A learning-based approach for metasurface design beyond the unit-cell approximation

State-of-the-art nanofabrication permits the realization of highly aligned multi-layered metasurfaces with high lateral resolution and wide areas. The exploitation of the vast degrees of freedom and material choice is hampered by the difficulty in the inverse design of metasurfaces. The prevalent design approach of unit-cell library creation and element juxtaposition is known to result in reduced efficiency owing to the inaccurate accounting of inter-element coupling. We report on our recent efforts in accelerated evolutionary optimization for designing metasurfaces with extended unit-cells using learned surrogate models. The difficulty in creating learned models with acceptable predictive capacity in higher dimensional parameter spaces arises from the need for extensive ground-truth generation. By a systematic study of network architectures and dataset sampling strategies, we uncover efficient ground-truth generation strategies. Specifically, we consider 2 and 3-nanoellipse titania metaatoms allowing full control over the elliptical parameters and with an optical response consisting of the spectral behavior of various transmission and reflection-mode diffracted orders for proof-of-concept demonstration. The systematic investigation reveals that densely connected neural architecture and judicious sampling strategies can allow learned model creation even with smaller ground-truth datasets.