Informed Deep Learning in Metamaterials

Deep neural networks have demonstrated capability to solve challenging forward and inverse problems in electromagnetic metamaterials. However, they often require large quantities of data to achieve a given level of accuracy, which poses a data bottleneck issue and an initial delay in progress. Here we demonstrate two informed deep learning approaches which address the data bottleneck issue in metamaterial design. We show that through direct inclusion of physics in deep neural networks the required network size as well as the size of the dataset can be reduced compared to a vanilla feed forward neural network. We classify our informed deep learning approaches using a recently proposed taxonomy and give an outlook of this exciting field.