▎ 摘　　要

NOVELTY - The activation layer comprises unit structures respectively comprising waveguide structure, first electric control phase shifter, graphene heterogeneous enhanced bismuth telluride (Bi2Te3) nonlinear material and second electric control phase shifter. The waveguide structure includes waveguide beam splitter, first upper and lower branch waveguides, directional coupler, and second upper and lower branch waveguides. The waveguide beam splitter has output end divided into two branch waveguides. The directional coupler has output end divided into two branch waveguides. The waveguide beam splitter has input end connected with output end of linear calculation unit waveguide to introduce optical nonlinear calculation in optical neural network on chip. USE - Optical neural network all-optical non-linear activation layer is used for integrated optical neural network on chip and for other integrated optical signal processing platform requiring non-linear calculation. ADVANTAGE - The optical non-linear activation layer provides high response speed, and satisfies low energy consumption and high speed calculation requirement. DETAILED DESCRIPTION - The optical neural network all-optical nonlinear activation layer comprises unit structures respectively comprising waveguide structure, first electric control phase shifter, graphene heterogeneous enhanced bismuth telluride (Bi2Te3) nonlinear material and second electric control phase shifter. The waveguide structure comprises waveguide beam splitter, first upper branch waveguide, first lower branch waveguide, directional coupler, second upper branch waveguide, and second lower branch waveguide. The waveguide splitter has output end divided into two branch waveguides with equal light intensity, preferably first upper and upper branch waveguides having tail coupling entering directional coupler. The directional coupler has output end divided into two branch waveguides, preferably second upper and lower branch waveguides. The first electric control phase shifter is arranged on first upper branch waveguide to cover single-layer graphene on second upper branch waveguide, and deposit layer of bismuth telluride on single-layer graphene to realize heterogeneous structure of single-layer graphene-bismuth telluride and to obtain bismuth telluride nonlinear material reinforced by graphene heterogeneity. The graphene heterogeneity reinforced bismuth telluride nonlinear material heterostructure of carrier relaxation to become faster than bismuth telluride to improve light absorption intensity, which is good for enhancing light nonlinear and improving modulation depth. The second electric control phase shifter is arranged on second lower branch waveguide. The second upper and lower branch waveguides have combined ends, so that whole unit structure forms Mach-Zehnder interferometer structure. The waveguide beam splitter of Mach-Zehnder interferometer structures has input end connected with output end of corresponding calculation unit waveguide to form optical neural network all-optical nonlinear activation layer. The optical signal input from input end of waveguide splitter and output end of waveguide splitter are divided into two paths of light intensity, in which one path through first upper branch waveguide is provided with first electric control phase shifter and one path through first lower branch waveguide. The coupling of first electric control phase shifter is adjusted in first upper branch waveguide to be coupled in respective phase difference two paths. The second upper branch waveguide provided with graphene heterogeneity reinforced bismuth telluride nonlinear material, while second lower branch waveguide is provided with second electric control phase shifter, such that by adjusting voltage of second electric control phase shifter, optical signal of second upper branch waveguide and optical signal of the second lower branch waveguide have phase difference to counteract linear absorption part of bismuth telluride to optical signal, thus amplifying nonlinear absorption part of bismuth telluride to optical signal and combining beam into one path to obtain nonlinear transmission part of optical signal. The input end of waveguide splitter of Mach-Zehnder interferometer structures is connected with output end of linear calculation unit waveguide as non-linear activation layer to introduce optical nonlinear calculation in optical neural network on chip. An INDEPENDENT CLAIM is included for implementation method of optical neural network all-optical nonlinear activation layer. DESCRIPTION OF DRAWING(S) - The drawing shows a schematic diagram of unit structure of optical neural network all-optical non-linear activation layer.