Deep learning approach for inverse design of metasurfaces with a wider shape gamut

Abstract

While the large design degrees of freedom (DOFs) give metasurfaces a tremendous versatility, they make the inverse design challenging. Metasurface designers mostly rely on simple shapes and ordered placements, which restricts the achievable performance. We report a deep learning based inverse design flow that enables a fuller exploitation of the meta-atom shape. Using a polygonal shape encoding that covers a broad gamut of lithographically realizable resonators, we demonstrate the inverse design of color filters in an amorphous silicon material platform. The inverse-designed transmission-mode color filter metasurfaces are experimentally realized and exhibit enhancement in the color gamut.

Deep-learning enabled photonic nanostructure discovery in arbitrarily large shape sets via linked latent space representation learning

“…Furthermore, curvature constraints assume signicance, as sharp or highly curved features may deviate from the intended geometry during fabrication. 62,63 For instance, the plus-shaped design obtained through inverse design may not align seamlessly with the fabrication process due to curvature constraints, as evidenced by Vashistha et al 64 Consequently, relying solely on a singular design becomes impracticable, necessitating the generation of multiple shapes akin to those derived through inverse design, albeit with slight variations to preserve desired spectral characteristics. Utilizing trained networks, such sensitivity analysis can be rapidly conducted.…”

Deep learning aids simultaneous structure–material design discovery: a case study on designing phase change material metasurfaces

“…Therefore, there is an increasing demand for the exploration of geometries that are within reach of nanofabrication. The recent repertoire has showcased complex multi-variable metasurface designs in the shape of polygon meta-atoms, 22 – 25 free-form geometries, 26 – 30 extended meta-atoms, 31 , 32 and volumetric structures 33 . Approaches for simultaneous design discovery in structure–material design space remain largely unexplored.…”

Machine learning enabled rational design for dynamic thermal emitters with phase change materials

“… 35 , 36 , 37 , 38 , 39 In the forward modeling process, the well-trained neural network functions as a fast prototyping tool with high accuracy comparable to full-wave simulations; whereas in the inverse design, the network is coupled with genetic algorithms to retrieve the optimized geometric parameters for the given spectral requirements. Although the machine learning approach has been widely applied to many fields, 18 , 19 , 40 , 41 its combination with the dynamic emitter has rarely been reported. By using an ensemble machine-learning toolkit, the implicit relationship between structural features and spectral responses is elucidated qualitatively.…”