Efficient prediction of interconnect crosstalk using neural networks

Interconnect crosstalk prediction has become increasingly important with deep submicron downscaling of IC's and wafer scale integration. Existing tools for management of the emi problem are computationally expensive and not very broad in application. The unique approach proposed involves the creation of parameterized models of primitive interconnect structures - wirecells - using modular artificial neural networks (MANN's). The finite element method, a circuit simulator and a neural network multi-paradigm prototyping system are coupled together to produce a library of re-usable MANN-based wirecell models. The method is especially attractive because it is capable of modeling the simultaneous effect of several uncorrelated variables such as interconnect length, width, thickness, separation, conductor and insulating medium characteristics on crosstalk and delay. Equi-coupling contours called isocouples are derived for noise characterization to guide design activities such as placement (e.g. matched devices placed on same isocouple). Experimental results from a transconductance amplifier demonstrate the viability of the approach.