Three-Stage Generative and Surrogate Modeling Framework for Efficient MIMO Antenna Decoupling in Vehicular Communications

This paper presents a data-efficient optimization framework for compact MIMO antenna decoupling in space constrained vehicular communications. To address the challenges of mutual coupling and limited training data, a three-stage data generation pipeline is proposed, including frequency-aware data partitioning, a multibranch encoder-assisted conditional generative adversarial network, and a dual-filtering strategy for quality assurance. High-fidelity generated data are used to train forward and inverse surrogate models, enabling fast and accurate antenna design optimization without repeated electromagnetic simulations. Applied to a broadband stacked patch antenna array, the optimized design achieves a measured isolation of –42.75 dB and meets bandwidth and reflection criteria. Compared with traditional simulation-based methods, the proposed framework reduces the full-wave simulation time for data generation by 83.43%. These results highlight the potential of combining generative modeling and surrogate learning for intelligent and efficient antenna design in next-generation vehicular MIMO systems.

Duke Scholars

Author William T. Joines Pierre R. Lamond Department of Electrical and Computer Engin ...