Simultaneous inverse design continuous and discrete parameters of nanophotonic structures via back-propagation inverse neural network

In this paper, we present an effective tandem neural network with back-propagation optimization strategy for simultaneous inverse design of the continuous and discrete optical parameters, such as the thickness of dielectric layers, the number of periods and materials types. In order to update the discrete parameters in the back-propagation, we use an inverse design network combined with a new constraint function to optimize these two types of parameters together. The forward network model which is trained firstly and is responsible for giving an approximate range of the parameters. The inverse design network then finds the exact parameters in this range by employing back-propagation. Our optimization strategy could find several sets of design parameters that well match the target spectrum, which is an effective method to tackle the problem of non-uniqueness in the inverse design. Furthermore, we demonstrate the inverse design of photonic crystal structure with a specific band gap and center frequency based on a small training dataset.