Having an artificial neural network that solves Maxwell’s equations in a general setting is an intellectual challenge and a great utility. Recently, there have been multiple successful attempts to use artificial neural networks to predict electromagnetic fields, given a specific source and interacting material distribution. However, many of these attempts are limited in domain size and restricted to object shapes similar to the learned ones. Here, we overcome these restrictions by using graph neural networks (GNNs) that adapt the propagation scheme of the finite-difference time-domain (FDTD) method to solve Maxwell’s equations for a distinct time step. GNNs yield a significant advantage, i.e., size invariance, over conventional neural network architectures, such as convolutional or linear neural networks. Once trained, a GNN can work on graphs of arbitrary size and connectivity. This allows us to train them on the propagation procedure of electromagnetic fields on small domain sizes and, finally, expand the domain to an arbitrary scale. Moreover, GNNs can adapt to any material shape and work not only on structured grids, such as FDTD, but also on arbitrary meshes. This work may be seen as the first benchmark for field predictions with graph networks and could be expanded to more complex mesh-based optical simulations, e.g., those based on finite elements.
We propose a device for subwavelength optical imaging based on a metal-dielectric multilayer hyperlens designed in such a way that only large-wavevector (evanescent) waves are transmitted while all propagating (small-wavevector) waves from the object area are blocked by the hyper-lens. We numerically demonstrate that as the result of such filtering, the image plane only contains scattered light from subwavelength features of the objects and is completely free from background illumination. Similar in spirit to conventional dark-field microscopy, the proposed dark-field hyperlens is shown to enhance the subwavelength image contrast by more than two orders of magnitude. These findings are essential for optical imaging of weakly scattering subwavelength objects, such as real-time dynamic nanoscopy of label-free biological objects.
We propose using deep neural networks for the fast retrieval of effective properties of metamaterials based on their angular-dependent reflection and transmission spectra from thin slabs. While we noticed that non-uniqueness is an issue for a successful application, we propose as a solution an automatic algorithm to subdivide the entire parameter space. Then, in each sub-space, the mapping between the optical response (complex reflection and transmission coefficients) and the corresponding material parameters (dielectric permittivity and permeability) is unique. We show that we can easily train one neural network per sub-space. For the final parameter retrieval, predictions from the different sub-networks are compared, and the one with the smallest error expresses the desired effective properties. Our approach allows a significant reduction in run-time, compared to more traditional least-squares fitting. Using deep neural networks to retrieve effective properties of metamaterials is a significant showcase for the application of AI technology to nanophotonic problems. Once trained, the nets can be applied to retrieve properties of a larger number of different metamaterials.