Home | english  | Impressum | KIT

Neural-Networks-Based FDTD-Backed Computational Schemes for Microwave Inverse Problem and Optimization

Neural-Networks-Based FDTD-Backed Computational Schemes for Microwave Inverse Problem and Optimization

Dipl.-Wi.-Ing.  Matthias Rother 
Dr.-Ing. Marc Regier





innovative challenge of the research work

In the conventional techniques of microwave optimization, the analysis part is based on methods employing equivalent lumped and transmission line element networks and resulting in quick generation of data for the optimized system. Unfortunately, these methods are not sufficient for representing many of the complex devices and electrically large systems. However, it is typically considered impractical to use full-wave modeling instead, because although it is capable of handling a much wider class of problems, it requires much more CPU time. Because of this, many practically important scenarios are not dealt with using optimization. Our goal in this project is to demonstrate that full-wave numerical solution backing numerical optimization could be viable. We will develop an innovative approach which relies on 3D conformal FDTD analysis, is based on artificial neural networks (ANN), and supported by special optimization strategies and methods of acceleration of ANN optimization. Then the approach will be implemented in a series of CAD tools designed, for the first time, for inverse and optimization problems in telecommunications, medicine, microwave thermal processing, and applied to a variety of valuable applications in these fields.

scientific objectives of the research work

In this project, we will develop an innovative approach in neural networks solution of inverse and optimization problems in high frequency electromagnetics. Traditionally, the techniques of ANN-based optimization work with specific microwave devices: the computational schemes are backed by simplified models, and properties and design of those devices are used to adjust the networks. Therefore, these methods are applicable only to relatively simple structures for which the models are available. Our objective is to make it possible to optimize complicated and electrically large systems of arbitrary configuration/design. This will be achieved by designing special ANN (1) capable of quickly finding an optimal solution, (2) employing the results of full-wave numerical analysis (3D conformal FDTD method), and (3) supported by special optimization strategies. Since accurate simulation of complicated scenarios may take notable CPU time, the key problem here is how to solve optimization/inverse problems with limited FDTD data (and hence within reasonable computational time). We will work on this problem by developing special algorithms of database forming and the network training/testing. To do this, rather than first build the entire database (which may become unnecessarily large) and then train the network, we shall instead add FDTD data to the database as long as it is needed for successful training – provided that this process is conditioned by the scenario being optimized. With appropriate specifications, the concept of ANN-FDTD microwave optimization can become the basis of two novel efficient methods: (1) Determination of dielectric properties of arbitrarily shaped materials – based on elementary measurement of S -parameters and placing minimal physical requirements on fixture and sample geometry, the method will remain practically simple, convenient, and accurate. (2) Non-invasive reconstruction of 2D/3D images of objects exposed to the microwave field – the technique characterized by resolution and accuracy exceeding those demonstrated so far.

We shall mathematically formulate the problem of control and optimization of the electric field and power dissipated in the lossy materials situated in closed cavities. Then we shall expand the formulation towards temperature, thus address a major challenge in microwave thermal processing. A particular issue here is an identification of 2D/3D goal functions

resembling the desirable field/temperature distributions in typical applications. Adequate modeling of microwave heating requires efficient numerical solution of the coupled electromagnetic-thermal problems. An additional challenge here is brought by the necessity to make numerical output of FDTD and FEM solvers compatible – we are determined to work with these methods since they are the best ones for modeling electromagnetic and thermal components, respectively.

contribution to the achievement of some of the global targets of the EADS Foundation

Theoretical aspects of this project will result in development of efficient techniques for information processing and thus contribute to the areas of computer technology dealing with data analysis, design methodology, image processing/recognition/visualization, pattern classification/association, and as such may make impact on applications in remote sensing (e.g., of earth from space), telecommunications (e.g., in antenna synthesis), microwave plasma, material sciences, etc. Design of computational tools assisting in image/permittivity reconstruction and in putting microwave heating under control may lead to further expansion of the related technologies on new applications in microwave chemistry, hyperthermia, microwave imaging for medical diagnostics, thermal processing of dielectrics including ceramics, composites, wood, and various technologies involving drying, melting, curing, thawing, sterilization, pasteurization, cooking, etc. In particular, efficient and controllable microwave processing of food products may stimulate more intense use of this technology in industry, aircrafts and space stations, and help make better and safer food