Method for controlling spatial wave field intensity distribution based on generative neural network

Abstract

The invention discloses a method for controlling spatial wave field intensity distribution based on a generative neural network, and the method comprises the steps: building a training network through employing the generative neural network, designing the phase distribution of a metasurface through the network through employing any electric field, modulating a target electric field through employing the metasurface built through the phase distribution, and achieving the control of the spatial wave field intensity distribution. The target electric field is any electric field input into the neural network, and the generative neural network is composed of a generator and a discriminator. The generator is composed of eight convolutional layers and eight transposed convolutional layers, the input data is 1 * 256 * 256, and the output data is 1 * 256 * 256; the generative neural network is trained by using phase images and electric field cross-section images of 4841 metasurfaces as data sets, the working frequency band of the metasurfaces is 0.8 THz, and the image size of the data sets is 256 * 256 gray-scale maps; the construction of the unknown electric field has high flexibility, and the unknown electric field is designed by using the existing electric field, or the unknown electric field is directly drawn as required.

Description

Technical Field

[0001] The present invention belongs to the field of physical electronics, and specifically relates to a method for indirectly manipulating an electric field by manipulating a super surface using a neural network. Background Art

[0002] With the rapid development of information technology, the operating frequency and integration of devices are constantly improving, and the electromagnetic environment is becoming more and more complex, posing new challenges to the high-degree-of-freedom control of electromagnetic waves.

[0003] The metasurface uses unit structures to control the phase and realize macroscopic electromagnetic properties that do not exist in nature. However, the traditional design method is mainly based on the generalized Snell's law, which determines the phase distribution of the metasurface through complex mathematical calculations. Although this method is effective, it is relatively fixed and lacks sufficient flexibility. In the field of electric field control, this method provides a reliable theoretical basis and calculation means on the one hand, but also limits the innovation and diversity of the design on the other hand. With the improvement of computing power and the development of machine learning technology, neural network control has become possible.

[0004] The existing neural network of metasurfaces can realize the phase prediction of unit structures on the one hand, and the inverse design of holographic images on the other hand. Although the phase of the unit structure can be accurately predicted, the electric field modulation still needs to rely on physical calculations. The inverse design of holographic images can customize images with fixed focal lengths, but this modulation is achieved in two-dimensional space. At present, there is a lack of methods to customize three-dimensional electric fields. Summary of the invention

[0005] Technical problem: The purpose of the present invention is to provide a method for controlling the distribution of spatial wave field intensity based on a generative neural network. The neural network uses arbitrary electric fields to design metasurfaces, thereby realizing arbitrary electric field manipulation. It can customize unknown electric fields through known electric fields, or directly customize unknown electric fields.

[0006] Technical solution: In order to achieve the above-mentioned purpose, the present invention proposes a method for controlling the spatial wave field intensity distribution based on a generative neural network as follows: A training network is constructed using a generative neural network. The network uses an arbitrary electric field to design the phase distribution of the metasurface. The metasurface constructed using the phase distribution modulates the target electric field, and the target electric field is the arbitrary electric field input into the neural network.

[0007] The generative neural network consists of a generator and a discriminator.

[0008] The generator consists of eight convolutional layers and eight transposed convolutional layers, the input data is 1*256*256, and the output data is 1*256*256.

[0009] The generative neural network has a learning rate of 0.0002, a batchsize of 256, activation functions including a LeakyReLU function and a ReLU activation function, and a loss function of BCEWithLogitsLoss.

[0010] The generative neural network is trained using a data set of 4841 phase images and electric field cross-section images of the metasurface, the working frequency band of the metasurface is 0.8 THz, and the image size of the data set is a grayscale image of 256*256.

[0011] In the data set, the metasurface is an all-silicon metasurface based on the PB phase, the substrate thickness is 600um, and the unit structure period is 110um.

[0012] In the eight convolutional layers, before the second convolution starts, the Leaky ReLU activation layer is entered, and after the second convolution, the batch normalization layer is entered. The third to seventh convolutional layers are consistent with the second layer, and the batch normalization layer is not entered after the last convolutional layer.

[0013] In the eight transposed convolutional layers, before entering the first transposed convolutional layer, the network enters the ReLU activation layer, after the transposed convolutional layer, the network enters the batch normalization layer, and after the last transposed convolutional layer, the network enters the Tanh activation function layer.

[0014] The discriminator has three convolutional layers, the input data is 2*256*256, and the output data is 1*256*256; The discriminator enters the Leaky ReLU activation layer after the first convolution layer, and enters the batch normalization layer and ReLU activation layer after the second convolution.

[0015] The target electric field is obtained by normalizing the electric field of the focal section to obtain an electric field matrix, and then multiplying the focal length by the electric field matrix to obtain the target electric field. At this time, the matrix is ​​the modified target electric field; a circle is drawn at any position in the grayscale image, and the circle is the focus, and the grayscale value of the circle is the focal length. At this time, the grayscale image is the drawn target electric field.

[0016] Beneficial effect: The generative neural network proposed in the present invention can generate a metasurface phase according to any input electric field, and the metasurface constructed using the phase can modulate the input electric field.

[0017] In the present invention, the input electric field data can be modified by using the existing electric field information, or can be directly drawn.

[0018] In the present invention, the neural network can manipulate the position of the focus in three-dimensional space. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1Schematic diagram of a method for controlling spatial wave field intensity distribution based on a generative neural network.

[0020] Figure 2 It is the randomly generated metasurface electric field and phase for training, which is based on the metasurface electric field obtained by Lumerical simulation; Figure 2 (a) is the phase and electric field cross-section diagram of a single focus. Figure 2 (b) in the figure is the phase and electric field cross-section diagram of the dual focus.

[0021] Figure 3 Schematic diagram of the unit structure of the metasurface. In the figure, L is the length of the unit structure, W is the width of the unit structure, H1 is the height of the unit structure, H2 is the thickness of the substrate, θ is the rotation angle of the unit structure, and Si is the material of the unit structure.

[0022] Figure 4 It is a schematic diagram of the generator of the network.

[0023] Figure 5 is a schematic diagram of the discriminator of the network.

[0024] Figure 6 is a modified arbitrary electric field and result diagram, where the modified electric field is modified from the existing 35mm focal length electric field to the 20mm focal length electric field. Figure 6 (a) in the figure is the modified target electric field, Figure 6 (b) is a hypersurface designed by a neural network. Figure 6 (c) in the figure is the electric field diagram of the metasurface modulation.

[0025] Figure 7 It is an arbitrary electric field and result diagram drawn. The electric field drawn is the focused electric field at 20mm directly drawn using drawing software. Figure 7 (a) in the figure is the target electric field drawn. Figure 7 (b) is a hypersurface designed by a neural network. Figure 7 (c) in the figure is the electric field diagram of the metasurface modulation. DETAILED DESCRIPTION

[0026] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The schematic implementation modes of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0027] A method for controlling the distribution of spatial wave field intensity based on a generative neural network of the present invention uses a generative neural network to construct a training network, the network uses an arbitrary electric field to design the phase distribution of a metasurface, and the metasurface constructed using the phase distribution modulates a target electric field, wherein the target electric field is an arbitrary electric field input into the neural network. The generative neural network is composed of a generator and a discriminator. The generative neural network has a learning rate of 0.0002, a batchsize of 256, an activation function including a Leaky ReLU function and a ReLU activation function, and a loss function of BCEWithLogitsLoss. The generative neural network is trained using a data set of 4841 phase images and electric field cross-sectional images of the metasurface, the working frequency band of the metasurface is 0.8THz, and the image size of the data set is a grayscale image of 256*256. In the data set, the metasurface is an all-silicon metasurface based on the PB phase, the substrate thickness is 600um, and the unit structure period is 110um. Use the electric field normalization of any focal point, multiply the normalized electric field by the focal length, and the data is the target electric field modified by the known electric field; draw a circle at any position in the grayscale image, the circle is the focus, the value of the circle is the focal length, and the data is the arbitrary electric field.

[0028] Figure 1 It is a schematic diagram of a method for controlling the distribution of spatial wave field intensity based on a generative neural network. This embodiment provides a 0.8THz electric field manipulation method, the method is: using a generative neural network to construct a training network to generate a metasurface phase according to an arbitrary target electric field, or the neural network designs a metasurface according to an arbitrary target electric field, and the metasurface can modulate the electric field to the target electric field.

[0029] Figure 2 is the randomly generated metasurface electric field and phase for training, which is the metasurface electric field obtained based on Lumerical simulation. The data set used for training in this implementation is mainly obtained using Lumerical, where the unit structure is as follows Figure 3 shown.

[0030] In the implementation, the periodic structure of the unit structure is 110um, W is 42um, L is 80um, H1 is 400um, H2 is 600um, and the material is high-resistance silicon.

[0031] In this implementation, the data set is used to train the neural network to obtain weight data. Figure 1 Flowchart of the method for controlling electric field by generative neural network. An arbitrary electric field (i.e., target electric field) is input into the generative neural network. The generative neural network can generate the phase distribution of the metasurface. The metasurface constructed by the phase distribution can modulate the target electric field. Figure 1From left to right in the figure are: arbitrary electric field (i.e., target electric field), generative neural network, constructed metasurface, and modulated target electric field. The generative neural network is composed of a generator and a discriminator.

[0032] like Figure 4 As shown, the generator consists of eight convolutional layers and eight transposed convolutional layers, the input data is 1*256*256, and the output data is 1*256*256. In the eight convolutional layers, before the second convolution begins, it enters the LeakyReLU activation layer, and after the second convolution, it enters the batch normalization layer. The third to seventh convolutional layers are consistent with the second layer, and the batch normalization layer is not entered after the last convolutional layer. In the eight transposed convolutional layers, before entering the first transposed convolutional layer, it enters the ReLU activation layer, after the transposed convolutional layer, it enters the batch normalization layer, and after the last transposed convolutional layer, it enters the Tanh activation function layer.

[0033] In the convolutional layer, before the second convolution begins, the Leaky ReLU activation layer is entered, and after the second convolution, the batch normalization layer is entered. The third to seventh convolutional layers are consistent with the second layer. The batch normalization layer is not entered after the last convolutional layer.

[0034] In the transposed matrix, before entering the first transposed convolution layer, it enters the ReLU activation layer, after the transposed convolution layer, it enters the batch normalization layer, and after the last transposed convolution layer, it enters the Tanh activation function layer.

[0035] like Figure 5 As shown, the discriminator has three convolutional layers, the input data is 2*256*256, and the output data is 1*256*256; after the first convolutional layer, the discriminator enters the Leaky ReLU activation layer, and after the second convolution, it enters the batch normalization layer and the ReLU activation layer.

[0036] In this implementation, the method for obtaining an arbitrary target electric field is as follows: The existing electric field is normalized to a data range of 0-1, and then the focal length is multiplied by the electric field matrix. The obtained electric field is consistent with the existing electric field in terms of the number of focal points and focal plane distribution, but the focal length is different.

[0037] To draw the electric field, use software to draw a grayscale image and draw a circle in a background of all zeros. The grayscale value of the circle is the focal length. At this time, the focal position of the electric field can be adjusted arbitrarily in three-dimensional space.

[0038] In this implementation, the weight information is loaded into the neural network architecture for generating the metasurface phase distribution.

[0039] The arbitrary target electric field is input into the neural network architecture, and the metasurface phase distribution generated by the neural network is as follows: Figure 6and Figure 7 shown.

[0040] Figure 6 and Figure 7 Electric field map of the metasurface constructed with a generated phase distribution based on the PB phase with the phase being half the rotation angle.

[0041] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for controlling spatial wave field intensity distribution based on a generative neural network, characterized in that: A training network is constructed using a generative neural network. The network uses an arbitrary electric field to design the phase distribution of the metasurface. The metasurface constructed using the phase distribution modulates the target electric field, and the target electric field is the arbitrary electric field input into the neural network.

2. The method for controlling spatial wave field intensity distribution based on a generative neural network according to claim 1, characterized in that: The generative neural network consists of a generator and a discriminator.

3. The method for controlling spatial wave field intensity distribution based on generative neural network according to claim 2, characterized in that: The generator consists of eight convolutional layers and eight transposed convolutional layers, the input data is 1*256*256, and the output data is 1*256*256.

4. The method for controlling spatial wave field intensity distribution based on a generative neural network according to claim 1, characterized in that: The generative neural network has a learning rate of 0.0002, a batchsize of 256, activation functions including a LeakyReLU function and a ReLU activation function, and a loss function of BCEWithLogitsLoss.

5. The method for controlling spatial wave field intensity distribution based on a generative neural network according to claim 1, characterized in that: The generative neural network is trained using a data set of 4841 phase images and electric field cross-section images of the metasurface, the working frequency band of the metasurface is 0.8 THz, and the image size of the data set is a grayscale image of 256*256.

6. The method for controlling spatial wave field intensity distribution based on generative neural network according to claim 5, characterized in that: In the data set, the metasurface is an all-silicon metasurface based on the PB phase, the substrate thickness is 600um, and the unit structure period is 110um.

7. The method for controlling spatial wave field intensity distribution based on a generative neural network according to claim 3, characterized in that: In the eight convolutional layers, before the second convolution starts, the Leaky ReLU activation layer is entered, and after the second convolution, the batch normalization layer is entered. The third to seventh convolutional layers are consistent with the second layer, and the batch normalization layer is not entered after the last convolutional layer.

8. The method for controlling spatial wave field intensity distribution based on generative neural network according to claim 3, characterized in that: In the eight transposed convolutional layers, before entering the first transposed convolutional layer, the layer enters the ReLU activation layer, after the transposed convolutional layer, the layer enters the batch normalization layer, and after the last transposed convolutional layer, the layer enters the Tanh activation function layer.

9. The method for controlling spatial wave field intensity distribution based on generative neural network according to claim 2, characterized in that: The discriminator has three convolutional layers, the input data is 2*256*256, and the output data is 1*256*256; after the first convolutional layer, the discriminator enters the Leaky ReLU activation layer, and after the second convolution, it enters the batch normalization layer and the ReLU activation layer.

10. The method for controlling spatial wave field intensity distribution based on generative neural network according to claim 1, characterized in that: The target electric field is obtained by normalizing the electric field of the focal section to obtain an electric field matrix, and then multiplying the focal length by the electric field matrix to obtain the target electric field. At this time, the matrix is ​​the modified target electric field; a circle is drawn at any position in the grayscale image, and the circle is the focus, and the grayscale value of the circle is the focal length. At this time, the grayscale image is the drawn target electric field.

Images