Computer aided design of microwave circuits with neural networks

提出的研究计划的目标是开发快速和准确的微波电路计算机辅助设计的神经模型。传统的微波滤波器/多路复用器建模方法是基于电磁场（EM）的方法，这种方法非常精确，但计算量很大;或者是基于经验或等效电路的方法，这种方法速度快得多，但精度有限。作为一种非传统的建模技术，神经网络最初受到人类大脑从观察中学习和抽象概括能力的研究的启发。近年来，有越来越多的兴趣在应用神经网络的微波设计问题。在本研究计画中，将类神经模型应用于微波滤波器与多工器的设计与模拟。通过收集基于EM的仿真数据作为训练数据来开发神经模型。在输出多路复用器应用的情况下，频率范围和带宽被很好地定义，使得信道滤波器成为神经建模的理想候选者。在训练好神经网络模型后，单个滤波器和多路复用器的建模速度可以比传统的建模方法快几百倍，同时保持相当的精度。微波滤波器/多路复用器广泛应用于空间和地面通信系统。建议的研究计划，预计将显着提高微波滤波器/多路复用器的仿真效率和产品设计能力，由于上级仿真速度和良好的精度开发的神经模型。

The objective of the proposed research program is to develop fast and accurate neural models for microwave circuit computer aided design. The emphasis will be on the application of neural network for microwave filter and multiplexer design and simulation.Conventional approaches for microwave filter/ multiplexer modeling are either electromagnetic (EM)-based, which is very accurate but computationally intensive, or empirical or equivalent-circuit-based, which is much faster but limited by the lack of accuracy. As a non-conventional modeling technique, neural networks are originally inspired by the studies of the ability of the human brain to learn from observations and to generalize by abstraction. In recent years, there have been increased interests in applying neural networks to microwave design problems. In this research program, neural modeling is applied to microwave filter and multiplexer design and simulation. Neural models are developed through collecting EM-based simulation data as training data. In the case of output multiplexer application, the frequency range and bandwidth are well defined, making channel filters ideal candidates for neural modeling. After the neural models are trained, modeling of single filters as well as multiplexers can be done hundred of times faster than conventional modeling technique while retaining comparable accuracy.Microwave filters/multiplexers are widely used in space and ground based communications systems. The proposed research program is expected to significantly enhance microwave filter/multiplexer simulation efficiency and product design capabilities due to the superior simulation speed and good accuracy of the developed neural models.