Bandpass filters set the achievable selectivity and the noise floor of practically every radio front end, yet the design cycle for millimetre-wave and multi-band devices is still dominated by repeated full-wave simulation. This paper reports a synthesis and optimization framework that couples a genetic algorithm to a convolutional surrogate model, and applies it to three realizations: a 28 GHz substrate-integrated waveguide filter for 5G base stations, a 2.4 GHz miniaturized microstrip filter for battery-powered IoT nodes, and a Ka-band cavity filter for satellite payloads. The surrogate is trained on 5000 previously simulated geometries spanning 1-100 GHz and screens an entire candidate population in under one hour, roughly twelve times faster than direct electromagnetic evaluation; only the top decile is passed to full-wave verification, and the surrogate is retrained on the augmented archive at every cycle. Convergence is reached in five cycles instead of fifteen, which cuts wall-clock optimization time from 72 h to 12 h and lowers the average passband insertion loss of the converged population from 2.1 dB to 1.8 dB. A dual-band variant built on Hilbert-curve defected ground structures occupies 15 mm × 10 mm, about 70 % smaller than comparable published designs, and suppresses the third harmonic at 4.8 GHz by 12 dB relative to hairpin and interdigital references. Measured insertion loss for the 28 GHz prototype is 2.1 dB against 1.8 dB simulated, a 0.3 dB gap consistent with the surface roughness of the fabricated copper. The obstacles that remain are the cost of low-loss laminate in volume IoT production and the scarcity of training data above 100 GHz, both of which limit how far the method extends towards terahertz designs.

Ashikur Mollik, Md Nadim Hosain, Md Imran Hossain, "Optimization of Bandpass Filters for Modern Communication Systems: A Hybrid Genetic Algorithm-Machine Learning Framework with Fractal Defected Ground Structures" International Journal of Scientific Research and Engineering Development, V9(5): Page(985-1010) September - October 2026. ISSN: 2581-7175. www.ijsred.com. Published by Scientific and Academic Research Publishing.