Machine learning–driven prediction of wideband slot-free FSS-based terahertz antenna performance for 6 G networks

The fabrication of slotted terahertz antennas is challenging due to their high sensitivity to dimensional tolerances, where small slot variations can cause severe resonance shifts and impedance mismatch. To overcome this limitation, this paper proposes a low-complexity, slot-free terahertz antenna operating over the 0.1–1.0 THz frequency band, integrated with a periodic cross-shaped frequency selective surface (FSS), a passive electromagnetic structure, that improves impedance matching through near-field coupling with the antenna, partially cancelling the reactive component of the input impedance at 0.5 THz. The proposed design achieves a wide bandwidth of 0.73 THz, a minimum reflection coefficient of −62 dB at 0.5 THz, a maximum gain of 11.2 dBi, and a peak radiation efficiency of 85%. To enable rapid performance prediction, seven supervised regression models are trained and evaluated on a dataset of 22,500 CST-generated samples. Among the evaluated models, Extra Trees Regression achieves the highest prediction accuracy with a Mean Absolute Error of 0.07%, a Root Mean Squared Error of 0.23%, and a near-unity coefficient of determination and Explained Variance Score of 99.96%. The combined fabrication-friendly antenna design and machine learning prediction framework support the suitability of the proposed solution for next-generation 6 G terahertz communication systems.