Sub-micrometer Bipolar Transistor Modeling Using Neural Networks

Abstract

An approach based on Artificial Neural Networks (ANN) for constructing models of high speed bipolar transistors is described. This method is proposed as an alternative for physical modeling for circuit simulation, when high frequency and small device size make classical models very complex or even unreliable. In the ANN here adopted, neurons are represented in terms of continuous-time differential equations, allowing the immediate application inside conventional circuit simulators. The most difficult task in this approach is the network training from the measurements on the real device, and the usual learning rules for ANN's easily lead to poor approximation or unacceptable slowness. A generative method has been developed, where a subset of the network parameters is trained inside an auxiliary static network, using measurements at fixed DC bias. The complete network is trained on the full set of measurements using this subset as a starting point. The learning rule is a combination of global optimization followed by a quasi-Newton conjugate-gradient iterative process.