A new second–order adaption rule and its application to electrical model synthesis

Abstract

Analog circuit simulation requires the knowledge of current and voltage behaviour and physical parameter dependencies of the devices used in the electrical circuit. Although a couple of nonlinear device models are implemented in the circuit simulator, the analog model of the hardware-realization of a standard-circuit device, as for example a digital Gate or an A/D converter from Texas Instruments, is seldom available. In this case, the terminal behaviour and parameter dependencies of voltages and currents have to be measured, and the results must be made usable for the circuit simulator.

We use neural networks to approximate the terminal behaviour of electrical devices, maintaining the parameter dependencies. We have improved the adaption rule by an adaptive evaluation of the learning parameters to accelerate the approximation time. The network paradigm can be automatically transformed either into a net list of an electrical subcircuit (for SPICE-simulation for example) or into a mathematical description language (for a behavioural simulator like SABER for example). Examples demonstrate the very accurate representation of nonlinear electrical devices for circuit simulation.