A Study on Machine Learning Assisted Accelerated Design of Microwave Structures

An increasing number of researchers devote to applying machine learning for ac-celerating design of microwave structures (e.g., antenna, metasurface, filter, etc.), inspired by the great potential that machine learning shows in many fields, such as image/speech/digits recognition, self-driving, text processing, etc. Despite the fact that machine learning based design has been widely validated to be accurate and well-behaved, machine learning based design methods are often doubted in terms of efficiency, because a large amount of simulation works are mandatory to be executed previously for preparing sufficient training data. In that sense, machine learning based design seems not to be efficient, as it takes more simulation works in total than conventional optimization algorithm based design methods. This paper investigates the efficiency of machine learning based design compared with typical optimization algorithm based design, and a generic solution is proposed for reducing the burden of data preparation to improve the efficiency of machine learning based design. By qualitatively analyzing the required simulation cycles during the whole design process, we propose efficiency measures to demonstrate and compare the efficiency of machine learning based design and typical optimization algorithm based design in the context of metasurface design. According to the comparison result, machine learning based design outperforms other methods in terms of efficiency when it comes to high-bit metasurface design, while optimization algorithm based design is more efficient for low-bit meta-surface. Based on the observation, we introduced an improved design approach that combines the advantages of optimization algorithms and machine learning. The qualitative analysis and improved design approach mayalso bring inspiration to the design of other microwave structures. Investigating on improved data acquisition method for reducing required simulation and training data is a promising direction for further boosting machine learning based accelerated design of microwave structures.