Data-Driven Surrogate-Assisted Design-Space Exploration of a 28-GHz Passive RIS Geometry Using CST-Generated Electromagnetic Data

Millimeter-wave (mmWave) systems can provide large bandwidth and high data rates, but are challenging to design electromagnetically at high frequencies. This work proposes a data-driven surrogate-assisted design-space exploration framework for a passive reflective RIS geometry at 28 GHz based on the full-wave CST Studio Suite simulations. The excitation source was determined to be a rectangular microstrip patch antenna (MPA) fabricated on Rogers RT/duroid 5880. The antenna had a resonating frequency of 28 GHz and minimum reflection coefficient of about −28.0 dB and VSWR of 1.376, main lobe gain of 9.17 dBi, main beam direction of 0°, 3 dB angular beam width of 59.5° and side lobe level of −15.3 dB. A 5×5 finite reflective array with 25 four-patch elements (100 metallic sub-patches) was parameterized to produce the electromagnetic design data. Data partitioning was done at the Geometry-ID level to minimise leakage between samples that belong to the same physical geometry. To test the performance of surrogate models for reflection phase prediction, three models were tested: Random Forest, Extreme Gradient Boosting (XGBoost), and Multilayer Perceptron (MLP) models. The best reported held-out performance was scored by XGBoost, which was MAE = 2.3°, RMSE = 3.4° and R² = 0.991. From the surrogate-guided candidate screening, a final selected candidate geometry was obtained, which had a sub-patch of 2.00mm, a central gap of 0.30mm, and a horizontal and vertical pitch of 7.00mm. The phase predicted by XGBoost of −179.0° was independently revalidated in CST at −177.3° with an absolute deviation of 1.7°. The mean of the absolute surrogate-to-CST discrepancy for each of the seven candidate designs revalidated in this study was 2.21°, with a maximum discrepancy of 2.8°. The results indicate that the supervised surrogate models can serve as a computational low cost screening tool in a full-wave EM workflow with CST as a final validated reference. Electronically switchable states are not considered in the study, nor is any attempt made to quantify the beam steering, received-power improvement, or link-budget gain under blockage.