DescriptionVirtual metrology (VM) in semiconductor manufacturing is the technique of predicting critical dimensions of wafer quality characteristics without direct measurement based on process data of production equipment. VM is important in semiconductor manufacturing since it enables engineers to monitor the quality of wafers in production without physical wafer metrology thereby increasing the throughput of the process. As the process information consists of a large number of process variables in the form of raw sensor signals, learning new useful features in a low dimensional space is a key to build accurate VM prediction models. Earlier efforts in VM modeling were carried out by employing linear dimensionality reduction techniques such as PCA. Autoencoder is a deep learning based feature extraction method that has the capability to explore the non-linearity in the modeling and to represent high dimensional input into a low dimensional space. In this thesis, we propose a new VM model by incorporating the autoencoder based feature learning. We apply the proposed model to the prediction of critical dimensions of wafers at a plasma etching process in semiconductor manufacturing and compare the predictive performance of the proposed model with conventional VM models. The experimental results show that the proposed model outperforms the existing models thus showing that autoencoder based feature learning is helpful in VM modeling with raw sensor signals.