Low-Order Rational Approximation of Interconnects Using Neural-Network Based Pole-Clustering Techniques

This paper presents a neural network approach of pole clustering techniques to construct compact models of high-order systems. The approach can be used to reduce the order of complex models of high-speed interconnect systems obtained from standard rational approximation. The reduction of the order and complexity of circuit models are essential to improve the efficiency and stability of the time-domain simulation in very large distributed systems. The proposed procedure uses the clustering capabilities of self-organizing maps of artificial neural networks. Self-clustering maps are very suitable to identify the pole distributions, and efficiently generate cluster centers and representative poles for the compact models. To illustrate the validity of the method, examples of frequency-domain simulation results of high-speed memory system are given