Neural-network-based transient modeling of nonlinear electronic circuits for high-speed applications.

摘要

Artificial neural networks (ANN) have recently emerged as useful computational tools for the computer-aided design (CAD) of nonlinear radio frequency (RF)/microwave devices and circuits. The overall objective of this thesis is to develop efficient and robust neural network methodologies for transient modeling and simulation of nonlinear electronic circuits. The first contribution of the thesis is the development of a novel neural-network-based transient modeling technique. Specifically, a state-space dynamic neural network (SSDNN) technique is proposed for modeling the transient behavior of essential nonlinear blocks in high-speed systems, such as drivers and receivers. The proposed SSDNN expands the existing time-domain ANN structure into a general and flexible formulation for transient-oriented modeling. Under this generalized framework, a set of stability criteria is presented for verifying both local/global stabilities of a trained SSDNN model. Furthermore, a new training algorithm is developed that incorporates the proposed stability criteria into model training as a series of constraints. Using the proposed constrained training, the trained SSDNN model can maintain high accuracy while preserving global stabilities. As a further contribution, this thesis presents a new training approach for robust transient modeling of nonlinear circuits based on the recurrent neural network (RNN) formulation. One of the main challenges of nonlinear transient modeling using neural networks is that the training information depends on the shapes of circuit waveforms and/or circuit terminations. To overcome this limitation, an internal space of the RNN is exploited to relate the information of training waveforms with that of test waveforms. Based on this concept, a new circuit block combining a generic load and a generic excitation is formulated as a general termination of the original circuit. By sweeping the coefficients of the proposed circuit block, a rich combination of training waveforms is obtained to effectively cover the region of interest in the internal space of the RNN. Through this systematic approach, the resulting RNN model maintains good accuracy for versatile test waveforms and test loads that are not known during training. The techniques presented in this thesis provide a further advancement beyond the existing methods, enabling accurate and robust transient modeling of nonlinear circuits for high-speed CAD.