Deep learning for virtual metrology: Modeling with optical emission spectroscopy data

Virtual Metrology is one of the most prominent Advanced Process Control applications in Semiconductor Manufacturing. The goal of Virtual Metrology is to provide estimations of quantities that are important for production and to assess process quality, but are costly or impossible to be measured. Virtual Metrology solutions are based on Machine Learning approaches. The bottleneck of developing Virtual Metrology solutions is generally the feature extraction phase that can be time-consuming, and can deeply affect the estimation performance. In particular, in presence of data with additional dimensions, such as time, feature extraction is typically performed by means of heuristic approaches that may pick features with poor predictive capabilities. In this work, we propose the usage of modern Deep Learning approaches to bypass manual feature extraction and to provide high-performance automatic Virtual Metrology modules. The proposed methodology is tested on a real industrial dataset related to Etching. The dataset at hand contains Optical Emission Spectroscopy data and it is paradigmatic of the feature extraction problem under examination.