The demand for semiconductor components in consumer products has in-creased rapidly in the 2020s which has led to a shortage of electrical components and thus breaks in availability of the products. For this reason, semiconductor manufacturers are motivated to deploy machine learning and automation solutions, such as virtual metrology, that increase the yield and quality of manufacturing. Regardless of the clear benefits, there is an absence of comprehensive comparisons of current state-of-the-art machine learning methods for virtual metrology.

This thesis presents and reviews multiple state-of-the-art machine learning methods for virtual metrology of chemical-mechanical polishing – a crucial process in semiconductor manufacturing. We compare typical virtual metrology pipelines consisting of hand-crafted features and tree ensemble models to the current state-of-the-art time series extrinsic regression models, such as Inception Time and recurrent neural networks. In addition, we propose and evaluate several approaches for including time-invariant extrinsic process variables to recurrent neural networks. Furthermore, we implement semi-supervised auto-encoder models for a prediction scenario where only a fraction of process runs are labeled. These autoencoders are compared to other machine learning methods in a limited labeled data setting. In addition, we evaluate the performance of tree ensemble models by analyzing their feature importance scores. The experimental work is conducted on a public dataset which allows simple comparison of our work to other recent publications utilizing the same dataset.

Our experiments and comparison of different models show that hand-crafted features combined with tree ensemble models are a strong choice for the virtual metrology of chemical-mechanical polishing. We discovered, however, that semi-supervised autoencoder models predict metrology results more accurately than supervised machine learning methods in a limited labeled data setting. Furthermore, our tree ensemble model feature importance analysis revealed that only a fraction of the measurements in the process data are sufficient for accurate virtual metrology of chemical-mechanical polishing tool. In fact, a gradient boosted trees model utilized only 19% of the measurements and achieved an R2 score of 0.99 in the testing set.