Gradient-based inverse problem methodology for design optimization in electromagnetic applications
Abstract
Computer Aided Optimum Design (CAOD) is employed in this dissertation to obtain optimum designs of some electromagnetic devices to satisfy the given performances. Real world problems are "inverse problems" in which a part of the device is unknown. In inverse problems, the performance of the device is given and the unknown part has to be identified. Inverse Problem Methodology (IPM), which is a CAOD technique, is a formal approach to solve inverse problems. In this approach, the device is described by variables; an object function is defined in such a way that the given performance is satisfied by the optimum design at the optimum point of the function. The optimum values of the variables have to be determined by optimizing the function. In the GIPM, gradient-based optimization techniques are combined with the Computer Aided Analysis (CAD) techniques. To obtain optimum designs of a capacitive transducer and an electromagnetic circuit, and to identify the dimensions of an electrostatic source, the optimization process was employed at multi-levels. The multi-level techniques are required to overcome the difficulties arose in identifying all the variables at the same time. The object functions corresponding to the transducer and the identification problem have several local minima and a narrow valley respectively. In this dissertation, Powell's conjugate directions method is modified and employed to find the global minima. The GIPM is applied first time to miniaturize a MOSFET device to avoid breakdown due to high electric fields. The pole face of an electromagnetic device is optimized using GIPM, but the optimum shape is not smooth enough. This is the first time some linear constraints are imposed to smoothen the pole face. The sensitivities of the magnetization vectors of PM devices are computed first time using the formulations derived from the finite element analysis and applied for the optimization of a circuit with a PM and a PM pole face.
This paper has been withdrawn.