Efficient modeling of complex Analog integrated circuits using neural networks

Abstract

This paper introduces a black-box method for automatically learning an approximate but simulation-time ef- ficient high-level abstraction of given analog integrated circuit (IC). The learned abstraction consists of a non-linear auto- regressive neural network with exogenous input (NARX), which is trained and validated from the input-output traces of the IC stimulated with particular inputs. We show the effectiveness of our approach on the power-up behavior and supply dependency of a CMOS band-gap reference (BGR) circuit. We discuss in detail the precision of the NARX abstraction, and show how this model can be used and implemented in testing of Analog ICs within the Cadence environment. By using our method one can automatically learn high-level abstractions of all the components of an Analog IC. This dramatically speeds up the transient simulation time of the Analog ICs.