DESIGN OF FREQUENCY SELECTIVE SURFACES USING A GENETIC ALGORITHM-DRIVEN FINITE-DIFFERENCE TIME-DOMAIN METHOD

Abstract

The design of Frequency Selective Surfaces (FSS) to meet a desired frequency response is an arduous task due to the complex nature of the variables involved. The Genetic Algorithm (GA) is thus a useful technique to solve the optimisation problem as it can handle complex combinatorial variables. Moreover, the GA has a wide range of adjustable parameters which make it a robust method for optimisation. Of the various crossover and mutation schemes tested, GA performed the best with a high degree of elitism, uniform crossover with mutation (probability of mutation 0.01) and single child from a single parent pair. The GA was able to find solutions that fitted the transmission and reflection coefficients of 2 test target FSS screens with a high level of accuracy, even though visually the designs looked starkly different. Both solutions were within 5.00 % of the global optimum. This verified the results from the experiments done without the PB-FDTD code. The number of function evaluations needed were 3,907 and 3,844, and each function evaluation took approximately 8 minutes on a Pentium 3 933 MHz 128 Mb RAM machine. Initial work has already been started on including other design parameters such as relative permitivities of the dielectric layers, thickness, and position of the FSS screen, apart from the FSS unit cell metallization. A major obstacle, however, is the timeframe needed for the GA to obtain a solution within a 5.00 % margin. Proposed solutions are porting the presently Windows-based PB-FDTD code into Linux, making use of Linux-based multi-processor parallel computers, or modifying the presently general PB-FDTD code to take further advantage of the problem's geometry. Other numerical methods, such as the method of moments, may be a better option because the FSS composites considered in this problem are generally thin and infinite in the x- and y- directions. The PB-FDTD method though versatile enough to handle a wide range of geometries, is too computationally expensive to use in this context.