Frequency-Query Enhanced Electromagnetic Surrogate Modeling with Edge Anti-Aliasing Pixelation for Bandpass Filter Inverse Design

This work presents a frequency-querying mechanism with Transformer architecture for electromagnetic (EM) surrogate modeling, demonstrated in radio frequency (RF) bandpass filter design represented in an edge anti-aliasing pixelated design space. Existing deep learning (DL)-based circuit design approaches using convolutional neural networks (CNNs) and multilayer perceptrons (MLPs) have succeeded in rapid EM prediction for pixelated structures/topologies. However, model size restrictions limit S-parameter predictions to a few to dozens of frequency points, making it challenging to fully characterize wideband EM characteristics, while the treatment of diagonal connections between microstrip pixels constrains design freedom and interpretability. To address those challenges, we developed a surrogate model that efficiently predicts 964 S-parameters across 241 frequency points with reduced training data requirements. The effectiveness of our approach was validated through a compact bandpass filter design with edge anti-aliasing pixelated configuration, achieving $0.138 \lambda_{g} \times 0.138 \lambda_{g}$ size with a -3dB fractional bandwidth (FBW) of 67.6% and operating at a passband from 1.70 to 2.84 GHz.