Fast retrieval of effective material parameters using artificial neural network

While further advancing the homogenization of metamaterials, nonlocal constitutive relations came in the focus of interest. However, the retrieval is quite tedious as the parameter space tends to be very large. A direct inversion anyhow is not possible and the identification of the effective properties by solving an optimization problem for each configuration of interest is quite time consuming. To mitigate this problem, we propose here a framework to use artificial neural networks to perform the parameter retrieval. Once trained, the networks can be applied many times to serve their purpose. To avoid issues arising from non-uniqueness of the optical response in the reflection and transmission, we present an automated approach to subdivide the parameter space into subspaces, where each of the subspaces is well-behaving with a suitably trained neural network.