multiple interconnect pitches to be sampled by a single measurement spot. To build a robust model unaffected by underlying feature variations, a machine learning model was built using Atomic Force Microscopy (AFM) images as ground truth training data to correlate to a span of pad recess depths ranging from -30 nm (protruded) to 40 nm (recessed), as well as dielectric shape varying from flat to over-polished states. 50 μm x 50 μm AFM scans were collected for a total of 456 sites across 6 wafers with varying pad recess skews, a process that took approximately 6 hours per wafer. Pad recess data from fewer than 20% of these sites were shared for building a machine learning inference model, while the remainder of site data was retained for validation. SCD spectra collection across all 6 wafers (456 sites) was completed in less than 75 seconds per wafer (more than 5 times faster than conventional white light interferometry techniques for the same number of sites).