Deep Generative Modeling and Inverse Design of Manufacturable Free-Form Dielectric Metasurfaces

Conventional approaches on design and modeling of metasurfaces employ accurate simulation methods. However, these methods require considerable computational power and time for every simulation, making them computationally expensive in the long run. To address this high computational cost and learn compact yet expressive design representations of high-dimensional meta-atoms for efficient design optimization, deep learning (DL) based approaches have emerged as an alternative solution and numerous applications have been demonstrated in recent years. However, there are still outstanding challenges in DL-assisted modeling and design that need to be overcome, such as limited degrees of design freedom, insufficient generalizability of models, and poor fabrication feasibility of final designs. Here, concurrently addressing these challenges, we propose an end-to-end framework for generative modeling and inverse design of dielectric free-form metasurfaces. The framework is generic, as it can accommodate a variety of physical scenarios including dispersion, incident polarization, and operation wavelength using a single data set and model. We develop a shape generation method to generate an inclusive, free-form, and feasible meta-atom library with manufacturability considerations. A forward model that exhibits improved generalizability in terms of material dispersion, polarization, and spectral window of operation is constructed using neural networks. In the final stage, an inverse design of free-form yet manufacturable metasurfaces is realized. As a proof-of-concept, forward design of a meta-lens and inverse optimization of a polarization filter and a quarter-wave plate are demonstrated.