Data‐driven model based design and analysis of antenna structures

Abstract

Data-driven models, or metamodels, offer an efficient way to mimic the behaviour of computation-intensive simulators. Subsequently, the usage of such computationally cheap metamodels is indispensable in the design of contemporary antenna structures where computationintensive simulations are often performed in a large scale. Although metamodels offer sufficient flexibility and speed, they often suffer from an exponential growth of required training samples as the dimensionality of the problem increases. In order to alleviate this issue, a Gaussian process based approach, known as Gradient Enhanced Kriging, is proposed in this work to achieve cost-efficient modeling of antenna structures. The Gradient Enhanced Kriging approach incorporates adjoint-based sensitivity data in addition to function data obtained from electromagnetic simulations. The approach is illustrated using a dielectric resonator and an ultra-wideband antenna structures. The method demonstrates significant accuracy improvement with less number of training samples over the Ordinary Kriging approach which utilises function data only. The discussed technique has been favorably compared with Ordinary Kriging in terms of computational cost.

Low-Cost Multi-Objective Optimization of Multiparameter Antenna Structures Based on the l1 Optimization BPNN Surrogate Model

“…Therefore, it is worthwhile improving the optimization speed at the cost of small degradation of the calculation accuracy. Fortunately, surrogate model techniques [28][29][30][31][32][33][34][35][36][37][38][39][40][41] have been proven to effectively avoid the huge computational cost of the EM-driven process. The use of surrogate models has been a recurrent approach adopted by the evolutionary computational community to reduce the fitness function evaluations required to produce acceptable results.…”

“…Once the estimation and uncertainty functions are defined, GP can be handled following rules of probability as applied to multivariate Gaussian distributions. In [37], the researchers constructed a gradient-enhanced GP-based surrogate model by exploiting gradient information from adjoint simulations to reduce the number of training sample points, and the method was used to design dielectric resonators and UWB antennas. ANNs are based on a nonlinear parametric model that can provide a 'universal' approximation capability for emulating the functions describing the behavior of complex systems.…”

“…In [39], multilayer perceptron NN (MLP-NN)-based model and differential evolution algorithm are used to optimize performance index for reconfigurable antenna. In [40], the BP algorithm-based MLP-NN as antenna analysis model is combined with the particle swarm optimization (PSO) algorithm to form a design program for making a custom fractal antenna, and a Sierpinski gasket and Koch monopole antennas were taken as the candidate antennas to verify effectiveness of the developed In [37], the researchers constructed a gradient-enhanced GP-based surrogate model by exploiting gradient information from adjoint simulations to reduce the number of training sample points, and the method was used to design dielectric resonators and UWB antennas. ANNs are based on a nonlinear parametric model that can provide a 'universal' approximation capability for emulating the functions describing the behavior of complex systems.…”

Data-Driven Model-Order Reduction for Magnetostatic Problem Coupled With Circuit Equations

“…With commercial FE softwares, the matrix system is not necessarily accessible. As alternative, nonintrusive approaches of MOR, like Data Driven (DD) methods, have been developed in the literature [2]. One of these approaches (DD-POD) [3], builds an approximation of the reduced model from the known inputs and outputs of the full model.…”

Constrained Algorithm for the Selection of Uneven Snapshots in Model Order Reduction of a Bearingless Motor

“…The matrix system, however, is not always convenient for access, especially when using commercial FE software. Data-driven methods(Ulaganathan et al, 2016;Gosea & Antoulas, 2018), which include system identification and machine learning methods, are alternatives for order reduction without extraction of the matrix system Pierquin et al (2018). combined the data-driven method with POD to build a reduced model of a magnetostatic problem from the known inputs-outputs of the system.…”