All-optical reconfigurable silicon-based photonic neural network chip based on wavelength division multiplexing

Abstract

The present invention relates to an all-optical reconfigurable silicon-based photonic neural network chip based on wavelength division multiplexing, which belongs to the field of optical computing. The chip adopts a new all-optical nonlinear control mechanism, which can realize the nonlinear activation of wavelength division multiplexing signals, and also has a multi-layer, reconfigurable photonic neural network with variable number of network layers, number of neurons in each layer, and adjustable nonlinear activation function. The single-layer network in the photonic neural network chip consists of a light splitting area, a linear operation module, a nonlinear operation module, a light extraction device module, and a light combining area. The light splitting / light combining area is used to realize the multi-path splitting / light combining of signal light and detection light; the linear operation module is used to realize the weight operation of signal light and regulate the number of neurons in a single layer; the nonlinear operation module is used to realize the operation of adjustable nonlinear activation function; the light extraction device module is used to output the photonic neural network chip results and realize the regulation of the number of network layers.

Description

Technical Field

[0001] The present invention belongs to the field of optical computing and relates to an all-optical reconfigurable silicon-based photonic neural network chip based on wavelength division multiplexing. Background Art

[0002] With the increase in communication speed and the development of the Internet of Everything, the demand for ultra-high-speed and energy-efficient computing power is also increasing, and the existing traditional electronic computing system under the von Neumann architecture is difficult to achieve ultra-high computing power and low energy consumption at the same time. To solve this problem, it is urgent to seek high-speed, low-energy technical methods and means to achieve the improvement of computing power. Photons have become the most promising computing carrier with their large capacity, ultra-low latency, low energy consumption and high parallel capabilities. With the continuous development of integration technology, micro photonic chips can realize function-rich signal generation, processing, transmission and other operations. In 2019, the US Department of Defense launched the "Future Computing System" project to carry out research on photonic computing chips. The German Aerospace Center started the optical quantum computer research and development project in 2022. Nippon Telephone and Telegraph Corporation has started research on coherent Ising machines in the past two years. The Ministry of Science and Technology's special application guide also makes optoelectronic hybrid AI acceleration computing chips and optical neuromorphic computing systems important contents. Therefore, photonic chips have excellent performance in improving computing power and reducing power consumption, and are expected to become one of the important representatives of new computing power processors in the future. The research on photonic neural network chips is not only the most important branch in the field of optical computing, but also the most valuable research direction due to its wide range of applications.

[0003] Most of the early neural networks realized by optical means focused on the use of semiconductor lasers to realize pulse neural networks (SNN) (Jha A, Huang C, Peng HT, et al. Photonic spiking neural networks and graphene-on-silicons piking neurons. Journal of Lightwave Technology, 2022). In some specific tasks, its power consumption and robustness are better than the artificial neural network (ANN) at that time, and it is very suitable for event-based tasks, such as time series prediction (Hao, SXHSGZRWMZMXDZRDLHA review: Photonics devices, architectures, and algorithms for optical neural computing. Journal of Semiconductors, 2021), etc., but it requires special training algorithms and is very limited in many applications. After years of development, ANN's architecture and ability to handle tasks have become very mature. Therefore, in recent years, a large number of solutions for realizing ANN using optical means have emerged, and have developed into a new research direction-photonic neural network (ONN). At present, ONN is divided into two categories: integrated ONN and non-integrated ONN. Non-integrated ONN is mainly based on the principle of spatial diffraction. It is also the most complete and largest type of network (Zuo Y, Li B, Zhao Y, et al. All-optical neural network with nonlinear activation functions. Optica, 2019). However, it has disadvantages such as large size and poor stability. In addition, it is difficult for this type of ONN to implement nonlinear operations in free space. Existing integrated ONNs also have many problems. For example, in linear operations, the most common is the coherent matrix-vector multiplication (MVM) scheme based on MZI (Shen Y, Harris NC, Skirlo S, et al. Deep learning with coherent nanophotonic circuits. Nature photonics, 2017). When implementing the weight matrix, matrix decomposition is first required. It takes N (N-1) basic units to construct an N×N unitary matrix. For large-scale networks, the difficulty of implementation can be imagined. Therefore, the scale of linear operations that can be achieved is relatively small.However, the simplest incoherent MVM scheme based on wavelength division multiplexing (Tait AN, De Lima TF, Zhou E, et al. Neuromorphic photonic networks using silicon photonic weightbanks. Scientific reports, 2017) does not have a matching all-optical nonlinear computing scheme. Nonlinear computing is currently usually implemented by computers, and some schemes use optoelectronic hybrid platforms. This not only requires a large number of opto-electric-optical conversion devices such as electro-optical modulators and photodetectors, which greatly increases the cost and power consumption, but also weakens the advantages of large capacity and high speed of optical computing. The existing all-optical nonlinear computing schemes mostly use the nonlinear effects of the materials themselves, resulting in a high nonlinear power threshold, high nonlinear computing power consumption, and are only applicable to coherent MVM schemes. In addition, once the chip is prepared, the number of ONN layers, the number of neurons in each layer, and the nonlinear activation function curves of each layer are fixed, and cannot be flexibly changed according to actual tasks and training sets to obtain higher computing accuracy. The network structure and function are relatively simple. In addition, although the existing all-optical nonlinear operation schemes can achieve an accuracy of more than 90% in simulations, to obtain this accuracy, the pixel scale of the general weight matrix needs to reach ~105, which is almost impossible to achieve under the existing on-chip linear operation network structure. Therefore, there are very few integrated solutions that include both linear and nonlinear operation modules to build a complete ONN and have been experimentally verified.

[0004] Through analysis, it can be seen that on the basis of existing integration technology, comprehensive consideration of the linear and nonlinear operation characteristics of ONN, and the use of a simple and compact structure to achieve an all-optical reconfigurable ONN chip based on wavelength division multiplexing, with variable number of neural network layers and number of neural networks in each layer, and adjustable nonlinear activation function curve, has groundbreaking research value. Summary of the invention

[0005] In view of this, the purpose of the present invention is to provide an all-optical reconfigurable silicon-based photonic neural network chip based on wavelength division multiplexing. This chip breaks through the defects of the existing wavelength division multiplexing ONN scheme that the all-optical nonlinear activation function cannot be used and the problem that the neuron scale of the linear operation module and the nonlinear operation module do not match each other. It adopts a new nonlinear control method to realize all-optical operation for wavelength division multiplexing signals, and achieves a calculation accuracy of 94% based on the single-layer neuron number of only 12. In addition, the chip can also realize the reconfigurable function of variable number of neural network layers and neurons in each layer, and adjustable nonlinear activation function curve, with a wide range of applications and strong usability.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] An all-optical reconfigurable silicon-based photonic neural network chip based on wavelength division multiplexing, wherein a single-layer network in the photonic neural network chip consists of a light splitting area, a linear operation module, a nonlinear operation module, a light extraction device module, and a light combining area;

[0008] The light splitting area is used to realize multi-path light splitting of signal light and detection light;

[0009] The linear operation module is used to realize the weight operation of signal light and regulate the number of single-layer neurons;

[0010] The nonlinear operation module is used to implement the operation of an adjustable nonlinear activation function;

[0011] The light extraction device module is used to output the photonic neural network chip results and realize the regulation of the number of network layers;

[0012] The light combining area is used to realize multi-path combining of signal light and detection light to the next layer as optical signal input of the network.

[0013] Furthermore, the light splitting area is a structure of a plurality of Y-branch power splitters connected in series, a plurality of 1×2-way multi-mode interference coupler MMI power splitters connected in series, or a 1×n-way MMI power splitter.

[0014] Furthermore, the linear operation module is a wavelength division multiplexing signal power modulation module implemented by a circular, racetrack-shaped, or special-shaped resonant cavity or grating array, and the basic units in the array are connected in series, parallel, or cascade series.

[0015] Furthermore, the nonlinear operation module is an array consisting of a single or multiple waveguide suspended structures, a ring resonant cavity suspended structure, or a ring resonant cavity and a straight waveguide coupling region suspended structure.

[0016] Furthermore, the light extraction device module is an array composed of Mach-Zehnder interferometer MZI structures. Path selection is achieved by regulating the micro-thermal electrode on one of the arms of the MZI, and a photoelectric conversion device is prepared or connected at one of the output ports to achieve the output of the neural network chip.

[0017] Furthermore, the light combining area is a structure of a plurality of Y-branch power dividers connected in series, a plurality of 1×2-way MMI power dividers connected in series, or a 1×n-way MMI power divider.

[0018] The beneficial effects of the present invention are: adopting a new nonlinear control method, realizing an all-optical nonlinear activation function based on the existing wavelength division multiplexing ONN solution, solving the problem of mismatching the neuron scale of the linear operation module and the nonlinear operation module, and realizing a calculation accuracy of 94% based on the number of neurons in a single layer of only 12. In addition, the chip can also realize the reconfigurable function of variable number of neural network layers and number of neurons in each layer, and adjustable nonlinear activation function curve, with a wide range of applications and strong usability.

[0019] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0021] Figure 1 It is a schematic diagram of the double waveguide suspended structure; Figure 1 (a) is a schematic diagram of the 3-D structure; Figure 1 (b) Top view and related parameter settings;

[0022] Figure 2 The calculation results of the dual-waveguide suspended optical gradient force unit are shown in Figure 2. Figure 2 (a) The effective refractive index and optical gradient force of the optical gradient force unit change with the waveguide spacing; Figure 2 (b) is the offset between the nanowire waveguide and the conventional waveguide when the total power in the waveguide is 0.7 mW; Figure 2 (c) for Figure 2 (b) Variation trend of effective refractive index along the waveguide direction under offset; Figure 2 (d) is the relationship between the phase change of the detection light caused by the optical gradient force structure and the total power of the transmitted light;

[0023] Figure 3 It is a schematic diagram of the structure of the nonlinear operation unit;

[0024] Figure 4 The nonlinear activation function curve graph implemented by the nonlinear operation unit. Figure 4 (a) is the change with Pd; Figure 4 (b) When Pd = 1.08mW, the change with φ;

[0025] Figure 5 It is a schematic diagram of the linear operation module structure;

[0026] Figure 6 This is a schematic diagram of the ONN chip structure;

[0027] Figure 7 It is a schematic diagram of the structure of the light extraction device unit;

[0028] Figure 8 This is a diagram of the fully connected neural network architecture and simulation results;

[0029] Fig. 9 This is a schematic diagram of the present invention.

[0030] Figure numerals: 1-light splitting area, 2-linear operation module, 3-non-linear operation module, 4-light extraction device module, 5-light combining area. DETAILED DESCRIPTION

[0031] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention 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 invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0032] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0033] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0034] In the ONN system based on wavelength division multiplexing, since the information is carried on optical signals of different wavelengths, it is impossible to use the nonlinear effects of the material itself, such as the Kerr effect, carrier absorption, etc. to realize the nonlinear activation function g(·). Therefore, a new regulation mechanism is needed.

[0035] The present invention uses an optical gradient force to achieve nonlinear regulation of the multi-wavelength signal optical power to the detection light phase, wherein the optical gradient force structure is a double-waveguide suspended structure, such as Figure 1 As shown, Figure 1 It is a schematic diagram of the double waveguide suspended structure; Figure 1 (a) is a schematic diagram of the 3-D structure; Figure 1 (b) is a top view and related parameter settings. One of the waveguides is a conventional waveguide and the other is a nanowire waveguide. Both waveguides use double-ended fixed beams. For this structure, the following formula can be used to calculate the optical gradient force:

[0036]

[0037] Where F represents the light gradient force, l g represents the length of the suspended waveguide, P t Represents the total power of the light transmitted in the waveguide, including the multi-wavelength signal light power and the detection light power. eff represents the effective refractive index of the dual waveguide structure as a whole, and g represents the spacing between the two waveguides. Using the Euler-Bernoulli beam theory, the deflection of the two waveguides can be calculated by the optical gradient force:

[0038]

[0039]

[0040] where u and u d represents the deflection of conventional waveguide and nanowire waveguide respectively; z-axis direction is the direction of light transmission and also the direction of waveguide; E = 1.71 × 10 11 represents the Young's modulus of silicon material, w and w d represent the widths of conventional waveguide and nanowire waveguide, A and A respectively. d represent the cross-sectional areas of conventional waveguide and nanowire waveguide, respectively.

[0041] Design appropriate structural parameters and bring them into equations (1), (2) and (3), and we can get Figure 2 result. Figure 2 The calculation results of the dual-waveguide suspended optical gradient force unit are shown in Figure 2. Figure 2 (a) The effective refractive index (blue solid line) and optical gradient force (red dashed line) of the optical gradient force unit change with the waveguide spacing; Figure 2 (b) is the offset between the nanowire waveguide (blue dashed line) and the conventional waveguide (red solid line) when the total power in the waveguide is 0.7 mW; Figure 2 (c) for Figure 2 (b) Variation trend of effective refractive index along the waveguide direction under offset; Figure 2 (d) The relationship between the phase change of the detection light caused by the optical gradient force structure and the total power of the transmitted light.

[0042] Figure 2 The blue solid line in (a) represents the curve of the effective refractive index of the system composed of conventional waveguide and nanowire waveguide as the distance between the two waveguides changes. Figure 2 The red dashed line in (a) is the curve obtained by substituting various structural parameters into equation (1) to show the variation of the optical gradient force with the spacing between the two waveguides under unit pump light input and unit waveguide length. Figure 2 (b) is the total power P of the light transmitted in the waveguide t The deflection of the conventional waveguide and the nanowire waveguide when the power is 0.7mW. In this case, the maximum offset of the conventional waveguide is only 0.21nm, which can be considered as no deflection, while the nanowire waveguide is close to the conventional waveguide, with a maximum offset of about 21nm. The change in the interval between the waveguides will cause the change in the effective refractive index. Figure 2 (c) is Figure 2 (b) The value of the effective refractive index along the waveguide direction when the waveguide is deflected. Furthermore, the change in the effective refractive index will cause the phase of the probe light transmitted in the waveguide to change, and the total amount of phase change is φ d It can be calculated by the following formula:

[0043]

[0044] where λ d is the wavelength of the detection light. Calculate different P t The phase change under Figure 2 (d). However, the change of light phase is difficult to detect, and in the photonic neural network, the output signal of the nonlinear activation function unit will be modulated by the weight matrix again as the input signal of the next layer of ONN. Therefore, it is necessary to convert the change of the detection light phase into the detection light power. By adding a reference waveguide and utilizing the interference of the detection light, the above effect can be easily achieved, as shown in the schematic diagram. Figure 3 As shown, assuming that the signal optical power is P s , the detection light power is P d , a micro-thermo-optic modulator is set in the reference waveguide, and the phase shift caused by the temperature change is assumed to be Then the output power P of the two detection lights after interference occurs at the Y branch out for:

[0045]

[0046] From equations (2) to (5), we can see that by adjusting the detection optical power P d It can effectively control the nonlinear activation function curve. Assuming the signal optical power P s Changing between 0 and 0.3 mW, α = 0.5, Figure 4 The nonlinear activation function curve graph implemented by the nonlinear operation unit. Figure 4 (a) shows different P d The nonlinear activation function curve achieved in this case. Similarly, the phase shift caused by regulating the micro-thermo-optic modulator It can also effectively adjust the nonlinear activation function curve. Figure 4 (b) shows P d =1.08mW, the nonlinear activation function curve follows It can be seen that without changing the chip structure, the nonlinear activation function curve can be flexibly adjusted to meet the requirements of different computing tasks and improve the chip calculation accuracy. In addition, the total input signal optical power required by the nonlinear activation function unit is only 0.3mW, which is much smaller than the power threshold of the existing nonlinear activation function scheme, and can significantly reduce the requirements for the ONN system light source.

[0047] The present invention uses a wavelength division multiplexing MRR array to implement the linear operation module in the ONN chip. Its structural diagram is shown in Figure 5 As shown in Figure 1, thermo-optical modulation is used to implement the weight operation in the linear operation unit. Assume that the input signal dimension is n and the weight matrix size is n×m. The input signal a i is loaded to a wavelength of λ i (i=1,2,…,n) signal light, through MRR ji (j=1,2,…,m) power modulation, the weight w ji Multiply it by the corresponding signal light. It is easy to see that the resonant wavelength of each column of MRR is the same, which is λ i The output signal of the jth row is Since the subsequent nonlinear operation unit needs the assistance of detection light to perform operations, a wavelength of λ is set in each row. pj The detection light and the corresponding MRR are also added to the MRR array pj In addition, the detection light does not perform any linear operations, so the corresponding MRR pj The power modulation coefficient w p Both are 1.

[0048] Combining the above two modules, taking a k-layer ONN network with n neurons in each layer as an example, the schematic diagram of the ONN chip structure of the present invention is as follows: Figure 6 As shown. The single layer of the all-optical reconfigurable silicon-based photonic neural network chip based on wavelength division multiplexing includes a light splitting area 1, a linear operation module 2, a nonlinear operation module 3, a light extraction device module 4 and a light combining area 5. The input light can be divided into two groups, one of which is the initial signal light with wavelengths of λ 10 ~λ n0 , loaded with the input signal of the ONN chip Another set of wavelengths is λ 11 ~λ nk , do not carry any information, and serve as the detection light for the 1st to kth layers respectively.

[0049] In the first layer, all the light is first divided into n paths. The wavelength of the detection light in the first path is λ 11 , is a resonant wavelength of the same 11 The MRR is downloaded to the waveguide below, and the power modulation coefficient Then, the initial signal light λ 10 ~λ n0 They are downloaded to the waveguide below by the corresponding MRR, and the power modulation coefficients are Implement the multiplication operation of the input signal and the weight (i.e. ). After that, the wavelength is λ 11 The probe light is MRR (the resonance wavelength is also λ 11 , through the thermo-optical modulator to achieve variable power modulation coefficient) is divided into two The upper branch transmits the detection light and signal light, and the lower branch transmits only the detection light. The two signals pass through the dual-waveguide suspension and single-waveguide suspension optical gradient force units respectively, and then are combined by the Y branch to realize the operation of the nonlinear activation function. At this time, the power of the detection light is the output signal of the first neuron in the first layer. In order to prevent the initial signal light from affecting the second layer network, a third resonant wavelength of λ is used. 11 MRR, download the detection light again And through the resonance wavelength λ 12 ~λ 1k The MRR will download the corresponding wavelength of the input (power modulation coefficient w (1) =1), and the Y branch and λ 11 The configuration of the 2nd to nth paths in the first layer is similar to that of the first path, and the detection light is λ 21 ,λ 31 ,…λ n1 , the signal light is also λ 10 ~λ n0 After the first layer, the wavelength is λ i1The detected light power is the output signal of the i-th neuron in the first layer.

[0050] The input wavelength of the second layer in ONN is λ 11 ~λ nk , where the wavelength is λ 11 ~λ n1 The light of becomes the signal light of this layer and is modulated by the corresponding MRR power to realize weighted operation. 12 ~λ n2 Become the probe light of the 1st to nth paths of the second layer (power modulation coefficient w (2) All are 1, In order to realize the nonlinear activation function operation of this layer. Similar to the first layer, after passing this layer, the wavelength is λ i2 The detected light power is the output signal of the i-th neuron in the second layer. The wavelength is λ 11 ~λ n1 The light will not enter the next layer. Similarly, the input light wavelength of the kth layer in ONN is λ 1(k-1) ~λ nk , where the wavelength is λ 1(k-1) ~λ n(k-1) The light becomes the signal light of this layer and is modulated by the corresponding MRR power to realize weighted operation. 1k ~λ pnk Become the detection light of the 1st to nth paths of the kth layer (power modulation coefficient w (k) All are 1, Its final output power is the output signal of the 1st to nth neurons in the kth layer. Finally, a photoelectric detector (PD) is used to convert the optical power into an electrical signal to complete the output of the ONN system.

[0051] In the ONN chip, in order to achieve reconfigurable ONN layers and the number of neurons in each layer, by adding optical switches and PDs at the output end of each layer, light can be extracted at any layer to change the number of ONN layers. The light extraction device unit is an MZI structure, in which a micro-thermal electrode is made on one arm, such as Figure 7As shown. By changing the voltage applied to the electrodes, light will enter either Channel 1 or Channel 2. A PD is fabricated in Channel 1. If light enters this channel, the reasoning process ends, and this layer is the last layer of the ONN network. Channel 2 is connected to the subsequent waveguides of the next layer network. If light enters this channel, it will continue to be input to the next layer network. By adjusting the power modulation coefficient of the corresponding MRR through a thermo-optic modulator, the number of neurons in each layer can be changed. For example, assume the number of neurons in the j-th layer is n. If you want to change it to q (q < n) neurons, then only need to set the from (q + 1) to n to 0, and at the same time keep also 0.

[0052] Use a conventional handwritten character recognition task (MNIST dataset) to verify the computing performance of this network. In the simulation, a 4-layer fully connected neural network structure is adopted, and its schematic diagram is as shown in Figure 8 (a). The number of neurons in the input layer is 784, the number of neurons in the first hidden layer is 12, the number of neurons in the second hidden layer is also 12, and the number of neurons in the output layer is 10. Among them, the weight range from the input layer to the first hidden layer is -1 to 1, and the non-linear activation function uses ReLU; the weight range from the first hidden layer to the second hidden layer is 0 to 1, and the non-linear activation function is generated by the optical gradient force unit; the weight range from the second hidden layer to the output layer is -1 to 1. The result finally obtained by this neural network is as shown in Figure 8 (b), and the calculation accuracy reaches 94%. It can be seen from the simulation results that the present invention has good results in small-scale neural networks and can break through the problem that the neuron scales of the linear operation module and the non-linear module based on wavelength division multiplexing do not match currently.

[0053] The principle of the present invention is as shown in Fig. 9 .

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the purpose and scope of the present technical solution, and they should all be covered by the scope of the claims of the present invention.

Claims

1. An all-optical reconfigurable silicon-based photonic neural network chip based on wavelength division multiplexing, wherein a single-layer network in the photonic neural network chip consists of a light splitting area, a linear operation module, a nonlinear operation module, a light extraction device module, and a light combining area; The light splitting area is used to realize multi-path light splitting of signal light and detection light; The nonlinear operation module is used to implement the operation of an adjustable nonlinear activation function; The light extraction device module is used to output the photonic neural network chip results and realize the regulation of the number of network layers; The light combining area is used to realize the multi-path combining of signal light and detection light to the next layer as the optical signal input of the network; Features: The nonlinear operation module is an array composed of a single or multiple waveguide suspended structures, a ring resonant cavity suspended structure, or a ring resonant cavity and a straight waveguide coupling region suspended structure; The nonlinear operation module uses an optical gradient force to perform nonlinear regulation of the multi-wavelength signal light power to the detection light phase; the optical gradient force structure is a double-waveguide suspended structure composed of two waveguides, one of which is a conventional waveguide and the other is a nanowire waveguide, and the two waveguides both adopt a double-end fixed beam structure; Calculate the light gradient force, the formula is as follows: Where F represents the light gradient force; l g represents the length of the suspended waveguide; P t Represents the total power of the light transmitted in the waveguide, including the multi-wavelength signal light power and the detection light power; n eff represents the effective refractive index of the dual-waveguide suspended structure as a whole; g represents the interval between the two waveguides; Using the Euler-Bernoulli beam theory and the optical gradient force, the deflection of the two waveguides is calculated as follows: Among them, u and u d represent the deflection of conventional waveguide and nanowire waveguide respectively; the z-axis direction represents the direction of light transmission, which is the direction of the waveguide; E = 1.71 × 10 11 , represents the Young's modulus of silicon material, w and w d represent the widths of conventional waveguide and nanowire waveguide, A and A respectively. d represent the cross-sectional areas of conventional waveguide and nanowire waveguide, respectively; Calculate the total amount of phase change φ of the probe light transmitted in the waveguide caused by the change in effective refractive index d , the formula is as follows: Among them, λ d is the wavelength of the detected light; calculate different P t The phase change under By adding a reference waveguide, the change of the detection light phase is converted into the detection light power by using the interference of the detection light. Let the signal light power be P s , the detection light power is P d , a micro-thermo-optic modulator is set in the reference waveguide, and the phase shift caused by the temperature change is assumed to be Then the output power P of the two detection lights after interference occurs at the Y branch out for: According to formulas (2) to (5), the detection optical power P is adjusted d To control the nonlinear activation function curve; or to control the phase shift caused by micro-thermo-optic modulator To adjust the nonlinear activation function curve; The linear operation module is used to realize the weight operation of signal light and regulate the number of single-layer neurons; The linear operation module in the ONN chip is realized by using the wavelength division multiplexing MRR array; the weight operation in the linear operation unit is realized by using thermal optical modulation; the input signal dimension is n, and the weight matrix size is n×m; the input signal a i Loaded to a wavelength of λ i On the signal light, i=1,2,…,n, through MRR ji , j = 1, 2, ..., m, power modulation, the weight w ji Multiply it by the corresponding signal light; the resonant wavelength of each column of MRR is the same, which is λ i ; Output signal of row j Set a wavelength λ in each row pj Probe light, which does not perform linear operations; corresponding MRRs are also added to the MRR array pj ;MRR pj The power modulation coefficient w p All are 1; An optical switch and a PD are added at the output end of each layer, and the light is led out at any layer to change the number of layers of the ONN; an optical lead-out device is used to implement path selection; the power modulation coefficient of the corresponding MRR is regulated by a thermo-optic modulator to change the number of neurons in each layer; let the number of neurons in the j-th layer be n, if the number of neurons needs to be changed to q, q < n, then the are all set to 0 and remain 0.

2. The all-optical reconfigurable silicon-based photonic neural network chip based on wavelength division multiplexing according to claim 1 is characterized in that: The light splitting area is a structure of a plurality of Y-branch power splitters connected in series, a plurality of 1×2-way multi-mode interference coupler MMI power splitters connected in series, or a 1×n-way MMI power splitter.

3. The all-optical reconfigurable silicon-based photonic neural network chip based on wavelength division multiplexing according to claim 1 is characterized in that: The linear operation module is a wavelength division multiplexing signal power modulation module implemented by a circular, racetrack-shaped, or special-shaped resonant cavity or grating array, and the basic unit connection mode in the array is series connection, parallel connection, or cascade series connection.

4. The all-optical reconfigurable silicon-based photonic neural network chip based on wavelength division multiplexing according to claim 1, characterized in that: The light extraction device module is an array composed of Mach-Zehnder interferometer (MZI) structures. Path selection is achieved by regulating the micro-thermal electrode on one arm of the MZI, and a photoelectric conversion device is prepared or connected at one of the output ports to achieve the output of the neural network chip.

5. The all-optical reconfigurable silicon-based photonic neural network chip based on wavelength division multiplexing according to claim 1 is characterized in that: The light combining area is a structure of a plurality of Y-branch power dividers connected in series, a plurality of 1×2-way MMI power dividers connected in series or a 1×n-way MMI power divider.

Images