Mach-zehnder interferometer network for optical real matrix computation

Abstract

The application discloses a Mach-Zehnder interferometer network for optical real matrix calculation, and belongs to the field of optical calculation. The Mach-Zehnder interferometer network comprises a non-coherent light source array, a modulator array, a real optical MZI matrix, a photodetector array and an FPGA module; the non-coherent light source array emits laser light sources with the same power and different wavelengths, the non-coherent light array loaded with input information is coupled into the real optical MZI matrix for interference, and the final output result is controlled by a thermode on each Mach-Zehnder interferometer unit. The N+1 light signals are converted into electrical signals after passing through the photodetector array, and then are converted into digital signals by an analog-to-digital conversion chip; finally, the FPGA chip calculates the result of matrix vector multiplication and feeds back the control of the thermode of the real optical MZI matrix to realize functions such as optical neural network training. The Mach-Zehnder interferometer network has the advantages of low energy consumption, ultra-compactness and scalability, and greatly improves the performance of optical matrix calculation.

Description

Technical Field

[0001] This invention belongs to the field of optical computing, and more specifically, relates to a Mach-Zehnder interferometer network for optical real matrix calculation. Background Technology

[0002] The advent of the era of big data and artificial intelligence has dramatically increased the demand for computing power and energy consumption in today's computers. However, Moore's Law for integrated circuits is coming to an end; the size and power consumption of electronic transistors are approaching their limits and will no longer be able to meet the ever-increasing demand for computing power, and the resulting power consumption will become unbearable. Therefore, new computing architectures are emerging. Photonic devices have advantages such as high bandwidth, low crosstalk, and low power consumption, and are therefore widely used in communication and computing fields. As a core component of big data processing and artificial intelligence algorithms, matrix-vector multiplication calculations account for a large portion of the computing power. Light, with its multi-dimensional resources and parallel transmission characteristics, is widely used in matrix-vector multiplication calculation applications such as optical neural networks and optical Ising machines. Integrated optical matrix computing will become the acceleration engine of future CPUs.

[0003] Currently, integrated optical matrix computation architectures mainly include Mach-Zehnder interferometer (MZI) networks, microring resonators (MRR) arrays, and phase change material crossbar network structures. Among these, MZI networks, with their advantages of large bandwidth and low insertion loss, are widely used in optical neural networks and solving the Ising problem. The current mainstream MZI networks are complex matrix architectures based on singular value decomposition; an N×N complex optical MZI matrix has 2N... 2 There are 1 degree of freedom, therefore 2N are needed. 2 A phase shifter. However, most applications only require matrix-vector multiplication in the real field, and an N×N real matrix has only N... 2 Therefore, a real MZI network theoretically requires at least N degrees of freedom. 2 A phase shifter. If it were possible to construct one that only requires N phase shifters. 2 The MZI real matrix computation network with phase shifters will greatly improve the energy efficiency of optical computing. Summary of the Invention

[0004] In view of the above-mentioned defects or improvement needs of existing optical computing technologies, this invention provides a Mach-Zehnder interferometer network for optical real matrix calculation, which aims to solve the problems of electrode redundancy, power consumption and excessive layout size in existing integrated MZI optical real matrix calculation.

[0005] To achieve the above objectives, the present invention provides a Mach-Zehnder interferometer network for optical real matrix calculations, which includes an incoherent light source array, a modulator array, a real optical MZI matrix, a photodetector array, and a field programmable gate array (FPGA) module.

[0006] The incoherent light source array is used to generate an incoherent light array. The modulator array loads the input vector information onto the intensity of the incoherent light array. The real optical MZI matrix performs matrix multiplication on the input incoherent light vector. The photodetector array is used to detect the intensity information of the output vector. The FPGA module is used to perform differential calculation on the output vector, calculate the loss function, execute the particle swarm optimization algorithm, and control the driving voltage of the thermoelectrode and modulator array in the real optical MZI matrix.

[0007] The real-number optical MZI matrix comprises a silicon-based thermo-optical Mach-Zehnder interferometer, interconnected waveguides, and couplers interconnected with optical fibers. The silicon-based thermo-optical Mach-Zehnder interferometer includes a 3dB coupler and a thermo-optical phase shifter. The 3dB coupler can be a multimode interference coupler or a directional coupler. The thermo-optical phase shifter is a silicon waveguide with a silicon dioxide cladding and heated metal electrodes. By changing the voltage applied to the heated electrodes, different amounts of heat are generated, causing a change in the effective refractive index of the underlying silicon waveguide, effectively controlling the interference state of light in the real-number optical MZI matrix. The silicon waveguides are fabricated using standard photolithography and etching processes to meet single-mode transmission requirements. The couplers can be grating couplers or end-face couplers.

[0008] Furthermore, the N×N real optical MZI matrix is ​​obtained by removing one of the two phase shifters from the two columns of MZI units on the left and right sides of the network, based on the (N+1)×(N+1) traditional optical MZI unitary matrix. When N is odd, it is also necessary to move one phase shifter from each column of MZI with N+1 electrodes to the bottom of the adjacent left column of MZI.

[0009] Furthermore, the implementation of the real optical MZI matrix relies on subtracting the electrical signal detected by the first output port from the electrical signals detected by the last N ports of the N+1 output ports of the real optical MZI matrix in the FPGA as an N×1 output vector.

[0010] Furthermore, of the N+1 input ports of the N×N real optical MZI matrix, only the first N ports are retained, and the other port is left floating without inputting an optical signal.

[0011] The incoherent light source array comprises multi-channel tunable lasers. The tunable lasers generate C-band wavelength-tunable light sources, and the wavelengths of different lasers need to be spaced to ensure incoherence, guaranteeing that the detector at the output end only detects the power superposition. The output optical power of the light source array should be uniform, providing only the carrier wave of the input vector.

[0012] The modulator array can be a discrete lithium niobate Mach-Zehnder modulator, an electroabsorption modulator, or a silicon-based Mach-Zehnder modulator integrated on a chip, including carrier injection / depletion modulators and thermo-optic modulators.

[0013] The detector array can be an on-chip germanium-silicon PIN photodetector or an off-chip discrete group III-V PIN photodetector, and its function is to convert the output light intensity information into voltage information.

[0014] The FPGA module includes an FPGA chip, a digital-to-analog (DAC) converter, and an analog-to-digital (ADC) converter chip. The FPGA chip controls the DAC chip and performs basic arithmetic operations and particle swarm optimization. The DAC chip outputs analog voltage and controls the thermal phase shifter and modulator. The ADC chip converts the analog electrical signals detected by the photodetector into digital electrical signals for subsequent processing by the FPGA.

[0015] Furthermore, in practical applications, the particle swarm optimization algorithm is used to optimize the voltage value of the thermoelectric electrode to minimize the loss function. Its applications are not limited to optical neural networks.

[0016] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0017] 1. The Mach-Zehnder interferometer network for optical real matrix calculation provided by this invention requires only N 2 One phase shifter, compared to the 2N required by traditional MZI complex networks. 2 The number of phase shifters is reduced to half, thus the power consumption is also reduced to half.

[0018] 2. The network depth of the Mach-Zehnder interferometer network for optical real matrix calculation provided by this invention is N+1, which is nearly half that of the traditional MZI complex network of 2N+1. Therefore, the size of the entire MZI network will be greatly reduced, and the corresponding network insertion loss will also be reduced.

[0019] 3. The Mach-Zehnder interferometer network for optical real matrix calculation provided by this invention realizes the real matrix through differential detection, avoiding the use of coherent detection to detect complex value output. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of a Mach-Zehnder interferometer network for optical real matrix calculation provided by an example of the present invention.

[0021] Figure 2 This is a schematic diagram illustrating the construction rules of a Mach-Zehnder interferometer network for optical real matrix calculations, provided by an example of the present invention.

[0022] Figure 3 This is a comparison diagram of a Mach-Zehnder interferometer network for optical real matrix calculation provided by an example of the present invention and a traditional complex optical MZI matrix.

[0023] Figure 4 This is an optical neural network construction diagram of a Mach-Zehnder interferometer network for optical real matrix calculation provided by an example of the present invention.

[0024] Figure 5 This invention provides an example of a Mach-Zehnder interferometer network for optical real matrix calculation, which is used to train an iris dataset in a constructed optical neural network. (a) shows the evolution of the loss function and accuracy of the training set during neural network training, (b) shows the evolution of the loss function and accuracy of the test set during neural network training, and (c) shows the confusion matrix of the training set after training. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0026] like Figure 1As shown, an example of the present invention provides a Mach-Zehnder interferometer network for optical real matrix calculations, comprising an incoherent light source array 1, a modulator array 2, a real optical MZI matrix 3, a photodetector array 4, and an FPGA module 5. The incoherent light source array emits lasers with the same power but different wavelengths. The FPGA loads the input vector onto the modulator array via a digital-to-analog converter module, thereby loading it onto the light intensity. N optical signals loaded with input information are coupled into the real optical MZI matrix via couplers, interfering within the matrix, and the final output is controlled by the thermoelectrodes on each MZI unit. The output N+1 optical signals are converted into electrical signals by the photodetector array, then into digital signals by the analog-to-digital converter chip. Finally, the FPGA chip calculates the matrix-vector multiplication result and feeds back to control the thermoelectrodes of the real optical MZI matrix to achieve functions such as optical neural network training.

[0027] Figure 2 This demonstrates the construction rules for real-valued optical MZI matrices. For an N×N real-valued matrix, we first construct an (N+1)×(N+1) MZI unitary matrix, at which point the number of phase shifters is (N+1)N. To reduce the number of phase shifters to N... 2 There are some redundant phase shifters that need to be removed. The first and last columns of MZI at the input have redundant global phase, so one of the two phase shifters in these two MZI columns can be removed first. For the case where N is even, all phase shifters are utilized, and the network construction is complete. However, for the case where N is odd, some columns of MZI have N+1 electrodes, while their adjacent columns only have N-1 electrodes. There is an unnecessary global phase shift in the column with N+1 electrodes, so there is electrode redundancy. One phase shifter can be moved to an adjacent column. Figure 1 The noncoherent real optical MZI matrix can be expressed as:

[0028]

[0029] Figure 3 This paper compares traditional complex optical MZI matrices with real optical MZI matrices to illustrate the superiority of this optical matrix architecture. In terms of layout area, the layout area of ​​a real optical MZI matrix is ​​reduced to half that of a complex optical MZI matrix. In terms of power consumption, the number of phase shifters in a real optical MZI matrix is ​​half that of a complex optical MZI matrix, thus halving the power consumption. From a theoretical perspective, incoherent optical matrices are more robust to phase errors than coherent optical matrices, making them more reliable in practical applications.

[0030] In practical applications, the MZI network provided by this invention can serve as a real matrix acceleration module, such as... Figure 4As shown, a 4×4 real optical MZI matrix and a 3×4 real optical MZI matrix are applied to an optical neural network. The task here is to classify the iris flower dataset, including an input layer, a hidden layer, and an output layer. The ReLU function is used as the nonlinear activation function in the hidden layer to ensure that the vector input to the 3×4 real optical MZI matrix is ​​positive. The network is trained using a particle swarm optimization algorithm, and this in-situ training method has good robustness to process errors. Figure 5 In the diagram, (a) and (b) are the evolution graphs of the loss function and accuracy of the training and test sets obtained by training this network, respectively. The final prediction accuracy can reach 98.1% and 100%, respectively. (c) is the confusion matrix of the training set after training. From the simulation results of optical neural network applications, the real-number MZI network architecture proposed in this invention is completely feasible.

[0031] This invention provides a Mach-Zehnder interferometer network for optical real matrix computation. This MZI network reduces the number of phase shifters, power consumption, and layout size of the original complex MZI network by half. The incoherent detection method further reduces the complexity of the on-chip structure and the difficulty of detection. This invention lays the foundation for large-scale optical neural networks based on integrated MZI networks.

[0032] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A Mach-Zehnder interferometer network for optical real matrix calculations, characterized in that, It includes an incoherent light source array (1), a modulator array (2), a real-number optical MZI matrix (3), a photodetector array (4), and an FPGA module (5). The incoherent light source array (1) is used to generate an incoherent light array. The modulator array (2) is used to load the input vector information onto the intensity of the incoherent light array to form an incoherent light vector. The real optical MZI matrix (3) is used to perform matrix multiplication on the input incoherent light vector. The photodetector array (4) is used to detect the intensity information of the output vector of the real optical MZI matrix (3). The FPGA module (5) is used to perform differential calculation on the intensity information of the output vector, calculate the loss function, execute the particle swarm optimization algorithm, and control the driving voltage of the thermoelectrode in the real optical MZI matrix (3) and the modulator array (2). The real-number optical MZI matrix (3) includes a silicon-based thermo-optical Mach-Zehnder interferometer arranged in a matrix, a waveguide for interconnecting the silicon-based thermo-optical Mach-Zehnder interferometer, and a coupler for interconnecting with an optical fiber. The silicon-based thermo-optical Mach-Zehnder interferometer includes a 3 dB coupler and a thermo-optical phase shifter. The thermo-optical phase shifter is a silicon waveguide with a silicon dioxide cladding and a thermoelectrode. The real number optical MZI matrix (3) is an N×N matrix obtained by removing one of the two thermo-optical phase shifters of the two columns of MZI units on the left and right sides of the network from the (N+1) × (N+1) optical MZI unitary matrix. When N is odd, it is also necessary to move one of the phase shifters of each column of MZI with N+1 thermoelectrodes to the bottom of the adjacent left column of MZI.

2. The Mach-Zehnder interferometer network for optical real matrix calculation according to claim 1, characterized in that, The N+1 input ports of the real optical MZI matrix (3) retain only the first N ports, and the other port is left floating and no optical signal is input.

3. The Mach-Zehnder interferometer network for optical real matrix calculation according to claim 1, characterized in that, In the FPGA module (5), the electrical signals detected by the last N ports of the N+1 output ports of the real optical MZI matrix are subtracted from the electrical signals detected by the first output port to form an N×1 output vector.

4. The Mach-Zehnder interferometer network for optical real matrix calculation according to claim 1, characterized in that, The incoherent light source array includes N-channel tunable lasers; the tunable lasers generate C-band light sources with the same power and adjustable wavelengths, and there is a certain interval between the wavelengths of different lasers.

5. A Mach-Zehnder interferometer network for optical real matrix calculation according to claim 1, characterized in that, The modulator array is a discrete lithium niobate Mach-Zehnder modulator, an electroabsorption modulator, a carrier injection / depletion modulator, or a thermo-optic modulator.

6. A Mach-Zehnder interferometer network for optical real matrix calculation according to claim 1, characterized in that, The detector array is an on-chip germanium-silicon PIN photodetector or an off-chip discrete group III-V PIN photodetector, used to convert intensity information into voltage information.

7. A Mach-Zehnder interferometer network for optical real matrix calculation according to claim 1, characterized in that, The FPGA module includes an FPGA chip, a digital-to-analog (D / A) converter, and an analog-to-digital (ADC) converter chip. The FPGA chip is used to control the D / A / A converter chip and perform basic arithmetic operations and particle swarm optimization. The D / A converter chip is used to output analog voltage and control the thermal phase shifter and modulator. The ADC chip is used to convert the analog electrical signals detected by the photodetector into digital electrical signals for subsequent processing by the FPGA.

8. A Mach-Zehnder interferometer network for optical real matrix calculation according to claim 1, characterized in that, The 3 dB coupler is either a multimode interference coupler or a directional coupler.

Images