Ultra-Compact and NonVolatile Nanophotonic Neural Networks

A nanophotonic neural network (N-PNN) architecture is proposed with compact nanophotonic scattering units and a hybrid structure of silicon and nonvolatile arrayed Sb2Se3. This PNN can execute deep neural networks (DNN) classification and identification tasks with a broad operation bandwidth and very compact footprint. The reconstruction of the convolutional kernel core is realized by digitally switching the phase state of the Sb2Se3 array. Based on a three-dimensional finite-difference time-domain analysis, the core unit received only 4.92 x 2.34 mu m(2) footprint. The convolution kernel unit weights are reconfigured with high-accuracy (7-bit) image processing and recognition in the wavelength C-band (1530-1570 nm). Furthermore, various deep-learning tasks (speech, digital patterns, and clothing patterns) are investigated. The accuracy of the classification and recognition efficiency reached almost the same level as that of a 64-bit computer. The size of the N-PNN is almost two orders of magnitude smaller than that of classic Mach-Zehnder interferometer meshes. It is conducive for scalability, high-radix DNN, and optoelectronic fusion of photonic integrated circuits and electronic integrated circuits.