Deep learning for inverse design of nanophotonic structures

The strong influence of wavelength-scale geometry on electromagnetic fields as well as the unintuitive nature of these responses make the inverse design of nanophotonic structures promising to improve the efficiency of nanophotonic components. However, these design processes are accompanied by challenges, such as their high sensitivity to initial conditions, computational expense, time-intense, and complexity in integrating multiple design constraints. Machine learning and deep learning approaches, however, show strengths, addressing these limitations allowing huge sample sets to be generated nearly instantaneously, and with transfer learning, allowing modifications in design parameters to be integrated with limited retraining. Therefore, in this thesis, we explore a variety of machine learning and deep learning approaches to improve the inverse design of nanophotonic structures. That includes randomly generated simple absorbing 2D nanophotonic structures, high-quality adjoint optimized 3D nanolens structures, and color splitter nanoscale structures to replace color filters in standard cameras. The forward design and inverse design performances are investigated for both interpolation and extrapolation performances. We introduce a hybrid deep learning approach, leveraging the accuracy and performance of adjoint-based topology optimization to produce a high-quality training data set to enhance the potential and performance of deep learning techniques than conventional techniques alone. Further, the transfer learning approach is utilized to retain networks on new design parameters with very few new training samples. This process can be used for general nanophotonic design and is particularly beneficial when a range of design parameters and constraints need to be applied. Moreover, tandem neural networks have seen success in the inverse design of nanophotonic structures compared to other commonly used techniques but they suffer from a significant limitation as they are single-input single output networks with zero ...