Fig. 2: TDNN architecture and analysis of lateral and fabrication error.
From: Holographic multiplexing metasurface with twisted diffractive neural network
a Schematic diagram of the TDNN operation. The structural twist produces a switching effect in the diffraction domain of NM1, allowing the construction of different functionalities on the output plane. b Schematic and comparative effectiveness of bilinear interpolation and nearest-neighbor interpolation. Taking the storage of Maxwell pattern as an example and comparing with theoretical calculation algorithms, both interpolation methods achieve high accuracy. c Flowchart of the TDNN algorithm. d Potential experimental error analysis. Varied degrees of Gaussian distribution error are induced into the phase distribution to observe its impact on the output plane. Here \(\alpha\) represents the rotation angle of NM1, \(\sigma\) represents the standard deviation of Gaussian distributed noise.
