SRAF printing prediction using artificial neural network

Sub-resolution assist features (SRAFs) are inserted in mask layout to improve the manufacturing process window of main patterns. SRAFs should be large enough to maximize their effect, but they are not intended to be printed on photoresist (PR). An accurate method of SRAF printing prediction is important to assure that no SRAFs are actually printed. We apply a machine learning model, specifically artificial neural network (ANN), for fast and accurate SRAF printing check (named ML- SPC). Polar Fourier transform signals and local layout densities are extracted from each SRAF pixel and its surroundings, and are provided to ANN. The area sum of member pixels that are predicted to be printed is used to determine the final printability of SRAF. Training data is carefully sampled so that similar number of printed and non-printed data are used for training; cost function is adjusted in such a way that missing predictions are treated more importantly than false alarms. When ML-SPC is applied to 10nm memory devices, it achieves 11% of false alarms while popular MTA method reports 24%. In addition, ML-SPC is faster than MTA by about 2.7 times.