Model management for low-computational-budget simulation-based optimization of antenna structures using nature-inspired algorithms

Antenna design has been increasingly reliant on computational tools, specifically, full-wave electromagnetic (EM) analysis. EM simulations are capable of rendering a reliable characterization of complex antenna architectures and quantify the effects (mutual coupling, dielectric losses, feed radiation, etc.) that cannot be accounted for using other methods. At the same time, it is CPU-intensive. Repetitive simulations incurred by numerical procedures, especially optimization, constitute a serious bottleneck of EM-driven design. Perhaps the most extreme example thereof is global tuning of antenna parameters, which is typically performed using soft computing methods, in particular, nature-inspired routines. Although these methods are generally recognized for the ability to handle multimodal problems, their computationally efficiency is poor; direct application to EM models is generally prohibitive. A viable alternative is the incorporation of surrogate-assisted frameworks, along the lines of efficient global optimization (EGO) paradigm, in which the surrogate model is refined in an iterative manner using aggregated EM data and serves as a prediction tool which facilitates finding the optimum design. Unfortunately, the scope of applicability of surrogate-assisted methods is encumbered by the curse of dimensionality, and also nonlinearity of antenna responses. The primary objective of this study is investigation of the possibilities of accelerating nature-inspired optimization of antenna structures using multi-fidelity EM simulation models. The primary methodology developed to achieve acceleration is a model management scheme in which the level of EM simulation fidelity is set using two criteria: the convergence status of the optimization algorithm, and relative quality of the individual designs within the solution pool. The search process is initiated using the lowest-fidelity (therefore, the fastest) EM model. The fidelity is step-by-step increased towards the conclusion of the process. At the same time, lower-quality designs are evaluated at lower resolution level as compared to the better ones. Our technique has been extensively validated using several microstrip antennas, and particle swarm optimization (PSO) algorithm as the search engine. The obtained results demonstrate that making the EM model fidelity dependent on just the convergence status of the algorithm allows for relative savings from forty to seventy percent, depending on the algorithm setup. At the same time, managing model fidelity as a function of both convergence status and relative design quality (within the population processed by the algorithm) allows for up to 85% savings, as compared to high-fidelity-based algorithms. Furthermore, the achieved acceleration is not detrimental to the optimization process reliability. Apart from the computational efficiency, the attractive feature of the proposed approach is implementation simplicity and versatility: the presented management scheme can be readily incorporated into most nature-inspired routines.