Semiconductor MEMS WaferBond Membrane Deflection Prediction from Infrared Images Using Deep Learning based Vision Models

Wafer bonding is critical for integrating smart systems, particularly in MEMS applications requiring high vacuum or wafer-level encapsulation. Often, finding and measuring the vacuum values is tedious. Accurate measurement of vacuum levels across the wafer, to monitor and evaluate the deflection is time-consuming. In this paper, we estimate the deflection of the normal membrane from the bossed membrane. The bossed membrane shows a fringe pattern under the infrared (IR) inspection. A normal membrane is surrounded by four bossed membranes. The fringes in the bossed membrane have varying fringe patterns and sizes which are useful for predicting the deflection of the normal membrane. For rapid analysis, these values can be determined via their correlation with fringes in infrared (IR) images of the corresponding membranes. This paper introduces a deep learning algorithm for evaluating and correlating fringe patterns in infrared images with vacuum values, offering a rapid and accurate approach to vacuum analysis in MEMS wafer bonding. This deep learning-based correlation-driven approach promises to significantly improve the efficiency and accuracy of wafer bonding analysis with a mean absolute error of around <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$12 \mu \mathrm{m}$</tex> with a dataset as narrow as 3 images.