Vector vortex light generation method based on hybrid diffraction deep neural network

Abstract

The present invention belongs to the field of optical light field control and nonlinear optical technology, and discloses a method for generating vector vortex light based on a hybrid diffraction deep neural network, including building a hybrid diffraction deep neural network constructed by a diffraction neural network layer and an anisotropic polarization layer. During training, the hybrid diffraction deep neural network processes the input light data through a forward propagation model, and optimizes the parameters of each diffraction layer and anisotropic polarization diffraction layer of the hybrid diffraction deep neural network through a backpropagation algorithm, and obtains the optimization of each parameter for the training data set. During testing, when Gaussian light is input at different spatial positions, the required vector vortex light can be obtained by the output layer. The present invention applies the hybrid diffraction deep neural network to the field of vector vortex light generation, realizing a new and rapid method for generating vector vortex light, which has great application potential in particle capture, object structure imaging, improving resolution, holographic display, and increasing communication capacity.

Description

Technical Field

[0001] The present invention belongs to the field of optical light field control and nonlinear optics technology, and specifically relates to a method for generating vector vortex light based on a hybrid diffraction deep neural network. Background Art

[0002] Vector vortex light is a type of structured beam with a unique spiral wavefront and spatially varying polarization state. This beam possesses both inseparable polarization and orbital angular momentum (OAM) states, and both polarization and OAM can be used to encode information, thus finding important applications in many fields. For example, in optical sensing, vector vortex beams can be used to measure parameters such as the rotation angle and angular velocity of an object. In optical communications, vector vortex beams can improve the signal transmission rate and capacity of optical communication systems. The photon orbital angular momentum (OAM) they carry, as a new physical dimension, possesses an infinite number of orthogonal modes, theoretically providing an infinite number of independent information channels, thereby significantly expanding the bandwidth of communication systems. In quantum computing, vector vortex beams can be used to construct quantum bits and achieve the creation of quantum superposition and entangled states, providing fundamental support for quantum computing. Vector vortex light can also be used in optical microscopy and laser imaging, where its unique phase distribution and polarization properties help improve the resolution of optical microscopy and the clarity and accuracy of imaging.

[0003] There are many methods for generating vector vortex beams. For example, by combining a plane wavefront analyzer and a folding mirror, vector vortex light can be generated by precisely controlling the reflection and phase change of the light beam. However, this method requires precise design and debugging of the optical system. In 2016, Liu et al. also proposed a method for generating arbitrary vector vortex beams on a mixed-order Poincare sphere by combining a Q-plate and a spiral phase plate. In 2017, Mamani et al. used a spatial light modulator and a vortex retarder in series to generate vector vortex beams. In recent years, with the emergence of metamaterials, Liu et al. used an all-dielectric metasurface platform in 2021 to generate broadband beams on a generalized perfect Poincare sphere. The above methods usually require complex optical paths and bulky optical elements. At the same time, the size of the sampling pixels is generally on the micron scale, which greatly limits the spatial density of the generated WB. Generating large-scale WBs will face great challenges and lack flexibility.

[0004] In 2018, Lin et al. proposed the Diffractive Deep Neural Network (DDNN), which is an optical network based on the Huygens-Fresnel diffraction theory. Once trained, it can realize a variety of complex machine learning functions at the speed of light and with low power consumption. However, in the existing DDNN, each neuron in each diffraction layer, that is, each pixel in the diffraction layer, is regarded as a wavelet source. According to the Rayleigh-Sommerfeld Equation, that is, the scalar light diffraction theory, the interconnection with the neurons in the next layer is realized. Therefore, the existing DDNN cannot be directly applied to the generation of vector vortex beams. For example, in 2021, Huang et al. proposed a vortex beam generation method based on a diffraction deep neural network. The polarization direction of the vortex light generated by this network cannot be controlled. Summary of the Invention

[0005] In order to solve the above problems, the present invention provides a method for generating vector vortex light based on a hybrid diffraction deep neural network. This method only uses a layer of anisotropic polarized diffraction layer and a multi-layer diffraction deep neural network diffraction layer to process the polarized light beam by utilizing the anisotropic characteristics of a single layer of anisotropic neurons, and successfully realizes the generation of vector vortex light with different orbital angular momentum modes, greatly reducing the material utilization rate and improving the adjustability and operability of the experiment. Compared with traditional vector vortex light generation methods, this method is more operable.

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

[0007] The present invention is a method for generating vector vortex light based on a hybrid diffraction deep neural network, which specifically includes the following steps:

[0008] Step 1: Build a hybrid diffraction deep neural network constructed by a diffraction neural network layer and an anisotropic polarization layer, wherein the hybrid diffraction deep neural network includes an input layer, a hidden layer, and an output layer. The hidden layer is composed of M diffraction deep neural network diffraction layers and 1 anisotropic polarization diffraction layer, where M>1. The hybrid diffraction deep neural network processes light data through a forward propagation model, and optimizes the phase parameters of the diffraction deep neural network diffraction layer and the phase and angle parameters of the anisotropic polarization diffraction layer through a backpropagation algorithm.

[0009] Step 2: Prepare data in the training dataset, provide different spatial position data in the input layer and corresponding different vector vortex beam data in the output layer;

[0010] Step 3. Pass the input light wave into the hybrid diffraction deep neural network built in step 1 for forward propagation: after the input light wave passes through the input layer, it is diffracted to reach the first diffraction deep neural network diffraction layer, and the input light wave is modulated by the first diffraction deep neural network diffraction layer to obtain the output light wave of the first diffraction deep neural network diffraction layer. After the input light wave is modulated by the a-th diffraction deep neural network diffraction layer, the output light wave of the a-th diffraction deep neural network diffraction layer is obtained. Similarly, after the light wave is modulated by the a-th diffraction deep neural network diffraction layer and diffracted, it reaches Reaching the anisotropic polarization diffraction layer, the input light wave of the anisotropic polarization diffraction layer is the sum of the output wave diffracted by the a-th layer and reaching the anisotropic polarization diffraction layer, and then the output light wave of the anisotropic polarization diffraction layer is diffracted again by the diffraction deep neural network diffraction layer of the b-th layer, and finally outputted by the light detector in the output layer, wherein a+b=M, 1≤a≤M, 1≤b≤M, the input light wave is a Gaussian beam at different positions on the input layer, and the output light wave in the output layer is the required vector vortex light. Gaussian beams at different input positions generate vector vortex light with different orbital angular momentum patterns at the output end;

[0011] Step 4. Optimize the phase parameters of the diffraction layer of the diffraction deep neural network, optimize the phase and angle parameters of the anisotropic polarization diffraction layer, and optimize the hybrid diffraction deep neural network. After multiple performance training of the data in the training data set in step 2, the optimal parameter values ​​of the phase parameters of the diffraction deep neural network diffraction layer and the optimal phase and angle parameters of the anisotropic polarization diffraction layer are obtained, and the optimal parameter values ​​of the hybrid diffraction deep neural network are optimized. During the test, Gaussian beams at different positions at the input plane pass through the hybrid diffraction deep neural network to obtain vector vortex light with different orbital angular momentum patterns.

[0012] A further improvement of the present invention is that in step 3, the input light wave is represented by the Jones matrix as E=[Ex, Ey] T , the forward propagation of the input light wave between the input layer, hidden layer and output layer satisfies the Huygens-Fresnel diffraction theory:

[0013]

[0014] in, represents the distance from the light beam to the i-th node, l represents the l-th layer of the hybrid diffraction deep neural network, l=1,2……M+1,i represents the i-th node of the l-th layer of the diffraction deep neural network, (x,y,z) represents the coordinates of the light beam, (x i ,y i , z i ) represents the coordinate of the i-th node, and λ is the wavelength of the light beam.

[0015] A further improvement of the present invention is that in step 3, when the input light wave reaches the diffraction layer of the diffraction deep neural network, the diffraction deep neural network diffraction layer modulates the input light wave, and the modulated output light wave is determined by the input light wave in each direction and the transmission coefficient in each direction. Since the diffraction deep neural network diffraction layer has no anisotropy, the same phase modulation is performed on the light field in the x-polarization direction and the y-polarization direction. The output of the i-th node in the l-th layer is expressed as:

[0016]

[0017] in, and They represent the x-polarization direction and y-polarization direction input waves of the i-th node in the l-th layer, respectively. g represents the light wave set of all the output waves of the l-1-th layer propagating through different paths to the node i in the l-th layer. is the diffraction layer modulation parameter of the diffraction deep neural network, The diffraction layer modulation parameters of the diffraction deep neural network are determined only by the phase composition.

[0018] The further improvement of the present invention is that: in the step 1, the phase and angle parameters of the anisotropic polarization diffraction layer are specifically: the angle ψ between the anisotropic neuron and the horizontal x-axis, the phase modulation parameter of the anisotropic neuron eigenmode in the s-axis direction and the phase modulation parameters of the anisotropic neuron eigenmode in the f-axis direction

[0019] A further improvement of the present invention is that in step 3, the process of polarized light propagating forward through the hybrid diffraction deep neural network is specifically as follows:

[0020] The input light wave is modulated and propagated by the a-layer diffraction deep neural network diffraction layer, and then reaches the anisotropic polarization diffraction layer.

[0021] Before the input light wave reaches the anisotropic polarization diffraction layer, it needs to be transformed into a superposition of anisotropic neuron eigenmodes, that is, the xy axis coordinates are transformed to the s-axis and f-axis directions of the neuron, which is obtained by coordinate rotation transformation:

[0022]

[0023] in, represents the x-polarization input light wave of the anisotropic polarization diffraction layer neuron, represents the input light wave of the anisotropic polarization diffraction layer neuron in the y-polarization direction, represents the input light wave component in the s-axis direction of the anisotropic neuron eigenmode after coordinate transformation, represents the input light wave component in the f-axis direction of the anisotropic neuron eigenmode after coordinate transformation, k indicates that the layer is an anisotropic polarization diffraction layer. After the input light wave reaches the anisotropic polarization diffraction layer, it is modulated by the anisotropic neurons, and each direction obtains an independent phase delay. The Jones vector of the outgoing light of the anisotropic neuron is:

[0024]

[0025] in, and is the phase modulation parameter in each direction, s and f represent the s-axis and f-axis of the anisotropic neuron eigenmode;

[0026] The outgoing light of anisotropic neurons should be expressed in the form of an xy coordinate system, so the coordinate rotation transformation matrix is ​​obtained:

[0027]

[0028] Therefore, the transformation of the incident light wave by the anisotropic neuron can be written as:

[0029]

[0030] Among them, the modulation coefficient of the diffraction layer of anisotropic polarization is:

[0031]

[0032] The outgoing light then passes through the b-layer diffraction deep neural network diffraction layer and reaches the output layer.

[0033] A further improvement of the present invention is that in step 4, the optimization of the phase parameters of the diffraction layer of the diffraction deep neural network and the optimization of the phase and angle parameters of the anisotropic polarization diffraction layer are achieved through the Adam back propagation algorithm.

[0034] The beneficial effects of the present invention are as follows: the present invention applies anisotropic neurons to the diffraction deep neural network, and through the combination of multiple layers of DDNN diffraction layers and one layer of anisotropic metasurface, it realizes the effective modulation of polarized light and successfully generates vector vortex light, thereby improving the operability of the experiment, obtaining a hybrid diffraction deep neural network, and avoiding the large demand for experimental materials, especially anisotropic metasurface materials.

[0035] Based on the adjustability of the phase modulation parameters of each diffraction layer and the flexibility of the angle between the anisotropic neurons and the horizontal direction, the network can generate vector vortex light with different orbital angular momentum patterns. At the same time, it reduces the requirements for the precise placement angle of the anisotropic neurons in each layer, simplifying the experimental operation difficulty of generating vector vortex beams. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Schematic diagram of the hybrid diffraction deep neural network of the present invention.

[0037] Figure 2 This is a schematic diagram of vector vortex light with different orbital angular momentum modes corresponding to different Gaussian light positions in the input area.

[0038] Figure 3 This is the simulation result of vector vortex light generated by a middle anisotropic polarization diffraction layer and two diffraction layers in front and behind.

[0039] Figure 4 It is the simulation result of vector vortex light generated by a middle anisotropic polarized diffraction layer and two diffraction layers in front and behind.

[0040] Figure 5 This is the simulation result of vector vortex light generated by a middle anisotropic polarized diffraction layer and three diffraction layers in front and behind. DETAILED DESCRIPTION

[0041] The following diagrams illustrate embodiments of the present invention. For clarity, many practical details are included in the following description. However, it should be understood that these practical details are not intended to limit the present invention. In other words, in some embodiments of the present invention, these practical details are not essential.

[0042] like Figure 1 As shown, the present invention is a method for generating vector vortex light based on a hybrid diffraction deep neural network, which specifically includes the following steps:

[0043] Step 1: Build a hybrid diffraction deep neural network constructed by a diffraction neural network layer and an anisotropic polarization layer, wherein the hybrid diffraction deep neural network includes an input layer, a hidden layer, and an output layer. The hidden layer is composed of M diffraction deep neural network diffraction layers and 1 anisotropic polarization diffraction layer, where M>1. The hybrid diffraction deep neural network processes light data through a forward propagation model, and optimizes the phase parameters of the diffraction deep neural network diffraction layer and the phase and angle parameters of the anisotropic polarization diffraction layer through a backpropagation algorithm.

[0044] Step 2: Prepare data in the training dataset, provide different spatial position data in the input layer and corresponding different vector vortex beam data in the output layer;

[0045] Step 3. Pass the input light wave into the hybrid diffraction deep neural network built in step 1 for forward propagation: after the input light wave passes through the input layer, it is diffracted to reach the first diffraction deep neural network diffraction layer, and the input light wave is modulated by the first diffraction deep neural network diffraction layer to obtain the output light wave of the first diffraction deep neural network diffraction layer. After the input light wave is modulated by the a-th diffraction deep neural network diffraction layer, the output light wave of the a-th diffraction deep neural network diffraction layer is obtained. And so on, after the input light wave is modulated by the a-th diffraction deep neural network diffraction layer and after diffraction propagation, , reaches the anisotropic polarization diffraction layer, the input light wave of the anisotropic polarization diffraction layer is the sum of the output wave diffracted by the a-th layer and reaching the anisotropic polarization diffraction layer, and then the output wave of the anisotropic polarization diffraction layer is diffracted again by the diffraction deep neural network diffraction layer of the b-th layer, and finally outputted by the light detector in the output layer, wherein a+b=M, 1≤a≤M, 1≤b≤M, the input light wave is a Gaussian beam at different positions on the input layer, and the output light wave in the output layer is the required vector vortex light. Gaussian beams at different input positions generate vector vortex light with different orbital angular momentum patterns at the output end;

[0046] Step 4. Optimize the phase parameters of the diffraction layer of the diffraction deep neural network, optimize the phase and angle parameters of the anisotropic polarization diffraction layer, and optimize the hybrid diffraction deep neural network. After multiple performance training of the data in the training data set in step 2, the optimal parameter values ​​of the phase parameters of the diffraction layer of the diffraction deep neural network and the optimal phase and angle parameters of the anisotropic polarization diffraction layer are obtained, and the optimal parameter values ​​of the hybrid diffraction deep neural network are optimized. During the test, after Gaussian beams at different positions at the input plane pass through the hybrid diffraction deep neural network, vector vortex light with different orbital angular momentum patterns can be obtained.

[0047] Example 1

[0048] When a=2, b=2, that is, an anisotropic polarization diffraction layer in the middle and two diffraction deep neural network diffraction layers in the front and back.

[0049] In the forward propagation model in step 1, the forward propagation of polarized light between layers satisfies the Huygens-Fresnel diffraction theory. Based on this, when polarized light reaches the diffraction layer, the diffraction layer modulates the polarized light. Because the neurons in the diffraction layer are non-anisotropic, the same phase modulation is performed on both the x- and y-polarization directions:

[0050]

[0051] in, and They represent the x-polarization direction and y-polarization direction input light waves of the i-th node in the l-th layer, respectively. g represents the light wave set of all output light waves of the l-1-th layer that propagate through different paths to the node i in the l-th layer. is the diffraction layer modulation parameter of the diffraction deep neural network, The diffraction layer modulation parameters of the diffraction deep neural network are determined only by the phase That is, the modulation of the diffraction layer only modulates the phase. In the actual simulation process, the network size of each layer is 28*28; It is a 28*28 phase matrix.

[0052] Before the polarized light reaches the anisotropic diffraction layer, it is first polarized according to the angle ψ between the anisotropic neurons and the horizontal direction. Then, the light fields in the two eigenmode directions of the anisotropic neurons are phase-modulated. Finally, the polarized light is represented as light fields in the x and y polarization directions through coordinate transformation:

[0053]

[0054] Among them, the value range of ψ is 0-2Π, and Both are 28*28 phase matrices.

[0055] The input plane in step 2 is divided into 16 small areas. Gaussian light can be input into each of the 16 different areas. The input Gaussian light at different regional positions will generate vector vortex light with different orbital angular momentum patterns on the output plane.

[0056] In the back-propagation model in step 3, the selected loss function is the MSE loss function, and the optimization method is Adam.

[0057] According to the above steps, a hybrid diffraction deep neural network was successfully designed to achieve the generation of vector vortex light with different orbital angular momentum patterns. The specific simulation results are as follows: Figure 4 As shown. Input Gaussian light at different positions can generate vector vortex light with different orbital angular momentum patterns, such as Figure 2 More specifically, when Gaussian light is in the left half of the input plane, the generated vector vortex light is in the y-polarization direction; when Gaussian light is in the right half of the input plane, the generated vector vortex light is in the x-polarization direction; when Gaussian light is in the upper half of the input plane, the generated vector vortex light has a positive topological charge, and when Gaussian light is in the lower half of the input plane, the generated vector vortex light has a negative topological charge. Using an anisotropic layer diffraction deep neural network, vector vortex light with 16 different orbital angular momentum patterns was generated.

[0058] Example 2

[0059] When a=3, b=3, it means an anisotropic polarization diffraction layer in the middle and three diffraction deep neural network diffraction layers in the front and back.

[0060] The present invention successfully designed a hybrid diffraction deep neural network and realized the generation of vector vortex light with different orbital angular momentum modes. The specific simulation results are as follows: Figure 5 As shown. Input Gaussian light at different positions can generate vector vortex light with different orbital angular momentum patterns, such as Figure 2 As shown. More specifically, when the Gaussian light is in the left half of the input plane, the vector vortex light generated is the vector vortex light in the y-polarization direction; when the Gaussian light is in the right half of the input plane, the vector vortex light generated is the vector vortex light in the x-polarization direction; when the Gaussian light is in the upper half of the input plane, the topological charge of the vector vortex light generated is positive, and when the Gaussian light is in the lower half of the input plane, the topological charge of the vector vortex light generated is negative. Through the anisotropic layer diffraction deep neural network, 16 vector vortex lights with different orbital angular momentum modes are generated. Compared with Example 1, the accuracy of the vector vortex light generated by Example 2 in the output plane will be higher, but it also consumes more resources, and the original 5-layer network has been reduced to 7 layers.

[0061] Figure 3 This is the simulation result of the vector vortex light generated by the middle anisotropic polarization diffraction layer and the front and back diffraction layers, such as Figure 3 As shown, the present invention can generate vector vortex light with different orbital angular momentum modes.

[0062] The present invention applies a hybrid diffraction deep neural network to the field of vector vortex light generation, realizing a new method for rapid generation of vector vortex light, which has great application potential in particle capture, object structure imaging, improving resolution, holographic display and increasing communication capacity.

[0063] The foregoing is merely an embodiment of the present invention and is not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.

Claims

1. A method for generating vector vortex light based on a hybrid diffraction deep neural network, characterized by: The vector vortex light generation method specifically comprises the following steps: Step 1: Build a hybrid diffraction deep neural network constructed by a diffraction deep neural network diffraction layer and an anisotropic polarization diffraction layer, wherein the hybrid diffraction deep neural network includes an input layer, a hidden layer, and an output layer, wherein the hidden layer is composed of M diffraction deep neural network diffraction layers and 1 anisotropic polarization diffraction layer, where M>1, and the hybrid diffraction deep neural network is used to process light data through a forward propagation model, and optimize the phase parameters of the diffraction deep neural network diffraction layer and the phase and angle parameters of the anisotropic polarization diffraction layer through a backpropagation algorithm; Step 2: Prepare data in the training data set to provide different spatial position data of the input layer and corresponding different vector vortex beam data in the output layer; Step 3. Pass the input light wave into the hybrid diffraction deep neural network built in step 1 for forward propagation: after the input light wave passes through the input layer, it is diffracted to reach the first diffraction deep neural network diffraction layer, and the input light wave is modulated by the first diffraction deep neural network diffraction layer to obtain the output light wave of the first diffraction deep neural network diffraction layer. After the input light wave is modulated by the a-th diffraction deep neural network diffraction layer, the output light wave of the a-th diffraction deep neural network diffraction layer is obtained. Similarly, after the light is modulated by the a-th diffraction deep neural network diffraction layer and diffracted, it reaches The input light wave of the anisotropic polarization diffraction layer reaches the anisotropic polarization diffraction layer, and the input light wave of the anisotropic polarization diffraction layer is the sum of the output wave of the a-th layer and reaches the anisotropic polarization diffraction layer. Then the output wave of the anisotropic polarization diffraction layer is diffracted again by the diffraction deep neural network diffraction layer of the b-th layer, and finally outputted by the light detector in the output layer, wherein a+b=M, 1≤a<M, 1≤b<M, the input light wave is the polarized Gaussian beam at different positions on the input layer, and the output light wave in the output layer is the required vector vortex light. Gaussian beams at different input positions generate vector vortex light with different orbital angular momentum patterns at the output end. Step 4. Optimize the hybrid diffraction deep neural network by optimizing the phase parameters of the diffraction layer of the diffraction deep neural network and optimizing the phase and angle parameters of the anisotropic polarization diffraction layer; after multiple performance training of the data in the training data set in step 2, the optimal parameter values ​​of the phase parameters of the diffraction deep neural network diffraction layer and the optimal phase and angle parameters of the anisotropic polarization diffraction layer are obtained, and the optimal parameter values ​​of the hybrid diffraction deep neural network are obtained. During the test, after Gaussian beams at different positions at the input plane pass through the hybrid diffraction deep neural network, vector vortex light with different orbital angular momentum patterns is obtained.

2. The method for generating vector vortex light based on a hybrid diffraction deep neural network according to claim 1, characterized in that: In step 3, the input light wave is represented by the Jones matrix as E=[Ex,Ey] T , the forward propagation of the input light wave between the input layer, hidden layer and output layer satisfies the Huygens Fresnel diffraction theory: in, represents the distance from the light beam to the i-th node, l represents the l-th layer of the hybrid diffraction deep neural network, l=1,2……M+1,i represents the i-th node of the l-th layer of the diffraction deep neural network, (x,y,z) represents the coordinates of the light beam, (x i ,y i ,z i ) represents the coordinate of the i-th node, and λ is the wavelength of the light beam.

3. The method for generating vector vortex light based on a hybrid diffraction deep neural network according to claim 1, characterized in that: In step 3, when the input light wave reaches the diffraction layer of the diffraction deep neural network, the diffraction deep neural network diffraction layer modulates the input light wave. The modulated output light wave is determined by the input light wave in each direction and the transmission coefficient in each direction. Since the diffraction deep neural network diffraction layer has no anisotropy, the light field in the x-polarization direction and the y-polarization direction is modulated with the same phase. The output of the i-th node in the l-th layer is expressed as: in, and They represent the x-polarization direction and y-polarization direction input waves of the i-th node in the l-th layer, respectively. g represents the light wave set of all the output waves of the l-1-th layer propagating through different paths to the node i in the l-th layer. is the diffraction layer modulation parameter of the diffraction deep neural network, The diffraction layer modulation parameters of the diffraction deep neural network are determined only by the phase composition.

4. The method for generating vector vortex light based on a hybrid diffraction deep neural network according to claim 1, characterized in that: In step 1, the phase and angle parameters of the anisotropic polarization diffraction layer are specifically: the angle ψ between the anisotropic neuron and the horizontal x-axis, the phase modulation parameter of the anisotropic neuron eigenmode in the s-axis direction and the phase modulation parameters of the anisotropic neuron eigenmode in the f-axis direction 5. The method for generating vector vortex light based on a hybrid diffraction deep neural network according to claim 4, characterized in that: In step 3, the process of the polarized Gaussian beam propagating forward through the hybrid diffraction deep neural network is specifically as follows: The input light wave is modulated and propagated by the a-layer diffraction deep neural network diffraction layer, and then reaches the anisotropic polarization diffraction layer. Before the input light wave reaches the anisotropic polarization diffraction layer, it needs to be transformed into a superposition of anisotropic neuron eigenmodes, that is, the xy axis coordinates are transformed to the s-axis and f-axis directions of the neuron, which is obtained by coordinate rotation transformation: in, represents the x-polarization input wave of the anisotropic polarization diffraction layer neurons, represents the y-polarization input wave of the anisotropic polarization diffraction layer neuron, represents the input wave component in the s-axis direction of the anisotropic neuron eigenmode after coordinate transformation, represents the input wave component in the f-axis direction of the anisotropic neuron eigenmode after coordinate transformation. The k layer is the anisotropic polarization diffraction layer. After the input light wave reaches the anisotropic polarization diffraction layer, it is modulated by the anisotropic neuron and obtains independent phase delay in each direction. The Jones vector of the outgoing light of the anisotropic neuron is: in, and is the phase modulation parameter in each direction, s and f represent the s-axis and f-axis of the anisotropic neuron eigenmode; The outgoing light of anisotropic neurons should be expressed in the form of an xy coordinate system, so the coordinate rotation transformation matrix is ​​obtained: Therefore, the transformation of the incident light by anisotropic neurons can be written as: Among them, the modulation coefficient of the anisotropic polarization diffraction layer is: The outgoing light then passes through the b-layer diffraction deep neural network diffraction layer and reaches the output layer.

6. The method for generating vector vortex light based on a hybrid diffraction deep neural network according to claim 1, characterized in that: In step 4, the optimization of the phase parameters of the diffraction deep neural network diffraction layer and the optimization of the phase and angle parameters of the anisotropic polarization diffraction layer are achieved through the Adam back propagation algorithm.

Images