Efficient modeling of RF CMOS spiral inductors using generalized knowledge-based neural network

Abstract

Efficient modeling of RF CMOS spiral inductors by virtue of a novel generalized knowledge-based neural network (GKBNN) is presented. Prior knowledge of on-chip inductors is used for constructing the GKBNN. This new modeling approach also exploits merits of the iterative multi stage algorithm. This GKBNN has much enhanced learning and generalization capabilities. Comparing with the conventional neural network or the knowledge-based neural network, this new GKBNN model can map the input---output relationships with fewer hidden neurons and has higher reliability for generalization. As a consequence, this GKBNN model can run as fast as an approximate equivalent circuit model yet generate results as accurate as detailed electromagnetic simulations. Experiments are included to demonstrate merits and efficiency of this new approach.