Answers are grounded in this paper and cite their sources.
This paper presents a surrogate-assisted study of the tolerance and of the identifiability of manufacturing deviations of a 28 GHz inset-fed microstrip patch antenna for 5G millimeter-wave applications. A Multilayer Perceptron (MLP) surrogate was trained on 500,625 (geometry, frequency) samples (625 full-wave HFSS geometries × 801 frequency points) and reaches R2 = 0.996 on unseen geometries. The geometry selected from the simulations achieves a simulated matching of −43 dB at 28 GHz, a 2.75 GHz bandwidth and a simulated gain of 4.4 dBi; the contribution of this work is not the antenna itself, which is a conventional benchmark. The fabricated prototype resonates at 28.07 GHz against 28.02 GHz in simulation, i.e., within 50 MHz (0.18%). The same full-factorial database (±100 µm around the fabricated design) shows that S11 at 28 GHz remains below −26 dB for every combination of deviations and that y1 is the geometrical parameter that most affects the resonance frequency, followed by x1, y2 and x2. A Bayesian inverse analysis based on the MLP shows, on a blind simulated test and on the measured curve, that S11 alone only partially identifies the individual deviations (only y1 is recovered, and the 90% intervals contain the true value in 50–67% of the cases). Moreover, additional ANSYS HFSS (version .13, Ansys Inc., Canonsburg, PA, USA) simulations show that a change of only −0.3% of the FR4 permittivity or −13 µm of its thickness reproduces the measured 50 MHz shift, whereas ±100 µm of any geometrical parameter shifts the resonance by at most ±40 MHz. The substrate therefore dominates over the geometrical deviations, and when a substrate nuisance parameter is included in the inverse analysis (approximate model), none of the four geometrical deviations is identified from S11. These results delimit what can and cannot be diagnosed from a single S11 measurement, and provide guidelines for the validation of fabricated mmWave antennas (direct dimensional measurements and substrate characterization).
No takes yet. Share an insight, caveat, or question.
Essid et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: