The choice of input data and its impact on a deep learning mask model

As fabs continue their effort to sustain Moore's Law and manufacture smaller and more complex features in lithographic masks, Mask Process Correction (MPC) becomes increasingly more relevant. An indispensable part of MPC is the ability to accurately predict the outcome of the printed mask. Machine learning beckons the possibility that lithographic masks can be modeled in an automated flow requiring little or no manual intervention. While some significant progress has been made towards this goal, it is important to come to terms with the two main items that impact the outcome of a modeling approach based on supervised Deep Learning: the first is the architecture of the model and its related internal settings (number of layers, activation function, etc.) and the second being the choice and quality of the input data used for training and validating the Deep Learning model which will thoroughly presented in this paper. The study focuses primarily on simulated SEM images of regular structures (e.g., lines and spaces, contact arrays) to allow full control on the quality of the input data as well as determine the ground truth with a very high degree of accuracy and presenting a comparison with the real data.