Inverse Design of Freely Partitioned Metasurface Antennas via Simulation-Driven Active Learning and Deep Neural Network

This letter presents a novel inverse design method for metasurface antennas (MSAs). Specifically, we propose a data generation strategy by employing a Voronoi-based partitioning algorithm, and construct a mapping model from performance parameters to structural topology using a deep neural network (DNN) trained with a simulation-driven active learning (SDAL) approach. Compared with passive learning, our method achieves a lower loss level after 990 training epochs. The trained SDAL-assisted DNN achieves training and validating accuracies of 95.59% and 90.84%, respectively. Using the trained DNN model, a wideband MSA and a dual-band MSA were inversely designed. Both simulation and experimental results prove that the generated structural topologies can fully achieve the desired design specifications. Therefore, our proposed inverse design approach enhances the design flexibility of MSAs and significantly reduces the time cost of design process, making it well-suited for a variety of design scenarios.