Surrogate Modeling of Nonlinear Components and Circuits
Abstract
Surrogate models are simplified approximations of functions that are described by complex equations. Surrogate modeling of physical systems have been used in various fields such as biology, fluid dynamics, climate modeling, and various other engineering disciplines. This data-driven approach is used to decrease computational cost, decrease computation time, or when the output of a system is difficult or impossible to measure, by using a "black-box" method to approximate the output given inputs. In regards to circuit analysis, surrogate models can be used to decrease computation time and computational load.
In this thesis, surrogate modeling is used to model various nonlinear components and circuits in fREEDA, a multi-physics circuit simulator, for the purpose of speeding up transient analysis. Neural networks are used in place of physics-based equations, resulting in a speedup of 5-18x for the evaluation of the components and 3x for the evaluation of entire circuits. The components and circuits tested in this work include: BJT (Bipolar Junction Transistor), MOSFET (Metal-Oxide-Semiconductor Field-Effect Transistor), common-emitter amplifier, and common-source amplifier.