Unified inverse lithography technology (ILT) and fracturing through recurrent neural network

ILT processes input layout patterns and yields curvilinear mask patterns. Fracturing follows to obtain a collection of rectangular shots, which are given to e-beam mask writer. Machine learning (ML) has been applied for quick prediction of ILT outcome. Since curvilinear mask patterns are very complex, such prediction requires large ML model with significant training effort. We are motivated by the observation that prediction of fracturing outcome, a collection of shots, from the layout patterns will require much simpler ML model. Such prediction will be toward unified ILT and fracturing. A large file to store ILT output is not necessary, which will be another benefit. A recurrent neural network (RNN), consisting of a number of neural network (NN) instances connected through recurrent hidden layers, is considered for the ML model. A layout is divided into a number of clips, and RNN is applied to each clip one by one. A clip is encoded into its feature vector, and the vectors from the neighbor clips are aggregated to it. Each NN instance receives an aggregated feature vector, shot coordinates generated by all previous instances, and hidden state information as inputs; two diagonal coordinates for a shot are extracted at the NN output. Some extracted shots are combined and some are refined for better quality of final shots. Experimental results demonstrate that the proposed approach achieves similar shot count and similar shot quality (as measured by the number of edge placement error (EPE) violations) compared to standard ILT with fracturing, while speedup is inherently significant, more than 150 times.