A method for creating behavioral models of oscillators using augmented neural networks

Abstract

This paper describes a novel technique to model the nonlinear time-domain behavior of oscillators using augmented neural networks. In the proposed method, a feed forward neural network with a periodic unit is used to capture the periodicity of the oscillatory output waveform. As opposed to the state space model, which is based on a system of differential equations, the output of the oscillator is generated explicitly using the neural network presented in this paper. The model is trained using the data obtained from the simulation of transistor-level circuit models. Examples applied to ring oscillators show the advantages using this method based on CPU time and accuracy. The proposed model is compatible with Verilog-A.