Neural networks application for OPC (optical proximity correction) in mask making

Abstract

In this paper a neural network method for optical proximity correction is presented. A non linear two-dimensional spatial inverse filter is proposed as deconvolutional factor over pattern plane. Test sets for the different correction strategies were prepared. Boltzmann machine neural network for training set preparation is proposed. For actual correction optimal autoassociative neural network configuration have been chosen. This strategy also opens the way to use a sub-resolution, non-printable correction structures. Additionally, such a system is available for interfacing with next step of correction - for electron proximity effects. Possibility of merging them in one unit were investigated. Some initial results concerning one step OPC and EPC (electron proximity correction) neurocorrector are presented. Different aggressiveness of the correction can be chosen straightforwardly by simple change of the correction kernels in neurocorrector. There is no structural difference of neurocorrector for attenuated, sized rim, Levenson's, outrigger or chromeless methods, thought different attitudes towards training tests generation have to be applied. Both, feature biasing and feature assisted techniques were investigated. Some attempts to quasi-analytical representation of the correction kernel to minimise or even avoid learning process is also presented. For the full power of artificial neural networks to be exploited from hardware implementation rather, initial considerations on VLSI architecture of the neurocorrector are also included.