This research work builds upon the work of Schrunner et al. [102] and Nartey [88] on the ensurance of the data and process integrity during the early stages of the semiconductor production process. The Wafer Health Factor (WHF) is a quality metric that combines the class probabilities from a process pattern recognition in analog wafer maps with a criticality score defined by domain experts. Motivated by a possible circumvention of an expensive feature engineering through the use of an end-to-end approach, this thesis explores the use of Deep Learning (DL) for the WHF and how it influences both the robustness and the performance of the pattern classifiers.
The small dataset of only 346 labeled wafer maps poses a major challenge in applying DL methods. To assess, if the use of additional unlabeled wafer maps can improve classification performance and robustness, recent self-supervised representation learning (SSL) methods are evaluated and compared against a supervised baseline. With the focus lying on the security and the reliability of the production process and product integrity, the robustness of these methods is measured through a novel metric based on adversarial attacks [81].
Due to the high effectiveness of a domain specific augmentation pipeline, the supervised baseline achieves very competitive results almost reaching the performance of an ensemble classifier build on a set of highly specialized hand-crafted features [88]. Two of the three evaluated SSL methods are able to outperform the supervised classifiers, especially when the dataset is reduced further.
While the supervised classifiers require stronger adversarial perturbations to change the classification output, the SSL classifiers stay very stable throughout the training process and demonstrate higher domain specific awareness as evident by the uncharacteristically high wafer related structure in the adversarial perturbations. The investigation of the robustness of the different DL classifiers yielded interesting results, that motivate further studies of novel training methods and strengthen the viability of ensemble approaches.