An Effective Antenna Shape Optimization Framework Employing Machine-Learning-Based Approach

A machine-learning-based framework for antenna shape optimization design is presented, which enables an accelerated design process for antennas. The employment of a self-supervised learning strategy has significantly reduced the construction cost of electromagnetic (EM) samples. Autoencoder (AE), as the core component of this framework, is trained to extract abstract latent features of antennas and perform shape reconstruction with low sample construction cost. Through the combination of optimization algorithms and self-supervised learning strategy, the proposed framework for antenna shape optimization is constructed. This framework achieves optimization with strong extrapolation capability, thereby greatly enhancing the degrees of freedom in optimization at a very low model training cost. Two design cases, involving a metasurface antenna and a transmitarray antenna design cases with distinct elements are utilized to demonstrate the effectiveness of the proposed framework. The optimization results indicate that antennas constructed from diverse element configurations exhibit superior bandwidth and gain performance compared to those based on single-shape patterns, thus highlighting the enhanced design flexibility offered by the framework.