Device and method for MZI (Mach Zehnder Interference) neural network based on balance detection

Abstract

The invention discloses an MZI (Mach Zehnder Interference) neural network device and method based on balance detection. The method comprises the following steps: S1, dividing a coherent optical signal of each wavelength in input multi-wavelength optical signals into two paths; s2, the signal light passes through an on-chip multiplexing module, a beam splitter and a light modulator, is input into the MZI-based interference neural network, and is output through a wavelength division demultiplexing module after calculation; s3, connecting the output signal light with different wavelengths and the intrinsic light to a frequency mixer, connecting the signal light with the same wavelength and the intrinsic light with the corresponding wavelength to the same frequency mixer, and generating two paths of light frequency mixing signals with the phase difference of pi by using the frequency mixer; and S4, inputting the signal output by the optical mixer into a balance detector corresponding to the mixer, carrying out photoelectric conversion on the two paths of optical mixing signals by the balance detector, and determining a calculation result based on the converted electric signal. According to the invention, high-precision optical calculation of the MZI interference neural network is realized through combination of balance detection and a self-supervised learning algorithm.

Description

Technical Field

[0001] The present invention relates to the field of optical computing and neural network technology, and in particular to a device and method for an MZI interferometric neural network based on balanced detection. Background Art

[0002] With the rapid development of artificial intelligence and next-generation communication technologies, the demand for high-performance computing continues to increase. This demand has driven the development of customized hardware designed to accelerate specific categories of computing. However, as the feature size and integration of electronic transistors are approaching the physical limits revealed by Moore's Law, processors based on electronic hardware have faced unsustainable performance growth bottlenecks. In emerging artificial intelligence applications, especially tasks involving massive matrix operations, such as the training and reasoning of deep learning models, the requirements for computing speed and energy efficiency are becoming increasingly stringent. Traditional electronic computing architectures are not only limited in speed when handling these tasks, but also have the problem of high energy consumption.

[0003] Existing technologies use ordinary detectors to receive output optical signals, but this will contain a lot of noise, including the relative intensity noise of the laser, the noise introduced by the optical waveguide transmission, and the thermal noise introduced by the electrical module, etc. In particular, after the signal light is transmitted and modulated by the waveguide, the power loss makes the signal weak. Due to the loss and noise in the optical link clock, the detection of the signal light becomes insensitive and inaccurate. For example, strong laser noise can cover the useful signal light, making the detected data erroneous, affecting the processing and conversion of large-scale optical computing data. Therefore, it is necessary to design a method to significantly improve the detection sensitivity of the signal light and effectively suppress the noise to achieve a high-precision optical computing system.

[0004] Optical computing can achieve ultra-low energy information processing in integrated photonic platforms or free space by utilizing the high speed and parallel characteristics of photons. This computing method can effectively transmit and process information through light waves, significantly improving computing speed and reducing energy consumption, thereby better meeting modern computing needs. For example, when performing matrix operations, optical computing can process multiple data streams simultaneously, greatly improving computing efficiency.

[0005] The integrated Mach–Zehnder interferometer (MZI) plays a vital role as a basic device unit in optoelectronic computing. The working principle of MZI is to split the optical signal into two paths and resynthesize them after interference, thereby achieving efficient phase modulation and interference effects. This structure makes MZI extremely advantageous in signal processing and computing, and can perform precise phase control and signal enhancement. In the optoelectronic computing architecture, MZI can be cascaded as a basic unit to form a more complex network structure, thereby achieving more advanced matrix operations and data processing. Through the flexible combination of multiple MZIs, optoelectronic networks that adapt to different application scenarios can be constructed, supporting a wide range of applications such as machine learning, image processing, and signal analysis. This flexible integration method not only improves the performance of optoelectronic networks, but also maintains high efficiency and low energy consumption when processing massive amounts of data.

[0006] Therefore, how to provide a device and method for an MZI interferometric neural network based on balanced detection is an urgent problem to be solved by those skilled in the art. Summary of the invention

[0007] One purpose of the present invention is to propose a device and method for an MZI interferometer neural network based on balanced detection. The present invention realizes high-precision signal detection and noise suppression of the MZI interferometer neural network in optical computing by combining balanced detection technology with a self-supervised learning algorithm based on a graph neural network. The device can accurately eliminate DC and common-mode noise, and improve signal detection sensitivity through dynamic calibration. It has the significant advantages of high signal-to-noise ratio, strong adaptability and high detection accuracy, and is suitable for efficient optical computing in the fields of big data and artificial intelligence.

[0008] A device of an MZI interferometric neural network based on balanced detection according to an embodiment of the present invention includes a multi-wavelength light source, a wavelength division multiplexing, a beam splitter, an optical modulator, a multiplexing module, an optical neural network, an optical mixer array, and a balanced detector array;

[0009] The multi-wavelength light source is used to generate coherent optical signals of multiple wavelengths, and the optical signal of each wavelength is divided into a signal light for optical calculation and an eigenlight for reference;

[0010] The wavelength division multiplexing is connected to the multi-wavelength light source, and is used to divide the coherent optical signal of each wavelength into one signal light and one intrinsic light, wherein the proportion of the signal light is 99% and the proportion of the intrinsic light is 1%;

[0011] The beam splitter, optical modulator and multiplexing module are configured to distribute, modulate and transmit the signal light to the optical neural network module;

[0012] The optical neural network module includes a linear calculation layer composed of an MZI interference neural network and an electro-optical nonlinear module for nonlinear activation, and the MZI interference neural network is used to perform a matrix-vector multiplication operation on the input signal light;

[0013] The optical mixer array includes a plurality of mixers, which respectively coherently mix each signal light with the intrinsic light of the corresponding wavelength to generate two optical mixing signals with a phase difference of π;

[0014] The balanced detector array includes a plurality of balanced photodetectors, which respectively receive two optical mixing signals output by the optical mixer array and eliminate direct current noise and common mode noise through differential detection.

[0015] A method for MZI interferometric neural network based on balanced detection according to an embodiment of the present invention comprises the following steps:

[0016] S1, splitting the coherent optical signal of each wavelength in the input multi-wavelength optical signal into two paths;

[0017] S2, the signal light passes through the on-chip multiplexing module, beam splitter and optical modulator, is input into the MZI interference neural network, and is output through the wavelength division multiplexing module after calculation;

[0018] S3, connecting the output signal light of different wavelengths and the intrinsic light to a mixer, connecting the signal light of the same wavelength and the intrinsic light of the corresponding wavelength to the same mixer, and using the mixer to generate two optical mixing signals with a phase difference of π;

[0019] S4. The signal output by the optical mixer is input into a balanced detector corresponding to the mixer. The balanced detector performs photoelectric conversion on the two optical mixing signals and determines the calculation result based on the converted electrical signal.

[0020] Optionally, the S1 specifically includes: a first path of 99% signal light, which is used for the transmission of on-chip photonic devices to support the calculation of MZI interference neural network; a second path of 1% intrinsic light, which is used as reference light and connected to an input port of a 2×2 mixer of corresponding wavelength in the output module.

[0021] Optionally, the signal light and the intrinsic light are phase-shifted and adjusted by a phase encoding module before mixing, so that the output signal after mixing contains a specific phase mark;

[0022] Optionally, the MZI interference neural network includes a programmable MZI network, and the programmable MZI network adjusts the phase of the signal light through an electric control module to achieve flexible regulation of optical computing.

[0023] Optionally, the two optical mixing signals with a phase difference of π output two beams of light with the same intensity.

[0024] Optionally, the balanced detector is automatically calibrated through a self-supervised learning algorithm based on a graph neural network, and intelligently adjusts detection parameters using historical detection data and current detection results.

[0025] The beneficial effects of the present invention are:

[0026] First, the present invention realizes high-precision signal detection and noise suppression in optical neural networks by introducing a Mach-Zehnder interferometer (MZI) neural network device based on balanced detection. Using balanced detection technology, the present invention can effectively eliminate DC and common-mode noise, thereby greatly improving the signal-to-noise ratio of the signal, so that weak signals can also be accurately identified. After the coherent optical signal generated by the multi-wavelength light source is processed by wavelength division multiplexing and beam splitter, the optical neural network performs matrix operations and nonlinear activation, ensuring fast and accurate calculation of multi-dimensional input signals.

[0027] Secondly, through the adaptive graph neural network self-supervised learning algorithm, the device can realize dynamic calibration based on historical and current detection data, intelligently adjust the detection parameters of the balance detector, and optimize the detection sensitivity and reliability of the system under different environments and signal changes.

[0028] In addition, by performing phase shift processing on the signal light and the intrinsic light before mixing, the present invention enables the output signal to contain a specific phase mark, increases the signal differentiation, and significantly improves the accuracy of the calculation. Combined with the programmable MZI network, the system can dynamically adjust the phase to adapt to different application requirements, making the MZI interferometric neural network device flexible and highly adaptable in the fields of big data, artificial intelligence, and optical computing. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0030] Figure 1 A flow chart of a device and method for an MZI interferometric neural network based on balanced detection proposed by the present invention;

[0031] Figure 2 This is a schematic diagram of the architecture of a device and method for an MZI interferometric neural network based on balanced detection according to Embodiment 1 of the present invention;

[0032] Figure 3 This is a schematic diagram of the architecture of Embodiment 2 of an apparatus and method for an MZI interferometric neural network based on balanced detection proposed by the present invention. DETAILED DESCRIPTION

[0033] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0034] refer to Figure 1-3 , a device of a MZI interferometric neural network based on balanced detection, comprising a multi-wavelength light source, a wavelength division multiplexing, a beam splitter, an optical modulator, a multiplexing module, an optical neural network, an optical mixer array and a balanced detector array;

[0035] The multi-wavelength light source is used to generate coherent optical signals of multiple wavelengths, and the optical signal of each wavelength is divided into a signal light for optical calculation and an eigenlight for reference;

[0036] The wavelength division multiplexing is connected to the multi-wavelength light source, and is used to divide the coherent optical signal of each wavelength into one signal light and one intrinsic light, wherein the proportion of the signal light is 99% and the proportion of the intrinsic light is 1%;

[0037] The beam splitter, optical modulator and multiplexing module are configured to distribute, modulate and transmit the signal light to the optical neural network module;

[0038] The optical neural network module includes a linear calculation layer composed of an MZI interference neural network and an electro-optical nonlinear module for nonlinear activation, and the MZI interference neural network is used to perform a matrix-vector multiplication operation on the input signal light;

[0039] The optical mixer array includes a plurality of mixers, which respectively coherently mix each signal light with the intrinsic light of the corresponding wavelength to generate two optical mixing signals with a phase difference of π;

[0040] The balanced detector array includes a plurality of balanced photodetectors, which respectively receive two optical mixing signals output by the optical mixer array and eliminate direct current noise and common mode noise through differential detection.

[0041] refer to Figure 1-3 , a method for MZI interferometric neural network based on balanced detection, comprising the following steps:

[0042] S1, splitting the coherent optical signal of each wavelength in the input multi-wavelength optical signal into two paths;

[0043] S2, the signal light passes through the on-chip multiplexing module, beam splitter and optical modulator, is input into the MZI interference neural network, and is output through the wavelength division multiplexing module after calculation;

[0044] S3, connecting the output signal light of different wavelengths and the intrinsic light to a mixer, connecting the signal light of the same wavelength and the intrinsic light of the corresponding wavelength to the same mixer, and using the mixer to generate two optical mixing signals with a phase difference of π;

[0045] S4. The signal output by the optical mixer is input into a balanced detector corresponding to the mixer. The balanced detector performs photoelectric conversion on the two optical mixing signals and determines the calculation result based on the converted electrical signal.

[0046] In this embodiment, S1 specifically includes: the first path is 99% signal light, which is used for the transmission of on-chip photonic devices to support the calculation of MZI interference neural network; the second path is 1% intrinsic light, and the intrinsic light is used as reference light and connected to an input port of a 2×2 mixer of the corresponding wavelength in the output module.

[0047] In this embodiment, the signal light and the intrinsic light are phase-shifted and adjusted by a phase encoding module before mixing, so that the output signal after mixing contains a specific phase mark;

[0048] In this implementation, the MZI interference-based neural network includes a programmable MZI network, and the programmable MZI network adjusts the phase of the signal light through an electric control module to achieve flexible regulation of optical computing.

[0049] In this implementation manner, the two optical mixing signals with a phase difference of π output two light beams with the same intensity.

[0050] In this embodiment, the balanced detector is automatically calibrated through a self-supervised learning algorithm based on a graph neural network, and the detection parameters are intelligently adjusted using historical detection data and current detection results.

[0051] Embodiment 1:

[0052] Figure 2 This is a schematic diagram of the architecture of a device and method for an MZI interferometric neural network based on balanced detection proposed by the present invention. The main steps are as follows:

[0053] (1) The input layer uses an on-chip multi-wavelength light source. First, the coherent optical signal of each wavelength in the input multi-wavelength light source signal is divided into two paths: one path is 99% of the signal light, which is used for transmission of on-chip photonic devices through optical fiber coupling to support the calculation of the MZI interferometric neural network; the other path is 1% of the intrinsic light, which can be separated by an off-chip optical module (wavelength division multiplexing, beam splitter) to connect to an input port of the 2×2 mixer of the corresponding wavelength in the output module.

[0054] (2) Wavelength division multiplexing treats each frequency / wavelength component as a coherent laser that performs its own tasks, and then couples all frequency components into an optical neural network (ONN) chip to achieve high-capacity signal processing and parallel high-dimensional computing. The flexible programmable nature of ONN ensures the feasibility of processing different data sets, and uses demultiplexing technology to divide the optical frequency components separately. By introducing dense wavelength division multiplexing technology, a single multi-wavelength light source is decomposed into multiple equidistant coherent lasers through demultiplexing to complete hundreds of tasks and achieve parallel high-dimensional computing.

[0055] (3) After the wavelength division multiplexing module passes through the beam splitter, each output is connected to an optical neural network unit. Signal lights of different wavelengths are encoded by an optical modulator (the specific implementation scheme can adopt thermo-optical modulation / electro-optical modulation / liquid crystal / mechanical methods, such as lithium niobate modulator or silicon-based lithium niobate modulator, etc.) to achieve intensity encoding of the optical signal and then input into the optical network.

[0056] (4) The neural network of optical computing is composed of a linear layer based on MZI and a nonlinear layer carrying an activation function. The linear optical network layer can be used to implement matrix-vector multiplication operations. Among them, the linear layer can be composed of a programmable (the programmable function of the matrix can be achieved by controlling the electrical signal through the FPGA module, using the thermo-optical effect or the electro-optical effect to change the refractive index of the waveguide and thus change the phase of the optical signal) MZI network to implement matrix-vector multiplication operations on the input signal.

[0057] (5) According to the classification results of the preset algorithm or optimization process, the power of the output signal is tested to characterize the potential identification information or calculation results. The output layer consists of a wavelength demultiplexer and a photoelectric detection module, which detects the intensity of the optical signal at different output ports and outputs the vector result after the matrix multiplication operation. Here, an optical mixer and a balanced detector are used to form a photoelectric detection module. The signal light output from the MZI network is input to the balanced photodetector through a single-mode optical fiber; the balanced photodetector is used to convert the optical signal into an electrical signal, and the two electrical signals are further differentiated and sent to the next-level electrical signal processing unit.

[0058] By introducing intrinsic laser for coherent detection and using a balanced detector at the signal light receiving end of the system, the influence of relative intensity noise, background light noise, dark current, etc. can be significantly eliminated. Balanced detection eliminates the DC component in the signal and suppresses the common mode signal, thereby improving the detection sensitivity of the MZI interferometric neural network.

[0059] Two photodiodes with the same performance are connected in parallel inside the balanced detector, which are used to detect two equally divided mixed signals. Since a 180° phase shift is introduced between the two optical signals, the DC component of the photocurrent after detection is eliminated by subtraction. In addition, the relative intensity noise of the two signals has a strong correlation. Their power and frequency are the same but the phase is opposite. Therefore, after subtraction, the relative intensity noise is also effectively eliminated. This processing greatly improves the signal-to-noise ratio (SNR) of the system, thereby enhancing the reliability of the signal and the detection accuracy.

[0060] Embodiment 2:

[0061] Figure 3 This is a schematic diagram of the architecture of Embodiment 2 of the apparatus and method for MZI interferometric neural network based on balanced detection proposed by the present invention. Compared with Embodiment 1, the following differences are made:

[0062] First, the off-chip multi-wavelength laser source can be retained or replaced with an on-chip / off-chip single-wavelength laser source (such as light source + waveguide optical amplifier + microcavity) to drive the on-chip soliton microcavity to generate a multi-wavelength coherent light source for operation. A high-quality microring resonator is used for light splitting to separate the wavelengths of the microcavity single soliton optical frequency comb for independent signal loading.

[0063] Second, the optical signals of different wavelengths generated by the demultiplexer are divided into multiple signal lights through the beam splitter array for parallel optical calculation. Here, a beam of light needs to be separated on the chip as an intrinsic optical signal. For this intrinsic light, the beam splitter needs to split it according to a certain ratio. A small part of the light is used as the intrinsic light, and the other output signal light for optical calculation is equally divided for each port (for example, each beam splitter divides the light into three beams: the intrinsic light intensity is 10%, and the two signal light decibels are 45% and 45%). The intrinsic light separated by the beam splitter is directly connected to the optical mixer module, and the intrinsic light of different wavelengths needs to correspond to the signal light.

[0064] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A device for MZI interferometric neural network based on balanced detection, characterized in that: It includes multi-wavelength light source, wavelength division multiplexing, beam splitter, optical modulator, multiplexing module, optical neural network, optical mixer array and balanced detector array; The multi-wavelength light source is used to generate coherent optical signals of multiple wavelengths, and the optical signal of each wavelength is divided into a signal light for optical calculation and an eigenlight for reference; The wavelength division multiplexing is connected to the multi-wavelength light source, and is used to divide the coherent optical signal of each wavelength into one signal light and one intrinsic light, wherein the proportion of the signal light is 99% and the proportion of the intrinsic light is 1%; The beam splitter, optical modulator and multiplexing module are configured to distribute, modulate and transmit the signal light to the optical neural network module; The optical neural network module includes a linear calculation layer composed of an MZI interference neural network and an electro-optical nonlinear module for nonlinear activation, and the MZI interference neural network is used to perform a matrix-vector multiplication operation on the input signal light; The optical mixer array includes a plurality of mixers, which respectively coherently mix each signal light with the intrinsic light of the corresponding wavelength to generate two optical mixing signals with a phase difference of π; The balanced detector array includes a plurality of balanced photodetectors, which respectively receive two optical mixing signals output by the optical mixer array and eliminate direct current noise and common mode noise through differential detection.

2. A method of MZI interferometric neural network based on balanced detection, characterized in that: The steps include: S1, splitting the coherent optical signal of each wavelength in the input multi-wavelength optical signal into two paths; S2, the signal light passes through the on-chip multiplexing module, beam splitter and optical modulator, is input into the MZI interference neural network, and is output through the wavelength division multiplexing module after calculation; S3, connecting the output signal light of different wavelengths and the intrinsic light to a mixer, connecting the signal light of the same wavelength and the intrinsic light of the corresponding wavelength to the same mixer, and using the mixer to generate two optical mixing signals with a phase difference of π; S4. The signal output by the optical mixer is input into a balanced detector corresponding to the mixer. The balanced detector performs photoelectric conversion on the two optical mixing signals and determines the calculation result based on the converted electrical signal.

3. The method of MZI interferometric neural network based on balanced detection according to claim 2, characterized in that: The S1 specifically includes: the first path is 99% of signal light, which is used for the transmission of on-chip photonic devices to support the calculation of MZI interference neural network; the second path is 1% of intrinsic light, which is used as reference light and connected to an input port of a 2×2 mixer of the corresponding wavelength in the output module.

4. The method of MZI interferometric neural network based on balanced detection according to claim 3, characterized in that: The signal light and the intrinsic light are phase-shifted and adjusted by a phase encoding module before mixing, so that the output signal after mixing contains a specific phase mark.

5. The method of MZI interferometric neural network based on balanced detection according to claim 2, characterized in that: The MZI interference-based neural network includes a programmable MZI network, and the programmable MZI network adjusts the phase of the signal light through an electric control module to achieve flexible regulation of optical computing.

6. The method of MZI interferometric neural network based on balanced detection according to claim 2, characterized in that: The two optical mixing signals with a phase difference of π have two light beams output with the same intensity.

7. The method of MZI interferometric neural network based on balanced detection according to claim 2, characterized in that: The balance detector is automatically calibrated through a self-supervised learning algorithm based on a graph neural network, and intelligently adjusts detection parameters using historical detection data and current detection results.

Images