Abstract
Integrated circuit (IC) delayering workflows are highly reliant on operator experience to determine processing end points. The current method of end point detection during IC delayering uses qualitative correlations between the thickness and color of dielectric films observed via optical microscopy. The goal of this work is to quantify this relationship using computer vision. As explained in the paper, the authors trained a convolutional neural network to estimate the thickness of dielectric films based on images and measurements recorded during processing. The trained vision model explained 39% of the variance in dielectric film thickness with a mean absolute error of approximately 47 nm. The paper describes the entire workflow, including verification testing, and addresses the primary sources of error.
