A method for controlling the spatial wave field intensity distribution based on a generative neural network

Abstract

The present invention discloses a method for controlling the spatial wave field intensity distribution based on a generative neural network. A training network is constructed by using the generative neural network. This network designs the phase distribution of the metasurface by using an arbitrary electric field, and the metasurface constructed by using this phase distribution modulates the target electric field, and the target electric field is the arbitrary electric field input into the neural network. The generative neural network consists of a generator and a discriminator. 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. For the generative neural network, the training uses a data set of phase images and electric field cross-section images of 4841 metasurfaces. The working frequency band of the metasurface is 0.8 THz, and the data set image size is a grayscale image of 256*256. There is high flexibility in constructing an unknown electric field, either designing an unknown electric field by using an existing electric field or directly drawing an unknown electric field as needed.

Description

Technical Field

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

[0002] With the rapid development of information technology, the operating frequency and integration degree of devices have been continuously improved, the electromagnetic environment has become increasingly complex, posing new challenges to the high-degree-of-freedom control of electromagnetic waves.

[0003] The metasurface performs phase modulation through unit structures to achieve macroscopic electromagnetic characteristics that do not exist in nature. However, traditional design methods are mainly based on the generalized Snell's law, and the phase distribution of the metasurface is determined 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, and also limits the innovation and diversity of design on the other hand. With the improvement of computing power and the development of machine learning technology, neural network manipulation has become possible.

[0004] The existing neural networks of metasurfaces can, on the one hand, achieve phase prediction of unit structures, and on the other hand, achieve inverse design of holographic images. Although the phase of the unit structure can be accurately predicted, electric field modulation still relies on physical calculations. The inverse design of holographic images can customize images with a fixed focal length, but this modulation is achieved in a two-dimensional space. Currently, there is a lack of a method for customizing 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 spatial wave field intensity distribution based on a generative neural network. The neural network designs a metasurface using an arbitrary electric field, and then realizes arbitrary electric field manipulation, which can customize an unknown electric field either through a known electric field or directly customize an unknown electric field.

[0006] Technical Solution: To achieve the above purpose, a method for controlling the spatial wave field intensity distribution based on a generative neural network proposed by the present invention is as follows:

[0007] A training network is constructed using a generative neural network. This network designs the phase distribution of a metasurface using an arbitrary electric field, and the metasurface constructed using this phase distribution modulates the target electric field, where the target electric field is the arbitrary electric field input into the neural network.

[0008] The generative neural network consists of a generator and a discriminator.

[0009] 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.

[0010] The generative neural network has a learning rate of 0.0002, a batch size of 256, activation functions including the LeakyReLU function and the ReLU activation function, and a loss function of BCEWithLogitsLoss.

[0011] The training dataset used for the generative neural network consists of phase images and electric field cross-section images of 4841 metasurfaces. The working frequency band of the metasurface is 0.8 THz, and the size of the dataset images is grayscale images of 256*256.

[0012] In the dataset, the metasurface is an all-silicon metasurface based on PB phase, with a substrate thickness of 600 um and a unit structure period of 110 um.

[0013] In the eight-layer convolutional layer, before the second convolution, it enters the Leaky ReLU activation layer, and after the second convolution, it enters the batch normalization layer. The third to seventh convolutional layers are the same as the second layer, and after the last convolutional layer, it does not enter the batch normalization layer.

[0014] In the eight-layer transposed convolutional layer, before entering the first transposed convolutional layer, it enters the ReLU activation layer, and after the transposed convolutional layer, it enters the batch normalization layer. After the last transposed convolutional layer, it enters the Tanh activation function layer.

[0015] The discriminator has three convolutional layers. The input data is 2*256*256, and the output data is 1*256*256.

[0016] After the first convolutional layer of the discriminator, it enters the Leaky ReLU activation layer, and after the second convolution, it enters the batch normalization layer and the ReLU activation layer.

[0017] The target electric field is obtained by normalizing the focal cross-section electric field 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; draw a circle at any position in the grayscale image. This 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.

[0018] Advantageous effects: The generative neural network proposed in the present invention can generate metasurface phases according to any input electric field, and the metasurface constructed using this phase can modulate the input electric field.

[0019] In the present invention, the input electric field data can be modified using existing electric field information or directly drawn.

[0020] In the present invention, the neural network can manipulate the position of the focus in three-dimensional space. Description of the Drawings

[0021] Figure 1 It is a schematic diagram of a method for controlling the spatial wave field intensity distribution based on a generative neural network.

[0022] Figure 2 They are randomly generated metasurface electric fields and phases for training, and the metasurface electric fields are obtained based on Lumerical simulations; Figure 2 In (a) of [reference], there is a phase and electric field cross-sectional view of a single focus, Figure 2 In (b) of [reference], there is a phase and electric field cross-sectional view of a double focus.

[0023] Figure 3 It is a 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.

[0024] Figure 4 It is a schematic diagram of the generator of the network.

[0025] Figure 5 It is a schematic diagram of the discriminator of the network.

[0026] Figure 6 It is a modified arbitrary electric field and result diagram, where the modified electric field is obtained by modifying the existing electric field with a 35 mm focal length to an electric field with a 20 mm focal length, Figure 6 In (a) of [reference], there is the target electric field to be modified, Figure 6 In (b) of [reference], there is the metasurface designed by the neural network, Figure 6 In (c) of [reference], there is the electric field diagram modulated by the metasurface.

[0027] Figure 7 It is a drawn arbitrary electric field and result diagram, where the drawn electric field is directly drawn using drawing software as a focused electric field with a focus at 20 mm, Figure 7 In (a) of [reference], there is the target electric field drawn, Figure 7 In (b) of [reference], there is the metasurface designed by the neural network, Figure 7 In (c) of [reference], there is the electric field diagram modulated by the metasurface. Detailed implementation mode

[0028] In order to make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below in combination with embodiments and the accompanying drawings. The illustrative embodiments of the present invention and their descriptions are only used to explain the present invention and are not intended to limit the present invention.

[0029] A method for controlling the spatial wave field intensity distribution based on a generative neural network. A training network is constructed using a generative neural network. This network designs the phase distribution of a metasurface using an arbitrary electric field, and the metasurface constructed using this phase distribution modulates the target electric field, where the target electric field is the arbitrary electric field input into the neural network. The generative neural network consists of a generator and a discriminator. For the generative neural network, the learning rate is 0.0002, the batch size is 256, the activation functions include the Leaky ReLU function and the ReLU activation function, and the loss function is BCEWithLogitsLoss. For the generative neural network, the training dataset used is the phase images and electric field cross-section images of 4841 metasurfaces. The operating frequency band of the metasurface is 0.8 THz, and the size of the dataset images is a grayscale image of 256*256. In the dataset, the metasurface is an all-silicon metasurface based on PB phase, the substrate thickness is 600 um, and the unit structure period is 110 um. Normalize using an arbitrary focal point electric field, multiply the normalized electric field by the focal length, and this data is the target electric field modified using the known electric field; draw a circle at an arbitrary position in the grayscale image, this circle is the focal point, and the value of the circle is the focal length, and this data is the drawn arbitrary electric field.

[0030] Figure 1 It is a schematic diagram of a method for controlling the spatial wave field intensity distribution based on a generative neural network. This implementation provides a method for manipulating an electric field at 0.8 THz. The method is as follows: Use a generative neural network to construct a training network to generate the metasurface phase according to an arbitrary target electric field, or in other words, the neural network designs a metasurface according to an arbitrary target electric field, and this metasurface can modulate the electric field into the target electric field.

[0031] Figure 2 They are randomly generated metasurface electric fields and phases for training, and are metasurface electric fields based on Lumerical simulation. In this implementation, the dataset used for training is mainly obtained using Lumerical, and the unit structure is as Figure 3 shown.

[0032] In the implementation, the periodic structure of the unit structure is 110 um, W is 42 um, L is 80 um, H1 is 400 um, H2 is 600 um, and the material is high-resistivity silicon.

[0033] In this implementation, the dataset is used for the training of the neural network to obtain weight data, Figure 1 It is a flowchart of the method for a generative neural network to manipulate an electric field. Input an arbitrary electric field (i.e., the target electric field) into the generative neural network. The generative neural network can generate the phase distribution of the metasurface, and the metasurface constructed using the phase distribution can modulate the target electric field. Figure 1From left to right are: an arbitrary electric field (i.e., the target electric field), a generative neural network, a constructed metasurface, and a modulated target electric field. The generative neural network consists of a generator and a discriminator.

[0034] As Figure 4 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, it enters the LeakyReLU activation layer, and after the second convolution, it enters the batch normalization layer. The third to seventh convolutional layers are the same as the second layer, and after the last convolutional layer, it does not enter the batch normalization 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.

[0035] In the convolutional layer, before the second convolution, it enters the Leaky ReLU activation layer, after the second convolution, it enters the batch normalization layer, and the third to seventh convolutional layers are the same as the second layer. After the last convolutional layer, it does not enter the batch normalization layer.

[0036] In the transposed matrix, 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.

[0037] As Figure 5 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 of the discriminator, it enters the Leaky ReLU activation layer, and after the second convolution, it enters the batch normalization layer and the ReLU activation layer.

[0038] In this implementation, the method for obtaining an arbitrary target electric field is as follows:

[0039] Normalize the existing electric field, with the data range being 0 - 1, and then multiply the focal length by the electric field matrix. The obtained electric field has the same number of focal points and focal plane distribution as the existing electric field, but different focal lengths.

[0040] For plotting the electric field, use software to draw a grayscale image, draw a circle in a background of all 0s, and the grayscale value of the circle is the focal length. At this time, the focal point position of the electric field can be adjusted arbitrarily in three-dimensional space.

[0041] In this implementation, load the weight information into the neural network architecture to generate the metasurface phase distribution.

[0042] Input any target electric field into the neural network architecture, and the metasurface phase distribution generated by the neural network is as Figure 6 and Figure 7 shown.

[0043] Figure 6 and Figure 7 are the electric field diagrams of the metasurface, which is constructed with the generated phase distribution. Based on the PB phase, the phase is half of the rotation angle.

[0044] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for controlling the spatial wave field intensity distribution based on a generative neural network, characterized in that, A training network is constructed using a generative neural network. This network designs the phase distribution of a metasurface using an arbitrary electric field, and the metasurface constructed using this phase distribution modulates the target electric field, which is the arbitrary electric field input into the neural network. The target electric field is obtained by normalizing the focal plane electric field 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 an arbitrary position in the grayscale image, which is the focal point, and the grayscale value of the circle is the focal length. At this time, the grayscale image is the drawn target electric field. The generative neural network consists of a generator and a discriminator. 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. 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 of the discriminator, it enters the Leaky ReLU activation layer, and after the second convolution, it enters the batch normalization layer and the ReLU activation layer.

2. The method for controlling the spatial wave field intensity distribution based on a generative neural network according to claim 1, wherein For the generative neural network, its learning rate is 0.0002, the batchsize is 256, the activation functions include the LeakyReLU function and the ReLU activation function, and the loss function is BCEWithLogitsLoss.

3. The method for controlling the spatial wave field intensity distribution based on a generative neural network according to claim 1, characterized in that For the generative neural network, the training dataset used is the phase images and electric field cross-section images of 4841 metasurfaces. The operating frequency band of the metasurface is 0.8 THz, and the size of the dataset images is a grayscale image of 256*256.

4. The method for controlling the spatial wave field intensity distribution based on a generative neural network according to claim 3, wherein In the dataset, the metasurface is an all-silicon metasurface based on PB phase, the substrate thickness is 600 um, and the unit structure period is 110 um.

5. The method for controlling the spatial wave field intensity distribution based on a generative neural network according to claim 1, wherein In the eight convolutional layers, before the second convolution, it enters the Leaky ReLU activation layer, and after the second convolution, it enters the batch normalization layer. The third to seventh convolutional layers are the same as the second layer, and after the last convolutional layer, it does not enter the batch normalization layer.

6. The method for controlling the spatial wave field intensity distribution based on a generative neural network according to claim 1, wherein 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.

Images