Abstract
With a short product cycle as we see today, fast and accurate modeling methods are becoming crucial for the development of new generation of electronics devices. Furthermore, increased complexity in circuitry and integration compounds design iteration and the associated, high-dimensional sensitivity analysis and performance optimization studies. Therefore, black-box surrogate models replacing the actual circuitry offer an attractive alternative for more efficient design iteration, optimization, and even direct Monte Carlo analysis. In this article, surrogate models built using nonparametric Gaussian process (GP) are presented. A robust framework based on probabilistic programming is used for training GP models. Other methods, such as partial least-square regression, support vector regression, and polynomial chaos, are used to compare with the performance of GP. Three design applications, a high-speed channel, a millimeter-wave filter, and a low-noise amplifier are used to demonstrate the robustness of the proposed GP-based surrogate models.
Highlights
Expedient design iteration and performance optimization and design verification of state-of-the-art electronic devices and systems are hindered by the ever-increasing functionality integration
Hanzhi Ma and Er-Ping Li are with the Zhejiang University, Hangzhou, China, e-mail: {mahanzhi, liep}@zju.edu.cn autoregressive network with exogenous input type of recurrent neural network (NARX-RNN), while [13]–[17] use Elman RNN (ERNN) to model electrostastic discharge (ESD) circuits, digital high-speed link, etc. without an explicit feedback connection in the model construction. [17] and references therein has a throrough review about different NN-based approaches for time-domain circuit simulation and detailed explanation about different type of RNN models and their advantages as well as disadvantages, [18] combines RNN models with system identification to improve the prediction accuracy up to 99%
Single-out models have more difficulties learning the shape factor than the center frequency and the bandwidth as shown in Figure 2b, Partial Least-square Regression (PLS) and Support Vector Regression (SVR) models require more training samples to converge than MOPLS or multioutput GP (MOGP)
Summary
Expedient design iteration and performance optimization and design verification of state-of-the-art electronic devices and systems are hindered by the ever-increasing functionality integration. In the quest for computationally efficient methods capable of handling the high-dimensional design space of such devices and systems, machine learning (ML) methods are being explored recently for modeling and design optimization applications. Using neural network for modeling, for frequency domain analysis on passive linear time-invariant (LTI) systems, [1]– [5] use feed-forward neural networks (FNN) to learn the mapping from geometry parameters to electrical measure such as S-parameter. For time-domain modeling, [7]–[12] present modeling methods using nonlinear. GP, multi-output GP, overall performs consistently well in both experiments; it offers an attractive option for input-output black-box modeling, and for efficient uncertainty propagation and sensitivity analysis.
Sign-in/Register to view highlights & summary Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callMore From: IEEE Transactions on Components, Packaging and Manufacturing Technology
Disclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
