Optical matrix vector multiplier realized based on polarization rotation beam splitter

Abstract

The invention discloses an optical matrix vector multiplier realized based on a polarization rotation beam splitter. Comprising a multi-wavelength light source module composed of a semiconductor laser array, an electro-optical conversion module composed of a modulator array, a signal weighting module based on a reconfigurable polarization rotation beam splitter array, and a signal accumulation module composed of a balance detector and a trans-impedance amplifier. The reconfigurable polarization rotation beam splitter is realized by adopting a mode of combining a phase change material and a metasurface structure, the reconfigurable polarization rotation beam splitter can separate a TE polarization mode optical signal and a TM polarization mode optical signal, and bias voltage is added as external excitation, so that the crystalline state of the phase change waveguide surface of the polarization rotation beam splitter can be changed, and the polarization rotation beam splitter can be used as a phase change waveguide. Therefore, the proportion of the TE polarization mode optical signal and the TM polarization mode optical signal is changed, the weight value is expressed according to the difference of the power components of the polarization modes with different proportions, accurate regulation and control of the weight state can be realized, and the problem of positive and negative expression of the weight is solved.

Description

technical field [0001] The invention relates to the technical field of optical computing, in particular to an optical matrix vector multiplier realized based on a polarization rotating beam splitter. Background technique [0002] In recent years, with the rapid development of information technology, today's society has entered the era of big data and artificial intelligence, and the data information has exploded. In this context, people's requirements for data processing and computing efficiency are getting higher and higher. The application in the field of artificial intelligence largely relies on the processing of a large number of large data sets, involving a large number of matrix multiplication operations, which consumes a lot of energy and slows down the training speed for traditional computer electronic hardware frameworks. There are two main factors that hinder the further improvement of computing efficiency. One is that the development of Moore's Law is approaching ...

AI Technical Summary

Problems solved by technology

One is the matrix multiplication unit based on the principle of singular value decomposition and based on the Mach-Zehnder interferometer (MZI) device. The cascaded MZI network has large processing feature size, good tolerance characteristics, small insertion loss, single wavelength, It has the advantages of convenient regulation and complete theory, and has good application feasibility, but at the same time, its characteristic size is large, resulting in high energy consumption; the other is to use the principle of wavelength division multiplexing, based on the microring resonator (MRR ), the microring device has the advantages of small size and low energy consumption, but it is very sensitive to temperature, and the process error has a great impact on its performance

The implementation of the above two schemes requires power-on to maintain the expression of the weight state, resulting in a large calculation energy consumption, and due to the physical characteristics of the device itself, it is still relatively difficult to achieve high-precision weighting operations

[0005] In the prior art, no matter whether an MZI network or an MRR array is used to weight the input optical signal, it is necessary to always power on the device to maintain the weight state, resulting in a large calculation energy consumption, and the use of the resonance of the microring resonator Method for mapping peak transmission curves as a weighted signal

Due to the device characteristics of the microring itself, its resonant peak transmission line sidebands are very steep, and its resonant wavelength is very sensitive to temperature changes, which is difficult in the actual adjustment process, which is not conducive to the realization of high-precision weighting calculations

Images