Evolutionary optimization for continuous, binary, mixed-variable and constrained electromagnetic problems

Evolutionary optimization algorithms are a type of computational intelligence. Their application to electromagnetic (EM) engineering problems open exciting possibilities of improving the performance of existing solutions, finding novel complex configurations and, ultimately, having the potential to fully automate the design of antennas, filters and other electronic devices. This thesis focuses on extending the functionality that is provided by evolutionary algorithms to the field of electromagnetics. A detailed study is undertaken to identify the needs of real-world EM engineering problems and propose methodologies that enhance the available capabilities. The EM problems optimized in this thesis include compact high-gain wide-band resonant cavity antennas (RCAs), wide band aperture-coupled microstrip patch antennas (ACMPAs), pixelated EM surfaces and microstrip filters. In order to obtain accurate performance prediction of the designed devices, all considered designs are simulated using a full-wave EM solver. The optimization process is automated by interfacing simulation software with evolutionary algorithms. Particle swarm optimization (PSO), the cross-entropy (CE) method and co-variance matrix adaptation evolutionary strategy (CMA-ES) are the evolutionary algorithms used in this thesis. By means of PSO, improved designs of a new class of compact RCAs are obtained that outperform the designs reported in the literature. Four antennas of only 1.7-2.2λ0 in diameter have a directivity 17.6-19.6dBi and a 3-dB radiation bandwidth of 24%, 50%, 55% and 70%. An ACMPAwith 53% bandwidth is proposed as a new planar feed solution for RCAs. The designed RCA with an ACMPA feed has a peak directivity of 19 dBi and a 3-dB radiation bandwidth of 40%. The most important outcome of this thesis are the proposed methodologies for the optimization of continuous, mixed-variable, binary and constrained EM problems. It is shown that various design requirements can be incorporated in optimization using the CE method, which is not possible with PSO or CMA-ES. In contrast to the previous CE applications, which only utilise the normal distribution, this thesis makes use of probability distribution functions that model the design space, such as the beta, Dirichlet, Bernoulli and discrete probability distribution families. Thus, the flexibility and broader potential of the CE method is exploited and demonstrated. By encoding the metallic patterns printed on thin pixelated EM surfaces into binary strings, single- and dual-frequency artificial magnetic conductors and phase-shifting metasurfaces are designed. The methodology of handling equality and inequality constraints is demonstrated on the design of a 5-pole microstrip low-pass filter, which has a selectivity of 60 dB and an 8.5 GHz rejection bandwidth. All the designs in this thesis are obtained by an automated optimization process. In comparison with the initial designs or existing alternative solutions, the optimized designs exhibit improved desired characteristics.