Parameter generation and manufacturing methods for optical neural network chips

Abstract

This disclosure provides a method for generating parameters of an optical neural network chip, a chip manufacturing method, an apparatus, and a medium. The method for generating parameters of the optical neural network chip includes: setting basic chip parameters of the optical neural network chip and constructing a chip neural network model based on the basic chip parameters, wherein the optical neural network chip includes at least one optical feature quantity mutation unit for modulating optical feature quantities of light in the optical neural network chip; obtaining a reference function of the optical field of the optical neural network chip based on the basic chip parameters; obtaining an initial training set and mapping the initial training set to the optical field to obtain an optical training set; and training the chip neural network model using the optical training set to obtain chip feature parameters, wherein the chip feature parameters characterize the structural characteristics of at least one optical feature quantity mutation unit, and wherein the basic chip parameters and the chip feature parameters are used to manufacture the optical neural network chip.

Description

Technical Field

[0001] This disclosure relates to the field of semiconductor technology, and more specifically to a method for generating parameters of an optical neural network chip, a chip manufacturing method, and an apparatus, medium, and computer program product. Background Technology

[0002] With the continuous improvement of machine learning algorithms, neural networks are being applied more and more widely in various fields, and in many areas, their performance has surpassed the current limits of human capabilities, such as in computing power. However, because traditional CPUs (Central Processing Units) use the von Neumann architecture, they cannot efficiently run large-scale parallel computing processes. When neural networks handle more complex tasks, they require enormous computing resources. Therefore, finding faster and more energy-efficient alternative computing methods is of great significance. Optical neural networks offer a way to solve this problem because, compared with digital neural network computing, they have advantages such as ultra-fast speed, ultra-low energy consumption, and the ability to process signals in large-scale parallelism. As research on optical neural networks deepens, how to design and manufacture optical neural network chips capable of performing neural network computing has become a problem that needs to be solved. Summary of the Invention

[0003] Therefore, this disclosure provides a method for generating parameters of an optical neural network chip, a chip manufacturing method, and an apparatus, medium, and computer program product.

[0004] According to one aspect of the present disclosure, a method for manufacturing an optical neural network chip is provided, comprising: setting basic chip parameters of the optical neural network chip and constructing a chip neural network model based on the basic chip parameters, wherein the optical neural network chip includes at least one optical feature quantity mutation unit for modulating optical feature quantities of light in the optical neural network chip; obtaining a reference function of the optical field of the optical neural network chip based on the basic chip parameters, wherein the reference function is an optical field function when the optical feature quantity is zero; obtaining an initial training set and mapping the initial training set to the optical field to obtain an optical training set; training the chip neural network model using the optical training set to obtain chip feature parameters, wherein the chip feature parameters characterize the structural characteristics of the at least one optical feature quantity mutation unit; and manufacturing the optical neural network chip based on the basic chip parameters and the chip feature parameters.

[0005] According to an example of an embodiment of this disclosure, the basic parameters of the chip include one or more of the following: the basic size of the optical neural network chip, the number of the at least one optical feature mutation unit, the number and location of mutation structures included in each optical feature mutation unit, and the number and location of the output detectors.

[0006] According to an example of an embodiment of this disclosure, constructing a chip neural network model based on the chip's basic parameters includes: mapping the number of the at least one optical feature mutation unit to the number of hidden layers in the chip neural network model, and mapping one or more mutation structures in each optical feature mutation unit to neurons in the chip neural network model.

[0007] According to an example of an embodiment of this disclosure, obtaining the reference function of the light field of the optical neural network chip based on the chip's basic parameters includes: obtaining the reference function of the light field of the optical neural network chip by optical theory calculation or optical simulation based on the chip's basic parameters.

[0008] According to an example of an embodiment of this disclosure, the initial training set includes initial input values ​​and initial label values, and mapping the initial training set to the light field to obtain an optical training set includes: mapping the initial input values ​​of the initial training set to optical features of the light field as optical input values ​​of the optical training set; and mapping the initial label values ​​of the initial training set to the output optical power of the light field as optical label values ​​of the optical training set.

[0009] According to an example of an embodiment of this disclosure, training the chip neural network model using the optical training set to obtain chip feature parameters includes: inputting the optical input values ​​from the optical training set into the chip neural network model and calculating the output value of the chip neural network model; calculating a loss function between the output value and the optical label value; and training the chip neural network model by minimizing the loss function to obtain the chip feature parameters.

[0010] According to an example of an embodiment of this disclosure, training the chip neural network model by minimizing the loss function to obtain the chip feature parameters includes: training the chip neural network model by minimizing the loss function to obtain a chip optical feature quantity sequence; and mapping the chip optical feature quantity sequence to the chip feature parameters.

[0011] According to an example of an embodiment of this disclosure, manufacturing the optical neural network chip based on the chip's basic parameters and chip's characteristic parameters includes: depositing a silicon dioxide thin film on a substrate material as a lower cladding layer of the optical neural network chip based on the chip's basic parameters; depositing an optical waveguide material on the lower cladding layer as an optical waveguide layer of the optical neural network chip; forming a network structure of the optical neural network chip on the optical waveguide layer based on the chip's basic parameters and chip's characteristic parameters, the network structure including the at least one optical characteristic quantity mutation unit; and depositing a thin film material on the optical waveguide layer as an upper cladding layer of the optical neural network chip.

[0012] According to an example of an embodiment of this disclosure, the network structure of the optical neural network chip formed on the optical waveguide layer based on the chip's basic parameters and chip characteristic parameters includes: sequentially forming an input coupler, an input straight waveguide, an input width-gradient waveguide, a wide waveguide, at least one optical characteristic quantity abrupt change unit, an edge anti-reflection module, an output interface anti-reflection module, an output width-gradient waveguide, an output straight waveguide, and an output coupler on the optical waveguide layer, wherein light entering the optical neural network chip propagates sequentially through the input coupler, the input straight waveguide, the input width-gradient waveguide, the wide waveguide, at least one optical characteristic quantity abrupt change unit, the edge anti-reflection module, the output interface anti-reflection module, the output width-gradient waveguide, the output straight waveguide, and the output coupler.

[0013] According to an example of an embodiment of this disclosure, the manufacturing method further includes: forming an optical feature modulator on the upper cladding of the optical neural network chip, the optical feature modulator being used to modulate the optical feature of initial light entering the optical neural network chip through the input coupler, so as to modulate the initial light into input light carrying the input data information by mapping the input data information to the optical feature of the initial light.

[0014] According to an example of an embodiment of this disclosure, the edge anti-reflection module is used to absorb light propagating to the side section of the wide waveguide of the optical neural network chip, and the output interface anti-reflection module is used to absorb light reflected back to the wide waveguide from the output interface of the wide waveguide other than the output width gradient waveguide.

[0015] According to an example of an embodiment of this disclosure, the optical characteristic quantity includes one of the light amplitude, phase, polarization, frequency, and angular momentum.

[0016] According to an example of an embodiment of this disclosure, the at least one optical feature quantity mutation unit is a rectangular groove array, and the chip feature parameter is the size of each rectangular groove in the rectangular groove array.

[0017] According to another aspect of the present disclosure, a method for generating parameters of an optical neural network chip is provided, comprising: setting basic chip parameters of the optical neural network chip and constructing a chip neural network model based on the basic chip parameters, wherein the optical neural network chip includes at least one optical feature quantity mutation unit for modulating optical feature quantities of light in the optical neural network chip; obtaining a reference function of the optical field of the optical neural network chip based on the basic chip parameters, wherein the reference function is an optical field function when the optical feature quantity is zero; obtaining an initial training set and mapping the initial training set to the optical field to obtain an optical training set; training the chip neural network model using the optical training set to obtain chip feature parameters, wherein the chip feature parameters characterize the structural characteristics of the at least one optical feature quantity mutation unit, wherein the basic chip parameters and the chip feature parameters are used to manufacture the optical neural network chip.

[0018] According to another aspect of the present disclosure, a method for manufacturing an optical neural network chip is provided, comprising: depositing a silicon dioxide thin film on a substrate material as a lower cladding layer of the optical neural network chip based on chip basic parameters; depositing an optical waveguide material on the lower cladding layer as an optical waveguide layer of the optical neural network chip, and forming a network structure of the optical neural network chip on the optical waveguide layer based on the chip basic parameters and chip characteristic parameters, the network structure including at least one optical characteristic abrupt change unit; and depositing a thin film material on the optical waveguide layer as an upper cladding layer of the optical neural network chip, wherein the chip basic parameters and the chip characteristic parameters are generated according to the parameter generation method of the optical neural network chip as described above.

[0019] According to another aspect of the present disclosure, a parameter generation apparatus for an optical neural network chip is provided, comprising: a setting unit configured to set basic chip parameters of the optical neural network chip and construct a chip neural network model based on the basic chip parameters, wherein the optical neural network chip includes at least one optical feature quantity mutation unit for modulating optical feature quantities of the optical neural network chip; an acquisition unit configured to obtain a reference function of the optical field of the optical neural network chip based on the basic chip parameters, wherein the reference function is an optical field function when the optical feature quantity is zero, acquire an initial training set, and map the initial training set to the optical field to obtain an optical training set; and a training unit configured to train the chip neural network model using the optical training set to obtain chip feature parameters, wherein the chip feature parameters characterize the structural characteristics of the at least one optical feature quantity mutation unit, wherein the basic chip parameters and the chip feature parameters are used to manufacture the optical neural network chip.

[0020] According to another aspect of the present disclosure, an optical neural network chip is provided, which is manufactured according to the manufacturing method of the optical neural network chip as described in the above aspects.

[0021] According to another aspect of the present disclosure, a computer-readable storage medium is provided that stores computer-readable instructions thereon, which, when executed by a processor, cause the processor to perform the method described in any one of the foregoing aspects.

[0022] According to another aspect of the present disclosure, a computer program product is provided, which includes computer-readable instructions that, when executed by a processor, cause the processor to perform the method described in any of the foregoing aspects.

[0023] By utilizing the parameter generation method, manufacturing method, apparatus, medium, and computer program product of the optical neural network chip according to the above aspects of this disclosure, the optical neural network chip can be designed based on neural network training. Furthermore, based on the chip's basic parameters and the chip's characteristic parameters obtained through training, the optical neural network chip can be manufactured using traditional chip processing technology. This process is simple, low-cost, and can be mass-produced. Compared to traditional CPU-based digital neural networks, the resulting optical neural network chip has advantages such as ultra-low power consumption, ultra-fast speed, and the ability to perform large-scale parallel processing when performing neural network calculations. Attached Figure Description

[0024] The above and other objects, features, and advantages of the present disclosure will become more apparent from the more detailed description of the embodiments thereof in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of the present disclosure and form part of the specification. They are used together with the embodiments of the present disclosure to explain the disclosure and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0025] Figure 1 A flowchart illustrating a parameter generation method for an optical neural network chip according to an embodiment of the present disclosure is shown.

[0026] Figure 2 A flowchart illustrating the training process of an example chip neural network model according to an embodiment of this disclosure is shown.

[0027] Figure 3 A flowchart illustrating a method for manufacturing an optical neural network chip according to an embodiment of the present disclosure is shown.

[0028] Figure 4 A schematic diagram of the structure of an optical neural network chip according to an embodiment of the present disclosure is shown.

[0029] Figure 5 A flowchart illustrating a method for manufacturing an optical neural network chip according to another embodiment of this disclosure is shown.

[0030] Figure 6 A schematic diagram of the structure of a parameter generation apparatus for an optical neural network chip according to an embodiment of the present disclosure is shown. Detailed Implementation

[0031] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without creative effort are within the scope of protection of this disclosure.

[0032] Because traditional central processing units (CPUs) are not ideal for implementing deep neural network algorithms, academia and industry have begun to focus on developing new hardware architectures suitable for artificial neural networks and deep learning applications. Optical neural network chips are an important research direction in this field. The present disclosure addresses the problem of how to design and manufacture an optical neural network chip that utilizes the propagation and detection of light within the chip to perform neural network calculations such as deep learning, while overcoming the high power consumption and slow speed of traditional digital neural networks.

[0033] Taking classic classification algorithms in machine learning as an example, existing digital neural networks based on traditional CPUs can already skillfully complete classification tasks on large-scale datasets. However, to enable optical neural network chips to perform such classification tasks, it is necessary to map the abstract data attributes of the dataset to the physical propagation characteristics of light, and to specifically design and train the structural characteristics and manufacturing parameters of the optical neural network chip. Therefore, this disclosure first provides a parameter generation method for optical neural network chips, which obtains the parameters used to manufacture the optical neural network chip based on neural network training.

[0034] Figure 1 A flowchart of a parameter generation method 100 for an optical neural network chip according to an embodiment of the present disclosure is shown. To generate parameters for the optical neural network chip through neural network training, a chip neural network model needs to be constructed. Therefore, some basic parameters of the chip need to be set first. In step 110, the basic parameters of the optical neural network chip are set. These basic parameters may include the basic dimensions of the optical neural network chip, such as the chip's total length, total width, and total thickness, as well as the dimensional parameters of chip structures such as input couplers and output couplers.

[0035] As described above, in order for the optical neural network chip based on the embodiments of this disclosure to perform machine learning calculations on large-scale datasets, it is first necessary to map the abstract data attributes of the dataset to the physical propagation characteristics of light in the chip. For example, for the Iris (IRIS) dataset commonly used in classification tasks, it contains 150 data samples of irises, with 50 data samples taken from one of three iris species. That is, these data samples can be divided into three categories, with 50 data samples in each category. Each data sample contains four attributes: sepal length, sepal width, petal length, and petal width. Using these four attributes, it is possible to predict which of the three categories the data sample belongs to.

[0036] In this embodiment of the disclosure, if an optical neural network chip is to be used to classify the IRIS dataset, the four attribute values ​​of each data sample in the IRIS dataset need to be mapped to the physical propagation characteristics of light. That is, the propagation characteristics of light are used to characterize the four attributes of each data sample in the IRIS dataset. Optical features that can reflect the physical propagation characteristics of light may include, for example, the amplitude, phase, polarization, frequency, angular momentum, and any other features of light. This embodiment of the disclosure does not impose specific limitations on these. Taking the phase feature of light as an example, for instance, the four attributes of each data sample in the IRIS dataset can be mapped to the phases of four light beams respectively. Then, the calculation of the IRIS dataset can be correspondingly mapped to the modulation of the phase of light in the chip. This leads to the key structure of the optical neural network chip—the optical feature mutation unit.

[0037] In this embodiment, the optical characteristic quantity mutation unit is used to modulate the optical characteristic quantity of light in the optical neural network chip. When light propagates through the specially designed optical characteristic quantity mutation unit, its optical characteristic quantity changes, thereby causing the distribution of light output power detected at the chip's output terminal to differ. For example, taking the optical characteristic quantity as the optical phase, the optical characteristic mutation unit can be called the optical phase mutation unit, and the phase of light changes when it propagates through the optical phase mutation unit.

[0038] An optical neural network chip may include at least one optical feature quantity unit, and each optical feature quantity mutation unit may consist of multiple mutation structures. These mutation structures may be, for example, grooves, protrusions, or any other shape; this disclosure does not impose specific limitations on this. Taking a groove as an example, each optical feature quantity mutation unit may be a groove array, such as a rectangular groove array or a groove array of any other shape; this disclosure does not impose specific limitations on this. The structural characteristics of the optical feature quantity unit are a key factor affecting optical feature quantity modulation; in other words, they are a key factor affecting optical neural network computation. Therefore, in this disclosure, it is necessary to train the neural network based on the structural characteristics of the optical feature quantity unit to ensure that the optical feature quantity unit has optimal structural characteristics, in order to obtain an optical neural network chip capable of optimally performing optical neural network computation.

[0039] In the examples of the embodiments of this disclosure, the structural characteristics of at least one optical characteristic quantity mutation unit are characterized by chip feature parameters. For example, the size of each mutation structure in the optical characteristic quantity mutation unit can be characterized. Taking a rectangular groove as an example, the length or width of the rectangular groove can be characterized. However, the embodiments of this disclosure are not limited to this. Chip feature parameters can also characterize other structural characteristics of the optical characteristic quantity mutation unit, such as the number and positional distribution of mutation structures.

[0040] According to an example of an embodiment of this disclosure, when setting the basic parameters of an optical neural network chip, the number of optical feature quantity mutation units in the chip, as well as the number and position of mutation structures included in each optical feature quantity mutation unit, can be preset. The size of each mutation structure in the optical feature quantity mutation unit can be obtained through neural network training, and the size of each mutation structure is the desired chip feature parameter. However, the embodiments of this disclosure are not limited to this. The number of optical feature quantity mutation units in the chip, as well as the size of the mutation structures included in each optical feature quantity mutation unit, can also be preset, and the number and position of mutation structures in each optical feature quantity mutation unit can be obtained through neural network training. In this case, the number and position of mutation structures are the desired chip feature parameters, and so on.

[0041] Furthermore, in this embodiment of the present disclosure, the pre-set basic parameters of the optical neural network chip may also include the number and position of the output detectors. The output detectors are used to detect the light output power at the output end of the chip, and the results of optical neural network calculations can be obtained based on the detection results, such as the classification results of input data samples.

[0042] In step S110, after setting the basic parameters of the optical neural network chip, a chip neural network model can be constructed based on the basic parameters. In this embodiment of the present disclosure, the number of optical feature mutation units in the optical neural network chip can be mapped to the number of hidden layers in the chip neural network model, and one or more mutation structures in each optical feature mutation unit can be mapped to neurons in the chip neural network model.

[0043] In one example, the optical neural network chip may be pre-configured to include three optical feature mutation units, and the corresponding chip neural network model may include three hidden layers. However, the embodiments disclosed herein are not limited to this. The optical neural network chip may include more or fewer optical feature mutation units, and the chip neural network model may include more or fewer hidden layers.

[0044] In one example, multiple mutation structures in each optical feature mutation unit can be mapped to a neuron in the chip neural network model. For example, in the case where the mutation structure is a rectangular groove, three rectangular grooves in each optical feature mutation unit can be mapped to a neuron in the chip neural network model. However, the embodiments of this disclosure are not limited to this, and more or fewer mutation structures in the optical feature mutation unit can be mapped to a neuron.

[0045] In step S120, a reference function for the optical field of the optical neural network chip is obtained based on the set basic chip parameters. Here, the reference function for the optical field of the optical neural network chip refers to the optical field function when the optical characteristic quantity is zero. According to the examples of embodiments of this disclosure, the reference function for the optical field of the optical neural network chip can be obtained based on the basic chip parameters through optical theory calculations or optical simulations. For example, the reference function for the chip optical field can be obtained by using a chip space diffraction light field propagation model based on the Rayleigh-Sommerfeld formula through optical theory calculations or optical simulations. However, embodiments of this disclosure are not limited to this, and the reference function for the chip optical field can also be obtained by any other suitable method.

[0046] Next, in step S130, an initial training set is obtained and mapped to a light field to obtain an optical training set. Here, the initial training set can be any dataset used for neural network computation, which includes initial input values ​​and initial label values ​​for result comparison, such as the IRIS dataset described above. This disclosure does not impose specific limitations on this.

[0047] As mentioned earlier, in order for the optical neural network chip according to the embodiments of this disclosure to perform neural network calculations, it is necessary to map the abstract data attributes of the dataset to the physical propagation characteristics of light in the chip. Therefore, in order to train the chip neural network model, the traditional initial training set must first be converted into an optical training set that reflects optical characteristics. Specifically, according to the examples of the embodiments of this disclosure, the initial input values ​​of the initial training set can be mapped to the optical features of the light field as the optical input values ​​of the optical training set; and the initial label values ​​of the initial training set can be mapped to the output optical power of the light field as the optical label values ​​of the optical training set. For example, when the initial training set is the IRIS dataset and the optical features are phase, the four attribute values ​​of each data sample in the IRIS dataset can be mapped to the four phase values ​​of the light field, and the label value of each data sample can be mapped to the optical output power of the light field, thereby obtaining an optical training set for training the chip neural network model.

[0048] After obtaining the optical training set, in step S140, the chip neural network model is trained using the optical training set to obtain chip feature parameters, wherein the chip feature parameters characterize the structural characteristics of at least one optical feature quantity mutation unit.

[0049] Specifically, in this step, for each optical input value in the optical training set, the optical input value is input into the chip neural network model, and the output value of the chip neural network model is calculated accordingly. For example, the output optical power can be calculated using a chip spatial diffraction light field propagation model based on the Rayleigh-Sommerfeld formula, and this output value is used as the output value of the chip neural network model. Then, the loss function between each output value and its corresponding optical label value is calculated. For example, the mean square error (MSE) and normalized mean square error (NMSE) between the output value and the optical label value can be calculated as the loss function. Afterwards, the chip neural network model is trained by minimizing the loss function to obtain chip feature parameters. For example, a predetermined threshold can be set, and training can end when the loss function is less than the predetermined threshold; or, a maximum number of iterations can be set, and training can end when the maximum number of iterations is reached, etc. This disclosure embodiment does not impose specific limitations in this regard.

[0050] After training, a sequence of chip optical features that results in a loss function less than a predetermined threshold or reaches the maximum number of iterations can be output. This sequence is then mapped to chip feature parameters, thus obtaining the optimal parameters characterizing the structural properties of the optical feature mutation units. Taking the optical features as phases, the mutation structure of the optical feature mutation unit as rectangular grooves, and the chip feature parameters as the dimensions of each rectangular groove as an example, a phase sequence can be output after training. This phase sequence includes the output phase value corresponding to each rectangular groove. Then, the following formula can be used to map each phase value in the output phase sequence to the corresponding rectangular groove size:

[0051]

[0052] Among them, L i L represents the size of the i-th rectangular groove. For example, given a preset width for each rectangular groove, L... i L can represent the length of the i-th rectangular groove, or, given that the lengths of each rectangular groove are preset, L i This can represent the width of the i-th rectangular groove, but the embodiments disclosed herein do not impose specific limitations on this. The phase delay is generated by the i-th rectangular groove, which can be determined based on the output phase value and input phase value corresponding to the i-th rectangular groove; n eff The effective refractive index of the unit cell representing the abrupt change in optical characteristic quantities; n slab This represents the effective refractive index of the wide waveguide of the chip where the optical characteristic quantity mutation unit is located. k0 = 2π / λ is the wavenumber of light propagating through the optical characteristic quantity mutation unit, where λ represents the wavelength of the input light source, which can be, for example, 1550nm, or any other suitable wavelength. This disclosure does not impose specific limitations on this.

[0053] After determining the chip's characteristic parameters using the methods described above, optical neural network chips can be manufactured based on the chip's basic parameters and characteristic parameters, as will be described in further detail below. To make the above steps clearer, the following will combine... Figure 2 The training process of the chip neural network model is further described.

[0054] Figure 2 A flowchart 200 illustrates the training process of an example chip neural network model according to an embodiment of this disclosure. Figure 2As shown, in step S210, parameter settings are first performed. In this step, the basic parameters of the optical neural network chip can be set, as described in step S110 above, and will not be repeated here. Furthermore, in this step, the initial training parameters of the chip's neural network model can also be set. For example, the type of loss function can be set, such as mean squared error (MSE), normalized mean squared error (NMSE), etc.; a predetermined threshold for the loss function can be set, for example, a predetermined threshold of 0; the maximum number of iterations can be set, for example, a maximum number of iterations of 3000; the type of optimizer can be set, for example, an Adam optimizer, a cross-entropy optimizer, etc.; and other initial parameters required for neural network training, etc. This embodiment of the disclosure does not impose specific limitations on these aspects.

[0055] In step S220, the training set is input into the chip neural network model. Here, the training set can be, for example, the optical training set obtained in step S130 above. In step S230, for each data sample in the training set, the output value of the chip neural network model is calculated. For example, the output optical power can be calculated using a chip-space diffraction light field propagation model based on the Rayleigh-Sommerfeld formula, and used as the output value of the chip neural network model. In step S240, it is determined whether all data samples in the training set have been traversed. If the determination result is no, the training set is traversed again and the output value of each data sample is calculated; if the determination result is yes, the next step is performed. In step S250, the loss function is calculated. Specifically, based on the output value of each data sample in the training set and its corresponding label value, the loss function of the training set is calculated. For example, the NMSE between the output value and the label value of each data sample can be calculated as the loss function. In step S260, it is determined whether the loss function is below a predetermined threshold or whether the maximum number of iterations has been reached. If the determination result is negative, the gradient is calculated based on the loss function to update the optical feature quantities (e.g., phase) of each neuron in the neural network, and the process returns to step S230 (step S290). If the determination result is positive, the iteration can be terminated, the optical feature quantity sequence is output (step S270), and the trained chip neural network model can be tested using a test set (step S280). At this point, the training of the chip neural network model is completed. The output optical feature quantity sequence can be further mapped to chip feature parameters that characterize the structural properties of the optical feature quantity mutation unit, for example, using the above formula (1) to map it to the size of each rectangular groove in the rectangular groove array, for use in manufacturing optical neural network chips.

[0056] Using the parameter generation method for optical neural network chips according to the above embodiments of the present disclosure, optical neural network chips can be designed based on neural network training. Based on the basic chip parameters and chip feature parameters obtained by the method, optical neural network chips can be manufactured using traditional chip processing technology. The manufacturing process is simple, low-cost, and can be mass-produced. When performing neural network calculations, the resulting optical neural network chips have the advantages of ultra-low power consumption, ultra-fast speed, and large-scale parallel processing compared to traditional CPU-based digital neural networks.

[0057] The following reference Figure 3 A method for manufacturing an optical neural network chip according to an embodiment of the present disclosure is described. Figure 3 A flowchart of a method 300 for manufacturing an optical neural network chip according to an embodiment of the present disclosure is shown.

[0058] In step S310, a silicon dioxide thin film is deposited on the substrate material as the lower cladding layer of the optical neural network chip based on the chip's basic parameters. Here, the chip's basic parameters can be parameters preset during the generation of the optical neural network chip parameters, as described in detail above with reference to step S110. According to an example of an embodiment of this disclosure, the substrate material can be silicon, or any other material on which a silicon dioxide thin film can easily be grown; this disclosure does not impose specific limitations on this. The thickness of the silicon dioxide thin film can be, for example, 5 micrometers, or any other suitable thickness; this disclosure does not impose specific limitations on this. Alternatively, if the silicon dioxide thin film serving as the lower cladding layer is sufficiently thick, the silicon dioxide thin film can be directly used as the substrate, omitting the process step of growing the silicon dioxide thin film on the substrate material.

[0059] In step S320, an optical waveguide material is deposited on the lower cladding layer as the optical waveguide layer of the optical neural network chip. Based on the chip's basic parameters and characteristic parameters, a network structure for the optical neural network chip is formed on the optical waveguide layer. This network structure refers to the chip structure on the optical neural network chip capable of performing optical neural network calculations. The network structure may include at least one optical characteristic quantity mutation unit. According to an example of an embodiment of this disclosure, the optical waveguide material may be a material with a higher refractive index than silicon dioxide and high light transmittance, such as silicon, silicon nitride (Si3N4), or any other suitable material, to minimize light loss during propagation in the optical waveguide material. The thickness of the optical waveguide material may be, for example, 220 nanometers, or any other suitable thickness; this embodiment of the disclosure does not impose specific limitations on this. The chip characteristic parameters may be obtained through the parameter generation method 100 for the optical neural network chip as described above, and characterize the structural characteristics of at least one optical characteristic quantity mutation unit. For example, it may characterize the size of each rectangular groove in the rectangular groove array constituting the optical characteristic quantity mutation unit.

[0060] In this step, based on the obtained chip feature parameters, such as the size of each rectangular groove in the rectangular groove array that constitutes each optical feature quantity mutation unit, a network structure including at least one optical feature quantity mutation unit can be formed in the optical waveguide layer using conventional semiconductor processing techniques such as cleaning, coating, pre-baking, exposure, development, hardening, etching, and photoresist removal. That is, a chip structure capable of optical neural network calculation is formed in the optical waveguide layer of the optical neural network chip.

[0061] In step S330, a thin film material is deposited on the optical waveguide layer as the upper cladding (or protective layer) of the optical neural network chip. According to an example of an embodiment of this disclosure, the thin film material can be a material with a higher refractive index than the optical waveguide material and high light transmittance, such as silicon dioxide or any other suitable material, to minimize light loss during propagation; the thickness of the thin film material can be, for example, 2 micrometers, or any other suitable thickness, and this disclosure does not impose specific limitations on this.

[0062] According to examples of embodiments of this disclosure, the optical neural network chip may further include more structures to realize various processes such as light input, propagation, and output within the chip. The following describes... Figure 4 The manufacturing method 300 for the optical neural network chip is further described in detail. Figure 4 A schematic diagram of the structure of an optical neural network chip 400 obtained by method 300 according to an example embodiment of this disclosure is shown. Figure 4 As shown, the optical neural network chip 400 may include a substrate material 410, a lower cladding layer 420, an optical waveguide layer 430, and an upper cladding layer 440, wherein at least one optical characteristic abrupt change unit 4305 is formed on the optical waveguide layer 430.

[0063] According to an example of an embodiment of this disclosure, the network structure of an optical neural network chip formed on an optical waveguide layer based on basic chip parameters and chip characteristic parameters may include an input coupler 4301, an input straight waveguide 4302, an input width-gradient waveguide 4303, a wide waveguide 4304, at least one optical characteristic quantity abrupt change unit 4305, an edge anti-reflection module 4306, an output interface anti-reflection module 4307, an output width-gradient waveguide 4308, an output straight waveguide 4309, and an output coupler 4310 sequentially formed on the optical waveguide layer 430. The input coupler 4301 is used to couple light from a light source (e.g., a laser source) into the optical neural network chip 400. For example, it can couple light horizontally into the chip 400 or vertically into the chip 400, and this embodiment does not impose any specific limitations on this. The input straight waveguide 4302 is used to conduct light coupled into the chip 400 through the input coupler 4301. The output straight waveguide 4309 is used to conduct light coupled out of the chip 400 through the output coupler 4310.

[0064] As mentioned earlier, before performing neural network calculations using the optical neural network chip 400, it is necessary to convert the data attributes of the traditional dataset into the physical propagation characteristics of light. For example, the four attribute values ​​of each data sample in the IRIS dataset are converted into four optical phase values. To achieve this, the manufacturing method 300 of the optical neural network chip also includes forming an optical feature modulator 4401 on the upper cladding 440 of the optical neural network chip. The optical feature modulator 4401 is used to modulate the optical feature quantities of the initial light entering the optical neural network chip 400 through the input coupler 4301, so as to modulate the initial light into input light carrying the input data information by mapping the input data information to the optical feature quantities of the initial light.

[0065] For example, when the optical feature quantity is phase, the optical feature quantity modulator 4401 can modulate the phase of the initial light to map the input data information to the phase of the initial light, thereby modulating the initial light into input light carrying the input data information. Taking the input data as an IRIS dataset and the optical feature quantity as phase as an example, each input data sample includes four attribute values. The optical feature quantity modulator 4401 can map the four attribute values ​​of each input data sample to the phase of four initial beams entering the chip 400 through the input coupler 4301, thereby modulating these four initial beams into input beams carrying the input data information. Figure 4 The diagram schematically shows that the input coupler 4301 includes four branches, which can couple four initial beams into the chip 400 respectively. However, this is only an example, and the input coupler 4301 may include more or fewer branches depending on the actual characteristics of the input data. This disclosure does not impose any specific limitations on this.

[0066] According to an example of an embodiment of this disclosure, in order to form an optical characteristic modulator 4401, a specific electrode can be fabricated on the upper cladding 440 directly above the input straight waveguide 4302 using conventional semiconductor processing techniques. For example, a metal compound material such as titanium nitride can be used to fabricate a thermo-optical device electrode as the optical characteristic modulator 4401. This optical characteristic modulator 4401 formed by the thermo-optical device electrode can modulate optical characteristic quantities (e.g., the phase of light) based on the thermo-optical effect. Alternatively, conductive leads can be fabricated on the upper cladding 440 to form an electro-optical device electrode as the optical characteristic modulator 4401. This optical characteristic modulator 4401 formed by the electro-optical device electrode can modulate optical characteristic quantities (e.g., the phase of light) based on changes in voltage or electric field.

[0067] According to an example of an embodiment of this disclosure, the input straight waveguide 4302 and the output straight waveguide 4309 enable light to propagate in a single mode, such as transverse electromagnetic O (TE0) mode light, thereby maximizing the coupling and propagation efficiency of the light. The input width-gradient waveguide 4303 is used to convert the single-mode light into multi-mode light on the light input side, for example, from TE0 mode light to approximately planar light, so that the light entering the optical feature mutation unit 4305 is multi-mode light. This allows for a more accurate mapping from at least one optical feature mutation unit to the neural network, further improving the accuracy of the optical neural network. Correspondingly, the output width-gradient waveguide 4308 is used to convert the multi-mode light into a single-mode light on the light output side, for example, from approximately planar light to TE0 mode light, so that the light continues to propagate in single-mode light in the output straight waveguide 4309 and is finally coupled out of the chip 400 through the output coupler 4310.

[0068] According to an example of an embodiment of this disclosure, at least one optical characteristic abrupt change unit 4305 is formed on a wide waveguide 4304, therefore the width of the wide waveguide 4304 depends on the width of the optical characteristic abrupt change unit 4305. After obtaining the basic chip parameters and chip characteristic parameters through the method 100 described above, the structural characteristics of the optical characteristic abrupt change unit 4305 can be obtained, such as the dimensions of each rectangular groove in the rectangular groove array of each optical characteristic abrupt change unit 4305, thereby determining the width of the wide waveguide 4304. Figure 4 The diagram schematically shows three optical feature quantity mutation units 4305, each of which is a rectangular groove array. However, this is merely an example. The optical neural network chip 400 may also include more or fewer optical feature quantity mutation units, and each optical feature quantity mutation unit may be any other shape or structure. This disclosure does not impose any specific limitations on this.

[0069] According to an example of an embodiment of this disclosure, the edge anti-reflection module 4306 is used to absorb light propagating to the side section of the wide waveguide 4304 of the optical neural network chip 400; the output interface anti-reflection module 4307 is used to absorb light reflected back to the wide waveguide 4304 from the output interface of the wide waveguide 4304, excluding the output width gradient waveguide 4308. That is, it ensures that light emitted from the output interface of the wide waveguide 4304 only enters the output width gradient waveguide 4308 and is not reflected back to the wide waveguide 4304 from other parts of the output interface. Therefore, the edge anti-reflection module 4306 and the output interface anti-reflection module 4307 can prevent light reflected back into the chip from the side section of the wide waveguide 4304 and the output interface (excluding the output width gradient waveguide 4308) from causing noise interference to the entire optical neural network system.

[0070] According to an example of an embodiment of this disclosure, light is coupled into the optical neural network chip 400 through the input coupler 4301, then enters the input straight waveguide 4302, and carries input data information under the modulation of the optical feature modulator 4401. It then propagates sequentially through the input width-gradient waveguide 4303, the wide waveguide 4304, at least one optical feature mutation unit 4305, the edge anti-reflection module 4306, the output interface anti-reflection module 4307, the output width-gradient waveguide 4308, the output straight waveguide 4309, and the output coupler 4310. When light passes through at least one optical feature mutation unit 4305, the optical feature changes under the action of a mutation structure such as a rectangular groove array. For example, if the optical feature is phase, a phase change occurs, ultimately altering the distribution of the output optical power.

[0071] According to an example of an embodiment of this disclosure, the output coupler 4310 may include one or more branches, and at each branch of the output coupler 4310, one or more output detectors for detecting the output power of the light may be arranged. Figure 4 (Not shown in the image), for example, a photodetector. By analyzing the light output power detected by each output detector, the computational results of the optical neural network chip can be obtained. For example, in the case where the optical neural network chip 400 performs a classification task on the IRIS dataset, the output coupler 4310 can include three branches, and an output detector can be arranged at each branch of the output coupler 4310. By comparing the intensity of the light output power detected by the three output detectors, the data samples in the IRIS dataset are classified, that is, each data sample is classified into one of the three categories of iris. It should be noted that in Figure 4In the present invention, the output coupler 4310 is shown as having three branches, but this is merely an example. The output coupler 4310 may include more or fewer branches depending on the actual application. This disclosure does not impose any specific limitations on this.

[0072] Reference above Figure 3 and Figure 4 A method 300 for manufacturing an optical neural network chip has been described. Using the manufacturing method 300 according to embodiments of this disclosure, an optical neural network chip can be manufactured using conventional chip fabrication processes. This method is simple to manufacture, low in cost, and can be mass-produced. Compared to traditional CPU-based digital neural networks, the resulting optical neural network chip exhibits advantages such as ultra-low power consumption, ultra-high speed, and the ability to perform large-scale parallel processing when performing neural network calculations.

[0073] The following reference Figure 5 A method for manufacturing an optical neural network chip according to another embodiment of the present disclosure will be described. Figure 5 A flowchart of a method 500 for manufacturing an optical neural network chip according to another embodiment of this disclosure is shown. The details of the steps of method 500 are the same as those described above. Figure 1 and Figure 2 Method 100 described, and references Figure 3 and Figure 4 The steps of method 300 are similar in detail, so for simplicity, repeated descriptions of the same content are omitted here.

[0074] like Figure 5 As shown, in step S510, the basic parameters of the optical neural network chip are set, and a chip neural network model is constructed based on the basic parameters. The optical neural network chip includes at least one optical feature quantity mutation unit for modulating the optical feature quantities of light in the optical neural network chip. In step S520, a reference function of the optical field of the optical neural network chip is obtained based on the set basic parameters, wherein the reference function is the optical field function when the optical feature quantity is zero. In step S530, an initial training set is obtained and mapped to the optical field to obtain an optical training set. In step S540, the chip neural network model is trained using the optical training set to obtain chip feature parameters, wherein the chip feature parameters characterize the structural characteristics of at least one optical feature quantity mutation unit. In step S550, the optical neural network chip is manufactured based on the basic parameters and the chip feature parameters.

[0075] The following reference Figure 6 A parameter generation apparatus for an optical neural network chip according to an embodiment of the present disclosure is described. Figure 6 A schematic diagram of the structure of a parameter generation apparatus 600 for an optical neural network chip according to an embodiment of the present disclosure is shown. Figure 6 As shown, the device 600 includes a setting unit 610, an acquisition unit 620, and a training unit 630. Besides these three units, the device 600 may also include other components; however, since these components are not relevant to the embodiments of this disclosure, their illustrations and descriptions are omitted here. Furthermore, since the function of the device 600 is the same as described above... Figure 1 and Figure 2 The details of the steps in method 100 are similar, so for simplicity, repeated descriptions of some parts are omitted here.

[0076] The setting unit 610 is configured to set the basic parameters of the optical neural network chip and construct a chip neural network model based on the basic parameters. The optical neural network chip includes at least one optical feature mutation unit for modulating the optical feature quantities of the optical neural network chip. The basic parameters may include the basic dimensions of the optical neural network chip, such as the total length, total width, and total thickness of the chip, as well as the dimensional parameters of chip structures such as input couplers and output couplers. The basic parameters may also include the number of optical feature mutation units in the optical neural network chip, and the number and location of mutation structures included in each optical feature mutation unit. Furthermore, the basic parameters may include the number and location of output detectors, which are used to detect the optical output power at the chip's output end. Based on the detection results, the results of the optical neural network calculations can be obtained, such as the classification results of input data samples.

[0077] According to an example of an embodiment of this disclosure, the setting unit 610 is further configured to map the number of optical feature mutation units in the optical neural network chip to the number of hidden layers in the chip neural network model, and to map one or more mutation structures in each optical feature mutation unit to neurons in the chip neural network model.

[0078] In one example, the setting unit 610 may pre-set the optical neural network chip to include 3 optical feature mutation units, and the corresponding chip neural network model includes 3 hidden layers. However, the embodiments of this disclosure are not limited to this. The optical neural network chip may include more or fewer optical feature mutation units, and the chip neural network model may include more or fewer hidden layers.

[0079] In one example, the setting unit 610 can map multiple mutation structures in each optical feature mutation unit to a neuron in the chip neural network model. For example, in the case where the mutation structure is a rectangular groove, three rectangular grooves in each optical feature mutation unit can be mapped to a neuron in the chip neural network model. However, the embodiments of this disclosure are not limited to this, and more or fewer mutation structures in the optical feature mutation unit can be mapped to a neuron.

[0080] The acquisition unit 620 is configured to obtain a reference function of the optical field of the optical neural network chip based on the chip's basic parameters; acquire an initial training set and map the initial training set to the optical field to obtain an optical training set. Here, the reference function of the optical field of the optical neural network chip refers to the optical field function when the optical feature quantity is zero. According to an example of an embodiment of this disclosure, the reference function of the optical field of the optical neural network chip can be obtained based on the chip's basic parameters through optical theory calculation or optical simulation. For example, the chip space diffraction light field propagation model based on the Rayleigh-Sommerfeld formula can be used to obtain the reference function of the chip's optical field through optical theory calculation or optical simulation calculation. However, this disclosure is not limited to this, and the reference function of the chip's optical field can also be obtained through any other suitable method. The initial training set can be any dataset used for neural network calculation, which includes initial input values ​​and initial label values ​​for result comparison, such as the IRIS dataset mentioned above. This disclosure does not impose specific limitations on this.

[0081] According to an example of an embodiment of this disclosure, the initial input values ​​of the initial training set can be mapped to optical features of the optical field as optical input values ​​of the optical training set; and the initial label values ​​of the initial training set can be mapped to the output optical power of the optical field as optical label values ​​of the optical training set. For example, when the initial training set is the IRIS dataset and the optical features are phase, the four attribute values ​​of each data sample in the IRIS dataset can be mapped to the four phase values ​​of the optical field, and the label value of each data sample can be mapped to the optical output power of the optical field, thereby obtaining an optical training set for training the chip neural network model.

[0082] Training unit 630 is configured to train a chip neural network model using an optical training set to obtain chip feature parameters, wherein the chip feature parameters characterize the structural characteristics of at least one optical feature quantity mutation unit. Specifically, for each optical input value in the optical training set, training unit 630 inputs the optical input value into the chip neural network model and calculates the output value of the chip neural network model accordingly. For example, the output optical power can be calculated using a chip spatial diffraction light field propagation model based on the Rayleigh-Sommerfeld formula, and used as the output value of the chip neural network model. Then, training unit 630 calculates the loss function between each output value and the corresponding optical label value. For example, the mean square error (MSE), normalized mean square error (NMSE), etc., between the output value and the optical label value can be calculated as the loss function. Then, training unit 630 trains the chip neural network model by minimizing the loss function to obtain the chip feature parameters. For example, a predetermined threshold can be set, and training can end when the loss function is less than the predetermined threshold; or, a maximum number of iterations can be set, and training can end when the maximum number of iterations is reached, etc. This embodiment of the present disclosure does not impose specific limitations in this regard.

[0083] After training, a sequence of chip optical features that results in a loss function less than a predetermined threshold or reaches the maximum number of iterations can be output. This sequence is then mapped to chip feature parameters, thus obtaining the optimal parameters representing the structural characteristics of the desired optical feature mutation unit. Taking the optical feature as the phase, the mutation structure of the optical feature mutation unit as a rectangular groove, and the chip feature parameters as the dimensions of each rectangular groove as an example, a phase sequence can be output after training. This phase sequence includes the output phase value corresponding to each rectangular groove. Then, the phase values ​​in the output phase sequence can be mapped to the dimensions of the corresponding rectangular groove using the formula (1) described above.

[0084] Using the parameter generation apparatus for the optical neural network chip according to the above embodiments of the present disclosure, the optical neural network chip can be designed based on neural network training. Based on the basic chip parameters and chip feature parameters obtained by the apparatus, the optical neural network chip can be manufactured using traditional chip processing technology. The manufacturing process is simple, low-cost, and can be mass-produced. When performing neural network calculations, the resulting optical neural network chip has the advantages of ultra-low power consumption, ultra-fast speed, and large-scale parallel processing compared to traditional CPU-based digital neural networks.

[0085] This disclosure also provides an optical neural network chip, which can be manufactured according to the manufacturing method 300 or 500 described above. This optical neural network chip is capable of performing neural network calculations and, compared to traditional CPU-based digital neural networks, has the advantages of ultra-low power consumption, ultra-fast speed, and the ability to perform large-scale parallel processing.

[0086] The embodiments of this disclosure can also be implemented as a computer-readable storage medium. A computer-readable storage medium according to embodiments of this disclosure stores computer-readable instructions. When executed by a processor, the computer-readable instructions can perform a parameter generation method for an optical neural network chip according to embodiments of this disclosure, or a method for manufacturing an optical neural network chip, as described with reference to the above figures. The computer-readable storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0087] According to embodiments of this disclosure, a computer program product or computer program is also provided, which includes computer-readable instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer-readable instructions from the computer-readable storage medium and execute the computer-readable instructions, causing the computer device to perform the parameter generation method for the optical neural network chip or the manufacturing method for the optical neural network chip described in the various embodiments above.

[0088] Those skilled in the art will understand that the contents disclosed herein can be varied and modified in many ways. For example, the various devices or components described above can be implemented in hardware, or in software, firmware, or a combination of some or all of the three.

[0089] Furthermore, as shown in this disclosure and the claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. The terms "first," "second," and similar terms used in this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "comprising" or "including" and similar terms mean that the element or object preceding the word covers the element or object listed following the word and its equivalents, without excluding other elements or objects. The terms "connected" or "linked" and similar terms are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect.

[0090] Furthermore, flowcharts are used in this disclosure to illustrate the operations performed by the system according to embodiments of this disclosure. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously. Additionally, other operations can be superimposed on these processes, or one or more steps can be removed from these processes.

[0091] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains. It should also be understood that terms such as those defined in a common dictionary shall be interpreted as having a meaning consistent with their meaning in the context of the relevant art, and not as having an idealized or highly formalized meaning, unless expressly defined herein.

[0092] The present disclosure has been described in detail above; however, it will be apparent to those skilled in the art that the present disclosure is not limited to the embodiments described herein. The present disclosure can be implemented in modified and altered ways without departing from the spirit and scope defined by the claims. Therefore, the description herein is for illustrative purposes only and is not intended to be restrictive.

Claims

1. A method for manufacturing an optical neural network chip, comprising: setting chip basic parameters of an optical neural network chip, and constructing a chip neural network model based on the chip basic parameters, wherein the optical neural network chip comprises at least one light feature quantity mutation unit for modulating a light feature quantity of light in the optical neural network chip; obtaining a reference function of a light field of the optical neural network chip based on the chip basic parameters, wherein the reference function is a light field function when the light feature quantity is zero; obtaining an initial training set, and mapping the initial training set to the light field to obtain an optical training set; training the chip neural network model using the optical training set to obtain chip feature parameters, wherein the chip feature parameters represent structural characteristics of the at least one light feature quantity mutation unit; and manufacturing the optical neural network chip based on the chip basic parameters and the chip feature parameters.

2. The method of claim 1, wherein, The chip basic parameters comprise one or more of a basic size of the optical neural network chip, a number of the at least one light feature quantity mutation unit, a number and a position of a mutation structure included in each light feature quantity mutation unit, a number and a position of an output detector.

3. The method of claim 2, wherein, Constructing a chip neural network model based on the chip basic parameters comprises: mapping the number of the at least one light feature quantity mutation unit as a number of hidden layers of the chip neural network model, and mapping one or more mutation structures in each light feature quantity mutation unit as neurons of the chip neural network model.

4. The method of claim 1, wherein, Obtaining a reference function of a light field of the optical neural network chip based on the chip basic parameters comprises: obtaining the reference function of the light field of the optical neural network chip based on the chip basic parameters by optical theory calculation or optical simulation.

5. The method of claim 1, wherein, The initial training set comprises initial input values and initial label values, And wherein mapping the initial training set to the light field to obtain an optical training set comprises: mapping the initial input values of the initial training set as light feature quantities of the light field as optical input values of the optical training set; and mapping the initial label values of the initial training set as output light powers of the light field as optical label values of the optical training set.

6. The method of claim 5, wherein, Training the chip neural network model using the optical training set to obtain chip feature parameters comprises: inputting the optical input values in the optical training set into the chip neural network model, and calculating output values of the chip neural network model; calculating a loss function of the output values and the optical label values; and training the chip neural network model by minimizing the loss function to obtain the chip feature parameters.

7. The method of claim 6, wherein, Training the chip neural network model by minimizing the loss function to obtain the chip feature parameters comprises: training the chip neural network model by minimizing the loss function to obtain a chip light feature quantity sequence; mapping the chip light feature quantity sequence as the chip feature parameters.

8. The method of claim 1, wherein, manufacturing the optical neural network chip based on the chip basic parameters and the chip characteristic parameters comprises: depositing a silicon dioxide film on a substrate material as a lower cladding layer of the optical neural network chip based on the chip basic parameters; depositing an optical waveguide material on the lower cladding layer as an optical waveguide layer of the optical neural network chip, and forming a network structure of the optical neural network chip in the optical waveguide layer based on the chip basic parameters and the chip characteristic parameters, the network structure comprising the at least one light characteristic quantity mutation unit; and depositing a film material on the optical waveguide layer as an upper cladding layer of the optical neural network chip.

9. The method of claim 8, wherein, forming the network structure of the optical neural network chip in the optical waveguide layer based on the chip basic parameters and the chip characteristic parameters comprises: forming, in the optical waveguide layer, an input coupler, an input straight waveguide, an input width tapered waveguide, a wide waveguide, the at least one light characteristic quantity mutation unit, an edge anti-reflection module, an output interface anti-reflection module, an output width tapered waveguide, an output straight waveguide, and an output coupler in sequence, wherein light entering the optical neural network chip propagates through the input coupler, the input straight waveguide, the input width tapered waveguide, the wide waveguide, the at least one light characteristic quantity mutation unit, the edge anti-reflection module, the output interface anti-reflection module, the output width tapered waveguide, the output straight waveguide, and the output coupler in sequence.

10. The method of claim 9, further comprising: forming a light characteristic quantity modulator on the upper cladding layer of the optical neural network chip, the light characteristic quantity modulator being configured to modulate a light characteristic quantity of initial light entering the optical neural network chip through the input coupler to modulate the initial light into input light carrying input data information by mapping the input data information into the light characteristic quantity of the initial light.

11. The method of claim 9, wherein, The edge anti-reflection module is configured to absorb light propagating to a side cross-section of the wide waveguide of the optical neural network chip, and the output interface anti-reflection module is configured to absorb light reflected from an output interface of the wide waveguide other than the output width tapered waveguide back to the wide waveguide.

12. The method of any one of claims 1-11, wherein, The light characteristic quantity comprises one of an amplitude, a phase, a polarization, a frequency, and an angular momentum of light.

13. The method of any one of claims 1-11, wherein, The at least one light characteristic quantity mutation unit is a rectangular groove array, and the chip characteristic parameter is a size of each rectangular groove in the rectangular groove array.

14. A parameter generation method of an optical neural network chip, comprising: setting chip basic parameters of an optical neural network chip, and constructing a chip neural network model based on the chip basic parameters, wherein the optical neural network chip comprises at least one light characteristic quantity mutation unit configured to modulate a light characteristic quantity of light in the optical neural network chip; obtaining a reference function of an optical field of the optical neural network chip based on the chip basic parameters, wherein the reference function is an optical field function when the light characteristic quantity is zero; obtaining an initial training set, and mapping the initial training set to the optical field to obtain an optical training set; training the chip neural network model with the optical training set to obtain chip feature parameters, wherein the chip feature parameters characterize structural properties of the at least one optical feature mutation unit, wherein the chip basic parameters and the chip feature parameters are used to manufacture the optical neural network chip.

15. The method of claim 14, wherein, The chip basic parameters include one or more of a basic size of the optical neural network chip, a number of the at least one optical feature mutation unit, a number and a position of a mutation structure included in each optical feature mutation unit, a number and a position of an output detector.

16. The method of claim 15, wherein, Constructing a chip neural network model based on the chip basic parameters includes: mapping the number of the at least one optical feature mutation unit to a number of hidden layers of the chip neural network model, and mapping one or more mutation structures in each optical feature mutation unit to neurons of the chip neural network model.

17. The method of claim 14, wherein, Obtaining a reference function of an optical field of the optical neural network chip based on the chip basic parameters includes: obtaining the reference function of the optical field of the optical neural network chip based on the chip basic parameters through optical theory calculation or optical simulation.

18. The method of claim 14, wherein, The initial training set includes initial input values and initial label values, and wherein mapping the initial training set to the optical field to obtain an optical training set includes: mapping the initial input values of the initial training set to optical feature quantities of the optical field as optical input values of the optical training set; and mapping the initial label values of the initial training set to output optical powers of the optical field as optical label values of the optical training set.

19. The method of claim 18, wherein, Training the chip neural network model with the optical training set to obtain chip feature parameters includes: inputting the optical input values in the optical training set into the chip neural network model, and calculating output values of the chip neural network model; calculating a loss function of the output values and the optical label values; and training the chip neural network model by minimizing the loss function to obtain the chip feature parameters.

20. The method of claim 19, wherein, Training the chip neural network model by minimizing the loss function to obtain the chip feature parameters includes: training the chip neural network model by minimizing the loss function to obtain a chip optical feature sequence; mapping the chip optical feature sequence to the chip feature parameters.

21. The method of any one of claims 14-20, wherein, The optical feature quantity includes one of an amplitude, a phase, a polarization, a frequency and an angular momentum of light.

22. The method of any one of claims 14-20, wherein, The at least one optical feature mutation unit is a rectangular groove array, and the chip feature parameter is a size of each rectangular groove in the rectangular groove array.

23. A method for manufacturing an optical neural network chip, comprising: depositing a silicon dioxide film on a substrate material as a lower cladding layer of the optical neural network chip based on chip basic parameters; depositing a light waveguide material on the lower cladding layer as a light waveguide layer of the optical neural network chip, and forming a network structure of the optical neural network chip on the light waveguide layer based on the chip basic parameters and chip characteristic parameters, the network structure comprising at least one light characteristic quantity mutation unit; and depositing a thin film material on the light waveguide layer as an upper cladding layer of the optical neural network chip, wherein the chip basic parameters and the chip characteristic parameters are generated according to the method of claim 14.

24. The method of claim 23, wherein, forming a network structure of the optical neural network chip on the light waveguide layer based on the chip basic parameters and chip characteristic parameters comprises: forming, in sequence, an input coupler, an input straight waveguide, an input width gradient waveguide, a wide waveguide, the at least one light characteristic quantity mutation unit, an edge anti-reflection module, an output interface anti-reflection module, an output width gradient waveguide, an output straight waveguide, and an output coupler on the light waveguide layer, wherein light entering the optical neural network chip propagates through the input coupler, the input straight waveguide, the input width gradient waveguide, the wide waveguide, the at least one light characteristic quantity mutation unit, the edge anti-reflection module, the output interface anti-reflection module, the output width gradient waveguide, the output straight waveguide, and the output coupler in sequence.

25. The method of claim 24, further comprising: forming a light characteristic quantity modulator on the upper cladding layer of the optical neural network chip, the light characteristic quantity modulator being configured to modulate a light characteristic quantity of initial light entering the optical neural network chip through the input coupler to modulate the initial light into input light carrying input data information by mapping the input data information into the light characteristic quantity of the initial light.

26. The method of claim 24, wherein, the edge anti-reflection module is configured to absorb light propagating to a side cross-section of the wide waveguide of the optical neural network chip, and the output interface anti-reflection module is configured to absorb light reflected from an output interface of the wide waveguide other than the output width gradient waveguide back to the wide waveguide.

27. An apparatus for generating parameters of an optical neural network chip, comprising: a setting unit configured to set chip basic parameters of an optical neural network chip, and construct a chip neural network model based on the chip basic parameters, the optical neural network chip comprising at least one light characteristic quantity mutation unit configured to modulate a light characteristic quantity of the optical neural network chip; an obtaining unit configured to obtain a reference function of a light field of the optical neural network chip based on the chip basic parameters, wherein the reference function is a light field function when the light characteristic quantity is zero, obtain an initial training set, and map the initial training set to the light field to obtain an optical training set; and a training unit configured to train the chip neural network model using the optical training set to obtain chip characteristic parameters, wherein the chip characteristic parameters represent structural characteristics of the at least one light characteristic quantity mutation unit, wherein the chip basic parameters and the chip characteristic parameters are used to manufacture the optical neural network chip. ​ 28. An optical neural network chip manufactured according to the method of any one of claims 1-13 or 23-26.

29. A computer readable storage medium having stored thereon computer readable instructions which, when executed by a processor, cause the processor to perform the method of any one of claims 1-26.

30. A computer program product comprising computer readable instructions which, when executed by a processor, cause the processor to perform the method of any one of claims 1-26.

Images