Lithography test pattern synthesis and PVB prediction using GANs

Abstract

The lithography process is critical for manufacturing chips. With the decrease of feature size, computational lithography became very important to keep Moor's law alive. Various computational lithography techniques such as optical proximity correction (OPC), etch proximity correction (EPC), and source mask optimization (SMO) are being used to enable the lithography process to print sub-wavelength lithography features. Recently, machine learning is getting popular in computational lithography applications as it provides a runtime and accuracy trade-off. A key in the accuracy of all computational lithography techniques is the availability of a large amount of diverse layout data, which is difficult to obtain. In this research, we explore generative adversarial networks (GANs) for generating diverse training layouts and for transforming layouts to another domain. We propose two layout synthesis methods. In the first method, we represent a layout pattern with its low-frequency DCT (discrete cosine transform) signals and generate a new pattern by generating new DCT signals. For generating DCT signals we use a GAN network. This first method is suitable for applications where a global representation of layout is required. In this method, a layout is represented using a large number of features, where visualizing the feature space is not possible. The second method is for selective layout synthesis, suitable for a relatively low-dimensional feature space. In this method, a layout pattern is divided into a grid and represented by an image parameter set (IPS), consisting of minimum intensity, maximum intensity, and maximum slope of intensity in each grid. New layout patterns are generated from given IPS values at the center of the pattern. We also study GANs for predicting process variation band (PVB) in the lithography process. We propose a GAN-based fast PVB prediction flow for full-chip. We show experimentally that GANs can be adopted for PVB prediction and related applications such as hotspot detection.