CNN-assisted inverse design of wavelength demultiplexer with digital metamaterials

Inverse design is one of the most important design methods of nanophotonic devices. In recent years, with the rapid development of deep learning technique and applications, deep learning assisted inverse design method has been introduced into the field of nanophotonic device design. In this work, by combining the direct binary search method with multilayer convolutional neural networks, we present the inverse design of a wavelength demultiplexer which has 1352 design variables. The dropout strategy has been employed to avoid overfitting in training the inverse design model. The simulation results indicate that the trained CNN can both efficiently forward predict the spectrum and inverse design the structure.