Enhancing Scalability of Deep Learning Based Approaches in Semiconductor Manufacturing

In this thesis the necessity, the possibilities but also the difficulties of scaling and generalization of machine learning models in the context of process control in semiconductor manufacturing is discussed. Thanks to growing research, a broad variety of methods and models are available to deal with automation and transfer learning and high efforts are spent to make them usable for industry 4.0 applications. Inspired by the need of manufacturing for a deep understanding, explainability of models and comparability of features in the context of scalability, the thesis presents DANN-based Alignment Model (DBAM) and its extended version DANN-based Alignment with Cyclic Supervision (DBACS) that unites an interpretable approach of domain adaptation with (equipment) matching. The application focus in this work is on process control, more precisely virtual metrology, predictive maintenance and defect classification. Relevant examples for process, failure patterns and data sets are selected to showcase all functionalities of the introduced methods but also discuss encountered challenges and shortcomings or limitations when dealing with the highest complexity of semiconductor production.