A comparison of generative AI models for the top-down SEM image of line and space patterns

SEM (Scanning Electron Microscope) is used to measure the dimensions of semiconductor patterns. Since it has become difficult to maintain measurement accuracy as structures become finer and more complex, AI technology is being introduced. To streamline the construction of training datasets, there is growing interest in generating SEM images as a surrogate model for Monte Carlo simulation. In this study, we compare and evaluate three representative generative diffusion models using Monte Carlo simulation data of line-and-space patterns. To train the generative model, we obtained an SEM simulation image dataset (48 x 48 pixels) calculated with 594 different parameters (6 levels of sidewall width, 11 levels of top height, and 9 types of observation conditions) using the Nebula simulator. We compared the applicability of three generative models: diffusion autoencoder, stable diffusion, and concept sliders. The models are given geometry and observation conditions as a numerical sequence or text and predict SEM images. Visual inspection and quantitative evaluation by NRMSE (Normalized Root Mean Square Error) showed that concept sliders reproduced images that most closely resembled the simulated images. The computation time required for image generation was only a few seconds, making it faster than simulation.