Deep Learning for Predicting CD-SEMS of NEMS Devices

This paper presents an AI model that predicts the process output from photolithography and plasma etching based on CD-SEM data. This contrasts with physics-based models that are used in conventional TCAD tools. A large dataset was generated consisting of nanostructure CD-SEMs (~150,000) from outcomes of an ASML DUV lithography stepper and an Oxford Cobra ICP plasma etcher. The AI model is an Image-to-Image Translation deep learning algorithm that learns from a training set of the CD-SEMs. This deep learning model enables an evolving TCAD model in which layouts can be actively modified as data from the cleanroom is collected continuously. This model can be helpful to improve yield and device homogeneity and performance, hence time to market, for advanced sub-micron NEMS and MEMS. This nature of this dataset ensures the applicability of the presented algorithms to academic and industrial cleanrooms.