Fig. 3: Training results of numerical network and optimization of microstructure parameters.
From: Anti-interference diffractive deep neural networks for multi-object recognition
a Different error sources in the experiment. b–d Comparison of training accuracy for networks with neuron sizes of 100 µm and 200 µm under interlayer transverse shift errors, z-axis displacement errors and rotational misalignments as the network depth increases. e Training accuracy and loss of dual-layer AI D2NN under 1–10 training epochs. f Complex amplitude response of microstructure of different diameters under 0.85 THz. g General layout of the fabricated metasurface and zoom-in microscopic images of meta-atoms
