Fig. 9
From: Intelligent designs in nanophotonics: from optimization towards inverse creation
(a) A TO-modified 1 × 2 beam splitter. Left: an SEM image (size: 2.8 × 2.8 μm2). Right: the measured transmissions. The S-parameters exhibit low insertion losses down to -1.8 dB and -2.4 dB in the 1300 nm and 1550 nm bands, respectively, and the crosstalk is suppressed < -11 dB throughout both bands. (b) An adjoint method-derived beam switch displaying different E-field distributions induced by the Kerr nonlinearity. Left: the light passes to the right port in case of a low power of 10 −9 W/ μm. Right: the light goes to the downward gate with a high input power of 0.057 W/ μm. (c) A DNN-based framework for fast prototyping of 1 × 2 power splitters. Left: the spatial configuration of a close-packed splitter (footprint: 2.6 × 2.6 μm2). The exported intensity division is given by T1:T2, and effective manipulations of arbitrary ratios count on manipulating binary positional sequence of 20 × 20 etch holes (white circles). Middle: the established DNN model for forward and inverse modeling of nanophotonic devices. Right: targeted transmision spectra of the beam splitter. Figures reproduced with permission: (a) Ref. [26], Copyright 2015 Nature Publishing Group; (b) Ref. [105], Copyright 2018 ACS; (c) Ref. [106], Copyright 2019 Nature Publishing Group