Optical neural network module based on phonon polariton regulation and design method

Abstract

The application provides an optical neural network module based on phonon polariton regulation and a design method, comprising a heterostructure of a waveguide, a phase change material and hexagonal boron nitride, layers of the neural network are connected through diffraction of phonon polaritons, each unit on the diffraction layer is a sub-wave source of a secondary spherical wave; input of a neuron of an arbitrary layer is output of all neurons of a previous layer, and the input is superimposed on the neuron after diffraction; weight of each neuron is defined as influence of a unit structure on the diffraction layer on phase and amplitude of the phonon polariton; input data is input from an input layer, propagation of waves between diffraction layers is calculated, and output results of an output layer are obtained; then, weight of a neuron of each diffraction layer is continuously optimized through an error back propagation algorithm, a trained neural network is obtained after several cycles, and the neural network result is written on the phase change material through laser with different wavelengths and powers.

Description

Technical Field

[0001] This invention relates to the field of optical neural network computing technology, and in particular to an optical neural network module and design method based on phonon polariton modulation. Background Technology

[0002] In recent years, deep learning has been increasingly widely applied in artificial intelligence, achieving significant progress in fields such as computer vision, speech recognition, and natural language processing. With the massive increase in data volume and advancements in deep neural network architectures, deep learning has revolutionized information processing in many scientific and engineering fields. However, as the capabilities of neural networks increase, traditional electronic computers are approaching their physical limits. Large-scale neural networks consume enormous amounts of computing power and energy when performing inference tasks, thus requiring hardware that offers low power consumption, scalability, and high-speed computing capabilities beyond what existing electronic systems can provide.

[0003] Optical computing, with its inherent parallelism, high efficiency, and high speed, can meet some of these requirements. Because light possesses parallelism and high speed, optical neural networks exhibit high parallel processing capabilities and very high computational speed. Optical neural networks can effectively reduce some of the computational load in software and electronic hardware, providing a promising alternative to artificial neural networks. The most energy- and time-consuming part of artificial neural networks is intensive matrix multiplication. However, in optical neural networks, matrix multiplication can be performed at the speed of light.

[0004] Optical computing is considered one of the most promising technologies of the future. Optical neural networks use light waves as the computing medium to realize the logical analysis and computation functions of neural networks, and have advantages such as low power consumption, low latency, resistance to electromagnetic interference, high bandwidth, high interconnectivity, and parallel processing. Because of the parallelism and high speed of light, optical neural networks have very high parallel processing capabilities and extremely high computational speed. Optical neural networks can effectively reduce some of the computational load in software and electronic hardware, providing a promising alternative to artificial neural networks.

[0005] Compared to light propagating in free space, the wavelength of phonon polaritons can be compressed to 1 / 10, operating in the subwavelength band. Phonon polariton-based neural networks are on-chip neural networks, offering higher integration density and easier interconnection with optoelectronic chips compared to ordinary optical neural networks. The main objective of this application is to provide a design for an optical neural network based on phonon polariton modulation, aiming to offer a neural network with faster computation speed and lower power consumption compared to traditional computer neural networks. Furthermore, phonon polariton-based neural networks have a smaller size and are easily integrated onto a chip. Summary of the Invention

[0006] To address the problems in the existing technology, this application proposes an optical neural network module based on phonon polariton modulation, comprising a heterostructure of waveguide, phase change material, and hexagonal boron nitride. The layers of the neural network are connected by phonon polariton diffraction, with each unit in the diffraction layer serving as a sub-source of a secondary spherical wave. The input to a neuron in any layer is the output of all neurons in the previous layer, superimposed on the output after diffraction. The weight of each neuron is defined as the influence of a specific unit structure in the diffraction layer on the phase and amplitude of the phonon polariton. Data is input from the input layer, wave propagation between diffraction layers is calculated, and the output of the output layer is obtained. Then, the weights of neurons in each diffraction layer are continuously optimized using an error backpropagation algorithm. After several iterations, a trained neural network is obtained. The trained neural network is then inscribed onto the phase change material using lasers of different wavelengths and powers to form a metasurface structure, thereby achieving a classification function.

[0007] Preferably, the neural network is represented in the form of a metasurface. By designing the size, shape and arrangement of the metasurface unit structure, the neural network can achieve the function of four-channel image classification. The output surface includes four observation regions, and the classification structure is represented by the observation region with the highest intensity.

[0008] Preferably, the spacing between diffraction layers is 10 micrometers, the neuron feature size is 10 nanometers, the spacing between the center points of every two neurons is 50 nanometers, and the number of neurons per layer is 1*900.

[0009] A method for designing optical neural networks based on phonon polariton modulation includes the following steps:

[0010] S1. Set the physical parameters of the neural network, including the wavelength of the light source, the size of the neurons, the number of neurons in each layer, and the distance between each layer;

[0011] S2. Calculate the output light field of a single neuron to obtain the light wave transmission formula between adjacent diffraction layers of the neural network;

[0012] S3. Using the results in step S2, obtain the calculation formulas for the input layer, hidden layer, and output layer of the neural network;

[0013] S4. Construct a forward propagation model of the neural network using the results of step S3;

[0014] S5. Optimize the parameters of the neural network using stochastic gradient descent.

[0015] S6. Set up the training and test sets, and start training the neural network;

[0016] S7. The optimized neural network parameters are converted into metasurface structural parameters, and a metasurface structure is fabricated using laser on a heterostructure composed of phase change material and hexagonal boron nitride.

[0017] Preferably, the calculation process in step S2 is as follows:

[0018] The amplitude and relative phase of the secondary wave generated by a node are determined by the incident wave at that node and the node's transmission coefficient t; the output of the i-th node in the l-th layer of the network is:

[0019]

[0020] in, This represents the sum of the diffraction waves from all neurons in layer l+1 to the i-th neuron in layer l; φ represents the projection coefficient, a represents the amplitude coefficient, and φ represents the phase.

[0021]

[0022] Where λ represents the wavelength of light, l represents the l-th layer of the network, and i represents the i-th node with coordinates (x, y). i ,y i ,z i ), Let represent the Euclidean distance between neuron i in the l-th layer and a point (x, y, z) in space.

[0023] Preferably, step S4 includes:

[0024] The input layer is represented as Where 0 represents the input layer, k represents the pixel in the input layer, and p represents the neuron in the hidden layer. Defined as input mode;

[0025] Assuming the neural network has M layers besides the input and output layers, the output layer can be represented as:

[0026]

[0027] Preferably, step S5 includes:

[0028] The network was trained using the backpropagation algorithm and stochastic gradient descent optimization method. The loss function was the output surface light intensity. and target MSE between:

[0029]

[0030] in, Let f(x) represent the true value at the i-th position in the (M+1)-th layer, i.e., the label value of the training set; MSE is the mean of the sum of squares of the differences between the predicted value f(x) and the target value y, and its formula is as follows:

[0031]

[0032] The training objective of a neural network is to make it... To minimize, i.e., to solve:

[0033]

[0034] The i-th node in the l-th layer network The gradient relative to the error can be calculated as follows:

[0035]

[0036] Where K is the number of measurement points on the output plane. The output surface light intensity.

[0037] Preferably, step S6 includes:

[0038] The training and test sets consist of 3*5 pixel patterns, and the output has four channels, with the channel having the highest observation intensity representing the classification result of the neural network. The training set contains 40,000 images, and the test set contains 4,000 images. The specific steps for training the neural network are as follows:

[0039] S61. Process the image data in the dataset and convert the two-dimensional image data into one-dimensional data;

[0040] S62. Set the relevant parameters of the neural network, as described above;

[0041] S63. Train using the training set and then validate using the validation set. The training cycle is 100, and the batch size for each cycle is 128.

[0042] Preferably, step S7 includes:

[0043] The parameters of the trained neural network are converted into the parameters of the metasurface using the formula Δh=λΦ / 2πΔn, where λ represents the wavelength of the incident wave, Φ represents the phase value obtained by training the neural network, π represents pi, and Δn represents the refractive index difference between the metasurface material and the background medium.

[0044] The above-mentioned technical features can be combined in various suitable ways or replaced by equivalent technical features, as long as the purpose of the present invention can be achieved.

[0045] The optical neural network module and design method based on phonon polaritrile modulation provided by this invention have at least the following advantages compared with the prior art:

[0046] 1. By combining deep learning neural networks and optics, the inference process of the neural network is executed at the speed of light, and the energy consumption during the inference process is greatly reduced;

[0047] 2. By utilizing phonon polaritons to modulate and optimize optical neural networks, a smaller size than that of ordinary optical neural networks can be achieved, making on-chip integration easier; Attached Figure Description

[0048] The invention will now be described in more detail with reference to embodiments and the accompanying drawings.

[0049] Figure 1 A schematic diagram showing the physical parameter settings of the neural network of the present invention is displayed;

[0050] Figure 2 This diagram illustrates the process of phonon polaritons propagating between diffraction layers according to the present invention.

[0051] Figure 3 This illustrates the forward propagation model of the neural network of the present invention;

[0052] Figure 4 A schematic diagram of the phonon polariton structure of the present invention is shown. Detailed Implementation

[0053] The invention will now be further described with reference to the accompanying drawings.

[0054] This invention provides an optical neural network based on phonon polariton modulation, which consists of incident light, a heterostructure composed of phase change material Ge3Sb2Te6 (GST) and hexagonal boron nitride (hBN), and a waveguide.

[0055] The layers of the neural network are connected by the diffraction of phonon polaritons; that is, each unit in the diffraction layer is a secondary spherical wavelet source. The input to a neuron in any layer is the output of all neurons in the previous layer, which is then superimposed on that neuron after diffraction. The weight of each neuron is defined as the influence of a certain unit structure in the diffraction layer on the phase and amplitude of the phonon polaritons. Data is input from the input layer, and the propagation of the wave between the diffraction layers is calculated using the diffraction formula to obtain the output result of the output layer. Then, the weights of the neurons in each diffraction layer are continuously optimized using the backpropagation algorithm. After several rounds of iteration, a trained neural network is obtained. This training process is completed by a computer. The trained neural network is then inscribed onto GST material using lasers of different wavelengths and powers to form a metasurface structure, enabling image classification.

[0056] Neural networks are represented in the form of metasurfaces. By designing the size, shape, and arrangement of the metasurface unit structures, the four-channel image classification function of the neural network can be achieved. The output surface is divided into four observation regions, and the classification structure is represented by the observation region with the highest intensity.

[0057] Phonon polaritons are a phenomenon in solid-state physics where phonons are quasiparticles representing the collective motion of atoms in a crystal lattice. Phonons propagate through materials due to the periodic motion of atoms. A phonon polariton structure, composed of Ge3Sb2Te6 (GST) and hBN, allows control over the crystal phase of GST by irradiating it with lasers of varying power and wavelength. The refractive indices of crystalline and amorphous GST differ; this principle enables the fabrication of metasurface structures on GST, achieving various functionalities such as focusing and refraction.

[0058] The present invention discloses a design method for an optical neural network based on phonon polariton modulation, comprising the following steps:

[0059] (1) Set the physical parameters of the neural network, including the light source wavelength, neuron size, number of neurons per layer, and distance between layers, such as... Figure 1 As shown;

[0060] (2) Using the diffraction formula, the output light field of a single neuron is calculated, and the light wave transmission formula between adjacent diffraction layers of the neural network is obtained, such as... Figure 2 As shown;

[0061] (3) Using the results in (2), we obtain the calculation formulas for the input layer, hidden layer and output layer of the neural network.

[0062] (4) Construct the forward propagation model of the neural network using the results in (3), such as Figure 3 As shown;

[0063] (5) Optimize the parameters of the neural network using stochastic gradient descent;

[0064] (6) Set up the training set and test set, and start training the neural network.

[0065] (7) The optimized neural network parameters are converted into metasurface structural parameters, and a metasurface structure is fabricated using laser on a heterostructure composed of GST and hBN. For example... Figure 4 As shown.

[0066] In one embodiment, a laser with a wavelength of 10.6 micrometers is selected as the light source and irradiated onto the heterostructure composed of GST and hBN, which can excite phonon polaritons with a wavelength of 220 nanometers. The spacing between the diffraction layers is selected to be 10 micrometers, the neuron feature size is 10 nanometers, the spacing between the center points of every two neurons is 50 nanometers, and the number of neurons per layer is 1*900.

[0067] In one embodiment, the output light field of a single neuron is calculated using a diffraction formula to obtain the light wave transmission formula between adjacent diffraction layers of the neural network. The diffraction formula used is:

[0068]

[0069] Where λ represents the wavelength of light, l represents the l-th layer of the network, and i represents the i-th node with coordinates . Let represent the Euclidean distance between neuron i in the l-th layer and a point (x, y, z) in space.

[0070] The amplitude and relative phase of the secondary wave generated by this node are determined by the incident wave at the node and the node's transmission coefficient t (both are complex functions). Therefore, the output of the i-th node in the l-th layer of the network is:

[0071]

[0072] in, This represents the sum of the diffraction waves from all neurons in layer l+1 to the i-th neuron in layer l; φ represents the projection coefficient, a represents the amplitude coefficient, a = 1 in this invention (ideally), and φ represents the phase.

[0073] In one embodiment, a forward propagation model of a neural network is constructed. The input layer is represented as... Where 0 represents the input layer, k represents the pixel in the input layer, and p represents the neuron in the hidden layer. Defined as input mode, It can be calculated using the diffraction formula.

[0074] Assuming the neural network has M layers (excluding the input and output layers), the output layer can be represented as:

[0075] In one embodiment, the network is trained using the backpropagation algorithm and stochastic gradient descent optimization method. The loss function is the output surface light intensity. and target MSE between:

[0076]

[0077] in, Let f(x) represent the true value at the i-th position in the (M+1)-th layer, i.e., the label value of the training set. MSE is the mean of the sum of squared differences between the predicted value f(x) and the target value y, and its formula is as follows:

[0078]

[0079] The training objective of a neural network is to make it... To minimize, i.e., to solve:

[0080]

[0081] The i-th node in the l-th layer network The gradient relative to the error can be calculated as follows:

[0082]

[0083] Where K is the number of measurement points on the output plane. To output surface light intensity, It is the true value at the i-th position in the (M+1)-th layer, which is the label value of the training set.

[0084] In one embodiment, the training and test sets are 3*5 pixel patterns, and the output has four channels, where the channel with the highest observation intensity represents the classification result of the neural network. The training set consists of 40,000 images, and the test set consists of 4,000 images. The specific steps for training the neural network are as follows:

[0085] (1) Process the image data in the dataset to convert the two-dimensional image data into one-dimensional data;

[0086] (2) Set the relevant parameters of the neural network, as described above;

[0087] (3) Use the training set for training and then use the validation set for validation. The training epoch is 100 and the batch size for each epoch is 128.

[0088] In one embodiment, the parameters of the trained neural network are converted into the parameters of the metasurface using the formula Δh = λΦ / 2πΔn, where λ represents the wavelength of the incident wave, Φ represents the phase value obtained by training the neural network, π represents pi, and Δn represents the refractive index difference between the metasurface material and the background medium.

[0089] In one embodiment, a metasurface structure is fabricated on a phonon polariton structure. The phonon polariton structure is composed of the phase change material Ge3Sb2Te6 (GST) and hexagonal boron nitride (hBN). By irradiating the GST material with lasers of different powers and wavelengths, the crystal phase of GST can be controlled. Since the refractive indices of GST differ between its crystalline and amorphous states, this principle allows for the fabrication of metasurface structures on GST to achieve various functions, such as focusing and refraction.

[0090] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. An optical neural network module based on phonon polariton modulation, characterized in that, The neural network consists of a heterogeneous structure composed of waveguides, phase change materials, and hexagonal boron nitride. The layers are connected by diffraction of phonon polaritons. Each unit on the diffraction layer is a sub-wave source of a secondary spherical wave. The input of a neuron in any layer is the output of all neurons in the previous layer, which is then superimposed on the neuron after diffraction. The weight of each neuron is defined as the influence of a certain unit structure on the phase and amplitude of the phonon polaritons. Data is input from the input layer, the propagation of the wave between the diffraction layers is calculated, and the output result of the output layer is obtained. Then, the weight of the neurons in each diffraction layer is continuously optimized through the backpropagation algorithm. After several rounds of iteration, a trained neural network is obtained. The results of the trained neural network are then inscribed on the phase change material by lasers of different wavelengths and powers to form a metasurface structure, so as to realize the classification function. The amplitude and relative phase of the secondary wave generated by the node are determined by the incident wave at the node and the node's transmission coefficient. Decision; The first layer of the network There are 1 node, and its output is: ; in, express All neurons in the layer Layer The summation of diffraction waves from each neuron; Indicates the projection coefficient. Indicates the amplitude coefficient. The phase is represented; the output light field of a single neuron is calculated using the diffraction formula, resulting in the light wave propagation formula between adjacent diffraction layers of the neural network. The diffraction formula used is: ; Where λ represents the wavelength of light. Indicates the first Layered network, Indicates the first There are nodes and the coordinates of that node are... , Indicates the first neurons in the layer With a point in space The Euclidean distance between them .

2. The optical neural network module based on phonon polariton modulation according to claim 1, characterized in that, Neural networks are represented in the form of metasurfaces. By designing the size, shape and arrangement of the metasurface unit structure, the four-channel image classification function of the neural network can be realized. The output surface includes four observation regions, and the classification structure is represented by the observation region with the highest intensity.

3. The optical neural network module based on phonon polariton modulation according to claim 1, characterized in that, The spacing between diffraction layers is 10 micrometers, the neuron feature size is 10 nanometers, the spacing between the center points of every two neurons is 50 nanometers, and the number of neurons per layer is 1*900.

4. The optical neural network design method based on phonon polariton modulation according to claim 1, characterized in that, Includes the following steps: S1. Set the physical parameters of the neural network, including the wavelength of the light source, the size of the neurons, the number of neurons in each layer, and the distance between each layer; S2. Calculate the output light field of a single neuron to obtain the light wave transmission formula between adjacent diffraction layers of the neural network; S3. Using the results in step S2, obtain the calculation formulas for the input layer, hidden layer, and output layer of the neural network; S4. Construct a forward propagation model of the neural network using the results of step S3; S5. Optimize the parameters of the neural network using stochastic gradient descent. S6. Set up the training and test sets, and start training the neural network; S7. The optimized neural network parameters are converted into metasurface structural parameters, and a metasurface structure is fabricated using laser on a heterostructure composed of phase change material and hexagonal boron nitride.

5. The optical neural network design method based on phonon polariton modulation according to claim 4, characterized in that, Step S4 includes: The input layer is represented as ,in Indicates the input layer. Represents the cells of the input layer. This represents neurons in the hidden layer. Defined as input mode; Assuming the neural network has M layers besides the input and output layers, then the output layer can be represented as: .

6. The optical neural network design method based on phonon polariton modulation according to claim 5, characterized in that, Step S5 includes: The network was trained using the backpropagation algorithm and stochastic gradient descent optimization method; the loss function was the output surface light intensity. and target MSE between: ; in, Indicates the first Layer The true value at each position is the label value of the training set; MSE is the predicted value. With target value The formula for the mean of the sum of squares of the differences between them is as follows: ; The training objective of a neural network is to make it... To minimize, i.e., to solve: , s.t. ; No. Layer network Nodes The gradient relative to the error is calculated as follows: ; in To output the number of measurement points on the plane, The output surface light intensity.

7. The optical neural network design method based on phonon polariton modulation according to claim 4, characterized in that, Step S6 includes: The training and test sets consist of 3*5 pixel patterns, and the output has four channels, with the channel having the highest observation intensity representing the classification result of the neural network. The training set contains 40,000 images, and the test set contains 4,000 images. The specific steps for training the neural network are as follows: S61. Process the image data in the dataset and convert the two-dimensional image data into one-dimensional data; S62. Set the relevant parameters for the neural network; S63. Train using the training set and then validate using the validation set. The training cycle is 100, and the batch size for each cycle is 128.

8. The optical neural network design method based on phonon polariton modulation according to claim 4, characterized in that, Step S7 includes: The parameters of the trained neural network are converted into the parameters of the metasurface using the formula ∆h = λΦ / 2π∆n, where λ represents the wavelength of the incident wave, Φ represents the phase value obtained by training the neural network, π represents pi, and ∆n represents the refractive index difference between the metasurface material and the background medium.

Images