Machine-learning-based Optimization Method for Large-Phase-Shift Metacells (Invited)

A machine-learning-based optimization method is proposed for the design of large-phase-shift metacells. The artificial neural network (ANN) algorithm is utilized to build accurate and efficient surrogate models. Both forward and inverse processes are proposed. In the forward process, dimensions of metacells are fed as the input of an ANN and the neural work outputs the transmission coefficients of these metacells. In the inverse process, the input of an ANN is the desired transmission coefficients, and the neural network predicts the dimensions of metacell that satisfy the targeted performance. A five-layer patch-based metacell is investigated as an example to validate the proposed method. With the machine-learning-based optimization method, the achieved −1-dB phase-shift range of the five-layer patch-based metacell increases from $270^{\circ}$ of existing solution to $420^{\circ}$.