Aim: To move beyond manual tuning and parameter sweeps inside full-wave solvers through automated refinement and multi-objective design at four times the pixel count. Methods: A dual-band T-shaped microstrip antenna on a 70 mm × 45 mm × 1.6 mm FR4 substrate is modelled in ANSYS HFSS. The radiating region is encoded as a 20 × 20 binary pixel matrix with 400 design variables. A hybrid surrogate uses convolutional layers to extract local geometric features and a Vision Transformer encoder to model long-range coupling across the patch. The surrogate jointly predicts reflection coefficient, gain, radiation efficiency, VSWR and bandwidth. Adaptive Quantum Binary Particle Swarm Optimization (AQBPSO), combined with NSGA-II non-dominated sorting, returns Pareto-optimal geometries. Physics-informed penalties for symmetry, connectivity and feed contact keep the synthesized layouts manufacturable. Results: The antenna resonates at 3.4125 GHz and 5.0125 GHz, with reflection coefficients of −19.36 dB and −22.45 dB, respectively. The −10 dB bandwidths are approximately 3.27–3.55 GHz and 4.81–6.64 GHz, with VSWR below 1.25 at both resonances. The framework achieves an estimated 35–45 fold reduction in design time relative to direct full-wave optimization.

Full text: IJEIR_3001_FINAL.pdf