Miniaturized Diffraction Grating Design and Processing for Deep Neural Network

Abstract

Researchers of UCLA reported the fully connected optical neural network, the high computational rate, and low power consumption was realized, while the diffraction grating system based on the Terahertz source is expensive and bulky. In this letter, a long-wave infrared source with a wavelength of 10.6 um is used to establish an optical neural network transfer model using the Sommerfeld diffraction theory. Diffraction grating design, processing, and error analysis applied to deep neural networks are carried out. The MNIST handwritten database is used as the data set to train and optimize the phase parameters by forward propagation and backpropagation. The neuron size is 5um; the number of neurons is 200*200; the entire grating area is 1mm. Compared with the existing light diffraction neural network, the feature size of the deep learning neural network is reduced by 80 times. The Ge (Germanium)-based diffraction grating of 5 layers of neurons with four relative step heights is engraved by semiconductor standard processing technology. The surface-infrared high-efficiency anti-reflection film increased the grating transmission efficiency to over 90%.