Wafer SEM Image Generation with Conditional Generative Adversarial Network

Abstract

Scanning Electron Microscope (SEM) images play an essential role in the analysis and evaluation of the defects of the circuit in advanced integrated circuit manufacturing. Currently, the image generation method is a useful means to solve the insufficiency of wafer SEM images due to the high cost of getting a large number of labeled SEM images. In this paper, an algorithm based on conditional generative adversarial network (cGANs) is proposed for SEM image generation. Firstly, a Sobel operator is used to calculate the gradient information of images to guide the discriminator. Then we apply two discriminators to discriminate images of different resolutions. Finally, Wasserstein distance and smooth L1 loss functions are applied to accelerate network convergence. Experimental results show that this method could learn to mimic the distribution of wafer SEM image data effectively. Compared with other image generating models, our model improves the quality of the generated image, and its 1-Nearest-Neighbour (1-NN) classification score is increased by 0.3. Therefore, this algorithm is more suitable for generating images to alleviate the shortage of wafer SEM images.