We develop novel antenna arrays in the entire sub-THz frequency regime, including ultra-wideband antenna arrays, high-gain antenna arrays, met surfaces, etc.
A Deep Learning Convolutional Neural Network for Antenna Near-Field Prediction and Surrogate Modeling
This study investigates the use of deep learning techniques for building a generalized surrogate model that can accurately and very efficiently predict antenna performance parameters. Notably, we focus on applications where a substantial amount of simulation time is required and prior data is available for deep learning use. Specifically, for these applications, we introduce deep learning models that efficiently and reliably model the near-field of the antenna. These models, in turn, accurately predict far-field properties and essential antenna metrics, such as the reflection coefficient. To demonstrate the efficiency of our method, the widely used rectangular patch antenna is considered, encompassing variations in several important geometrical parameters, dielectric constant, and frequency. Based on our results, the proposed model, once trained, is over 200 times faster than conventional full-wave simulations with a nominal average root mean square error (RMSE) of 0.0174 in predicting all the necessary antenna parameters, such as resonant frequency, radiation pattern, and directivity.
M. R. Khan, C. L. Zekios, S. Bhardwaj and S. V. Georgakopoulos, “A Deep Learning Convolutional Neural Network for Antenna Near-Field Prediction and Surrogate Modeling,” in IEEE Access, vol. 12, pp. 39737-39747, 2024, doi: 10.1109/ACCESS.2024.3377219.
A Finite Element-Based Characteristic Mode Analysis
A novel Green’s function-free characteristic modes formulation is introduced in this work. The desired impedance or admittance matrix is obtained utilizing and appropriately modifying the versatile finite element method. For this purpose, the generalized eigenvalue problem of the electric or magnetic field vector wave equation is formulated. In the case of the electric field wave equation, using the Schur complement, the system is reformulated and expressed only in terms of the tangential electric field over the radiating apertures, retaining the equivalent magnetic currents. Similarly, in the case of the magnetic field wave equation, the electric current density on radiating metallic surfaces is isolated using the Schur complement. In both cases, the obtained matrix is split into its real and imaginary part to yield the characteristic modes eigenvalue problem. Key advantage of the proposed formulation is that it does not require the evaluation of Green’s function, thereby the study of any arbitrarily shaped, multilayered geometry loaded with anisotropic and inhomogeneous materials is feasible. To prove the validity of the proposed methodology various classical structures, with both homogeneous, and inhomogeneous and anisotropic materials, published in the bibliography are studied. Both the eigenvalues and eigenvectors compared with the published results show good agreement.
D. Paschaloudis, C. L. Zekios, S. V. Georgakopoul2022.3150594os and G. A. Kyriacou, “A Finite Element-Based Characteristic Mode Analysis,” in IEEE Open Journal of Antennas and Propagation, vol. 3, pp. 287-303, 2022, doi: 10.1109/OJAP.
Multifidelity Surrogate Modeling Based on Analytical Eigenfunction Expansions
In this work, a new method is proposed to derive the initial approximate model for a multifidelity (MF) surrogate optimization. Specifically, the proposed method is trained using a set of eigenfunction expansions that characterize the solution domain of the desired geometry and high-fidelity (HF) full-wave simulations. To demonstrate and validate the proposed method, an array of loops, a pyramidal horn antenna, and patch antennas of arbitrary shapes are studied. Notably, the proposed MF method is applied and tested in single- and multiobjective optimization settings to achieve two or three design goals. Our studies illustrate that the proposed eigenfunction expansion-based method can create approximate models needed in MF optimizations up to 243 times faster than the conventional coarse mesh low-fidelity (LF) approaches. This in turn makes the total training time of our MF models up to 2.8 times shorter than the conventional MF models.
D. Paschaloudis, C. L. Zekios, S. V. Georgakopoul2022.3150594os and G. A. Kyriacou, “A Finite Element-Based Characteristic Mode Analysis,” in IEEE Open Journal of Antennas and Propagation, vol. 3, pp. 287-303, 2022, doi: 10.1109/OJAP..