Deep Learning based Sequence Modeling for Optical response retrieval of photonic nanostructures

State of the art research in nanophotonics focuses on the development of compact and efficient on-chip devices that are compatible with integrated photonics. Optical Metasurfaces consisting two-dimensional array of subwavelength featured nanostructures have come forth as a perfect candidate to realize such compact photonics chips. They exhibit a unique capability of controlling and manipulating the electromagnetic waves to achieve novel optical responses. Regardless of the vast capabilities and immense potential of nanoscale optical components, their design procedure still suffers from some major drawbacks such as extreme time consumption and high computational resources requirement. A lot of manual work, numerical solutions and multiple simulations are required to achieve metasurfaces having desired responses. Here we present a unique extremely time efficient methodology for finding the optical response of photonic nanostructures using a hybrid deep neural network model that combines convolutional neural networks(CNN), Sequence modeling, and transfer learning. This model incorporates deep residual CNN (ResNet) to extract the spatial information form the images and other geometrical parameters of nanostructures, followed by a gated recurrent unit (GRU) based Sequence Model to map the feature vector to the output and predict the optical absorption spectrum of these structures. Later we utilize transfer learning to extend the model to accommodate nanostructures made up of different materials having diverse physical properties. Our experiments indicate that the proposed methodology accurately predicts the optical response with in a fraction of seconds, which makes it a potential alternative for the conventional time consuming and computationally exhaustive numerical simulations and EM software's.