Neural network computing applied in design for manufacturability

Abstract

A DFM (design for manufacturability) method that applies artificial intelligence techniques and robust design ideas and can be used in many semiconductor design and manufacturing areas is described. The method includes both controllable design and uncontrollable variation factors early in the design stage. It uses neural network computing techniques to find an optimal design for controllable design factors of maximizing the yield of the semiconductor product design and manufacturing process. The controllable design factors include the device component values, device physical variables, and fabrication parameters such as deposition time, rate, and doses. The uncontrollable factors are environmental condition variations such as temperature and humidity, as well as fabrication inaccuracies in alignment and diffusion. The method consists of sample, relate, and optimize (SR) stages. At the sample stage, a set of confined random design points are chosen from an initial design space. At the relate stage, the relationship between these points and their responses are used to train an artificial neural network (NN) based on a backpropagation model. At the optimize stage, acceptable inputs can be predicted and therefore the manufacturing yield increased by using this NN computational scheme. This method, when applied to a semiconductor processing-dependent IC design example, showed satisfactory results. >