A frequency band generalizable antenna design method based on deep reinforcement learning

Abstract

This invention belongs to the field of artificial intelligence-assisted electromagnetic structure design, specifically involving deep reinforcement learning (DRL). It provides a frequency-band generalizable antenna design method based on deep reinforcement learning to address problems in existing antenna design methods, such as strong reliance on manual intervention, high simulation overhead, weak cross-frequency band generalization ability, difficulty in embedding process constraints, and insufficient multi-task reuse capability. In this invention, firstly, an electromagnetic simulation proxy model is constructed to replace full simulation, outputting the antenna's frequency information and its corresponding S11 prediction value. Secondly, a deep reinforcement learning agent is constructed, outputting compensation values ​​for the structural parameters to be optimized. Based on these compensation values, the structural parameters are updated and iteratively optimized until the design expectation is achieved. In summary, this invention learns the commonalities between antenna structural changes and electromagnetic responses under different frequency bands through a reinforcement learning mechanism, thereby efficiently optimizing antenna structural parameters and possessing frequency-band generalization capability.

Images