Layered Unsupervised Learning-based Identification and Quantification of Voids in Package Thermal Interface Materials

Abstract

Electronics packaging thermal interface materials (TIM) are typically compliant materials with high thermal conductivity placed between a heat-generating chip and package lid or integrated heat spreader, and are used for effective heat dissipation. Shrinking thermal budgets have necessitated effective TIM properties and low-defect assembly. Dispensing TIM is a very exacting process and voids can become trapped during assembly, causing detrimental effects on the thermal resistance, reliability, and adhesion strength of TIM. Accurate detection of the presence of voids, their geometric parameters, and their growth are needed in order to assess TIM quality. Automated identification and measurement of these voids would create standardized tools and metrics that could seamlessly be integrated into the current manufacturing line for faster and more accurate detection of TIM defects. This also lends itself to the creation of large data sets for understanding the parametric space of material properties that can cause voids of certain sizes, in certain locations, and of a certain number.In this paper, a fully automated, artificial intelligence-Assisted method of identifying and quantifying voids in confocal scanning acoustic microscopy (CSAM) images of TIM was developed using unsupervised machine learning techniques and implemented as a custom code base in MATLAB. Images are pre-processed using a variety of two-dimensional signal processing methods such as mean/median filtering, Gabor filter banks, pseudo-planarization, and equalization. Modified k-means texture-and intensity-based image segmentation methods were implemented to separate the image iteratively into void and non-void regions. These void regions are each quantified by their absolute area, percent area of chip, major axis length, and minor axis length.