Expedited Machine Learning Antenna Optimization Using Dimensionality Reduction, Multifidelity EM Models, and Rapid Final Tuning Through Principal Directions

The design of modern antenna systems has increasingly relied on numerical optimization techniques. Their use is mainly pertinent to geometry parameter adjustment, which often must be realized in a global sense. The latter is computationally expensive, especially when carried out using electromagnetic (EM) simulation models. Acceleration is possible through the employment of surrogate modeling methods; however, rendering reliable surrogates is questionable for challenging scenarios (many design parameters, nonlinear system responses). In this work, we propose an alternative global antenna optimization framework. It is a machine learning approach utilizing kriging metamodels and a particle swarm optimizer to yield subsequent candidate solutions. The dimensionality of the search region is reduced using fast global sensitivity analysis, which facilitates surrogate model setup and handling. The global search stage is further expedited by incorporating multi-fidelity EM models. Whereas the final parameter tuning conducted in the design space of full dimensionality is sped up by sparse sensitivity updates, with gradient estimation using numerical gradients reserved for directions corresponding to the highest antenna response variability. Extensive verification experiments demonstrate consistent performance of our methodology, which is superior to several benchmark methods. Comparative data also corroborates the importance of applied algorithmic solutions from the standpoint of enhancing the search process efficacy and reliability.