Accurate Design of Topological Rainbow Concentrator Using Deep Neural Networks

Topological rainbow concentrator based on synthetic dimensionality becomes an emerging topic as its bandwidth and operation frequency can be tuned by additional dimensions. Controlling geometric parameters and lattice translational parameters of the photonic crystal is usually designed by intuition and experience. In this letter, two deep neural networks are trained specially targeting at topological rainbow based on synthetic dimensionality. The first network is a tandem neural network that solves the non-uniqueness problem of inverse networks by being able to predict the geometric parameters of a photonic crystal with a given bandwidth as input, and another inverse network can predict the lattice translation translational parameter with geometric parameters and a given operation frequency, both of which can design the structure over 99% accuracy. Our results show that deep neural networks can quickly and accurately build topological rainbow concentrators with better results than linear modulation-based topological rainbow concentrators.