Computational direct writing lithography utilizes U-Net convolutional neural network for controlling the morphology of three-dimensional microstructures

In this paper, a correction method based on U-Net convolutional neural network to correct the morphology of conical and spherical microlens with period from 30 μm-100 μm is proposed. Maximizing the utilization of the nonlinear characteristics of the convolution neural network and the encoder/decoder to capture the characteristics of the different scales, we employ the U-net network to describe the inverse process of lithography. By constructing and training the UNet convolutional neural network, on the basis of learning a large amount of data, we can obtain the inverse process model of direct writing lithography. After optimization, the average absolute error of conical microlens is decreased from 0.68 μm to 0.22 μm. For spherical microlens array, the full-width-half-maximum of intensity distribution is improved from 18 μm to 10 μm.