Journal of Electronic Systems Ricardo Rodríguez Jorge 16 1 2025 https://doi.org/10.6025/jes/2026/16/1/1-16 https://www.dline.info/jes/fulltext/v16n1/jesv16n1_1.pdf This paper presents an AI-assisted inverse design methodology for a compact, ultra wideband microstrip bandpass filter (BPF) that eliminates the need for conventional resonator topologies. Unlike traditional approaches relying on stepped-impedance resonators, stubs, or defected ground structures, the proposed design employs a pixelated conductive region optimized directly from electromagnetic performance objectives using artificial intelligence. The filter architecture consists of two 50- microstrip feedlines coupled through a 10 mm × 10 mm square domain discretized into a 20 × 20 binary pixel grid, fabricated on a Rogers RO4350B substrate. A convolutional neural network surrogate model accelerates the optimization process by over two orders of magnitude, enabling efficient exploration of the topology space while enforcing fabrication constraints such as minimum feature size and metal continuity. Full-wave simulations demonstrate a continuous passband from 3 to 10 GHz (fractional bandwidth >100%), in-band insertion loss below 1.2 dB, return loss better than 15 dB, and stopband rejection exceeding 40 dB. Parametric sensitivity analyses confirm robustness against ±5% variation in pixel fill ratio, ±3% deviation in substrate permittivity, and typical feedline width tolerances highlighting practical manufacturability. The AI-discovered geometry intrinsically integrates resonance generation, impedance matching, and transmission zero formation within a single planar region, achieving high performance in an electrically compact footprint (<0.25g × 0.25 g). This work establishes inverse electromagnetic design as a scalable and powerful alternative to classical synthesis techniques, with potential extensions to multi band and reconfigurable RF front end components.