Successive training of a generative adversarial network for the design of an optical cloak

Abstract

At the nanoscale level, optical properties of materials depend greatly on their shape. Finding the right geometry for a specific property remains a fastidious and long task, even with the help of modelling tools. In this work, we overcome this challenge by using artificial intelligence to guide a reverse engineering method. We present an optimization algorithm based on a deep convolution generative adversarial network for the design a 2-dimensional optical cloak. The optical cloak consists in a shell of uniform and isotropical dielectric material, and the cloaking is achieved via the geometry of this shell. We use a feedback loop from the solutions of this generative network to successively retrain it and improve its ability to predict and find optimal geometries. This generative method allows to find a global solution to the optimization problem without any prior knowledge of good cloaking geometries.