Optical simulation operation device, optical simulation operation method and related equipment

Abstract

The embodiment of the invention provides an optical simulation operation device, an optical simulation operation method and related equipment, and the device comprises a metasurface structure which comprises a plurality of metasurface units disposed on a substrate, and the phase distribution of the plurality of metasurface units corresponds to a weight matrix; the light emitting structure is arranged on one side of the metasurface structure and used for emitting a plurality of light signals vertically incident on the metasurface unit, and wavelength distribution of the light signals corresponds to an input vector; the light operation structure is arranged on the other side of the metasurface structure and used for calculating sine values of light deflection angles after the light signals pass through the corresponding metasurface units, an output vector is obtained according to the sine values, and the output vector is the product of an input vector and a weight matrix, the optical simulation operation can be completed by using the optical signal wavelength and the optical deflection angle corresponding to the metasurface unit phase, the operation efficiency is effectively improved, and the operation acceleration effect when the method is applied to the artificial intelligence accelerator is improved.

Description

Technical Field

[0001] This application relates to the field of optical computing technology, and in particular to an optical simulation computing device, an optical simulation computing method, and related equipment. Background Technology

[0002] Currently, in-memory computing (IMC) units are widely used in artificial intelligence accelerators to improve data processing speed and reduce power consumption. IMC units are usually integrated into digital signal processors (DSPs) or other types of computing chips. By integrating data storage and computing functions on the same chip, data handling is reduced, thereby reducing power consumption and increasing computing speed.

[0003] However, most existing in-memory computing units perform computational tasks based on changes in the current and voltage of electronic devices. This results in high power consumption of the corresponding electronic devices during high-frequency and large-scale parallel operations, and the computing speed of the electronic devices is limited by the speed at which electrons move in the circuit. These issues affect the computing efficiency of the in-memory computing units and their acceleration effect when applied to artificial intelligence accelerators. Summary of the Invention

[0004] This application provides an optical simulation computing device, which aims to use the wavelength of the light signal as the input vector and the phase of the metasurface unit as the weight matrix to obtain the output vector according to the light deflection angle, thereby completing the optical simulation computing and solving the problems of high power consumption and limited computing speed of electronic devices that affect computing efficiency, and effectively improving the computing acceleration effect.

[0005] To achieve the above objectives, a first aspect of this application provides an optical simulation computing device, comprising: a metasurface structure including a plurality of metasurface units disposed on a substrate, wherein the phase distribution of the plurality of metasurface units corresponds to a weight matrix; a light emitting structure disposed on one side of the metasurface structure for emitting a plurality of light signals perpendicularly incident on the metasurface units, wherein the wavelength distribution of the plurality of light signals corresponds to an input vector; and a light computing structure disposed on the other side of the metasurface structure for calculating the sine value of the light deflection angle of the light signals after passing through the metasurface units, and obtaining an output vector based on the plurality of sine values, wherein the output vector is the product of the input vector and the weight matrix.

[0006] To achieve the above objectives, a second aspect of this application provides an optical simulation computing method applied to an optical computing structure in an optical simulation computing device. The device further includes a metasurface structure and a light emitting structure. The metasurface structure includes multiple metasurface units disposed on a substrate. The light emitting structure is disposed on one side of the metasurface structure, and the optical computing structure is disposed on the other side of the metasurface structure. The method includes: when the light emitting structure emits multiple light signals perpendicularly incident on the metasurface units, calculating the sine value of the light deflection angle of the light signals after passing through the metasurface units, and obtaining an output vector based on the multiple sine values; wherein the phase distribution of the multiple metasurface units corresponds to a weight matrix, the wavelength distribution of the multiple light signals corresponds to an input vector, and the output vector is the product of the input vector and the weight matrix.

[0007] To achieve the above objectives, a third aspect of the present application provides an artificial intelligence accelerator, including the optical simulation computing device described in any one of the first aspects above.

[0008] To achieve the above objectives, a fourth aspect of the present application provides an optical device for performing an optical simulation operation method as described in any of the second aspects above.

[0009] The optical simulation computing device provided in this application integrates a metasurface structure comprising multiple metasurface units on a silicon-based substrate. Utilizing the anomalous refractive properties of these metasurface units, a novel optical computing paradigm is achieved. This device is suitable for matrix operation units in artificial intelligence accelerators, effectively improving computational efficiency and reducing energy consumption. The optical simulation computing device comprises three parts: a metasurface structure, a light-emitting structure, and a light-operating structure. Specifically, the metasurface structure consists of multiple metasurface units disposed on the substrate. The phase distribution of each unit corresponds to the weight values ​​of the weight matrix. The design of the metasurface units allows for the simulation of different weights in the input vector by adjusting their phases, thereby achieving modulation of the optical signal. The light-emitting structure, located on one side of the metasurface structure, is responsible for emitting multiple optical signals. The wavelength distribution of these signals corresponds to the values ​​of the input vector. The multiple optical signals are incident perpendicularly onto the metasurface units, using wavelengths as input quantities, so that each wavelength can represent an input value. The light-operating structure, located on the other side of the metasurface structure, is used to calculate the sine of the deflection angle of the optical signal after passing through the metasurface units. The sine value is obtained through optical means, and then an output vector can be obtained based on the sine value. The output vector is the product of the input vector and the weight matrix to realize the optical analog multiply-accumulate (MAC) operation. It can be understood that since the operation process of the device in this application embodiment is based on light rather than electrons, the energy consumption caused by electron migration can be reduced. It is particularly suitable for power-sensitive application scenarios. Moreover, since the propagation speed of light signals is much higher than that of electrons, the operation speed of the device is no longer limited by the movement speed of electrons in the circuit, which can effectively improve the operation speed. This solves the problem of high power consumption and limited operation speed of electronic devices affecting the operation efficiency, and effectively improves the operation acceleration effect of optical analog operation device when applied to artificial intelligence accelerator. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of an optical simulation computing device provided in an embodiment of this application;

[0011] Figure 2 This is a schematic diagram of the operation process of an optical simulation computing device provided in an embodiment of this application;

[0012] Figure 3 This is a schematic diagram of another operation process of an optical analog computing device provided in an embodiment of this application;

[0013] Figure 4 Another schematic diagram of an optical analog computing device provided in an embodiment of this application;

[0014] Figure 5 A schematic diagram of simulation analysis of a metasurface unit provided in an embodiment of this application;

[0015] Figure 6 A flowchart of an optical simulation calculation method provided in an embodiment of this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0017] In some embodiments, although functional modules are divided in the system diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the system or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the foregoing drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0018] Furthermore, unless otherwise explicitly specified and limited, the term "connection / linkage" should be interpreted broadly, for example, it can be a fixed connection or a movable connection, a detachable connection or a non-detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection or a connection that can communicate with each other; it can be a direct connection or an indirect connection through an intermediate medium.

[0019] In the description of the embodiments in this application, the terms "one embodiment / implementation," "another embodiment / implementation," or "some embodiments / implementations," "in the above embodiments / implementations," etc., refer to specific features, structures, materials, or characteristics described in conjunction with embodiments or examples that are included in at least two embodiments or implementations disclosed in this application. In this application disclosure, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or implementation. It should be noted that although a logical order is shown in the flowcharts, in some cases, the steps shown or described may be performed in a different order than that shown in the flowcharts.

[0020] Currently, in-memory computing (IMC) units are widely used in artificial intelligence accelerators to improve data processing speed and reduce power consumption. IMC units are typically integrated into digital signal processors (DSPs) or other types of computing chips. By integrating data storage and computing functions on the same chip, data transfer is reduced, thereby lowering power consumption and increasing computing speed. However, most existing IMC units perform computing tasks based on changes in the current and voltage of electronic devices. This results in high power consumption of the corresponding electronic devices during high-frequency and large-scale parallel operations, and the computing speed of the electronic devices is limited by the speed of electron movement in the circuit. These issues affect the computing efficiency of the IMC units and their acceleration effect when applied to artificial intelligence accelerators.

[0021] Based on this, embodiments of this application provide an optical simulation computing device, an optical simulation computing method, and related equipment. The device includes: a metasurface structure, comprising multiple metasurface units disposed on a substrate, wherein the phase distribution of the multiple metasurface units corresponds to a weight matrix; a light emitting structure disposed on one side of the metasurface structure, used to emit multiple light signals perpendicularly incident on the multiple metasurface units, wherein the wavelength distribution of the multiple light signals corresponds to an input vector; and a light computing structure disposed on the other side of the metasurface structure, used to calculate the sine values ​​of the light deflection angles of the multiple light signals after passing through the metasurface units, and to obtain an output vector based on the multiple sine values, wherein the output vector is the product of the input vector and the weight matrix. This application can utilize the light deflection angles corresponding to the wavelengths of the light signals and the phases of the metasurface units to complete optical simulation computing, effectively improving computing efficiency and enhancing the computing acceleration effect when applied to artificial intelligence accelerators.

[0022] The embodiments of this application will be further described below with reference to the accompanying drawings.

[0023] refer to Figure 1 , Figure 1 This is a schematic diagram of an optical simulation computing device provided in an embodiment of this application. A first aspect of this application provides an optical simulation computing device, comprising: a metasurface structure including multiple metasurface units disposed on a substrate, the phase distribution of the multiple metasurface units corresponding to a weight matrix; a light emitting structure disposed on one side of the metasurface structure for emitting multiple light signals perpendicularly incident on the multiple metasurface units, the wavelength distribution of the multiple light signals corresponding to an input vector; and a light computing structure disposed on the other side of the metasurface structure for calculating the sine values ​​of the light deflection angles of the multiple light signals after passing through the corresponding metasurface units, and obtaining an output vector based on the multiple sine values, the output vector being the product of the input vector and the weight matrix.

[0024] This application embodiment integrates a metasurface structure comprising multiple metasurface units on a silicon-based substrate, utilizing the anomalous refractive properties of the metasurface units to achieve a novel optical computing paradigm. The device described herein is suitable for matrix operation units in artificial intelligence accelerators, effectively improving computational efficiency and reducing energy consumption. The optical simulation computing device comprises three parts: a metasurface structure, a light-emitting structure, and a light-operating structure. Specifically, the metasurface structure consists of multiple metasurface units disposed on the substrate. The phase distribution of each unit corresponds to the weight values ​​of the weight matrix. The design of the metasurface units allows for the simulation of different weights in the input vector by adjusting their phases, thereby achieving modulation of the optical signal. The light-emitting structure, located on one side of the metasurface structure, is responsible for emitting multiple optical signals. The wavelength distribution of the multiple optical signals corresponds to the values ​​of the input vector. The multiple optical signals are incident perpendicularly onto the metasurface units, using wavelengths as input quantities, so that each wavelength can represent an input value. The light-operating structure, located on the other side of the metasurface structure, is used to calculate the sine of the deflection angle of the optical signal after passing through the metasurface units. The sine value is obtained through optical means, and then an output vector can be obtained based on the sine value. The output vector is the product of the input vector and the weight matrix to realize the optical analog multiply-accumulate (MAC) operation. It can be understood that since the operation process of the device in this application embodiment is based on light rather than electrons, the energy consumption caused by electron migration can be reduced. It is particularly suitable for power-sensitive application scenarios. Moreover, since the propagation speed of light signals is much higher than that of electrons, the operation speed of the device is no longer limited by the movement speed of electrons in the circuit, which can effectively improve the operation speed. This solves the problem of high power consumption and limited operation speed of electronic devices affecting the operation efficiency, and effectively improves the operation acceleration effect of optical analog operation device when applied to artificial intelligence accelerator.

[0025] refer to Figure 2 , Figure 2 This is a schematic diagram of the operation process of an optical simulation computing device provided in one embodiment of this application. In some embodiments, the optical computing structure includes multiple multiplication units arranged at a preset distance from multiple metasurface units, and an accumulation unit communicatively connected to the multiple multiplication units. The multiplication units are used to obtain the optical signal offset distance of the corresponding metasurface unit and calculate the sine value of the optical deflection angle based on the preset distance and the optical signal offset distance. The accumulation unit is used to group and accumulate multiple sine values ​​to obtain an output vector. Figure 2 , Figure 2 In this context, λ0, λ1 to λm represent the wavelengths of multiple optical signals, and the wavelength distribution of these signals corresponds to multiple input values ​​in the input vector. This refers to the phase of multiple metasurface units. The phase distribution of multiple metasurface units corresponds to multiple weight coefficients in the weight matrix. 0 to m corresponds to the number of rows in the metasurface array formed by multiple metasurface units, and 0 to n corresponds to the number of columns in the metasurface array formed by multiple metasurface units. The multiplication process of vector multiplication into matrix is ​​realized through multiple multiplication units Δ and accumulation units.

[0026] In some embodiments, the multiplication unit is used to obtain the optical signal offset distance corresponding to each metasurface unit and calculate the sine value of the optical deflection angle using a preset spacing and the optical signal offset distance. This calculation process is actually performing a multiplication operation, because the calculation of the sine value involves the ratio of the optical signal offset distance to the preset spacing. The accumulation unit is responsible for accumulating the sine values ​​calculated by all the multiplication units to obtain the final output vector. The process of the accumulation unit accumulating multiple sine values ​​corresponds to the accumulation operation in matrix multiplication. By setting multiple multiplication units in parallel, the offset distances of multiple optical signals can be processed simultaneously, thereby accelerating the calculation process of the sine value. Furthermore, using the preset spacing as a measurement unit can improve the accuracy of the calculation process and ensure the accuracy of the sine value.

[0027] In some embodiments, the weighting matrix includes M*N weighting coefficients, and the phase value of the metasurface unit is related to the weighting coefficients. Specifically, the phase value of the metasurface unit is equal to the product of the corresponding weighting coefficient in the weighting matrix and the complete phase period, where the complete phase period is 2π. The input vector includes M input values, and the wavelength of the optical signal is the corresponding input value in the input vector, where M and N are non-zero natural numbers. Figure 2 M = m + 1, N = n + 1, where the weight matrix has M × N weight coefficients. The phase of each metasurface unit is set as the product of the corresponding weight coefficient and a complete phase period. This means that the metasurface unit can encode the weight values ​​in the weight matrix through phase changes. The input vector consists of M input values, each corresponding to a specific optical signal wavelength. In this way, the weight matrix and input vector are converted into a form that can be processed by optical means. By converting the weight coefficients into phase values, the optical encoding of the weight values ​​is realized.

[0028] In some embodiments, the number of input values ​​in the input vector is equal to the number of rows in the weight matrix. It can be understood that the device of this application is a multiply-accumulate unit for vector-matrix multiplication. The input vector and the weight matrix are the vector and matrix input to the multiply-accumulate unit for calculation. When applied to convolution calculation, the matrix coefficients of the weight matrix can be the weight coefficients corresponding to the input values ​​in the input vector, so that the device of this application can complete the convolution calculation process through optical simulation.

[0029] In some embodiments, multiple metasurface units form an M x N metasurface array on a substrate, with phase distribution corresponding to the weight matrix. M optical signals emitted by the light-emitting structure are perpendicularly incident on the corresponding M rows of metasurface units in the array. It can be understood that the metasurface units form an M x N metasurface array on the substrate, and the phase values ​​of the M x N metasurface units in the array are equal to the product of the M x N weight coefficients in the weight matrix and 2π. The M optical signals emitted by the light-emitting structure are perpendicularly incident on the corresponding rows of the metasurface array. This arrangement allows for parallel processing of the weight matrix, with each metasurface unit corresponding to a specific weight value, ensuring accurate mapping and processing of the weight values. Based on the above, utilizing the fact that the incident angle of different wavelengths directly onto the metasurface is 0, the following derivation process can be obtained according to the generalized Snell's law:

[0030] The generalized Snell's law is an optical law that extends the classical Snell's law. It describes the relationship between the angle of incidence and the angle of refraction when light refracts between different media. In optics, when a light wave travels from one medium (with refractive index n1) to another (with refractive index n2), the relationship between the angle of incidence (θ1) and the angle of refraction (θ2) is given by the following formula: n1sin(θ1) = n2sin(θ2). This formula is Snell's law, applicable to most cases of light refraction. The generalized Snell's law adds consideration of phase to this, and can be expressed as: n2sin(θ2) - n1sin(θ1) = φ / (2π / λ).

[0031] According to the generalized Snell's law, when the incident angle θ1 is 0, n2sin(θ2)=φ / (2π / λ). The limiting condition of this application is perpendicular incidence, and the optical operation structure includes multiple multiplication units set at a predetermined distance from multiple metasurface units. The emission medium is air, and n2=1. Therefore, the relationship between the sine value sinθ of the light deflection angle θ after each light signal passes through a metasurface unit, the wavelength λ of the light signal, and the phase φ of the metasurface unit is as follows:

[0032]

[0033] Since the wavelengths λ of multiple optical signals are equal to multiple input values ​​in the input vector, the phase φ of the metasurface unit divided by the complete phase period 2π is the multiple weight coefficients in the weight matrix. Therefore, the sine value sinθ of the light deflection angle θ can be determined as the product of a single input value and a weight coefficient. Based on this, the accumulator unit is used to group and accumulate multiple sine values ​​according to the vector matrix multiplication rules to obtain the output vector, as shown in the following expression:

[0034]

[0035] The output vector includes N output values, which are the sum of the sine values ​​of the M light deflection angles θ corresponding to each of the N columns of metasurface units. Therefore, the following vector-matrix multiplication operation can be achieved through the above process:

[0036]

[0037] It is understandable that the output vector consists of N output values, each of which is the sum of M sine values ​​corresponding to each metasurface unit in the N columns of the metasurface array. For each column, the sine values ​​calculated by all metasurface units in that column are accumulated to obtain the final output value. By accumulating the sine values ​​of each column, the calculation process of the output vector is simplified, so that the multiplication and accumulation operation can be realized in an optical manner.

[0038] In some embodiments, it is understood that, compared to common matrix multiplication and addition operations implemented using optical multiplexing and routing arrays, this invention employs a metasurface array to pre-store weight values. This allows for multiplication and addition operations on activation signals of different wavelengths, achieving high-speed, low-latency, and low-power multiplication and addition operations. It also effectively solves the problem of traditional optical computing's inability to store weights. Specifically, this invention sets up a specific metasurface optical path, using wavelength as the input quantity to quantize different wavelengths into different input values, achieving highly efficient encoding of the input signal. The phase change of the metasurface array serves as the weight value, and its magnitude and rotation can be externally controlled, providing the system with the ability to dynamically adjust computational parameters. This design allows for real-time updates of weight values ​​without altering the physical structure, thus greatly increasing the system's flexibility and applicability.

[0039] refer to Figure 3 , Figure 3 This is a schematic diagram of another operation process of the optical simulation computing device provided in one embodiment of this application; in some embodiments, the multiplication unit Δ includes a photodetector and a light baffle. The light baffle is set at a preset distance from the metasurface unit, and the center position of the light baffle is located in the emission direction of the light signal corresponding to the metasurface unit. The photodetector is used to obtain the landing position of the light signal on the light baffle, and the distance between the landing position and the center position is determined as the light signal offset distance. Figure 3 , Figure 3 The solid arrow in the middle corresponds to the landing point of the actual light signal after refraction on the light baffle. The two dashed lines represent the preset distance between the light baffle and the metasurface unit, and the light signal offset distance determined by the distance between the landing point and the center position.

[0040] In some embodiments, a light baffle is disposed at a predetermined spacing between the metasurface units to block and reflect the light signal. A photodetector is used to capture the landing point of the light signal on the light baffle. By measuring the distance between the landing point and the center of the light baffle, the offset distance of the light signal is determined. That is, the offset of the light signal is determined by using optical principles and geometric relationships. The cooperation between the light baffle and the photodetector simplifies the measurement process of the light signal offset distance and improves the reliability of the measurement. It is understood that the design of the light baffle and the photodetector allows for adaptation to metasurface units of different sizes and shapes, increasing the applicability of the system.

[0041] In some embodiments, the multiplication unit uses a preset spacing as the measurement unit to accurately measure the offset distance of the optical signal. For example... Figure 3 As shown, the distance between the fixed multiplication unit and the metasurface unit is set to a unit distance of 1. The distance x between the landing point and the center position can then be obtained through a photodetector, and the corresponding light deflection angle θ can be calculated accordingly. i The sine value is then obtained, and the addition operation is performed using the accumulation unit to finally complete the multiplication and accumulation operation. The multiplication unit uses a preset spacing as the measurement unit to measure the offset distance of the optical signal. Therefore, regardless of the actual offset distance of the optical signal, it will be calculated based on the ratio with the preset spacing to obtain the sine value. Using the preset spacing as the measurement unit unifies the measurement standard of all multiplication units in the system, simplifies the system calibration process, and reduces errors in the manufacturing process through the standardization of the preset spacing.

[0042] In some embodiments, it is understood that the present invention obtains the multiplication value by calculating the sine of the light deflection angle, and performs the accumulation operation through an array of multiple metasurface units, thus realizing matrix multiplication. This process is entirely based on optical principles, avoiding the energy consumption and speed limitations of electronic computing. Compared with traditional electronic multiply-accumulate operations, the present invention significantly improves the computation speed; and compared with traditional optical multiply-accumulate units, the present invention has the ability to pre-store weight values ​​and achieve real-time adjustment, effectively enhancing the flexibility and scalability of the design.

[0043] In some embodiments, the metasurface cells can achieve weight control by adjusting the deflection angle or size, including dielectric and non-dielectric types. In the case of using dielectric metasurface cells, this application can be compatible with traditional CMOS processes. At the same time, the metasurface material can also be used as a storage cell to simulate in-memory computing operations, thereby reducing the need for data transfer.

[0044] refer to Figure 4 , Figure 4This is another schematic diagram of an optical simulation computing device provided in one embodiment of this application; in some embodiments, the light emitting structure includes a wavelength modulation unit and a light intensity adjustment unit. The light intensity adjustment unit is disposed between the wavelength modulation unit and the metasurface structure. The wavelength modulation unit is used to adjust the light signals of multiple wavelengths corresponding to multiple input values ​​in the emission input vector. The light intensity adjustment unit is used to adjust the light intensity of the light signal before the light signal is perpendicularly incident on the metasurface unit. The schematic diagram shows the optical path of a single-wavelength light signal incident on the metasurface unit. Figure 4 As shown, firstly, wavelength modulation units such as electro-optic modulators change the wavelength, and then light intensity is adjusted by light intensity adjustment units such as attenuators. The light is then incident perpendicularly onto a metasurface array formed by multiple metasurface units. Finally, the deflection angle is determined by a photosensitive detector to complete the multiplication calculation and realize optical simulation operation.

[0045] The wavelength modulation unit is responsible for adjusting the optical signals of multiple wavelengths corresponding to multiple input values ​​in the emission and input vectors, while the light intensity adjustment unit adjusts the intensity of the optical signal before it is incident on the metasurface unit. The wavelength modulation unit can realize the emission of optical signals of different wavelengths, which increases the flexibility and applicability of the system. The light intensity adjustment unit allows for fine adjustment of the intensity of the optical signal to reduce the light intensity and prevent the metasurface from being unable to distinguish each optical signal.

[0046] refer to Figure 5 , Figure 5 This is a simulation analysis diagram of a metasurface unit provided in one embodiment of this application. The horizontal axis, PILLAR DIAMETER (NM), represents the diameter (nanometers) of the nanopillar, and the vertical axis, PHASE (RAD), represents the phase value in radians (RAD) corresponding to the diameter (nanometers) of the nanopillar. In some embodiments, the metasurface unit includes multiple nanopillars disposed on a substrate. The diameter of the nanopillars is set according to the phase of the metasurface unit. Therefore, by adjusting the diameter of the nanopillars, the refraction and phase modulation of the light signal by each metasurface unit can be precisely controlled, thereby achieving fine control of the weight values. It can be understood that the metasurface unit simulates abnormal refraction phenomena by arranging cylinders, and by changing the diameter of the cylinders, different angles of deflection are achieved. Specifically, corresponding to... Figure 4 Simulation analysis was performed using software, fixing the height of the Si cylinder at 280nm, the circumference at 250nm, and the incident wavelength at 633nm. Figure 4 It can be seen that the diameter ranges from 80nm to 180nm, and the phase is adjustable from 0 to 2π(π). Therefore, the ratio of the metasurface unit phase to 2π(π) can be used to correspond to the weight coefficients from 0 to 1, and optical simulation calculations can be completed according to the light deflection angle, which can effectively improve the computational efficiency and the computational acceleration effect when applied to artificial intelligence accelerators.

[0047] In some embodiments, the method of setting the diameter of the nanopillar according to the phase of the metasurface unit includes: static weighted modulation, which selects only two phases, representing 0 or 1 values; and flexible, variable, dynamic adjustment of the metasurface phase to achieve full phase coverage and simulate the actual value. Specifically, static weighted modulation is a simple binary modulation method in which the phase of the metasurface unit takes only two values, typically corresponding to 0 or 1. In this method, the diameter of the nanopillar is set to two specific sizes, corresponding to these two phase values ​​respectively. When a light signal is incident on the metasurface unit, the light signal will undergo different refractions according to the phase change caused by the diameter of the nanopillar, thereby achieving weighted modulation. Heavy binary encoding, due to its simple design, allows static weighted modulation to reduce the complexity and cost of the manufacturing process, and the reduced number of modulation states helps improve the stability and reliability of the system. Flexible variable dynamic control of metasurface phase is a more complex and flexible method that allows the phase of metasurface units to change continuously to achieve full phase coverage. In this method, the diameter of the nanopillars can be dynamically adjusted as needed to simulate the actual values ​​in the calculation. This continuous phase modulation can be achieved by stretching through external control (such as electric field, temperature or mechanical stress), thereby dynamically changing the optical properties of the metasurface units to adapt to different computing needs and application scenarios.

[0048] refer to Figure 6 , Figure 6 This is a flowchart of an optical simulation operation method provided in an embodiment of this application. A second aspect of this application provides an optical simulation operation method applied to an optical operation structure in an optical simulation operation device. The device further includes a metasurface structure and a light-emitting structure. The metasurface structure includes multiple metasurface units disposed on a substrate. The light-emitting structure is disposed on one side of the metasurface structure, and the optical operation structure is disposed on the other side of the metasurface structure. The method includes, but is not limited to, the following steps:

[0049] Step S610: When the light emitting structure emits multiple light signals that are perpendicularly incident on the metasurface unit, calculate the sine value of the light deflection angle after the light signal passes through the metasurface unit, and obtain the output vector based on the multiple sine values.

[0050] In this method, the phase distribution of multiple metasurface units corresponds to the weight matrix, the wavelength distribution of multiple optical signals corresponds to the input vector, and the output vector is the product of the input vector and the weight matrix. It should be noted that the specific implementation details and beneficial effects of this optical simulation calculation method can be reflected in the embodiments of the above-mentioned optical simulation calculation devices, and will not be elaborated here.

[0051] A third aspect of this application provides an artificial intelligence accelerator, including an optical simulation computing device of any one of the first aspects described above. Since the artificial intelligence accelerator of this application has the optical simulation computing device of the above embodiments, the specific implementation and technical effects of the artificial intelligence accelerator of this application can be referred to the specific implementation and technical effects of the optical simulation computing device of any of the above embodiments.

[0052] A fourth aspect of this application provides an optical device for performing an optical simulation operation method as described in any of the second aspects above, for example, performing the above-described... Figure 6 Regarding the method step S610, it should be noted that the specific implementation details and beneficial effects of the optical device can be reflected in the embodiments of the above-mentioned optical simulation computing devices, and will not be elaborated here.

[0053] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network nodes. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0054] The above is a detailed description of the preferred embodiments of this application. However, this application is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. An optical analog computing device, comprising: A metasurface structure includes multiple metasurface units disposed on a substrate, wherein the phase distribution of the multiple metasurface units corresponds to a weight matrix; A light-emitting structure is disposed on one side of the metasurface structure for emitting multiple light signals that are perpendicularly incident on the metasurface unit, wherein the wavelength distribution of the multiple light signals corresponds to the input vector; An optical computation structure is disposed on the other side of the metasurface structure. It is used to calculate the sine value of the light deflection angle after the light signal passes through the metasurface unit, and to obtain an output vector based on multiple sine values. The output vector is the product of the input vector and the weight matrix.

2. The optical analog computing device according to claim 1, characterized in that, The optical computing structure includes multiple multiplication units spaced at a preset distance from the multiple metasurface units, and an accumulation unit communicatively connected to the multiple multiplication units. The multiplication units are used to obtain the optical signal offset distance of the corresponding metasurface unit and calculate the sine value of the optical deflection angle based on the preset distance and the optical signal offset distance. The accumulation unit is used to group and accumulate the multiple sine values ​​to obtain the output vector.

3. The optical analog computing device according to claim 2, characterized in that, The multiplication unit includes a photodetector and a light baffle. The light baffle is disposed at a preset distance from the metasurface unit. The center position of the light baffle is located in the emission direction of the light signal corresponding to the metasurface unit. The photodetector is used to obtain the landing position of the light signal on the light baffle and determine the distance between the landing position and the center position as the offset distance of the light signal.

4. The optical analog computing device according to claim 2, characterized in that, The multiplication unit measures the optical signal offset distance using the preset spacing as the measurement unit.

5. The optical analog computing device according to claim 1, characterized in that, The weight matrix includes M*N weight coefficients, the phase value of the metasurface unit is related to the weight coefficients, the input vector includes M input values, and the wavelength of the optical signal is the corresponding input value in the input vector, where M and N are non-zero natural numbers.

6. The optical analog computing device according to claim 5, characterized in that, Multiple metasurface units form an M-row, N-column metasurface array on the substrate, with a phase distribution corresponding to the weight matrix. M optical signals emitted by the light-emitting structure are respectively incident perpendicularly on the M-row metasurface units of the metasurface array.

7. The optical analog computing device according to claim 6, characterized in that, The output vector includes N output values, where the N output values ​​are the sum of M sine values ​​corresponding to each of the N columns of metasurface units.

8. The optical analog computing device according to claim 1, characterized in that, The metasurface unit includes a plurality of nanopillars disposed on the substrate, the diameter of which is set according to the phase of the metasurface unit.

9. The optical analog computing device according to claim 1, characterized in that, The light emitting structure includes a wavelength modulation unit and a light intensity adjustment unit. The light intensity adjustment unit is disposed between the wavelength modulation unit and the metasurface structure. The wavelength modulation unit is used to adjust the light signal of multiple wavelengths corresponding to multiple input values ​​in the input vector. The light intensity adjustment unit is used to adjust the light intensity of the light signal before the light signal is perpendicularly incident on the metasurface unit.

10. An optical simulation calculation method, characterized in that, An optical computing structure is applied in an optical analog computing device, the device further comprising a metasurface structure and a light-emitting structure, the metasurface structure comprising a plurality of metasurface units disposed on a substrate, the light-emitting structure being disposed on one side of the metasurface structure, and the optical computing structure being disposed on the other side of the metasurface structure, the method comprising: When the light emitting structure emits multiple light signals that are perpendicularly incident on the metasurface unit, the sine value of the light deflection angle after the light signal passes through the metasurface unit is calculated, and an output vector is obtained based on the multiple sine values; The phase distribution of the plurality of metasurface units corresponds to the weight matrix, the wavelength distribution of the plurality of optical signals corresponds to the input vector, and the output vector is the product of the input vector and the weight matrix.

11. An artificial intelligence accelerator, characterized in that, Includes the optical analog computing device as described in any one of claims 1 to 9.

12. An optical device, characterized in that, The optical device is used to perform the optical simulation calculation method as described in claim 10.

Images