Smart Modeling of Microwave Devices

Abstract

Modeling and computer-aided design (CAD) techniques are essential for microwave design, especially with our drive towards first-pass design success. In the past few decades, tremendous progress in microwave CAD has led to a large variety of microwave models for passive and active devices and circuit components. The high quality and the availability of these models have enabled us to design circuits efficiently. These models have also allowed us to design larger and more complicated circuits than ever before.At the same time, new technologies and materials, emerging and non-traditional devices continue to evolve. Although the existing models are good for modeling mature technologies and existing devices, they are often inadequate or unsuitable when new devices are needed in system design. Conventional approaches to create or modify models are heavily based on slow trial-and-error processes.As new technologies and devices continue to evolve, we need not only new models, but also computeraided modeling algorithms such that model development becomes fast and systematic.At high frequencies, equivalent circuit models often lack fidelity. Detailed electromagnetic (EM) based simulations become essential to achieve design accuracy. However, EM simulations are computationally expensive especially when physical or geometrical parameters have to be repeatedly adjusted during design cycle. With the increasing design complexities, coupled with tighter component tolerances and shorter design cycles, there is a demand for design methodologies that are both accurate and fast at the same time. These are contradictory requirements and difficult to satisfy with conventional CAD techniques. The problem becomes even more severe in yield optimization and statistical validation where process variations and manufacturing tolerances of components are required to be taken into account. In addition, accurate parametric modeling techniques have become increasingly necessary, where we strive to describe not only the behavior of the microwave device, but also the change of the behavior against physical or geometrical parameters of the device.In recent years, neural network (NN) or artificial neural network (ANN) techniques have been recognized as a useful alternative to conventional approaches in microwave modeling [1]- [2]. Artificial neural networks can be used to develop new models or to enhance the accuracy of existing models. Neural networks learn device data through an automated training process, and the trained neural networks are then used as fast and accurate models for efficient high-level circuit and system design. These models have the ability to capture multi-dimensional arbitrary nonlinear relationships. The theoretical basis of neural network is based on the universal approximation theory [3], which states that a neural network with at least one hidden layer can approximate any nonlinear continuous multidimensional function to any desired accuracy. This makes neural networks a useful choice for device modeling where a mathemati...

A novel quasi-Newton-based optimization for neural network training incorporating Nesterov's accelerated gradient

“…The trained neural networks can be used as models of microwave devices in place of CPU-intensive EM/physics models to significantly speed up circuit design while maintaining EM/physics-level accuracies [2,4]. Neural network based modeling has been used to model a variety of microwave circuit components at both device and circuit levels.…”

“…QN is one of the most efficient gradient-based optimization algorithms [10,11] and commonly-used training method for highly nonlinear function problems [2][3][4][5][6]. Note that, the secant equation,…”

“…For example, the techniques are useful for microwave modeling and design in which neural networks can be trained from Electro-Magnetic (EM) data over a range of geometrical parameters and trained neural networks become models providing fast solutions of the EM behavior it learned [2][3][4]. This is useful for modeling where formulas are not available or original model is computationally too expensive.…”

“…However these modelings have strong nonlinearity to itself and need lower training error. Therefore the second-order training methods are more suitable than the first-order algorithms [4].…”

“…On the other hand, quasi-Newton (QN) method which is one of the most efficient optimization techniques [10] is widely utilized as the robust training algorithm for highly nonlinear function approximation [2][3][4][5][6]. Furthermore several improvements for QN have been done [3,5].…”