A method and system for realizing dual-channel near-field dynamic control using optical computing metasurface

Abstract

The present invention discloses a method for realizing dual-channel near-field control by using optical computing metasurfaces, which is applied to the field of optical neural network technology, constructs an optical diffraction neural network, inputs the information of the input field and the information of the target field into the optical diffraction neural network for training, and the entire training process is iterated multiple times. The imaging efficiency and noise of the image are used to determine whether the iterative training is finished, and the optimal optical diffraction neural network is obtained; the incident wave with control is input to the optimal optical diffraction neural network, and a phase distribution diagram is output. The present invention discloses a method and system for realizing dual-channel near-field control by using optical computing metasurfaces, so as to increase the number of independent channels of optical computing, and can generate independently controllable electric field amplitude distribution for incident waves at different angles.

Description

Technical Field

[0001] The present invention relates to the field of optical neural network technology, and more specifically to a method and system for realizing dual-channel near-field dynamic control by using optical computing metasurface. Background Art

[0002] With the rapid development of communication technology and artificial intelligence technology, as well as the popularization and application of cloud computing and the Internet of Things, human current data networks are facing unprecedented challenges in massive data computing and processing in the information age. At the same time, with the large-scale commercial use of 5G technology and the continuous evolution of the new generation of 6G technology, a large number of new services, new applications and new scenarios have emerged. The exponential growth of information flow has had a huge impact on the existing classical computer system under the von Neumann architecture. Classical electronic computer systems are mainly composed of large-scale electronic components such as transistors. The size and performance of transistors determine the integration, processing speed and stability of the entire computer system. However, due to the limitations of atomic effects and power consumption issues, the compression of transistor volume has been very close to the physical limit. Therefore, high-speed information processing capabilities have become the main bottleneck of current computer systems.

[0003] At present, quantum computing and photon computing, as a new generation of computing methods, have received extensive attention from academia and industry, and have become a research hotspot in the current information technology field. The parallel computing power of quantum computing greatly improves data processing efficiency, but currently quantum computers can only operate in extremely low temperature environments and require specific equipment and expensive electricity costs. Photon computing systems that use photons instead of electrons as information carriers are becoming an important research direction with their unique advantages such as the speed of light, natural parallelism, lower energy consumption, and better operating costs. In the development of recent decades, all-optical computing solutions have provided an effective means to solve the problem of large-scale real-time parallel computing. It can perform specific operations on electromagnetic field amplitude, phase, frequency, polarization and other parameters used for information representation at the speed of light, and basically consumes no energy. Secondly, the high-throughput characteristics of light provide parallel signal processing capabilities for optical computing such as differential operations, convolution and Fourier transform. In addition, existing nanolithography technology provides a process basis for the processing and integration of all-optical computing chips.

[0004] Traditional electronic computing builds large-scale computing systems by increasing the density of semiconductor devices such as transistors within a limited chip area. However, there is a lack of such interaction between photons, making it difficult to increase computing power by increasing density.

[0005] Therefore, how to provide a method and system for realizing dual-channel near-field dynamic control using optical computing metasurfaces to increase the number of independent channels of optical computing and generate independently controllable electric field amplitude distribution for incident waves at different angles has become a technical problem that technical personnel in this field urgently need to solve. Summary of the invention

[0006] In view of this, the present invention provides a method and system for realizing dual-channel near-field dynamic control by using optical computing metasurface, so as to solve the problems in the background technology.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] On the one hand, the present invention discloses a method for realizing dual-channel near-field dynamic control using an optical computing metasurface, and the specific steps are as follows:

[0009] Construct an optical diffraction neural network. The incident wave is incident to the input layer, where it is encoded. The incident wave carrying information is quantized into a matrix. Each element in the matrix is ​​regarded as a new Huygens source, which is connected to each unit on the next layer of optical computing metasurface through diffraction. The cascaded diffraction layer simulates the working mode of a fully connected neural network, and complex electromagnetic responses are achieved by training the parameters of the units on the metasurface.

[0010] Input the information of the input field and the target field into the optical diffraction neural network for training. The entire training process is iterated multiple times. The end of the iterative training is determined based on the imaging efficiency and noise of the image to obtain the optimal optical diffraction neural network.

[0011] Input the regulated incident wave to the optimal optical diffraction neural network and output the phase distribution map.

[0012] Preferably, in the above-mentioned method for realizing dual-channel near-field dynamic control using optical computing metasurface, the specific steps of constructing the optical diffraction neural network are as follows:

[0013] In the forward propagation step, the forward propagation model is constructed using the diffraction coefficient of the i-th neuron in the layer to the corresponding neuron in the next layer, the sum of the inputs of all neurons in the previous layer to the i-th neuron in the l-th layer, and the transmission coefficient of the i-th neuron in the l-th layer;

[0014] Back propagation step: define the loss function as the mean square error between the target field and the actual output field, use gradient descent to minimize the loss function, and use the discrete gradient descent solution process of back propagation to obtain the optimal solution through continuous iteration.

[0015] Preferably, in the above-mentioned method for realizing dual-channel near-field dynamic control using optical computing metasurface, the forward propagation step is specifically as follows:

[0016] According to the Rayleigh-Sommerfeld diffraction equation, the position in the computational layer of the optical diffraction neural network is (x i ,yi ,z i )'s i-th neuron to the neuron at position (x, y, z) in the next layer is:

[0017]

[0018] Where l represents the current layer number, λ is the wavelength of the incident wave, represents the distance between two neurons; the amplitude and phase of the i-th neuron as a secondary Huygens source at (x, y, z) are determined by the product of the incident wave, the transmission coefficient and the diffraction coefficient, written as:

[0019]

[0020] definition represents the sum of the inputs of all neurons in the previous layer to the i-th neuron in the l-th layer, represents the transmission coefficient of the i-th neuron in layer l, is the transmission amplitude, is the transmission phase; the entire forward propagation model is:

[0021]

[0022] Wavefront Information Encoding Layer Located at layer 0, since it is a plane wave incident, the output of layer 0 is:

[0023]

[0024] Assume that the entire all-optical diffractive neural network consists of M computing layers, and the M+1th layer is recorded as the output layer.

[0025]

[0026] Preferably, in the above-mentioned method for realizing dual-channel near-field dynamic control using optical computing metasurface, the back propagation step is specifically as follows:

[0027] In the back propagation model, the loss function is defined as the mean square error between the target field and the actual output field.

[0028]

[0029] Where K is the number of sampling points on the entire output plane, so the back propagation optimization goal is:

[0030]

[0031] In order to minimize the loss function, gradient descent is used to solve it;

[0032]

[0033] Will Written but,

[0034]

[0035] In order to solve the reciprocal containing the absolute value, assume a complex number A = a-bj, then,

[0036]

[0037] Formula (10) is finally simplified to:

[0038]

[0039] When the number of layers is l≤M, the phase gradient is obtained using the chain derivative rule;

[0040] For layer M,

[0041]

[0042]

[0043]

[0044] For layer M-1,

[0045]

[0046] Similarly, for the M-2 layer and the ML layer,

[0047]

[0048]

[0049] Formula (8) to Formula (17) give the discrete gradient descent solution process of the entire back propagation, and the optimal solution is obtained through continuous iteration.

[0050] On the other hand, the present invention discloses a system for realizing dual-channel near-field dynamic control using optical computing metasurface, comprising: an incident wave beam, a metasurface and an image processor;

[0051] The incident wave is incident on the input layer of the metasurface, where it is encoded and quantized into a matrix. Each element of the matrix is ​​regarded as a new Huygens source, which is connected to each unit on the next layer of optical computing metasurface through diffraction. The cascaded diffraction layers simulate the working mode of a fully connected neural network, and complex electromagnetic responses are achieved by training the parameters of the units on the metasurface to output a phase distribution map. The image processor observes the light intensity of different incident object light beams in the detection area to display the imaging efficiency and noise of the image.

[0052] Preferably, in the above-mentioned system for realizing dual-channel near-field dynamic control using optical computing metasurface, the metasurface is composed of 30×30 units, each unit period is 0.6mm, and the size of the entire device is 180×180mm 2 .

[0053] Preferably, in the above-mentioned system for realizing dual-channel near-field dynamic control using optical computing metasurface, the center of the upper layer of the unit of the metasurface is composed of two vertical metal strips and one horizontal metal strip; the periphery is a circle of metal frame, and the bottom layer medium is F4B material with a thickness of h=0.25mm, a dielectric constant of 3, and a loss tangent of 0.002.

[0054] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a method and system for realizing dual-channel near-field dynamic adjustability by using optical computing metasurfaces. Under the conditions of limited integration and volume, based on the principle of all-optical diffraction neural network, incident waves at different angles are regarded as independent channels. Through the propagation and interconnection of diffraction waves between metasurfaces, all-optical matrix calculation of electromagnetic signals between planes is performed, and multi-dimensional light field information is multiplexed in parallel to increase the number of independent channels of optical computing, and independently controllable electric field amplitude distribution can be generated for incident waves at different angles. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0056] FIG1(a) is a schematic diagram of a unit of the present invention; FIG1(b) is a curve showing the variation of the transmittance and phase response of the unit of the present invention with the rotation angle φ; FIG1(c) is the transmittance of the present invention at different incident angles and rotation angles; FIG1(d) is the phase response of the present invention at different incident angles and rotation angles;

[0057] FIG2(a) is a schematic diagram of the working principle of the angle multiplexing metasurface based on a single-layer all-optical diffractive neural network of the present invention; FIG2(b) is a -30° incident target field Z of the present invention; FIG2(c) is a phase plane of a -30° incident wave of the present invention; FIG2(d) is a 30° incident target field U of the present invention; FIG2(e) is a phase plane of a 30° incident wave of the present invention;

[0058] FIG3(a) is the electric field during the training process of the present invention; FIG3(b) is the variation of the error and energy proportion of the present invention with the number of training times; FIG3(c) is the optimal phase distribution of the present invention;

[0059] Figure 4 The robustness analysis effect diagram of the present invention; wherein, (a) the electric field distribution when the incident angle increases from -15° to -45°; (b) the electric field distribution when the incident angle increases from 15° to 45°; (c) the electric field distribution when the incident angle is 30° and the frequency increases from 13 GHz to 17 GHz; (d) the electric field distribution when the incident angle is -30° and the frequency increases from 13 GHz to 17 GHz;

[0060] FIG5(a) is a model in the full-wave simulation software of the present invention; FIG5(b) is a normalized tangential electric field distribution at different positions of the present invention;

[0061] Figure 6 This is the tangential electric field distribution measured in the experiment of the present invention. DETAILED DESCRIPTION

[0062] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0063] The embodiment of the present invention discloses a method for realizing dual-channel near-field dynamic control by using an optical computing metasurface, and the specific steps are as follows:

[0064] Construct an optical diffraction neural network. The incident wave is incident to the input layer, where it is encoded. The incident wave carrying information is quantized into a matrix. Each element in the matrix is ​​regarded as a new Huygens source, which is connected to each unit on the next layer of optical computing metasurface through diffraction. The cascaded diffraction layer simulates the working mode of a fully connected neural network, and complex electromagnetic responses are achieved by training the parameters of the units on the metasurface.

[0065] Input the information of the input field and the target field into the optical diffraction neural network for training. The entire training process is iterated multiple times. The end of the iterative training is determined based on the imaging efficiency and noise of the image to obtain the optimal optical diffraction neural network.

[0066] Input the regulated incident wave to the optimal optical diffraction neural network and output the phase distribution map.

[0067] In order to further optimize the above technical solution, the specific steps of constructing the optical diffraction neural network are as follows:

[0068] In the forward propagation step, the forward propagation model is constructed using the diffraction coefficient of the i-th neuron in the layer to the corresponding neuron in the next layer, the sum of the inputs of all neurons in the previous layer to the i-th neuron in the l-th layer, and the transmission coefficient of the i-th neuron in the l-th layer;

[0069] Back propagation step: define the loss function as the mean square error between the target field and the actual output field, use gradient descent to minimize the loss function, and use the discrete gradient descent solution process of back propagation to obtain the optimal solution through continuous iteration.

[0070] In order to further optimize the above technical solution, the forward propagation steps are as follows:

[0071] According to the Rayleigh-Sommerfeld diffraction equation, the position in the computational layer of the optical diffraction neural network is (x i ,y i ,z i )'s i-th neuron to the neuron at position (x, y, z) in the next layer is:

[0072]

[0073] Where l represents the current layer number, λ is the wavelength of the incident wave, represents the distance between two neurons; the amplitude and phase of the i-th neuron as a secondary Huygens source at (x, y, z) are determined by the product of the incident wave, the transmission coefficient and the diffraction coefficient, written as:

[0074]

[0075] definition represents the sum of the inputs of all neurons in the previous layer to the i-th neuron in the l-th layer, represents the transmission coefficient of the i-th neuron in layer l, is the transmission amplitude, is the transmission phase; the entire forward propagation model is:

[0076]

[0077] Wavefront Information Encoding Layer Located at layer 0, since it is a plane wave incident, the output of layer 0 is:

[0078]

[0079] Assume that the entire all-optical diffractive neural network consists of M computing layers, and the M+1th layer is recorded as the output layer.

[0080]

[0081] In order to further optimize the above technical solution, the back propagation steps are as follows:

[0082] In the back propagation model, the loss function is defined as the mean square error between the target field and the actual output field.

[0083]

[0084] Where K is the number of sampling points on the entire output plane, so the back propagation optimization goal is:

[0085]

[0086] In order to minimize the loss function, gradient descent is used to solve it;

[0087]

[0088] Will Written but,

[0089]

[0090] In order to solve the reciprocal containing the absolute value, assume a complex number A = a-bj, then,

[0091]

[0092] Formula (10) is finally simplified to:

[0093]

[0094] When the number of layers is l≤M, the phase gradient is obtained using the chain derivative rule;

[0095] For the Mth layer,

[0096]

[0097]

[0098]

[0099] For layer M-1,

[0100]

[0101] Similarly, for the M-2 layer and the ML layer,

[0102]

[0103]

[0104] Formula (8) to Formula (17) give the discrete gradient descent solution process of the entire back propagation, and the optimal solution is obtained through continuous iteration.

[0105] Another embodiment of the present invention discloses a system for realizing dual-channel near-field dynamic control using an optical computing metasurface, comprising: an incident wave beam, a metasurface, and an image processor;

[0106] The incident wave is incident on the input layer of the metasurface, where it is encoded and quantized into a matrix. Each element of the matrix is ​​regarded as a new Huygens source, which is connected to each unit on the next layer of optical computing metasurface through diffraction. The cascaded diffraction layers simulate the working mode of a fully connected neural network, and complex electromagnetic responses are achieved by training the parameters of the units on the metasurface to output a phase distribution map. The image processor observes the light intensity of different incident object light beams in the detection area to display the imaging efficiency and noise of the image.

[0107] Specifically, according to the Pancharatnam-Berry principle, the embodiment of the present invention designs a unit as shown in FIG1(a), and the center of the upper layer of the unit is composed of two vertical metal strips and one horizontal metal strip. In order to reduce the mutual coupling after the unit, a circle of metal frame is added around it, and the specific dimensions after optimization are: l = 3.9mm, w = 0.75mm, s = 1.575mm, p = 6mm. The bottom dielectric is F4B material, with a thickness of h = 0.25mm, a dielectric constant of 3, and a loss tangent of 0.002.

[0108] To verify the performance of the unit, CST Microwave Studio software was used for simulation, with the unit cell boundary conditions set in the x and y directions and the Floquet port set in the z direction. The simulation results are shown in Figure 1(b). Since the unit only has electrical resonance, when a right-handed circularly polarized wave is incident, the cross-polarization transmittance is close to the theoretical limit of 0.5. When the rotation angle increases from 0° to 180°, the phase achieves 360° modulation.

[0109] In order to analyze the impact of incident waves at different angles on the performance of the designed unit, as shown in Figure 1(c), when the incident angle changes from -45° to 45°, at the operating frequency of 15 GHz, the transmittance basically remains in the range of 0.48-0.5. At the same time, it can be seen from Figure 1(d) that the transmission phase changes linearly with the change of the rotation angle.

[0110] The schematic diagram of the working principle of the angle-multiplexing metasurface based on a single-layer all-optical diffractive neural network is shown in Figure 2(a). The metasurface consists of 30×30 units, each unit period is 0.6mm, and the size of the entire device is 180×180mm. 2 The target field on the imaging plane corresponds to the output layer in the all-optical diffraction neural network. Figure 2(b) and 2(d) The target field distributions formed by -30° incident waves and 30° incident waves are also discretized into a 30×30 matrix. The white area indicates that the electric field value is 1, and the black area indicates that there is no field. Since the right-hand circularly polarized plane wave is incident obliquely, the amplitudes on all units are equal. The phase surfaces of the -30° incident and 30° incident waves on the metasurface are shown as follows: Figure 2(c) and 2(e) shown.

[0111] The information of the input field and the target field is input into the single-layer all-optical diffraction neural network for training, and the training results are shown in Figure 3. The entire training was iterated 50 times. When the iteration increased to 10, the basic target outline appeared, but the image itself had a large number of breakpoints and the surrounding noise was large. Combined with the error in Figure 3 (b), it can be seen that when the training was from the beginning to the 10th iteration, the error quickly decreased from 1 to 0.4, and the energy of the "Z" and "U" images accounted for 56.9% and 56.8% of the entire imaging surface, respectively. When the training was carried out for 25 times, the error was further reduced to 0.3, and the energy of the "Z" and "U" images increased to 62.7% and 63.7% of the entire imaging surface. The image is also clearer, and the general outline can be clearly seen, but there is still strong noise around the image. As the training iteration reached 50 times, the error gradually converged to 0.25, and the energy of the "Z" and "U" images accounted for 67.7% and 66.6% of the entire imaging surface, respectively. The overall image is very clear and the amplitude is balanced, and the surrounding noise is small. After the training is completed, the required phase distribution is shown in Figure 3(c).

[0112] In order to further study the influence of different incident angles on the electric field distribution generated by the optimized phase, Figure 4(a) and 4(b) show the electric field distribution at different incident angles of -30° to -45° and 30° to 45°. It can be seen that the image quality is not greatly affected within the 5° range of the incident angle set by the bias, and it has good angle robustness. The offset of the incident angle will cause a certain offset in the imaging position. However, when the incident angle is offset to the set angle of 15°, the image begins to blur and the surrounding noise begins to increase. And as the offset of the incident angle increases, the position of the image is further offset. Figure 4 (c) and 4(d) further show the impact of incident waves of different frequencies on the field distribution. It can be seen that within the frequency offset range of 1 GHz, the image is still very clear and has a certain degree of robustness to frequency. However, when the frequency offset reaches 2 GHz, the image becomes further blurred and a large amount of noise appears around it.

[0113] The model shown in Figure 5(a) was established in the full-wave simulation software CST, and the time domain solver was used for simulation. Perfectly-Matched-Layer (PML) boundary conditions were applied in the x and y directions, and free space boundary conditions were used in the z direction. Right-hand circularly polarized plane wave excitation was used. Figure 5(b) shows the normalized tangential electric field distribution obtained by simulation. It can be seen that on the plane at the preset 90mm, the electric field amplitude distribution is as expected, showing clear "Z" and "U" distributions. There is still a clear image distribution at a deviation of 5mm from the target plane, but the image noise and breakpoints will be larger.

[0114] Since the tangential electric field contains two components, x and y, the x and y polarization probes are used to scan the electric field distribution of the two components respectively during the test. The amplitude of the tangential electric field is obtained. The test results are shown in Figure 6 As shown, it can be seen that when a right-hand circularly polarized plane wave is incident on the metasurface at 30°, the electric field amplitude presents a clear "Z" distribution on a plane 90 mm away from the aperture. Similarly, when a right-hand circularly polarized plane wave is incident on the metasurface at -30°, the electric field amplitude presents a clear "U" distribution on a plane 90 mm away from the aperture. The calculation, simulation and experimental results all show that the metasurface designed in this embodiment can achieve independently controllable near-field amplitude regulation for incident waves at different angles.

[0115] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0116] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for realizing dual-channel near-field dynamic control using optical computing metasurfaces, characterized in that: The specific steps are as follows: Construct an optical diffraction neural network. The incident wave is incident on the input layer, and the incident wave is encoded on the input layer. The incident wave carrying the encoded information is quantized into a matrix. Each element on the matrix is ​​regarded as a new Huygens source, which is connected to each unit on the next layer of optical computing metasurface through diffraction. The cascaded diffraction layer simulates the working mode of the fully connected neural network, and the complex electromagnetic response is achieved by training the parameters of the units on the metasurface. Input the information of the input field and the target field into the optical diffraction neural network for training. The entire training process is iterated multiple times. The imaging efficiency and noise of the image are used to determine whether the iterative training is completed, and the optimal optical diffraction neural network is obtained. The incident wave with regulation is input into the optimal optical diffraction neural network, and a phase distribution map is output. The center of the upper layer of the unit is composed of two vertical metal strips and one horizontal metal strip; in order to reduce the mutual coupling after the unit, a circle of metal frame is added around it. The specific dimensions after optimization are: l = 3.9mm, w = 0.75mm, s = 1.575mm, p = 6mm; the bottom layer medium is F4B material, with a thickness of h = 0.25mm, a dielectric constant of 3, and a loss tangent of 0.002; Afterwards, the performance of the unit was verified by simulating it using CST Microwave Studio software, setting the unit cell boundary conditions in the x and y directions and the Floquet port in the z direction. Since the unit only has electrical resonance, when a right-handed circularly polarized wave is incident, the cross-polarization transmittance is close to the theoretical limit of 0.

5. When the rotation angle increases from 0° to 180°, the phase achieves 360° modulation. Since the tangential electric field contains two components, x and y, the x and y polarization probes are used to scan the electric field distribution of the two components respectively during the test. The amplitude of the tangential electric field is obtained; when a right-handed circularly polarized plane wave is incident on the metasurface at 30°, the electric field amplitude presents a clear "Z" distribution on the plane 90 mm away from the aperture; Similarly, when a right-hand circularly polarized plane wave is incident on the metasurface at -30°, the electric field amplitude presents a clear "U" distribution on the plane 90 mm away from the aperture.

2. The method for realizing dual-channel near-field dynamic control using optical computing metasurface according to claim 1, characterized in that: The specific steps of constructing the optical diffraction neural network are as follows: In the forward propagation step, the forward propagation model is constructed using the diffraction coefficient of the i-th neuron in the previous layer to the neuron at the corresponding position in the next layer, the sum of the inputs of all neurons in the previous layer to the i-th neuron in the l-th layer, and the transmission coefficient of the i-th neuron in the l-th layer; Back propagation step: define the loss function as the mean square error between the target field and the actual output field, use gradient descent to minimize the loss function, and use the discrete gradient descent solution process of back propagation to obtain the optimal solution through continuous iteration.

3. The method for realizing dual-channel near-field dynamic control using optical computing metasurface according to claim 2, characterized in that: The forward propagation steps are as follows: According to the Rayleigh-Sommerfeld diffraction equation, the position in the computational layer of the optical diffraction neural network is (x i ,y i ,z i )'s i-th neuron to the neuron at position (x, y, z) in the next layer is: Where l represents the current layer number, λ is the wavelength of the incident wave, represents the distance between two neurons; the amplitude and phase of the i-th neuron as a secondary Huygens source at (x, y, z) are determined by the product of the incident wave, the transmission coefficient and the diffraction coefficient, written as: definition represents the sum of the inputs of all neurons in the previous layer to the i-th neuron in the l-th layer, represents the transmission coefficient of the i-th neuron in layer l, is the transmission amplitude, is the transmission phase; the entire forward propagation model is: Wavefront Information Encoding Layer Located at layer 0, since it is a plane wave incident, the output of layer 0 is: Assume that the entire all-optical diffractive neural network consists of M computing layers, and the M+1th layer is recorded as the output layer.

4. The method for realizing dual-channel near-field dynamic control using optical computing metasurface according to claim 3, characterized in that: The back propagation steps are as follows: In the back propagation model, the loss function is defined as the mean square error between the target field and the actual output field. Where K is the number of sampling points on the entire output plane, so the back propagation optimization goal is: In order to minimize the loss function, gradient descent is used to solve it; Will Written but, In order to solve the reciprocal containing the absolute value, assume a complex number A = a-bj, then, Formula (10) is finally simplified to: When the number of layers is l≤M, the phase gradient is obtained using the chain derivative rule; for the Mth layer, For layer M-1, Similarly, for the M-2 layer and the ML layer, Formula (8) to Formula (17) give the discrete gradient descent solution process of the entire back propagation, and the optimal solution is obtained through continuous iteration.

5. A system for realizing dual-channel near-field dynamic control using optical computing metasurfaces, using a method for realizing dual-channel near-field dynamic control using optical computing metasurfaces as described in any one of claims 1 to 4, characterized in that: include: Incident wave beam, metasurface and image processor; the incident wave is incident on the input layer of the metasurface, where it is encoded and the information-carrying incident wave is quantized into a matrix. Each element on the matrix is ​​regarded as a new Huygens source, which is connected to each unit on the next layer of optical computing metasurface through diffraction. The cascaded diffraction layer simulates the working mode of a fully connected neural network, and realizes complex electromagnetic response by training the parameters of the units on the metasurface, and outputs a phase distribution map; the image processor observes the light intensity of different incident object beams in the detection area to display the imaging efficiency and noise of the image; The metasurface consists of 30×30 units, each unit period is 0.6mm, and the size of the entire device is 180×180mm 2 ; The upper center of the unit of the metasurface is composed of two vertical metal strips and one horizontal metal strip; The outer periphery is a circle of metal frame, and the bottom dielectric is F4B material with a thickness of h = 0.25mm, a dielectric constant of 3, and a loss tangent of 0.002; Since the tangential electric field contains two components, x and y, the x and y polarization probes are used to scan the electric field distribution of the two components respectively during the test. The amplitude of the tangential electric field is obtained; when a right-handed circularly polarized plane wave is incident on the metasurface at 30°, the electric field amplitude presents a clear "Z" distribution on the plane 90 mm away from the aperture.

Images