Deep Learning Based Modulation Parameters Extraction of Non-Reciprocal Bandpass Filter

Modulation parameters are crucial for achieving non-reciprocity in the spatio-temporal modulated non-reciprocal bandpass filters (NR-BPFs). However, no analytical method exists to extract these parameters, and traditional harmonic-balance simulation approach is time-consuming and prone to convergence issues. To solve this challenge, this paper proposes a data-driven framework that leverages a lightweight convolutional neural network to predict NR-BPFs modulation parameters from S-parameter magnitude and phase responses. The proposed deep learning model incorporates an inverted residual with linear bottleneck to reduce computational cost while maintaining high prediction accuracy. The effectiveness of the proposed framework is demonstrated by extracting modulation parameters of microstrip line NR-BPF and comparing with harmonic balance simulated results.