MZI optical neural network construction method based on polarization control and set port

Abstract

The invention belongs to the field of silicon-based photonics, and particularly relates to an MZI optical neural network construction method based on polarization control and a set port, singular value decomposition is carried out on a weight matrix W of a traditional neural network model to obtain a unitary matrix and a diagonal matrix, a mode of combining a plurality of MZI units is utilized, a unitary matrix or diagonal matrix structure is adopted, and the MZI optical neural network is constructed. Constructing an optical neural network; the method has the advantages that a larger-scale optical matrix is constructed, a higher-dimension data processing task is supported, and then the processing capacity and accuracy of the whole network are improved.

Description

Technical Field

[0001] The present application belongs to the field of silicon-based photonics, and specifically relates to a method for constructing an MZI optical neural network based on polarization control and a collection port. Background Art

[0002] With the development of information technology, the demand for high-speed, low-power data processing is increasing, and traditional electronic computing methods are gradually facing bottlenecks. At the same time, photonics, as an emerging means of information processing, is gradually becoming an important candidate for solving the above challenges due to its advantages of high speed, large bandwidth and low energy consumption. In traditional ONN design, the area and energy consumption of the chip usually increase quadratically with the increase of the input matrix dimension. By combining the design of MZI and diffraction elements, the relationship between the area and energy consumption of the ONN chip and the input matrix dimension is transformed into a linear relationship, which significantly improves the integration and reduces energy consumption.

[0003] Currently, optical computing based on ONN faces many key challenges in terms of function realization, device size, working environment, reconfigurability, etc. These challenges restrict the development and industrialization of optical computing. There is still a long way to go to realize optical computing devices with practical application value.

[0004] Mach-Zehnder Interferometer (MZI) is the basic unit for building optical neural networks. It modulates optical signals by adjusting the optical path difference to achieve linear transformation. However, existing MZI optical neural networks face challenges in scalability and integration. As the scale of the system increases, the optoelectronic synergy problem becomes more prominent, leading to increased system complexity and cost. In addition, the application of existing technologies in incoherent light environments is limited, affecting the practicality and flexibility of ONN. The polarization properties of light provide an additional control dimension for optical computing, but the existing technology does not make sufficient use of polarization, which limits the further improvement of optical computing performance.

[0005] Traditional optical processors based on Mach-Zehnder interferometers (MZIs) face several key challenges when implementing complex matrix operations. First, the limited number of ports and insufficient polarization control capabilities restrict the dimensionality of processable data, making it difficult to fully utilize all the characteristics of optical signals. Summary of the invention

[0006] In view of the defects and deficiencies in the above-mentioned background technology, the present invention proposes an innovative solution, that is, by setting the polarization state of the MZI and expanding the number of its ports, a multi-port, multi-functional optical neural network platform is constructed. This improvement not only increases the flexibility and adaptability of the system, but also provides the possibility of realizing more complex linear transformations. The technical solution is as follows:

[0007] A method for constructing an MZI optical neural network based on polarization control performs singular value decomposition on a weight matrix W of a traditional neural network model to obtain a unitary matrix and a diagonal matrix, and constructs an optical neural network by combining multiple MZI units and adopting a unitary matrix or a diagonal matrix structure.

[0008] Preferably, each MZI unit includes at least two couplers and one phase shifter, and a polarization controller is set at the input end of the front-end coupler to achieve port expansion. Assuming that an N*N unitary matrix architecture is constructed, it is required MZI units, N is a natural number greater than 2.

[0009] Preferably, the weight matrix W of the neural network model is subjected to singular value decomposition to decompose it into a unitary matrix U, a diagonal matrix Σ and a conjugate of a unitary matrix V Right now The unitary matrix U and the diagonal matrix are constructed by topological cascading of multiple MZI units, thereby realizing the optical on-chip mapping structure of the neural network weights W.

[0010] Preferably, the transmission matrix of the MZI unit structure of two couplers and one phase shifter is as follows:

[0011] The angle of the phase shifter is θ, and the phase shift values ​​of the two couplers are ω1 and ω2 respectively.

[0012] Preferably, the transmission matrix of the MZI unit structure of two couplers and two phase shifters is as follows:

[0013] The angles of the two phase shifters are θ and φ respectively, and the phase shift values ​​of the two couplers are ω1 and ω2 respectively.

[0014] Preferably, the optical neural network model uses the Adam optimizer and uses the mean square error to calculate the loss function:

[0015] N is the number of input data, x i is the input vector, y i is the output vector.

[0016] Preferably, multiple MZI units are combined to form a multi-port programmable optical processor, and the code defines a function to generate the position of the MZI required by the matrix and set the network layer. This structure can not only effectively control the phase difference of the optical signal, but also achieve more degrees of freedom by adjusting the polarization state, thereby allowing a more complex optical neural network to be built on a single chip.

[0017] A collection port based on a polarization-controlled MZI unit includes a plurality of MZI units, which are combined to form a multi-port cascade MZI optical processor; each MZI unit includes at least two couplers and a phase shifter; a polarization controller is added to the input end of each coupler, and the port is expanded to 2M under the premise that the total number M of the MZI units remains unchanged; the polarization controller changes the polarization direction of the light so that light of different polarization states can propagate on the same path without interference, and each unit can independently perform phase and polarization regulation.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] 1. High-precision polarization control: Using the polarization state as an additional control dimension can significantly increase the information carrying capacity of each MZI unit without increasing the physical size, build a larger optical matrix, support higher-dimensional data processing tasks, and thus improve the processing power and accuracy of the entire network.

[0020] 2. Automatically construct the MZI optical neural network matrix: By introducing an algorithm, the automatic layout of the MZI unit is realized, thereby simplifying the design process.

[0021] 3. Improve scalability and integration: This technical solution supports large-scale integration of multi-port MZI units, making it possible to build larger-scale and more complex optical neural networks.

[0022] In summary, this patent solves multiple bottleneck problems existing in the existing technology through technological innovation, provides a more efficient, flexible and easy-to-manufacture optical neural network platform, and greatly promotes the development and technological progress in this field. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Two couplers and two phase shifters form a basic MZI unit structure;

[0024] Figure 2 It is a basic unit structure of MZI with two couplers and one phase shifter;

[0025] Figure 3 It is a basic unit structure of MZI with polarization structure;

[0026] Figure 4 It is a basic unit structure of MZI with two couplers and one phase shifter;

[0027] Figure 5 Input and output data for the activation function of this application;

[0028] Figure 6 is the accuracy and loss of the neural network when no polarization is added;

[0029] Figure 7 The accuracy and loss of the neural network when polarization control is added. DETAILED DESCRIPTION

[0030] The technical solution of the present application is described in detail below through specific embodiments and drawings. It should be understood that the embodiments of the present application and the specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations on the technical solution of the present application, and the specific technical features may be combined with each other.

[0031] The present invention introduces polarization states into the traditional MZI architecture and expands the number of its ports to build a multi-port, multi-functional optical neural network platform. This network combines multiple MZI units to form a multi-port cascaded MZI programmable optical processor, which can perform linear transformations such as matrix multiplication after network training, and introduces nonlinear activation functions to enhance the system's expression capabilities.

[0032] A method for constructing an MZI optical neural network based on polarization control performs singular value decomposition on a weight matrix W of a traditional neural network model to obtain a unitary matrix and a diagonal matrix, and constructs an optical neural network by combining multiple MZI units and adopting a unitary matrix or a diagonal matrix structure.

[0033] 1. MZI unit structure with polarization control

[0034] To construct an optical neural network with a programmable MZI topology cascade, it is first necessary to perform a singular value decomposition (SVD) on the weight matrix W of the traditional artificial neural network to decompose it into a unitary matrix U, a diagonal matrix Σ, and a conjugate of the unitary matrix V Right now The unitary matrix and diagonal matrix after matrix decomposition can be realized by silicon-based MZI topological cascading, thereby realizing the optical on-chip mapping structure of the neural network weight W.

[0035] The traditional Mach-Zehnder interferometer (MZI) unit consists of two 3dB directional couplers and two adjustable phase shifters, such as Figure 1 As shown. These components together form a 2x2 reconfigurable MZI that can adjust the phase difference of the internal path as needed to achieve a specific transmission matrix. This structure allows linear transformations to be performed simultaneously between multiple channels, making MZI a powerful tool in photonics and has important applications in optical neural networks. Each MZI unit contains two adjustable phase shifters and two 3dB directional couplers. This design ensures that nonlinear operations can be implemented directly at the hardware level without relying on additional digital post-processing steps, thereby improving the overall efficiency of the system.

[0036] The transmission matrix of the MZI basic unit structure composed of two couplers and two phase shifters without adding a polarization controller is as follows:

[0037]

[0038] By precisely adjusting the angle θ of the phase shifter and The distribution ratio of the optical signal between the two paths can be changed, thereby affecting the intensity distribution of the final output.

[0039] The transmission matrix of the MZI unit composed of two couplers and a phase shifter without a polarization controller is as follows:

[0040]

[0041] To further increase the system's degree of freedom, we incorporate a polarization controller in each MZI unit, allowing the polarization state of the beam entering each branch to be adjusted independently. Figure 3 As shown, the transmission matrix of two couplers and one phase shifter is as follows:

[0042]

[0043] The transmission matrix of two couplers and two phase shifters is as follows:

[0044]

[0045] Where θ and φ are the values ​​of the adjustable phase shifter, and w1 and w2 represent the phase shift values ​​of the polarizer. By changing the polarization direction of light, light of different polarization states can propagate on the same path without interference. This polarization control not only improves the flexibility of the system and provides additional operating space for subsequent nonlinear activation, but also allows the construction of larger-scale optical matrices to support higher-dimensional data processing tasks.

[0046] 2. Constructing MZI optical neural network matrix

[0047] For example, in a 6x6 MZI topology, there are a total of 15 MZI units and 6 ports. When polarization control is added, it can become a 12x12 MZI topology, which expands the number of ports to 12 while keeping the total number of MZI units unchanged. Each unit can independently perform phase and polarization control. This design not only enhances the system's processing capabilities, but also reduces the need for additional light sources, reducing system complexity and costs.

[0048] To construct an N*N unitary matrix architecture, N*(N / 2-1) / 4 MZI units are required. Let's take a 12*12 rectangular unitary matrix as an example. Its architecture is as follows Figure 4As shown in the figure, the MZI optical neural network matrix is ​​constructed. This architecture requires 15 MZI basic units, where each number represents an MZI unit, and the dashed boxes of different colors represent different sections U1-U6. R(n) is the transfer function of the nth MZI. By multiplying by certain rules, a 12*12 rectangular unitary matrix U6 can be obtained. In the code, a function is defined to generate the position of the MZI required by the matrix, and the network layer is set.

[0049]

[0050] The network model uses an activation function that can perform complex operations, and its input and output data are as follows Figure 5 As shown, the Adam optimizer is used and the mean square error loss function is adopted:

[0051]

[0052] N is the number of input data, x i is the input vector, y i The output vector is trained for 200 epochs. The linear part uses a 64*64 rectangular unitary matrix architecture. Without polarization, the accuracy and loss of the neural network are as follows: Figure 6 As shown in Figure 2, this architecture requires 8062 devices (4032 DC and 4032 phase shifters) to build one layer. The accuracy and loss of the neural network with polarization control are shown in Figure 2. Figure 6 As shown in the figure, this architecture only needs 2976 devices (992 polarization controllers, 992 DCs, and 992 phase shifters) to build one layer. It can be seen that this architecture can still maintain a relatively stable accuracy while reducing more devices.

[0053] A collection port based on a polarization-controlled MZI unit includes multiple MZI units, which are combined to form a multi-port cascade MZI optical processor; each MZI unit includes at least two optoelectronic couplers and a phase shifter; a polarization controller is added to the input end of each coupler; multiple MZI units are combined to form a multi-port programmable optical processor, which can not only effectively control the phase difference of optical signals, but also achieve more degrees of freedom by adjusting the polarization state, thereby allowing a more complex optical neural network to be built on a single chip.

[0054] By using the polarization properties of light as an additional control dimension, the amount of information in each MZI unit is increased without increasing the physical size, thereby improving the processing power and accuracy of the entire network, while also improving the precision and flexibility of optical computing.

[0055] The architecture cascades multiple reconfigurable Mach-Zehnder interferometer (MZI) units to build a multi-port optical matrix that supports high-dimensional data processing tasks. Each MZI unit can not only independently perform phase control, but also introduce polarization control capabilities. This design allows more complex optical neural network structures to be implemented on a single chip, significantly improving the computing power of the system.

[0056] In summary, this patent solves multiple bottleneck problems existing in the existing technology through technological innovation, provides a more efficient, flexible and easy-to-manufacture optical neural network platform, and greatly promotes the development and technological progress in this field.

[0057] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for constructing an MZI optical neural network based on polarization control, characterized in that: The weight matrix W of the traditional neural network model is subjected to singular value decomposition to obtain a unitary matrix and a diagonal matrix. The optical neural network is constructed by combining multiple MZI units and using the unitary matrix or diagonal matrix structure.

2. The method for constructing a MZI optical neural network based on polarization control according to claim 1, characterized in that: Each MZI unit includes at least two couplers and a phase shifter. A polarization controller is set at the input end of the front-end coupler to achieve port expansion. Assuming that an N*N unitary matrix architecture is constructed, MZI units, N is a natural number greater than 2.

3. The method for constructing a MZI optical neural network based on polarization control according to claim 1, characterized in that: Perform singular value decomposition on the weight matrix W of the neural network model and decompose it into the conjugate of the unitary matrix U, the diagonal matrix Σ and the unitary matrix V Right now The unitary matrix U and the diagonal matrix are constructed by topological cascading of multiple MZI units, thereby realizing the optical on-chip mapping structure of the neural network weights W.

4. The method for constructing a MZI optical neural network based on polarization control according to claim 2, characterized in that: The transmission matrix of the MZI unit structure with two couplers and one phase shifter is as follows: The angle of the phase shifter is θ, and the phase shift values ​​of the two couplers are ω1 and ω2 respectively.

5. The method for constructing a MZI optical neural network based on polarization control according to claim 2, characterized in that: The transmission matrix of the MZI unit structure with two couplers and two phase shifters is as follows: The angles of the two phase shifters are θ and φ respectively, and the phase shift values ​​of the two couplers are ω1 and ω2 respectively.

6. The method for constructing a polarization-controlled MZI optical neural network according to claim 1, characterized in that: The optical neural network model uses the Adam optimizer and the mean square error to calculate the loss function: N is the number of input data, x i is the input vector, y i is the output vector.

7. The aggregate port based on polarization-controlled MZI unit according to claim 6, characterized in that: Multiple MZI units are combined to form a multi-port programmable optical processor. In the code, a function is defined to generate the position of the MZI required by the matrix and set the network layer. This structure can not only effectively control the phase difference of the optical signal, but also achieve more degrees of freedom by adjusting the polarization state, thereby allowing a more complex optical neural network to be built on a single chip.

8. A collection port based on polarization control MZI unit, using the method for constructing MZI optical neural network based on polarization control as described in any one of claims 1 to 7, characterized in that: It includes multiple MZI units, which are combined to form a multi-port cascade MZI optical processor; each MZI unit includes at least two couplers and a phase shifter; a polarization controller is added to the input end of each coupler to expand the port to 2M while keeping the total number M of the MZI units unchanged; the polarization controller changes the polarization direction of the light so that light of different polarization states can propagate on the same path without interference, and each unit can independently perform phase and polarization control.

Images