Mask Near Field Computation Based On Generative Model

Computational lithography,as one of the core technologies for Very Large Scale Integration(VLSI)manufacturing,plays an important role in a series of simulations for IC manufacturing.The near-field calculation of the mask is used as one of the basic tasks in the simulation of Extreme Ultra-violet(EUV)lithography for process modeling and simulation verification.Therefore,the calculation of the near-field of the mask is an essential and critical part of the simulation and modeling of computational lithography.However,with the continuation of Moore's law and the continuous improvement of lithography,the feature size of semiconductors continues to decrease and the thickness of the mask is no longer negligible in the calculations.These changes have resulted in optical diffraction effects becoming very severe when extreme ultraviolet light passes through the mask.Therefore,the scalar diffraction theory is no longer accurate,which requires us to find more accurate calculation method to calculate the diffraction near field of the mask.In some traditional mask near-field calculation methods,such as the strict electromagnetic field calculation method,the calculation of the mask near-field is less efficient due to problems such as complicated calculation formulas or too large calculation volume.And the calculation is easily affected by the mask pattern with random edges and feature sizes,which also makes the accuracy of the calculation poor.However,with the rapid development of deep learning in recent years and the popularity of Generative Adversarial Network(GAN)as one of its most interesting ideas,we have seen some new opportunities.After many simulation experiments attempts,this thesis selected a method to compute the mask near-field using the Cycle-consistent Generative Adversarial Network(Cycle GAN).Compared with the existing mask near-field calculation methods,the calculation method in this thesis overcomes the problem that the traditional strict electromagnetic field calculation method does not have any approximation constraint in the calculation process,which leads to a more complicated calculation process,and makes it possible to improve the calculation speed.In this thesis,based on the generative model framework of generative adversarial networks in deep learning,two generators are constructed to generate the near-field and mask imaging patterns of the mask forward and reverse,respectively,using end-to-end and pixelto-pixel data conversion methods,to learn feature transformations between different data domains.Res Net is used in the generator for the combined extraction of the feature vectors given by the encoder,and the learning of the transformation of the feature vectors from the source domain to the target domain is achieved using the extracted features.In addition,the cycle consistency loss is then added as a constraint for the two generated models to ensure that the imaging pattern of the mask near field before and after the generation of the obtained model can correspond to the imaging pattern of the mask one by one,so as to avoid problems such as pattern collapse.Finally,this thesis compares the simulation experimental results with some results from thesis in the field of computational lithography.The comparison results show that the mask near-field calculation method based on the generative model used in this paper is more than 4 times more efficient in terms of calculation and more than 300 times more accurate in terms of calculation compared with the FCN method.