Deep learning for robust optical overlay metrology in semiconductor device manufacturing

Fast and high-precision metrology is critical for production processes in the expanding high-tech industry sector. We explore the use of simple and cost-effective optical sensors in combination with a novel learning-based algorithm to simultaneously improve the accuracy of the parameter extraction, cost of the metrology sensor and time required for the measurement. We report sub-nanometer measurement accuracy and precision for our algorithm in the presence of strong optical aberrations and noise, which enables the use of affordable and simple optical measurement systems and faster measurement times. To contextualize our findings, we also performed Fisher information analysis.