Physics-Driven Inverse Design of Broadband Circulators Using Neural Network and Simplified Real-Frequency Technique

This article presents an inverse design framework for broadband ferrite circulators, which integrates analytical data generation, sparse full-wave calibration, lightweight neural-network prediction, and matching network (MN) synthesis based on the simplified real-frequency technique (SRFT). A physics-inspired calibration method (PICM) uses a small full-wave subset to correct the systematic discrepancy of the analytical ferrite-junction model and to construct a high-fidelity dataset for broadband three-port response prediction. The predicted junction response is then converted into the driving-point impedance required for SRFT synthesis, enabling automated co-design of the junction and MNs without repeated full-wave optimization in the loop. A broadband microstrip prototype achieves measured return loss and isolation better than 15 dB from 6.57 to 13.85 GHz, with insertion loss below 1.26 dB. The results show that the proposed framework avoids repeated full-wave optimization within the design loop while maintaining good agreement with full-wave simulation and measurement for the investigated ferrite circulator implementation.