Fig. 1. Comparison diagram of chip computing power

Fig. 2. Hybrid photoelectric convolutional neural network^[7]. (a) Imaging principle of optical 4f system; (b) optical convolutional layer design including input image, convolutional kernel stack, and output layer

Fig. 3. Photonic diffraction deep learning computing system^[8]. (a) Basic principle of photonic diffraction deep learning computing system; (b) 3D printing D2NN network test

Fig. 4. Princeton University of USA proposed the first photonic neural network, using wavelength division multiplexing (WDM) based optical patch scheme to input data, and using microloop groups for large-scale weighting and summation^[22-25]

Fig. 5. Programmable nanophotonic processor^[26]

Fig. 6. Phase transition conditions and lattice structure of PCM

Fig. 7. Principle of multiplication and addition in neural networks^[28]. (a) Principle of scalar multiplication; (b) principle of 1×2 matrix and 2×1 vector multiplication

Fig. 8. Refractive indices and extinction coefficients of phase-change materials^[29-30]. (a), (b) Refractive indices and extinction coefficients of GST and GSST in crystalline and amorphous states; (c), (d) refractive indices and extinction coefficients of SbS and SbSe in crystalline and amorphous states

Fig. 9. Microring element device structure^[38-41]. (a) Control port couples light to microring resonator and influences GST state by light heating; (b) silicon microring integrated switch with GST; (c) based on alumina encapsulated GST silicon nitride switch, when GST is amorphous, input light is coupled to microring, and when GST is crystalline, input light is decoupled from microloop; (d) wavelength selective photonic switch based on GST

Fig. 10. Straight waveguide element device structure^[43-46]. (a) State storage information of top GST. Memory reads and writes can be performed by light pulses. Readout of data reads efferent amount of optical waveguide by difference of light absorption between two crystal states of GST; (b) on-chip photon synapses; (c) eigenmode distribution and propagation distribution of GST on Si and SiN; (d) 1×2 and 2×2 switches based on GST-ON-SOI implementation

Fig. 11. Hybrid element device structure^[47-49]. (a) Incident light is transmitted through waveguide to phase-change material, and gold contacts act as electrodes to modulate phase-change material; (b) non-volatile reconfigurable photonic switch based on silicon PIN diode heater; (c) erasable optical memory switch based on GST

Fig. 12. MZI structure^[50-51]. (a) Schematic diagram of Si-SbS integrated optical switch; (b) phase state change of MZI during laser annealing

Fig. 13. Interconnected array of microloops which is an all-optical spike neural synaptic network with self-learning capabilities^[56]. (a) Microloop interconnected array consists of four input neurons and one output neuron; (b) neural network successfully identifies four patterns

Fig. 14. Cross interconnected array^[57]. (a) Integrated photonic structure convolution computing array structure; (b) schematic diagram of structure of photon tensor kernel used for convolution operations; (c) processing results of 128 pixel×128 pixel images; (d) processing results of parallel image processing method