Micro-ring resonator-based optical convolution computing structure and computing method
By using a single micro-ring structure and a special preprocessing method, the problems of chip size and complexity in high-dimensional OCNN computation are solved, achieving efficient optical convolution computation. Performance tests show significant computational advantages.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-01-13
- Publication Date
- 2026-05-29
Smart Images
Figure CN122114031A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of photonic computing and integrated optics technology, and relates to optical convolution computation based on microring resonators. Background Technology
[0002] In recent years, with the rapid development of artificial intelligence technology, large-scale convolutional neural networks (CNNs) have made breakthroughs in fields such as computer vision, natural language processing, and speech recognition. However, the exponential growth in model complexity and parameter size has posed severe challenges to traditional electronic computing architectures in terms of computing power, energy consumption, and latency. As Moore's Law gradually approaches its limits, the room for improvement in transistor density and transmission bandwidth of electronic devices is limited, and the computing bottleneck has become a key obstacle restricting the further development of artificial intelligence.
[0003] Optical computing, with its inherent advantages in high-speed parallelism, low power consumption, and broadband transmission, offers a new approach to overcoming the energy efficiency bottleneck of electronic computing. Utilizing the propagation, interference, and diffraction properties of light, optical systems can implement core operators such as matrix multiplication and convolution at the physical level, thus far exceeding the speed of electronic solutions. With the rapid development of photonic integration technology and programmable optical devices, optical neural networks (ONNs) have gradually become an important research direction in the field of intelligent computing, demonstrating excellent performance in tasks such as image recognition, signal classification, and data processing.
[0004] In ONNs, convolutional calculations can be performed using optical components such as spatial light modulators and microring resonator arrays, thus realizing Optical Convolutional Neural Networks (OCNNs). Compared to electronic CNNs, OCNNs theoretically offer hundreds to thousands of times faster computation speeds, which is particularly crucial for applications such as real-time video analysis, autonomous driving perception, and high-speed image recognition. Furthermore, OCNNs significantly reduce power consumption. In electronic CNNs, most energy consumption is concentrated in data transport and matrix multiplication processes; while in OCNNs, the propagation of light signals requires almost no additional energy, with only a small amount of energy consumed in the modulation and detection stages. Experimental studies show that optical computation can reduce the energy consumption per operation to the picojoule level, saving several orders of magnitude of energy compared to electronic implementations.
[0005] However, most existing OCNN schemes have many limitations in performing high-dimensional matrix calculations. For example, they use beam splitters and phase shifters to implement optical signal interference structures to complete convolution calculations, but this scheme requires an n×n interference structure to achieve n-dimensional operations, which contains n units. 2In addition, there is a scheme that uses micro-ring arrays to perform convolution calculations. Similarly, this scheme also requires n to achieve n-dimensional operations. 2 Each micro-ring unit. This shows that in existing OCNNs, as the computational dimensionality increases, the number of required structural units increases quadratically, leading to a sharp increase in chip size. This significantly limits the computational power of integrated OCNNs, and the manipulation of large-scale structural units also significantly increases the complexity of chip fabrication. Analysis reveals an urgent need for a compact OCNN scheme with fewer units to improve the computational dimensionality of on-chip OCNNs, reduce process requirements, and minimize the increase in chip size as the computational dimensionality increases.
[0006] A search revealed application publication number CN109639359B, which discloses a photonic neural network convolutional layer chip based on a microring resonator. This chip is applicable to all deep learning technologies involving convolutional computation. The chip utilizes wavelength division multiplexing (WDM) to load vectorized signals to be computed onto different light wavelengths. The microring resonator and balanced photodetector form a weight matrix, enabling convolution computation between the signal to be computed and the weight matrix, outputting the convolution result. Leveraging the tunability of the integrated microring resonator, convolution computation with arbitrary numerical values can be achieved. Furthermore, by using light as the numerical computation medium, the speed of convolution computation is increased to the constant level (i.e., the speed of light), while also possessing a higher energy efficiency ratio.
[0007] This patented solution is similar to the scheme proposed in the last paragraph of the background art of this patent, which uses a micro-ring array to perform convolution calculations. This scheme achieves N-dimensional convolution operations, requiring N... 2 The micro-ring structure will lead to a sharp increase in chip size when calculating high-dimensional matrices.
[0008] This invention employs a unique preprocessing method for data and a convolution result extraction method (similar to the pooling process in convolutional neural networks). It achieves convolution calculations of N-dimensional convolution kernels using only a single micro-ring structure, and can also achieve N... 2 The parallel computation of positive convolution kernels greatly reduces the chip size. Summary of the Invention
[0009] This invention aims to solve the problems of the prior art mentioned above. It proposes an optical convolution calculation structure and method based on a micro-ring resonator. The technical solution of this invention is as follows:
[0010] An optical convolution computation structure based on a microring resonator, comprising:
[0011] A single ring resonant cavity; n signal transmission waveguides, where n = 2 2 3 2…; n waveguide coupling regions; wherein, the waveguide coupling regions realize signal exchange between the ring resonant cavity and the signal transmission waveguide, and each coupling region is provided with one or more phase shifters to adjust the coupling coefficient to achieve different convolution kernels;
[0012] The ring resonant cavity mixes the input optical signal and transmits it through the coupling region to different signal transmission waveguides;
[0013] The input to the signal transmission waveguide is an optical signal carrying information to be convolutionally calculated. Before being transmitted to the waveguide, the optical signal undergoes preprocessing, including slicing and one-dimensional flattening, and there is a fixed delay between adjacent transmission waveguides. The output of the transmission waveguide is the optical signal after convolution calculation.
[0014] Furthermore, the signal transmission waveguide is a silicon-based single-mode waveguide or multi-mode waveguide, and its cross-section is ridge-shaped or strip-shaped.
[0015] Furthermore, the waveguide coupling region employs one of the following coupling methods:
[0016] Lateral or vertical coupling between silicon-based straight waveguides and ring resonators; lateral or vertical coupling between silicon-based bent waveguides and ring resonators.
[0017] Coupling forms include: coupling between two straight waveguides, single-point or multi-point coupling between a straight waveguide and a curved waveguide, and single-point or multi-point coupling between two curved waveguides.
[0018] Furthermore, the phase shifter in the waveguide coupling region is implemented using thermo-optical tuning or electro-optical tuning.
[0019] Furthermore, the ring resonant cavity has a circular, racetrack-shaped, or irregularly shaped structure, and its waveguide is a silicon-based single-mode waveguide or multi-mode waveguide with a strip-shaped or ridge-shaped cross-section.
[0020] Furthermore, when the ring resonant cavity is coupled at multiple points, the lengths of the waveguide sections at different coupling points are different.
[0021] A method for calculating optical convolution of any of the structures described in any one of the claims, comprising the following steps:
[0022] The information to be convolutionally calculated is converted into an optical signal; a fixed delay is set between the optical signals of adjacent signal transmission waveguides; the coupling coefficient is adjusted by a phase shifter in the waveguide coupling region to realize the convolution kernel calculation; the input optical signal is mixed in the ring resonant cavity, and the signal is then redistributed to different signal transmission waveguides through the coupling region according to the adjusted coupling coefficient; the optical signals output through each signal transmission waveguide are acquired and processed to complete the optical convolution calculation.
[0023] Furthermore, the working principle of convolution calculation is as follows: four incoherent optical signals of different wavelengths are input into the chip, and the relationship between the chip's input and output can be derived using the transfer matrix method as follows:
[0024]
[0025] Among them, E ij and E oj Wavelength λ j The input and output light fields, with a transmission matrix value of t jk Instead, j, k=1,2,3,4, different signals undergo different delays before entering the chip, so the total output power is the sum of the powers of the four wavelengths; retaining the m-th bit of the output data y enables convolution calculation, where m is selected according to the following rules:
[0026]
[0027] In the above formula, p represents the extraction index of the result obtained by each slice after each preprocessing, and q represents the slice number.
[0028] If you need to implement a convolution kernel that contains negative values, you can use any one row as the base and subtract that row from the other rows to form an (n-1) dimensional convolution kernel.
[0029] The advantages and beneficial effects of this invention are as follows:
[0030] This invention's system achieves complete convolution computation using a single micro-ring structure. Furthermore, when increasing the convolution computation dimension, the chip size increases linearly, significantly reducing both chip structure and complexity compared to the quadratic growth in traditional solutions. It enables higher-dimensional convolution computation within a limited size, greatly reducing fabrication complexity and providing a feasible path for large-scale photonic OCNN integration. Performance testing of this system on the MNIST handwritten character recognition task yielded a classification accuracy of 96.82%, demonstrating its highly competitive performance and representing a significantly advantageous optical convolution computation scheme.
[0031] The innovations mainly lie in the structure corresponding to right 1, as well as the preprocessing and the steps for implementing convolution.
[0032] For hardware structure, conventional techniques involve one microring corresponding to one input wavelength and one value in the convolution kernel, using a microring array to achieve high-dimensional convolution operations. We utilize a multi-port single microring structure, where a single microring can achieve a complete N-dimensional convolution operation, and can also achieve N... 2 Parallel computation.
[0033] For preprocessing, conventional techniques for flattening data typically flatten it in row-major or column-major order. However, this invention does not flatten the data in a conventional order. Instead, it first performs a slicing operation. After flattening, the input vector is not in a sequential structure, ensuring that the correct convolution result can be obtained in the output. Attached Figure Description
[0034] Figure 1 This is a schematic diagram of a preferred embodiment of the present invention;
[0035] Figure 2 This is a schematic diagram illustrating the principle of the OCNN constructed in this invention;
[0036] Figure 3 Flowchart for flattening the input image;
[0037] Figure 4 This is a simplified structural diagram of an embodiment of the optical convolution calculation based on a microring resonator according to the present invention;
[0038] Figure 5 This is a schematic diagram illustrating the working principle of the convolution calculation in this invention;
[0039] Figure 6 This is a schematic diagram illustrating the working principle of the present invention when performing parallel convolution with four convolution kernels.
[0040] Figure 7 The confusion matrix diagram for performing the MNIST handwritten character recognition task. Detailed Implementation
[0041] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.
[0042] The technical solution of the present invention to solve the above-mentioned technical problems is:
[0043] Figure 1 This is a schematic diagram of the present invention. Figure 2 This is a schematic diagram illustrating the principle of the OCNN constructed in this invention. The OCNN comprises an input layer, a convolutional layer, and an output layer. Taking MNIST handwritten digit recognition as an example, the input image is first preprocessed to generate a 1×1512 data set to form the input layer. Then, convolution calculations are performed using a convolutional kernel (taking a 2×2 kernel as an example). The output signal of the convolutional layer is then passed through a fully connected layer to achieve the final OCNN output, which recognizes the digits 0-9. The optical convolution calculation is entirely implemented by the micro-ring resonator-based optical convolution chip constructed in this invention. It should be noted that although the chip has multiple output ports, a single output is sufficient to complete the convolution calculation function. The input image preprocessing process is also shown in the figure. Image preprocessing consists of three steps: image information flattening, modulation, and delay. Figure 3 The flowchart below shows the input image flattening process. In this embodiment, performance testing uses 28×28 MNIST handwritten digit grayscale image data. The (28×28) image grayscale data is divided into 27 (2×28) slices. Each slice is sequentially connected end-to-end and stitched together. All slices are then processed in the same way, ultimately stitching together (1×1512) one-dimensional data. The flattened data is input into a Mach-Zehnder modulator and modulated onto four different wavelengths (λ1~λ4). Finally, the different wavelengths are delayed by 0~3∆t (λ4 is not delayed, λ3 is delayed by ∆t, λ2 is delayed by 2∆t, and so on, where ∆t is the reciprocal of the modulated signal baud rate).
[0044] The structure of the embodiments of the present invention is as follows Figure 4 As shown, the chip contains a single ring resonant cavity, n coupling regions, and n input / output waveguides (taking n=4 as an example). Optical signals of different wavelengths are input through the ports shown in the figure. It should be noted that the wavelengths of the input optical signals all satisfy the ring cavity resonance condition. The optical signals enter the ring cavity through the coupling regions (taking a two-point coupling structure as an example) and undergo mixing and resonance. The coupling efficiency can be changed by changing the voltage of the thermo-optic electrodes loaded in the coupling regions. Finally, the signal is output from the output waveguide. Low-loss, low-crosstalk crossover structures are set at the waveguide crossover positions to ensure efficient input and output. The waveguide crossover is achieved through multimode waveguide crossover, that is, the width of a section of tapered waveguide is increased first to achieve crossover, and then the multimode waveguide is gradually transformed into a single-mode waveguide by reverse tapering. The insertion loss is about 43mdB at each crossover, and the crossover is less than 50dB (Johnson M, et al. Optics Express, 2020, 28(9):12498-12507).
[0045] Figure 5 This is a schematic diagram illustrating the working principle of convolution calculations in this invention. Four incoherent optical signals of different wavelengths are then processed according to... Figure 4 The port shown is connected to the chip. Using the transfer matrix method, the relationship between the chip's input and output can be derived as follows:
[0046]
[0047] Among them, E ij and E oj Wavelength λ j The input and output light fields, with a transmission matrix value of t jkInstead of (j, k=1,2,3,4), its magnitude is related to the coupling coefficient of each coupling region and can be effectively adjusted by the thermo-optical electrode. Since different signals undergo different delays before entering the chip, the output of different wavelength signals at output port 1 varies with time as follows: Figure 5 As shown in the upper right figure, the total output power is the sum of the powers of the four wavelengths. Convolution calculation can be achieved by retaining the m-th bit of the output data y, where m is selected according to the following rules:
[0048]
[0049] In the above formula, p represents the extraction index of the result obtained by each slice after each preprocessing, and q represents the slice number.
[0050] Since a single port can perform a convolution calculation, this chip has the capability for parallel convolution calculations. Figure 6 This diagram illustrates the working principle of parallel convolution with four convolutional kernels. Furthermore, this chip can also implement negative convolutional kernels by reducing the number of parallel computations. For example, using the last row as the baseline and subtracting it from the other three rows, the three convolutional kernels containing negative values are shown below:
[0051] (3)
[0052] , , These represent the convolution kernel data containing complex values after performing the above operations.
[0054] To verify the performance of this convolutional computation system, we tested it using the MNIST handwritten digit recognition task. The test employed parallel computation with four convolutional kernels. The data processed by the convolutions was then fed into two fully connected layers for classification, with the weights trained offline using the backpropagation algorithm. The final result of this OCNN is shown below. Figure 7 As shown in the confusion matrix, each category of handwritten characters exhibits a high classification accuracy. The overall classification accuracy reaches 96.82%, demonstrating the system's highly competitive performance.
[0055] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions.
[0056] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0057] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.
Claims
1. An optical convolution computation structure based on a micro-ring resonator, characterized in that, include: A single ring resonant cavity; n signal transmission waveguides, where n = 2 2 3 2 …; n waveguide coupling regions; wherein, the waveguide coupling regions realize signal exchange between the ring resonant cavity and the signal transmission waveguide, and each coupling region is provided with one or more phase shifters to adjust the coupling coefficient to achieve different convolution kernels; The ring resonant cavity mixes the input optical signal and transmits it through the coupling region to different signal transmission waveguides; The input of the signal transmission waveguide is an optical signal carrying information to be convolutionally calculated. Before being transmitted to the waveguide, the optical signal undergoes preprocessing, including slicing and one-dimensional flattening, and there is a fixed delay between adjacent transmission waveguides. The output of the transmission waveguide is the optical signal after convolution calculation. The complete convolution layer function can be realized using only a single ring resonant cavity optical convolution calculation structure.
2. The optical convolution calculation structure based on a micro-ring resonator according to claim 1, characterized in that, The signal transmission waveguide is a silicon-based single-mode waveguide or multi-mode waveguide, and its cross-section is ridge-shaped or strip-shaped.
3. The optical convolution calculation structure based on a micro-ring resonator according to claim 1, characterized in that, The waveguide coupling region employs one of the following coupling methods: Lateral or vertical coupling between silicon-based straight waveguides and ring resonators; lateral or vertical coupling between silicon-based bent waveguides and ring resonators. Coupling forms include: coupling between two straight waveguides, single-point or multi-point coupling between a straight waveguide and a curved waveguide, and single-point or multi-point coupling between two curved waveguides.
4. The optical convolution calculation structure based on a micro-ring resonator according to claim 1, characterized in that, The phase shifter in the waveguide coupling region is implemented using thermo-optical tuning or electro-optical tuning.
5. The optical convolution calculation structure based on a micro-ring resonator according to claim 1, characterized in that, The ring resonant cavity is circular, racetrack-shaped, or irregularly shaped, and its waveguide is a silicon-based single-mode waveguide or multi-mode waveguide with a strip-shaped or ridge-shaped cross-section.
6. The optical convolution calculation structure based on a micro-ring resonator according to claim 1, characterized in that, When the ring resonant cavity is coupled at multiple points, the length of the waveguide section is different between different coupling points.
7. A method for calculating optical convolution of the structure according to any one of claims 1-7, characterized in that, Includes the following steps: The information to be convolutionally computed is converted into an optical signal; a fixed delay is set between the optical signals of adjacent signal transmission waveguides. Convolution kernel calculation is achieved by adjusting the coupling coefficient using a phase shifter in the waveguide coupling region; the input optical signal is mixed in the ring resonant cavity, and the signal is then redistributed to different signal transmission waveguides through the coupling region according to the adjusted coupling coefficient. The optical signals output from each of the signal transmission waveguides are acquired and processed to complete the optical convolution calculation.
8. The optical convolution calculation method according to claim 7, characterized in that, The working principle of convolution calculation is as follows: Four incoherent optical signals of different wavelengths are input into the chip. The relationship between the chip's input and output can be derived using the transfer matrix method as follows: ； Among them, E ij and E oj Wavelength λ j The input and output light fields, with a transmission matrix value of t jk Instead, j, k=1, 2,… n, different signals undergo different delays before entering the chip, so the total output power is the sum of the powers of n wavelengths; retaining the m-th bit of the output data y enables convolution calculation, where m is selected according to the following rules: ； In the above formula, p represents the extraction index of the result obtained by each slice after each preprocessing, and q represents the slice number.