Machine Learning-Based High-Isolation Dual-Band MIMO Antenna with Gain Prediction for 5G Networks at 28/38 GHz

Abstract

This paper proposes an ML-driven compact dual-band four-port MIMO antenna for next-generation 5G mm-wave communications at 28/38 GHz. The challenge of integrating ML-based gain prediction with a high-isolation four-element MIMO architecture is discussed. The antenna is designed on a Rogers RT/duroid 6002 substrate with a small footprint of 15.82 × 15.82 mm2 (1.48λ₀ × 1.48λ₀) and is systematically evolved from a single element to 2-port and 4-port MIMO configurations with the maximum gains of 8.5 dBi and 7.25 dBi with the radiation efficiencies of 95% and 96% at 28 GHz and 38 GHz, respectively. The 4-port MIMO system shows excellent isolation over 30 dB, envelope correlation coefficient below 0.001, diversity gain over 0.95, and channel capacity loss of only 0.01 bps/Hz at 28 GHz, which confirms the excellent spatial diversity and spectral efficiency for 5G deployment. To accelerate the design cycle, a comprehensive ML framework was developed using six advanced regression algorithms, with inputs including key geometrical parameters like slot dimensions, impedance transformer geometry and feed line specifications. Among the models evaluated, the Extra Trees Regressor model achieved the best predictive performance with over 98% accuracy in the estimation of the bandwidth at 38 GHz and approximately 94% accuracy in the prediction of the gain at 28 GHz. This provides a scalable design paradigm for future mm-wave MIMO antennas and enables a fast simulation-free performance prediction method. Extra Trees predictions, validated through test-set evaluation and K-fold cross-validation, confirm effectiveness of the proposed four-port MIMO antenna for 28/38-GHz 5G applications.