On-chip modulation optical chip based on micro-ring resonator and optical neural network system

By using on-chip modulation technology of micro-ring resonator arrays in optical chips, combined with wide spectral light source and photoelectric conversion module, the problems of low integration and thermal crosstalk sensitivity of micro-ring resonators are solved, and high-density information processing and efficient calculation are realized , enhances the performance of optical neural networks.

CN119045123APending Publication Date: 2024-11-29SHANGHAI TECH UNIV
View PDF 0 Cites 6 Cited by

Patent Information

Application Number
CN202411168069.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-23
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

The existing optical chips based on microring resonators have low integration and are sensitive to thermal crosstalk, making it difficult to meet the high-density and efficient optical neural network computing requirements.

Method used

The on-chip modulation technology based on micro-ring resonators is adopted, combined with a wide spectrum light source and photoelectric conversion module, and the modulation and multiplication of optical signals are realized through the micro-ring resonator array, and the resonance conditions are adjusted by the computer module through the programming language control heating unit. Implement high-density information processing.

Benefits of technology

It improves the integration and computing efficiency of micro-ring resonators, reduces chip size and manufacturing costs, enhances the compactness and performance of optical neural networks, adapts to the characteristics of micro-ring resonators, and reduces the impact of hardware errors.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119045123A_ABST
    Figure CN119045123A_ABST
Patent Text Reader

Abstract

According to the on-chip modulation optical chip based on the micro-ring resonator and the optical neural network system, the system ingeniously integrates the wavelength division multiplexing technology of a broadband light source and the on-chip light intensity modulation technology, the multi-wavelength characteristic of light is fully utilized, and high density and large capacity of information processing are achieved. The advanced on-chip modulation technology greatly enhances the integration level of an optical calculation chip, integrates the traditional off-chip modulation part to a silicon optical chip, remarkably reduces the cost, and shows the wide prospect of the photon calculation technology. The micro-ring resonator is subjected to partial Euler bending optimization, and transmission loss is reduced in small-radius waveguide bending, so that the size of a chip is remarkably reduced, the manufacturing cost is effectively reduced, and meanwhile, the energy efficiency ratio is improved. In addition, the optical neural network architecture in the application better adapts to the physical characteristics of the micro-ring resonator through special optimization, so that the performance of the micro-ring resonator is fully played in the optical neural network. A quantitative neural network and quantitative perceptual training (QAT) are introduced, the adaptability of the system is further improved through fusion of the technologies, the error influence possibly occurring in the manufacturing process based on a micro-ring resonator chip is reduced, and a solid technical basis is provided for high-precision photon calculation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the intersection of optical computing chips, semiconductor optoelectronics and deep learning, and in particular to an optical chip and an optical neural network system based on on-chip modulation of a microring resonator. Background Art

[0002] With the development of the times, global artificial intelligence technology continues to advance. But at the same time, the excellent performance of artificial intelligence comes from complex network models and huge computing resource consumption. The demand for computing power of artificial intelligence is growing rapidly, while the development of electronic chips has gradually reached the bottleneck of Moore's Law and can no longer keep up with the growth rate of the demand for computing power of artificial intelligence. Optical computing is the use of optical devices to simulate optical neural network operations. In optical neural networks, matrix multiplication and accumulation can be performed at the speed of light, and the optical signal operation process has low energy consumption. Photons have the characteristics of light speed propagation, anti-electromagnetic interference, and arbitrary superposition. Optical computing has natural parallel computing characteristics, so the computing speed is extremely fast and very suitable for parallel computing. Optical neural networks have become a promising architecture that relies on the low latency, high bandwidth, inherent parallelism and low power consumption of photonic integrated circuits.

[0003] Silicon photonics platforms can form interconnections between highly integrated, low-energy computing modules. Optical chips can quickly transmit large amounts of information, with the characteristics of ultra-high speed, low latency and high parallelism, and their energy consumption is much lower than that of electronic chips. An important function of an optical computing system is to perform matrix multiplication and addition operations on on-chip optical devices to realize the process of multiplying and accumulating input signals with weight matrices. Several current classic optical computing architectures for matrix multiplication and weighted interconnection are planar light conversion, Mach-Zehnder interferometer cascade, micro-ring arrays based on wavelength division multiplexing, and phase change materials. Mach-Zehnder interferometers are large in size, have slightly higher modulation power consumption, and rely on singular value decomposition to indirectly express the weight matrix. Relatively speaking, micro-ring resonators have the advantages of small size, low power consumption, and can directly express weights. Although micro-ring resonator arrays have the advantage of being easy to adapt to neural networks, there are still some problems that need to be solved for micro-ring resonators. The main challenge they face is that the chip integration based on micro-ring resonators needs to be improved. Traditionally, optical chips based on microring resonators use external optical modulators to modulate the input beam, which is contrary to the trend of integrating multiple optical components on a single chip. At the same time, since microring resonators are very sensitive to thermal crosstalk, this limits the reduction of the effective radius of the microring resonator, which in turn affects the integration and accuracy of chips based on microring resonators.

[0004] In recent years, in WG (microring resonator optical computing architecture), microring resonators are regarded as single neurons. Gu et al. reduced the footprint of optical neural networks and improved their scalability by using multi-operand microring resonators and fine-grained structure pruning algorithms. These studies improved the compactness and efficiency of optical neural networks on microring resonator arrays. Summary of the invention

[0005] In view of the shortcomings of the prior art described above, the purpose of the present application is to provide a solution for how to improve the integration of microring resonators and solve their sensitivity to thermal crosstalk, adapt to the needs of neural networks and improve the compactness and efficiency of optical neural networks.

[0006] To achieve the above-mentioned purpose and other related purposes, the first aspect of the present application provides an on-chip modulated optical neural network system based on a microring resonator, comprising: a wide-spectrum light source module, an optical neural network computing module, a photoelectric conversion module and a computer module; the wide-spectrum light source module is used to provide a continuous wide-spectrum optical signal; the optical neural network computing module is located on the emission path of the optical signal generated by the wide-spectrum light source module, and comprises an optical signal modulation module and an optical signal multiplication and addition operation module; the optical signal modulation module and the optical signal multiplication and addition operation module both comprise a microring resonator array; wherein the modulation of the optical signal modulation module and the weighted operation of the optical signal multiplication and addition operation module are The calculation is controlled by the computer module through a programming language; the photoelectric conversion module is connected to the optical neural network calculation module, and is used to convert the modulated and weighted optical signal into an electrical signal and transmit it to the computer module; the computer module performs simulation through the simulation system loaded thereon, and obtains the adaptive weight parameters after model pre-training of the pre-collected data set; the adaptive weight parameters are loaded into the optical neural network calculation module; and the computer module performs nonlinear activation processing on the electrical signal sent by the photoelectric conversion module, wherein the data obtained through the nonlinear activation processing is used to update the weight parameters for iterative optimization of the model.

[0007] In some embodiments of the first aspect of the present application, the structures of the optical signal modulation module and the optical signal multiplication and addition operation module include: a microring resonator array, and a heating unit provided on each microring resonator in the microring resonator array; each microring resonator utilizes its resonance characteristics for light of a specific wavelength to modulate the intensity of the optical signal, and the heating unit is used to change the refractive index of the corresponding microring resonator to adjust the resonant wavelength.

[0008] In some embodiments of the first aspect of the present application, the optical signal modulation module is implemented by spatially superimposing optical signals of different wavelengths in a microring resonator array, wherein each microring independently modulates an optical signal of a specific wavelength, and changes the resonance condition by adjusting the heating resistor unit, and finally completes the summation by superimposing the light intensity at the output end, thereby realizing the addition operation of multiple signal components in parallel.

[0009] In some embodiments of the first aspect of the present application, each microring resonator in the microring resonator array of the optical signal multiplication and addition operation module corresponds to a weight in a matrix and independently regulates the optical signal, and simulates the dot product of the weight and the input vector by adjusting the resonant frequency, and finally obtains the result of the multiplication operation by spatially superimposing the output light intensity of each microring resonator.

[0010] In some embodiments of the first aspect of the present application, the heating unit includes a heating resistor unit made of titanium nitride.

[0011] In some embodiments of the first aspect of the present application, the photoelectric conversion module includes a photodetector.

[0012] In some embodiments of the first aspect of the present application, the computer module obtains adaptation weight parameters after performing model pre-training on the pre-collected data set, and the process includes: inputting the pre-collected data set into a pre-selected machine learning model for model pre-training; after the pre-training is completed, encoding the obtained weights into light domain parameters adapted to the optical neural network; and after encoding the weights into corresponding light domain parameters, loading them into the optical neural network calculation module.

[0013] In some embodiments of the first aspect of the present application, the computer module performs nonlinear activation processing on the electrical signal sent by the photoelectric conversion module through any of the following functions: Sigmoid function, tanh function, ReLU function.

[0014] In some embodiments of the first aspect of the present application, the computer module uses Python programming language to control the modulation of the optical signal modulation module and the weight of the optical signal multiplication and addition operation module.

[0015] To achieve the above-mentioned purpose and other related purposes, the second aspect of the present application provides an optical chip based on a microring resonator, which is applied to the on-chip modulated optical neural network system based on the microring resonator, including: an optical signal modulation module, an optical signal multiplication and addition operation module; the optical signal modulation module and the optical signal multiplication and addition operation module both include a microring resonator array; wherein the modulation of the optical signal modulation module and the weighted operation of the optical signal multiplication and addition operation module are both controlled by a computer module through a programming language.

[0016] As described above, the optical chip and optical neural network system based on on-chip modulation of microring resonators of the present application have the following beneficial effects:

[0017] (1) By combining wavelength division multiplexing of broadband light sources with on-chip light intensity modulation, this scheme can effectively utilize the multi-wavelength characteristics of light to achieve high-density and high-capacity information processing. This on-chip modulation technology enables optical neural networks to achieve very fast computing speeds, showing the great potential of photonic computing.

[0018] (2) The optimized partial Euler bending microhole reduces the transmission loss through the small radius waveguide bending. This structure not only reduces the chip size, but also reduces the manufacturing cost and improves the overall energy efficiency.

[0019] (3) The optical neural network architecture provided in this application is optimized according to the physical characteristics of the microring resonator, so that it can better adapt to the characteristics of the microring resonator and can better exert the performance of the microring resonator in the optical neural network.

[0020] (4) The quantized neural network and quantization-aware training (QAT) introduced in this application are more adaptable to the current chip manufacturing level based on microring resonators and reduce the impact of errors on the hardware of the microring resonator array. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Shown is an architecture diagram of an optical neural network system based on on-chip modulation of a microring resonator in one embodiment of the present application.

[0022] Figure 2a It is a schematic diagram showing how to directly select resonance peaks of different wavelengths using micro-ring resonators with different effective radii in one embodiment of the present application.

[0023] Figure 2b Shown is a diagram of an optimized design embodiment of a single microring resonator in one embodiment of the present application.

[0024] Figure 2c Shown is a spectrum diagram generated by simulating a partially designed Euler bent microring resonator using a three-dimensional finite difference time domain method in an embodiment of the present application.

[0025] Figure 3 Shown is a schematic diagram of the process in which a computer module performs model pre-training on a data set to obtain adaptive weight parameters in one embodiment of the present application.

[0026] Figure 4 Figures a to d in the figure show a specific architecture embodiment diagram of an on-chip modulated optical neural network system based on a microring resonator in one embodiment of the present application.

[0027] Figure 5aThe graph shows the change of photoelectric power at the drop end and through end of the microring array as a function of voltage in one embodiment of the present application.

[0028] Figure 5b The figure shows the difference curve of the through end of the microring array minus the drop end in one embodiment of the present application.

[0029] Figure 6 Shown is a schematic diagram of the accuracy of the training data set in the handwriting data set classification in one embodiment of the present application. DETAILED DESCRIPTION

[0030] The following describes the embodiments of the present application through specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. The present application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0031] The present invention provides a structural design for an optical computing chip based on a micro-ring array and a quantized neural network suitable for the structure. Through the proposed on-chip modulated micro-ring modulator array, a wide-spectrum light source can be directly input to realize wavelength division multiplexing and on-chip light intensity modulation, without the need for a bulky off-chip light intensity modulator, and on-chip integration of optical components is performed, thereby effectively improving the integration of the optical computing chip. By adopting an optimized partial Euler bent micro-ring, the problem of large loss in small radius waveguide bending is overcome, and the transmission loss is reduced. This structure not only reduces the chip size, but also reduces the manufacturing cost. In addition, by adopting an advanced quantized perception training algorithm to perform training optimization of the optical neural network, the simulation time is accelerated, the bottleneck of the hardware device manufacturing level is overcome, and the influence of system errors can be reduced in the simulation, thereby optimizing the chip computing performance.

[0032] Figure 1 The architecture diagram of an on-chip modulated optical neural network system based on a microring resonator in an embodiment of the present invention is shown. The architecture is an integrated optoelectronic computing platform that combines optical chips and electronic components to achieve signal modulation, processing and detection. The optical computing system based on a microring resonator includes a wide spectrum light source module 1, an optical neural network computing module 2, an optoelectronic conversion module 3, and a computer module 4.

[0033] The wide-spectrum light source module 1 can use a wide-spectrum laser light source to generate a wide-spectrum laser output. Its main function is to provide the system with continuous optical signals, which will be used in subsequent modulation and calculation processes. It should be understood that a wide-spectrum laser light source is a laser that can produce an extremely wide spectral range, and its spectrum covers a wide range of wavelengths from ultraviolet light to visible light and even infrared light. Taking the laser-driven white light source (LDLS) as an example, it uses a continuous laser to directly heat xenon ions to a specific temperature to produce wide-spectrum light; the wavelength range of LDLS is from 170nm to more than 2100nm, with a long life of more than 9000 hours, and provides high brightness output.

[0034] The optical neural network calculation module 2 is located on the emission path of the optical signal generated by the wide spectrum light source module 1, and includes an optical signal modulation module 21 and an optical signal multiplication and addition operation module 22; the optical signal modulation module 21 and the optical signal multiplication and addition operation module 22 both include a micro-ring resonator array; wherein the modulation of the optical signal modulation module and the weighted operation of the optical signal multiplication and addition operation module are controlled by the computer module 4 through a programming language. It should be understood that each micro-ring resonator in the micro-ring resonator array represents a single neuron, which is used to express different weights and perform corresponding operations; the micro-ring resonator array performs weighted summation on the output signal to obtain a target calculation value, and performs matrix multiplication and addition operations accordingly.

[0035] In the embodiment of the present application, the structure of the microring resonator array in the optical signal modulation module 21 and the optical signal multiplication and addition operation module 22 includes a plurality of microring resonators arranged in a matrix, and a heating unit is provided on each microring resonator. The microring resonator modulates the intensity of the optical signal by using its resonance characteristics for light of a specific wavelength, and the heating unit is used to change the refractive index of the microring resonator to adjust the resonant wavelength.

[0036] It should be understood that a microring resonator is an integrated photonic device that uses the resonant properties of light to modulate the intensity of an optical signal. Its operating principle is based on the cyclic propagation of light waves in a microring, and when the frequency of the light wave matches the resonant frequency of the microring resonator, a resonance phenomenon occurs. The resonant characteristics of a microring resonator are closely related to factors such as its physical size, material properties, and coupling conditions. By precisely designing the geometric parameters of the microring and controlling the coupling strength, precise modulation of the optical signal can be achieved. In addition, the high quality factor (Q value) of the microring resonator means that light energy can be stored in the microring for a longer time, which helps to improve modulation efficiency and device performance.

[0037] In a microring resonator, the refractive index of the microring resonator can be changed by a heating unit, thereby adjusting the resonant wavelength. The change in refractive index affects the propagation constant of light in the microring, causing the resonance condition to change, thereby adjusting the resonant wavelength. This adjustment mechanism can be applied to devices such as tunable filters, optical switches, and optical sensors. Exemplarily, the heating unit in this embodiment can be a heating resistor unit or a heating electrode composed of titanium nitride, which is a golden crystalline powder with a high melting point, high hardness, good high temperature strength and chemical corrosion resistance. Its thermal conductivity is 29.1W / (m·K), the thermal expansion coefficient is 9.35×10^-6 / ℃, and the resistivity is 25μΩ·cm. These characteristics make titanium nitride very suitable for the heating resistor unit on the microring resonator, which can generate heat by applying current, thereby adjusting the optical properties of the microring resonator.

[0038] The modulation of the optical signal modulation module 21 is controlled by the computer module 4 through the Python programming language. For example, the computer module 4 controls the multi-channel data acquisition card through the Python programming language to modulate the voltage value of the heating electrode, thereby controlling the optical signal modulation module 21 to modulate the intensity of the optical signal through the thermo-optical effect. Thermo-optical modulation is a special optical modulation method that modulates the optical signal through temperature changes. In this system, the computer controls the multi-channel data acquisition card through the Python programming language, and the data acquisition card adjusts the voltage value of the heating electrode according to the computer's instructions, thereby achieving precise control of the intensity of the optical signal through the thermo-optical effect.

[0039] The optical signal multiplication and addition operation module 22 also includes a microring resonator array and a heating unit arranged on each microring resonator in the microring resonator array; each microring resonator in the microring resonator array represents a single neuron, which is used to express different weights and perform multiplication and addition operations of optical signals.

[0040] The multiplication of optical signals is accomplished by simulating the dot product process of matrices and vectors. Each microring resonator corresponds to a weight element in the matrix. By adjusting the heating resistor unit on the microring, the resonant frequency of the resonator can be changed, thereby changing the phase and intensity of the optical signal passing through the microring, which is equivalent to adjusting the weight. The input vector is represented by optical signals of different wavelengths. When these signals pass through the microring array, each microring modulates the light of a specific wavelength according to its resonance characteristics, realizing the product of the corresponding weight and the input value. Finally, the output optical signals of all microrings are superimposed in space to form four new optical signals, whose intensities represent the results of the original matrix and vector multiplication (the entire calculation unit is equivalent to performing a 1*4 vector and a 4*4 matrix multiplication, and the output result is a 1*4 vector with four values). This process effectively realizes the multiplication of matrices and vectors, converting the multiplication in linear algebra into phase and intensity modulation in the optical domain.

[0041] The addition operation of optical signals is completed by spatially superimposing the intensities of optical signals of different wavelengths. Each microring resonator independently modulates the optical signal passing through it, and these signals represent different components of the addend. By precisely controlling the heating resistor unit, the resonance condition of each microring can be adjusted so that each microring modulates the optical signal of a specific wavelength in a specific way. When these modulated optical signals meet at the output end, their light intensities are superimposed in space, simulating the summation process in the addition operation. The normalized input value is represented by the output wavelength intensity of the microring resonator, and the superposition sum of optical signals of different wavelengths at the output end is the result of the addition operation. This method takes advantage of the parallelism and high speed of optical signals to achieve fast summation of multiple signal components.

[0042] It should be understood that the optical neural network computing module combines the wavelength division multiplexing of optical signals with on-chip optical intensity modulation using an on-chip modulated microring resonator array, thereby realizing the modulation of input signals of multiple different wavelengths in one optical fiber. Wavelength division multiplexing (WDM) technology is a technology that simultaneously transmits multiple optical signals of different wavelengths in a single optical fiber. Its principle is to use the difference in optical wavelengths to multiplex multiple optical signals at the transmitting end, and then separate these optical signals through a demultiplexer at the receiving end to achieve parallel transmission of multiple signals. The input single-wavelength signal optical intensity modulation array is used to modulate the intensity of a single-wavelength optical signal. Here, "modulation" refers to the selection and processing of optical signals of different wavelengths, which refers to wavelength-selective modulation, that is, allowing optical signals of specific wavelengths to pass through, while blocking or weakening optical signals of other wavelengths. In short, wavelength division multiplexing technology allows multiple optical signals to be transmitted simultaneously in a single optical fiber, and the input single-wavelength signal optical intensity modulation array is used to adjust the intensity of optical signals of specific wavelengths.

[0043] To facilitate those skilled in the art to better understand that the on-chip modulated microring resonator array in the present application combines wavelength division multiplexing with on-chip light intensity modulation, the following is combined with Figure 2a to 2c The embodiments in the present invention are further explained.

[0044] Figure 2a It is shown in the figure that, compared with the common off-chip modulation and light source intensity modulation element solutions, the structure of the present application uses microring resonators with different effective radii to directly select resonance peaks of different wavelengths. The microring resonator array proposed in the present application can directly modulate the light intensity of different wavelengths on the chip to express different input data, such as the light intensity of wavelength λ1, the light intensity of wavelength λ2, the light intensity of wavelength λ3, and the light intensity of wavelength λ4. The compact structure of the microring resonator can be as small as a few microns in radius, which greatly reduces the space occupied by photonic devices and improves the integration density.

[0045] Figure 2b The figure shows an optimized design embodiment of a single microring resonator. Partial Euler bending is used in the microring resonator structure to achieve higher device integration on the silicon optical chip. It should be understood that partial Euler bending uses the Euler bending principle to minimize the transmission loss of light in the curved waveguide while maintaining the compactness of the waveguide, which allows light to turn in a smaller radius in the waveguide without mode coupling loss. Taking the Add-Drop type microring resonator as an example, it includes a drop port, an add port, an input port, and a through port. The Drop port is a download port, which is one of the output ports of the microring resonator and is used to extract optical signals of specific wavelengths that match the resonant mode of the microring resonator, that is, to allow selective optical signals to be "downloaded" from the microring (separated from the main optical path). The Add port is used to inject optical signals of specific wavelengths into the microring resonator. The port allows an external light source or signal to be coupled into the microring resonator and interfere with the optical signal passing through the port to achieve dynamic optical signal processing. The input port is the initial entrance for the optical signal to enter the microring resonator. All optical signals passing through the microring resonator first enter through this port, and then the optical signals are distributed to different ports according to the wavelength and the resonance conditions of the microring. The through port is a direct port that allows optical signals that are not coupled to the resonant mode of the microring resonator to pass directly without obvious energy exchange. Figure 2b The "V" in the figure represents the electric heating generated by applying voltage to the resistor in thermo-optical modulation.

[0046] Figure 2cThe spectrum generated by the simulation of the designed partial Euler bent micro-ring resonator (micro-ring simulation) using the three-dimensional finite difference time domain method is shown in the figure. The horizontal axis in the figure is the wavelength in nanometers (wavelength / nm), and the vertical axis is the variation of transmission efficiency with wavelength in decibels (transmission / dB). It can be seen from the figure that the transmission through the resonant wavelength is almost lossless. The simulation results verify that the partial Euler bending design can significantly reduce the transmission loss in the micro-ring resonator.

[0047] The photoelectric conversion module 3 is used to convert the modulated and calculated optical signal into an electrical signal and transmit it to the computer module 4. The photoelectric conversion module 3 can use a photodetector, whose working principle is based on the photoelectric effect. When light irradiates the semiconductor material, the photons hit the electrons in the semiconductor, causing the electrons to transition to the conduction band, forming holes and electron pairs, and these carrier pairs form current under an external bias. Specific types of photodetectors include, but are not limited to: photodiodes, photomultiplier tubes, phototransistors, avalanche photodiodes, etc., and the embodiments of the present application do not limit this.

[0048] The computer module 4 can be a computer (personal computer, laptop, tablet computer, etc.), a server, a workstation, etc. The computer module 4 performs simulation through the simulation system loaded thereon, and obtains the adaptation weight parameters after pre-training the model of the pre-collected data set. The obtained weight parameters are loaded onto the optical neural network calculation module 2. For example, based on the flexibility of the Python language and the rich libraries and frameworks, Python is used to simulate the system and devices, and a powerful simulation system can be constructed. The constructed simulation system includes, for example, SimuPy, Python Robotics, Elecpy, etc., which is not limited in the embodiments of the present application.

[0049] In addition, the computer module 4 is also used to perform nonlinear activation processing after receiving the electrical signal from the photoelectric conversion module 3. The data after the nonlinear activation processing is used to update the network weights, and then perform iterative optimization of the model.

[0050] In some examples, the computer module 4 can use the Python programming language to control the optical signal modulation module 21, that is, to control the modulation and weighting process of the signal through a Python script or program. The Python programming language provides a rich interface and library to interact with hardware devices (such as data acquisition cards, signal generators, etc.). For example, in the modulation process, a specific Python library (such as NumPy and SciPy) is used to generate and process signals, such as using nump.sin to generate a sine wave as a carrier, and then AM modulation is realized by signal and product. For another example, in the weighting process, Python can realize weighting by multiplying the signal array by a weight coefficient, and the weight coefficient can be adjusted according to the training process of the network. However, it should be understood that in addition to the Python programming language, other methods and tools can also be used to control the modulation and weighting process of the signal, including but not limited to: MATLAB and Simulink, professional software and toolboxes (such as DSP system toolboxes, phased array system toolboxes, communication toolboxes, etc.), hardware description languages ​​(HDL), C / C++ programming, application-specific integrated circuits (ASICs), MCU programming, etc., and the embodiments of the present invention do not limit this.

[0051] In some examples, the process in which the computer module 4 performs model pre-training on the data set to obtain the adaptive weight parameters includes: Figure 3 The steps shown in:

[0052] Step S31: input the pre-collected data set into a pre-selected machine learning model for model pre-training.

[0053] In this step, using optimization algorithms such as gradient descent, through forward propagation and back propagation processes, the weights in the network are continuously adjusted to minimize the loss function, thereby obtaining near-optimal network parameters. The machine learning model referred to herein may be, for example, any of the following: linear regression model, logistic regression model, decision tree model, support vector machine model, naive Bayes model, K-means clustering model, principal component analysis model, neural network model (such as convolutional neural network, feedforward neural network, recurrent neural network, long short-term memory network, etc.), deep learning model, reinforcement learning model, generative model, etc., and the embodiments of the present application are not limited to this.

[0054] Step S32: After the pre-training is completed, the obtained weights are encoded into light domain parameters adapted to the optical neural network.

[0055] It should be understood that the weights in the computer are converted into parameter forms suitable for the optical neural network, including converting the weights in electronic form into parameters in the optical domain, such as optical parameters such as phase, amplitude or polarized light. The schemes that can be implemented include but are not limited to any of the following: (1) Using diffractive optical elements: By designing a specific diffraction grating or phase plate, the weights can be encoded as phase changes of light waves. (2) Using Mach-Zehnder interferometer (MZI): In an all-optical neural network, MZI can be used to adjust the weights; by changing the phase of the phase modulator in the MZI, the light intensity ratio in the optical path can be adjusted, thereby simulating the weights in the neural network. (3) Using nonlinear optical effects: Nonlinear optical elements, such as saturated absorbers or Kerr effect materials, can implement nonlinear activation functions. These elements nonlinearly modulate the intensity or phase of light, thereby introducing nonlinear characteristics in the optical neural network. (4) Using phase change materials: Phase change materials can transform between amorphous phase and crystalline phase. This transformation can change the ability of the material to absorb light, thereby realizing weight adjustment. (5) Using MEMS variable optical attenuators (VOAs): MEMS VOAs can precisely control the attenuation of light, thereby adjusting the weights of photon pulse neurons. (6) Wavelength division multiplexing technology: Wavelength division multiplexing technology can be used to achieve parallel processing of weights. In this way, different wavelengths can represent different weights, achieving efficient weight encoding and calculation.

[0056] Step S33: after encoding the weights into corresponding light domain parameters, they are loaded into the optical neural network calculation module.

[0057] After encoding, the weights in the form of optical domain parameters are transmitted to each node of the optical neural network through waveguides, optical fibers or other optical transmission media, using the high speed and parallelism of light to accelerate the loading of weights.

[0058] In optical neural networks, nonlinear activation functions are indispensable, as they can accelerate the convergence speed of the network and improve the recognition accuracy. The activation functions that can be used in this embodiment include, but are not limited to, Sigmoid function, tanh function, ReLU function, etc. For example, the Sigmoid function can map the input value to between 0 and 1, while the ReLU function is more suitable for processing tasks such as natural language due to its performance efficiency and sparsity.

[0059] Therefore, in an embodiment of the present application, the computer module 4 uses the Python programming language to perform nonlinear activation processing on the electrical signal received from the photoelectric conversion module 3. Specifically, deep learning libraries such as PyTorch can be used to implement nonlinear activation functions, such as using functions such as torch.nn.ReLU() or torch.nn.Sigmoid(), so that the data processed by nonlinear activation can be used to update the network weights, and then iterative optimization is performed. Exemplarily, in an application scenario of the present application, only the inference stage of the neural network model is performed on the chip. The inference stage is the stage after training, which is used to use the trained model for prediction or classification. In this stage, the weights are not iteratively optimized, only the weights are loaded, multiplication and addition operations are performed, and the results are output. The data output after nonlinear processing continues to be used as the input value of the next layer of data.

[0060] Figure 4 (a) to (d) show the specific architecture implementation diagram of the on-chip modulated optical neural network system based on microring resonator proposed in the present invention.

[0061] like Figure 4 As shown in (a), the optical signal generated by the wide-spectrum light source module 1 (wide-spectrum laser light source) is input into the microring resonator array, and four single wavelengths (wavelengths λ1, λ2, λ3, and λ4) are obtained through wavelength division multiplexing, and the light intensity is modulated to achieve vector coding.

[0062] It is understandable that in wavelength division multiplexing technology, high-speed and large-capacity optical communication can be achieved by using optical signals of different wavelengths to transmit multiple signals in the same optical fiber. Vector coding refers to encoding the input signal using different wavelengths (such as wavelengths λ1, λ2, λ3, λ4), each wavelength corresponding to a specific component of the input signal, in this way to achieve parallel transmission and processing of signals, thereby improving the system's information processing capabilities and efficiency. In the specific process of vector coding, optical signals of different wavelengths are used to represent the various components of a vector, and each component corresponds to a specific wavelength. In optical communication systems, vector data can be encoded by modulating the intensity, phase or frequency of optical signals of different wavelengths. Then, these encoded optical signals can be transmitted simultaneously in the optical fiber to achieve high-density and high-speed data transmission.

[0063] like Figure 4As shown in (b), the overall system structure of the designed on-chip microring resonator array uses a 2-row 4-column microring resonator array to re-implement on-chip wavelength division multiplexing and intensity modulation of a specific wavelength. It uses a 4×4 array of adjustable microring resonators to implement matrix calculations in neural networks. It should be understood that the first two rows implement the functions of wavelength division multiplexing and input vector modulation, with a 1*4 vector as input. The 4*4 microring is the computing core that loads the 4*4 weights. The structure of the entire chip is Figure 4 From the schematic diagram in (b) and the microscopic image of the chip, it can be seen that the scale is 6*4.

[0064] Since different microring resonators in the array have different resonant wavelengths, each microring resonator can affect optical signals of different wavelengths. Voltage can be applied to the heating resistor unit composed of titanium nitride for thermal modulation, thereby changing the weight of each signal. The input value is then normalized and represented by the resonant wavelength intensity output by the first two rows of microring resonators. In this case, the superposition of the light intensity output by the microring resonators realizes matrix addition operations, while the microring resonators in the n×n array realize matrix multiplication operations. Finally, the microring resonator array can be combined with differential detection technology to realize the calculation of y=Wx+b in a single neuron implemented by the microring resonator.

[0065] Figure 4 In (c), the four laser wavelengths obtained by the broadband spectrum light source are the input signals of the microring resonator array computing core. The horizontal axis is the wavelength in nanometers (wavelength / nm), and the vertical axis is the optical power (Optical Power), which is in decibel milliwatts (dBm). Optical power refers to the energy of light passing through a unit area per unit time, and is usually used to describe the intensity of a light source or optical signal. The unit of optical power is usually watt (W), but in the field of optics, since the value of optical power is usually small, decibel milliwatt (dBm) is often used as a unit. It is a relative unit, indicating the ratio to 1 milliwatt.

[0066] Figure 4 (d) is a microscope image of the packaged chip and an image of the microring resonator array based on the overall system structure. The image is for reference only and is not intended to show the precise circuit design details.

[0067] Figure 5a , 5b It is a spectrum diagram of the output of the microring resonator array in the optical chip submitted by this application, and it is a schematic diagram of a single neuron operation based on a microring resonator implemented in the embodiment of this application. The experimental test image shows the spectrum of the final output of the broadband light source after passing through the microring array, which is the normalized photoelectric power curve obtained by the photodetector. Figure 5aThe graphs in the figure are the curves of the photoelectric power of the microring array download end (drop end) and the through end (through end) as a function of voltage. The horizontal axis is voltage in volts (Voltage / V), and the vertical axis is optical power (Normalized power). Figure 5b The horizontal axis is voltage, the unit is volt (Voltage / V), and the vertical axis is light power (Normalized power). Figure 5b The difference between the through end and the drop end can give the required [-1,1] weight expression function curve.

[0068] Figure 6 It is the accuracy of the training data set in the classification of the handwriting data set (Modified National Institute of Standards and Technology, MNIST) provided by the present invention. The present invention designs a photonic neural network (LightONN) that matches a chip based on a microring resonator (MRR), and proposes a universal quantization simulation scheme to reduce the performance loss of chip reasoning. The single neuron implemented by the optical computing chip obtained by experimental testing is introduced as a matrix multiplication operation unit into the constructed three-layer optical neural network of LightONN for MNIST data set recognition. The neural network includes two layers of 3×3 convolutional layers and one fully connected layer.

[0069] In the embodiments of the present application, words such as "first" and "second" are used to distinguish the same or similar items with substantially the same functions and effects. For example, the first XX and the second XX are only used to distinguish different XXs, and do not limit their order. Those skilled in the art can understand that words such as "first" and "second" do not limit the quantity and execution order, and words such as "first" and "second" do not necessarily limit them to be different.

[0070] It should be noted that in the embodiments of the present application, words such as "exemplary" or "for example" represent examples, illustrations or descriptions. Any embodiment or design described as "exemplary" or "for example" in the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific way.

[0071] In the embodiments of the present application, "at least one" refers to one or more, and "plurality" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone, where A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can represent: a, b, c, ab, ac, bc or abc, where a, b, c can be single or multiple.

[0072] The present application also provides an optical chip based on a microring resonator, which is applied to the optical neural network system based on on-chip modulation of a microring resonator provided in the above embodiment, including: an optical signal modulation module, an optical signal multiplication and addition operation module; the optical signal modulation module and the optical signal multiplication and addition operation module both include a microring resonator array; wherein the modulation of the optical signal modulation module and the weighted operation of the optical signal multiplication and addition operation module are both controlled by a computer module through a programming language. Since the structure and principle of the optical chip have been described in detail above, they will not be repeated here.

[0073] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0074] In summary, the present application provides an optical chip and an optical neural network system based on on-chip modulation of a microring resonator. By combining the wavelength division multiplexing of a broadband light source with on-chip light intensity modulation, the present solution can effectively utilize the multi-wavelength characteristics of light to achieve high-density and high-capacity information processing. This on-chip modulation technology enables optical neural networks to achieve higher integration and shows great potential for photonic computing. The optimized partial Euler bent microring reduces transmission loss through small radius waveguide bending. This structure not only reduces the chip size, but also reduces manufacturing costs; it improves the overall energy efficiency ratio. The optical neural network architecture provided by the present application is optimized for the physical properties of the microring resonator, so it can better adapt to the characteristics of the microring resonator and can better exert the performance of the microring resonator in the optical neural network. The quantized neural network and quantized perceptual training (QAT) introduced in the present application are more adaptable to the current chip manufacturing level based on microring resonators and reduce the error impact on the hardware of the microring resonator array. Therefore, the present application effectively overcomes the various shortcomings of the prior art and has a high industrial utilization value.

[0075] The above embodiments are merely illustrative of the principles and effects of the present application and are not intended to limit the present application. Anyone familiar with the technology may modify or change the above embodiments without violating the spirit and scope of the present application. Therefore, all equivalent modifications or changes made by a person of ordinary skill in the art without departing from the spirit and technical ideas disclosed in the present application shall still be covered by the claims of the present application.

Claims

1. An optical neural network system based on on-chip modulation of microring resonators, It is characterized in that include: Wide spectrum light source module, optical neural network calculation module, photoelectric conversion module and computer module; The wide spectrum light source module is used to provide a continuous wide spectrum light signal; The optical neural network calculation module is located on the emission path of the optical signal generated by the wide spectrum light source module, and includes an optical signal modulation module and an optical signal multiplication and addition operation module; the optical signal modulation module and the optical signal multiplication and addition operation module both include a microring resonator array; wherein the modulation of the optical signal modulation module and the weighted operation of the optical signal multiplication and addition operation module are both controlled by the computer module through a programming language; The photoelectric conversion module is connected to the optical neural network calculation module, and is used to convert the modulated and weighted optical signal into an electrical signal and transmit it to the computer module; The computer module performs simulation through the simulation system loaded thereon, obtains adaptation weight parameters after model pre-training of the pre-collected data set; the adaptation weight parameters are loaded into the optical neural network calculation module; and the computer module performs nonlinear activation processing on the electrical signal sent by the photoelectric conversion module, wherein the data obtained through the nonlinear activation processing is used to update the weight parameters for iterative optimization of the model.

2. The optical neural network system based on on-chip modulation of microring resonator according to claim 1, It is characterized in that The optical signal modulation module and the optical signal multiplication and addition operation module both include a microring resonator array, and the structure thereof includes: The microring resonator array comprises a plurality of microring resonators arranged in a matrix, and a heating unit is disposed on each microring resonator. The microring resonator uses its resonance characteristics for light of a specific wavelength to modulate the intensity of the optical signal, and the heating unit is used to change the refractive index of the silicon optical waveguide of the microring resonator to adjust the resonance wavelength.

3. The optical neural network system based on on-chip modulation of microring resonator according to claim 2, It is characterized in that The optical signal modulation module is realized by spatially superimposing optical signals of different wavelengths in a microring resonator array, wherein each microring independently modulates an optical signal of a specific wavelength, changes the resonance condition by adjusting a heating resistor unit, and finally completes the summation by superimposing the light intensity at the output end, thereby realizing the addition operation of multiple signal components in parallel.

4. The optical neural network system based on on-chip modulation of microring resonator according to claim 2, It is characterized in that Each microring resonator in the microring resonator array of the optical signal multiplication and addition operation module corresponds to a weight in a matrix and independently regulates the optical signal, and simulates the dot product of the weight and the input vector by adjusting the resonance frequency, and finally obtains the result of the multiplication operation by spatially superimposing the output light intensity of each microring resonator.

5. The optical neural network system based on on-chip modulation of microring resonator according to claim 2, It is characterized in that The heating unit includes a heating resistor unit made of titanium nitride.

6. The optical neural network system based on on-chip modulation of microring resonator according to any one of claims 1 to 5, It is characterized in that Partial Euler bending is used in the structure of the microring resonator to optimize the local geometry of the waveguide.

7. The optical neural network system based on on-chip modulation of microring resonator according to any one of claims 1 to 5, It is characterized in that After receiving the optical signal from the wide spectrum light source module, the microring resonator array obtains multiple single wavelength optical signals through wavelength division multiplexing technology, and modulates each single wavelength optical signal to complete vector coding and perform optical fiber transmission.

8. The on-chip modulated optical neural network system based on microring resonator according to claim 1, It is characterized in that The computer module performs model pre-training on the pre-collected data set to obtain the adaptation weight parameters, and the process includes: Input the pre-collected data set into the pre-selected machine learning model for model pre-training; After the pre-training is completed, the obtained weights are encoded into optical domain parameters adapted to the optical neural network; After the weights are encoded into corresponding light domain parameters, they are loaded into the optical neural network calculation module.

9. The optical neural network system based on on-chip modulation of microring resonator according to claim 1, It is characterized in that The computer module performs nonlinear activation processing on the electrical signal sent by the photoelectric conversion module through any of the following functions: Sigmoid function, tanh function, ReLU function.

10. An optical chip based on microring resonator, It is characterized in that Applied to the on-chip modulated optical neural network system based on microring resonators as claimed in claim 1, the optical chip comprises: an optical signal modulation module, an optical signal multiplication and addition operation module; the optical signal modulation module and the optical signal multiplication and addition operation module both comprise a microring resonator array; wherein the modulation of the optical signal modulation module and the weighted operation of the optical signal multiplication and addition operation module are both controlled by a computer module through a programming language.

Citation Information

Cited By

  • Reconfigurable vector-vector multiplication structure based on micro-ring resonators and implementation method

    CN119828292A

  • Optical neural network reconstruction-inhibition neuron model and architecture

    CN119886245A

  • Optical neural network general operator model, architecture and system

    CN119962603A

  • Optical neural network general operator model, architecture and system

    CN119962603B

  • Intelligent optical computing chip cluster architecture and system

    CN120146128A