| Summary: | This thesis proposes four methods to improve macro-inspection capability of defects on wafers at a semiconductor wafer fab. First, an investigation into the performance of current inspection tools is done, revealing results that are not reliable nor reproducible. Tool maintenance procedures and specification adjustments are recommended. Second, a software upgrade to the current inspection software is developed, including enhanced features that address pain points of reviewing wafer images. The image processing and loading time is reduced by over 50%. Third, three binary classification machine learning models are trained to isolate spin-on-glass defects, edge type defects, and center defects. Each of the models exhibits an area under curve (AUC) of over 0.90 on out-of-distribution test sets. Finally, a proof-of-concept for an in-line inspection system is designed and tested on the fab floor. New images from this system appear to be of sufficient quality for inspection. The results of each part of this study can be used to inform investment decisions required to move towards a more automated process.
Relevant to the machine learning community are the methods developed to address class imbalance in neural network training. Methods for preparing data to be trained in a meaningful way such as spitting, transforming, and creating synthetic data are proposed. The effect of generating data in such a fashion is shown to be positive, increasing the AUC of the specified model by up to 65%.
|
|---|