Neural Network Based Inverse Design For Nanophotonic Devices

The inverse design of nanophotonic has wide applications in such fields as metamaterials,nano-optical antennas,optical beam splitters,mode multiplexing or demultiplexers,and so on,which poses a great challenge to the inverse design algorithm for nanophotonic devices.Neural network,as a data-driven method,plays an important role in accelerating the inverse design of nanophotonic devices.But it encounters robustness and stability of the input target electromagnetic response in directly solving the inverse design problems.One cannot get the accurate information of the design target electromagnetic response in advance.Therefore,the neural network should be robust to the input target’s electromagnetic response,that is,it can output a reasonable and corresponding design response to the target response that does not physically exist.In this paper we propose a modified denoising autoencoder network,which includes(1)a pre-trained forward design network to replace the numerical simulation and(2)an antiinterference inverse design network.We introduce the idea of dual learning to expand the data set and improve the accuracy of inverse designing in the proposed network.And adding some random noises to the input of the inverse design network can make the modified denoising autoencoder network more accurate and robust than normal fully connected neural network.In the text,we prove the accuracy and robustness of our proposed network through three specific examples,and realize the inverse scattering problem of the core-shell structures.Finally,we also propose to combine the neural network with traditional optimization algorithm such as particle swarm optimization algorithm for inverse designing,which can integrate the neural network’s advantages of high-speed parallelism and the powerful ability of particle swarm optimization algorithm in searching global and local optimization.We also verify this method by solving the inverse scattering problem of core-shell structures,and prove the efficiency of introducing neural network into particle swarm optimization algorithm.