Research On Wafer Scanning Electron Microscope Image Generation Based On Generative Adversarial Network

With the guide of Moore's Law,the feature size of integrated circuit is continually shrinking.As the dimensions of microelectronic device enter the sub-100 nm region,the dominant lithography technology will be greatly affected by the factors of optical proximity effect(OPE)and line-edge roughness(LER),which makes the difference between the transferred pattern on the substrate and the mask pattern.These deviations can cause defects on the wafer surface.To ensure the quality of the wafer fabrication,some methods are usually utilized to find defects on the surface of wafer by analyzing the scanning electron microscope(SEM)images of wafer after lithography.However,it is indeed difficult to obtain a large number of actual SEM images,which limits the validation and comparison of real data.Therefore,many wafer defect detection algorithms are tested recently most based on simulated wafer data.According to the demand of a fab,this study attempts to generate more effective SEM images using its actual wafer SEM images of advanced lithography and circuit layout.In the subsequent process,researchers can use these images directly or after a little modification of them.The details are as follows:1)A deep learning method based on pix2 pix network is proposed to generate wafer SEM images.First,a pattern recognition algorithm is designed for searching a group of polygon patterns of SEM image from a chip layout combined with the characters of supplied SEM images.Then,the collected layout and SEM images are used to train generative adversarial network.The experiment shows that the image generated by this method can simulate the background noise,edge roughness and the change of secondary electrons number on the wafer surface.2)To solve the problem of insufficient resolution of the SEM images generated by the general image generation methods,this thesis proposes a wafer SEM image generation network model based on the conditional generative adversarial network.The edge information of SEM is first calculated using Sobel operator model,and the obtained information is added to the discriminator as guides.Second,two discriminators of different receptive fields are applied in this model for discriminating images of various resolutions.Finally,Wasserstein distance and smooth L1 loss functions are added to the objective function to accelerate network convergence.The experimental results show that the accuracy of the SEM images generated by this model is higher than that of the generated by the currently popular image generation algorithms.Using the 1-NearestNeighbour(1-NN)classifier to evaluate algorithms,the score of this algorithm is increased by 0.3.