FDTD-Backed RBF Neural Network Technique for Efficiency Optimization of Microwave Structures

A novel approach to computational microwave optimization is presented, namely, a neural network procedure controlling full-wave FDTD simulation of the structure. A radial-basis-function network is trained by geometrical parameters and the computed frequency responses of S-parameters. Computational efficiency of the procedure is illustrated for a waveguide T-junction with a post, a slotted waveguide radiating in free space, and a microwave oven with a sausage on the shelf. Introduction Although traditionally extensive experimentation was the major technique exploited in the development of microwave (MW) applicators, with appearance of new generation of modeling techniques, it has been recently realized that advanced computer simulation could make the design of the MW heating systems more intelligent and thoughtful, shorten the development time, and reduce the project’s cost. Modern modeling tools allow one to get valuable characteristics of the structure prior to constructing a physical prototype. However, routine analysis of the system may not result in many direct instructions for better design. The emerging practicability of inclusion of resourceful full-wave simulators in optimization and automated design of MW structures has been emphasized in [1], but this has not become a standard practice yet as traditionally considered unfeasible due to the high computational cost. Meanwhile, recent extraordinary growth of productivity and capabilities of computer hardware has made possible a comprehensive, fast and reliable numerical modeling of microwave circuits. Microwave Studio (MWS) by CST, the EM code based on Finite Integration Method, and QuickWave-3D (QW3D) by QWED, the conformal FDTD 3D electromagnetic (EM) simulator, have been recently identified among the most efficient full-wave simulators in the market [2]. The present paper proposes, for the first time, a simple yet efficient optimization technique based on artificial neural networks (NN) controlling 3D full-wave FDTD analysis performed by QW3D (http://www.qwed.com.pl). We show that, given the resources of today’s computers, such an approach can be reasonably productive and serve as a competent optimization tool in designing various MW systems. Background In this analysis we understand optimization of MW structure as traditional circuit optimization. While possible approaches to improving MW heating may include special findings in the applicator design, manipulation of the heat cycle, ingredient formulation, design of the package, etc. (see, e.g., [3, 4]), here we deal with parameters computable in the EM modeling. The use of time-domain simulation in design of practical systems has recently made it evident that the applicator’s efficiency (usually associated with energy coupling) can be controlled when one can compute a frequency characteristic of the magnitude of the reflection coefficient |S11| in the range adjacent to the operating frequency f0 [5]. So we interpret the problem of efficiency optimization as that of an appropriate optimization of the frequency response of |S11|. Neural networks were introduced in the computational electromagnetics in the 1990s, and, since then, their typical application has been associated with the networks representing (or directly imitating) the modeled devices and dealing with their Proceedings of the 9 AMPERE Conf. on Microwave & RF Heating, Loughborough, U.K., September 2003