What question did this study set out to answer?

This study aims to develop a framework for optimizing and predicting the performance of flower-shaped antennas using data augmentation and machine learning models.

July 14, 2026Open Access

Physics-informed-GPR data augmentation framework for flower-shaped antenna performance optimization and prediction using machine learning models

LALijaddis Getnet AyalewBahir Dar University TKTsion Yigzaw KumlachewBahir Dar University DCDemessu Kebede ChakaBahir Dar University

Key Points

This study aims to develop a framework for optimizing and predicting the performance of flower-shaped antennas using data augmentation and machine learning models.
Developed a physics-informed Gaussian process regression data augmentation framework.
Created predictive models including RSM, ANN, RR, RF, and GB with cross-validation.
Utilized a CST-simulated dataset expanded to 2,014 samples via analytical modeling and electromagnetic constraints.
RF model achieved an R^2 of 0.999, RMSE of 0.031 for resonant frequency, and 0.001 for MSE of impedance bandwidth.
Optimized single-element antenna has a gain of 3.0 dBi at 28 GHz and an impedance bandwidth of 22–41.378 GHz.
Quad-port 2x2 MIMO antenna shows a bandwidth of 19.833 GHz and a peak gain of 6.22 dBi with ECC below 0.0035.

Abstract

This paper presents a physics-informed Gaussian process regression (PI-GPR) data augmentation framework, combined with standalone machine-learning predictive models, for the design, optimization, and performance prediction of a flower-shaped wideband millimeter-wave antenna and its extension to a compact multiple-input multiple-output (MIMO) configuration. The framework expands an initial CST-simulated dataset to 2, 014 physics-consistent samples by integrating closed-form analytical modeling, electromagnetic constraints, and data-driven GPR, eliminating the need for exhaustive full-wave simulations. Following the data augmentation, five predictive modeling techniques including response surface methodology (RSM), artificial neural network (ANN), ridge regression (RR), random forest (RF), and gradient boosting (GB) are developed and cross-validated, where the RF model achieved improved performance of \ (R² = 0. 999\), RMSE \ (= 0. 031\) and 0. 028, and MSE \ (= 0. 001\) and 0. 001 for resonant frequency (\ (Fᵣ\) ) and impedance bandwidth (BW), respectively. The optimized single-element antenna, designed on a Rogers RT/5880 substrate with overall dimensions of \ (0. 7 ₀ 0. 7 ₀ 0. 069 ₀~mm³\), achieves a gain of 3. 0 dBi at 28 GHz and an impedance bandwidth spanning 22–41. 378 GHz. This element antenna is subsequently transformed into a compact quad-port \ (2 2\) MIMO antenna with overall dimensions of \ (1. 563 ₀ 1. 563 ₀ 0. 069 ₀\) mm \ (³\), yielding a bandwidth of 19. 833 GHz, a peak element gain of 6. 22 dBi, an envelope correlation coefficient (ECC) below 0. 0035, and a diversity gain (DG) exceeding 9. 983 dB. All results are validated by independent CST and HFSS electromagnetic simulations, which show close agreement. The proposed methodology demonstrates a scalable and simulation-efficient approach to mmWave MIMO antenna design for interference-prone, multi-sensor Internet of Things (IoT) applications.

Perguntar à IA

Bookmark

View Full Paper