Approaching Performance Bound of Microstrip Antennas Using Machine Learning

A method of approaching the performance bound of microstrip antennas is proposed by using generative machine-learning approach. Different from conventional optimization techniques of a determinant target, the proposed method explores the performance bound in a broader space. Specifically, a training dataset is built first to involve the geometrical and radiative characteristics of 3456 microstrip antennas, with which a multilayer perception (MLP) mapping network is trained for generating a significantly larger number of new designs from an expanded space so that innovative designs of higher performance can be found. In demonstration, hundreds of microstrip antennas of various gain-bandwidth products are predicted accurately in a minute, with one dual-band design measured for validation purpose, showing a larger gain-bandwidth product than that of the training set in the 2.4-/3.3-GHz bands for indoor Wi-Fi and 5G applications. The proposed method may offer a new efficient and low-cost solution to exploring and approaching the unknown performance bound of existing designs.