A method, system and related components for obtaining device parameters based on neural network

Abstract

The present application discloses a method, system, device and computer-readable storage medium for obtaining device parameters based on a neural network. The method for obtaining device parameters based on a neural network optimizes corresponding device parameters for a target electromagnetic spectral response through a neural network composed of an inverse network, a discriminant network and a forward prediction network. In the present application, the inverse network and the discriminant network are combined for adversarial training to optimize the generated device parameters to make them close to the real parameters. At the same time, the training data set required for the generated adversarial network is small, saving hardware resources and improving the robustness of the neural network. Then, the device parameters that meet the real conditions generated by the inverse network are input into the forward prediction network to realize the combined training of the inverse network and the forward prediction network. When the actual electromagnetic spectral response matches the target electromagnetic spectral response, the device parameters generated by the inverse network are output to optimize the generated device performance, and device parameters with higher performance are output under the input of the desired target electromagnetic spectral response.

Description

Technical Field

[0001] This application relates to the field of device design, and particularly to a method, system and related components for obtaining device parameters based on a neural network. Background Art

[0002] In nanophotonics, optical devices composed of sub-wavelength periodic or aperiodic geometric arrays have attracted extensive attention due to their special properties of changing dielectric constant and magnetic permeability. The design of sub-wavelength optical devices includes two aspects. One is the forward problem, that is, given the device size structure, calculating its electromagnetic spectral response. The other is the inverse problem, that is, given the ideal electromagnetic spectral response, calculating the corresponding device size structure.

[0003] In solving the inverse problem, since sub-wavelength optical devices involve multiple interdisciplinary fields such as optics, physics and materials science, mastering complex underlying principles is required when designing devices, which largely depends on the rich experience of nanophotonics researchers. In addition, since the solution space is non-convex, that is, there are many local optimal solutions, in most cases the solution cannot be directly obtained, and its solution process is very challenging, which greatly limits the popularization of optical device design. Neural networks have the ability to simulate non-linear physical relationships, so they provide a solution idea for solving the relationship between the geometry of the photon system and its electromagnetic spectral response. Although neural networks have been applied in the design of sub-wavelength optical devices, because general neural networks require tens of thousands of data sets for training, a lot of resources are required in practical applications. In addition, existing neural networks do not consider the device structure and electromagnetic spectral response at the same time, and their robustness is poor.

[0004] Therefore, how to provide a solution to the above technical problems is an issue that those skilled in the art need to solve currently. Summary of the Invention

[0005] The purpose of this application is to provide a method, system, device and computer-readable storage medium for obtaining device parameters based on a neural network, which can optimize the generated device parameters to make them close to the real parameters, while the training data set required for the generative adversarial network is small, saving hardware resources, improving the robustness of the neural network, and at the same time being able to optimize the performance of the generated device, and output device parameters with higher performance when the input is the desired target electromagnetic spectral response.

[0006] To solve the above technical problems, this application provides a method for obtaining device parameters based on a neural network. The neural network includes an inverse network, a discriminant network and a forward prediction network. The method for obtaining device parameters includes:

[0007] Using the inverse network to generate device parameters according to the target requirements, where the target requirements include the target electromagnetic spectral response;

[0008] Input the device parameters and the real samples into the discriminant network, and make the discriminant network perform adversarial training using the device parameters and the real samples to obtain a first adjustment parameter. Based on the first adjustment parameter, perform a first parameter update operation on the inverse network and the discriminant network until the device parameters generated by the inverse network according to the target electromagnetic spectrum response meet the real conditions;

[0009] Input the device parameters that meet the real conditions into the forward prediction network to obtain the actual electromagnetic spectrum response. When the actual electromagnetic spectrum response matches the target electromagnetic spectrum response, output the device parameters generated by the inverse network.

[0010] Optionally, the target requirement is a one-dimensional vector including the target electromagnetic spectrum response, the target incident angle, the refractive index of the target device material, and the simulated random noise.

[0011] Optionally, the process of using the inverse network to generate device parameters according to the target requirement includes:

[0012] Input the target requirement into the inverse network, and make the inverse network obtain reconstruction parameters based on the target requirement, and perform feature extraction on the reconstruction parameters to generate device parameters.

[0013] Optionally, the inverse network includes a transposed convolutional layer and an encoding-decoding module;

[0014] The process of inputting the target requirement into the inverse network, making the inverse network obtain reconstruction parameters based on the target requirement, and performing feature extraction on the reconstruction parameters to generate device parameters includes:

[0015] Input the target requirement into the inverse network;

[0016] Perform upsampling based on the target requirement through the transposed convolutional layer to obtain reconstruction parameters;

[0017] Extract feature parameters from the reconstruction parameters through the encoding-decoding module, and generate device parameters using the feature parameters.

[0018] Optionally, the process of making the discriminant network perform adversarial training using the device parameters and the real samples to obtain a first adjustment parameter, and performing a first parameter update operation on the inverse network and the discriminant network based on the first adjustment parameter includes:

[0019] Make the discriminant network perform adversarial training using the device parameters and the real samples to obtain the probability that the device parameters are real parameters;

[0020] Judge whether the probability is the target value;

[0021] Otherwise, obtain a first adjustment parameter, and perform a first parameter update operation on the inverse network and the discriminant network based on the first adjustment parameter;

[0022] If so, determine that the device parameters generated by the inverse network according to the target electromagnetic spectrum response meet the real conditions.

[0023] Optionally, after inputting the device parameters that meet the real conditions into the forward prediction network to obtain the actual electromagnetic spectrum response, the device parameter acquisition method further includes:

[0024] Obtain a loss function between the actual electromagnetic spectrum response and the target electromagnetic spectrum response;

[0025] When the loss function converges, determine that the actual electromagnetic spectrum response matches the target electromagnetic spectrum response.

[0026] Optionally, after obtaining the loss function between the actual electromagnetic spectrum response and the target electromagnetic spectrum response, the device parameter acquisition method further includes:

[0027] When the loss function does not converge, obtain a second adjustment parameter;

[0028] Based on the second adjustment parameter, perform a second parameter update operation on the inverse network and the forward prediction network until the actual electromagnetic spectrum response matches the target electromagnetic spectrum response.

[0029] Optionally, both the forward prediction network and the discriminant network include BN layers.

[0030] To solve the above technical problems, the present application also provides a device parameter acquisition system based on a neural network. The neural network includes an inverse network, a discriminant network, and a forward prediction network. The device parameter acquisition system includes:

[0031] A first processing module, configured to generate device parameters according to a target requirement by using the inverse network, where the target requirement includes a target electromagnetic spectrum response;

[0032] An adjustment module, configured to input the device parameters and real samples into the discriminant network, and enable the discriminant network to perform adversarial training using the device parameters and the real samples to obtain a first adjustment parameter, and perform a first parameter update operation on the inverse network and the discriminant network based on the first adjustment parameter until the device parameters generated by the inverse network according to the target electromagnetic spectrum response meet the real conditions;

[0033] A second processing module, configured to input the device parameters that meet the real conditions into the forward prediction network to obtain an actual electromagnetic spectrum response, and output the device parameters generated by the inverse network when the actual electromagnetic spectrum response matches the target electromagnetic spectrum response.

[0034] To solve the above technical problems, the present application also provides a device parameter acquisition device based on a neural network, including:

[0035] A memory, configured to store a computer program;

[0036] A processor, configured to implement the steps of the device parameter acquisition method based on a neural network as described in any one of the above when executing the computer program.

[0037] To solve the above technical problems, the present application also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the device parameter acquisition method based on a neural network as described in any one of the above are implemented.

[0038] The present application provides a device parameter acquisition method based on a neural network, which optimizes corresponding device parameters for a target electromagnetic spectrum response through a neural network composed of an inverse network, a discriminant network, and a forward prediction network. In the present application, the inverse network and the discriminant network are combined for adversarial training to optimize the generated device parameters to make them close to the real parameters. At the same time, the training data set required for the adversarial network is small, saving hardware resources and improving the robustness of the neural network. Then, the device parameters that meet the real conditions generated by the inverse network are input into the forward prediction network to realize the combined training of the inverse network and the forward prediction network. When the actual electromagnetic spectrum response matches the target electromagnetic spectrum response, the device parameters generated by the inverse network are output again to optimize the performance of the generated device, and under the input of the desired target electromagnetic spectrum response, device parameters with higher performance are output. The present application also provides a device parameter acquisition system, device, and computer-readable storage medium based on a neural network, which have the same beneficial effects as the above device parameter acquisition method. Description of the Drawings

[0039] To more clearly illustrate the embodiments of the present application, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0040] Figure 1 It is a flowchart of the steps of a device parameter acquisition method based on a neural network provided by the present application;

[0041] Figure 2A structural schematic diagram of a neural network provided by this application;

[0042] Figure 3 A schematic diagram of an encoding and decoding process provided by this application;

[0043] Figure 4 A flowchart of discriminant network training provided by this application;

[0044] Figure 5 A flowchart of joint network training provided by this application;

[0045] Figure 6 A flowchart of dataset establishment provided by this application;

[0046] Figure 7 A cross-sectional view of a subwavelength metal wire grid provided by this application;

[0047] Figure 8 Schematic diagrams of six reference subwavelength metal wire grid styles provided by this application;

[0048] Figure 9 A flowchart of pre-training of a forward prediction network provided by this application;

[0049] Figure 10 A structural schematic diagram of a device parameter acquisition structure based on a neural network provided by this application;

[0050] Figure 11 A structural schematic diagram of a device parameter acquisition device based on a neural network provided by this application;

[0051] Figure 12 A structural schematic diagram of another device parameter acquisition device based on a neural network provided by this application. Detailed implementation manners

[0052] The core of this application is to provide a method, system, device and computer-readable storage medium for acquiring device parameters based on a neural network, which can optimize the generated device parameters to make them close to the real parameters. At the same time, the training dataset required for the generative adversarial network is small, saving hardware resources, improving the robustness of the neural network, and also being able to optimize the generated device performance. Under the input of the desired target electromagnetic spectral response, device parameters with higher performance are output.

[0053] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of this application. Apparently, the described embodiments are some, but not all, of the embodiments of this application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in this application without creative efforts shall fall within the scope of protection of this application.

[0054] Please refer to Figure 1 , Figure 1 which is a flowchart of the steps of a method for obtaining device parameters based on a neural network provided by this application. First, an introduction to the neural network provided by this application will be given. This neural network includes an inverse network, a discriminant network, and a forward prediction network. The device parameters in this application may specifically refer to optical device parameters. The method for obtaining device parameters based on a neural network includes:

[0055] S101: Use the inverse network to generate device parameters according to the target requirements, where the target requirements include the target electromagnetic spectral response;

[0056] As an optional embodiment, the process of using the inverse network to generate device parameters according to the target requirements includes: inputting the target requirements into the inverse network, enabling the inverse network to obtain reconstruction parameters based on the target requirements, and performing feature extraction on the reconstruction parameters to generate device parameters.

[0057] Specifically, refer to Figure 2As shown in the figure, the inverse network includes an input layer, a fully connected layer, a transposed convolutional layer, and an encoding-decoding module. The encoding-decoding module includes an encoder and a decoder, and the inverse network also includes an output layer. Among them, the input layer of the inverse network is used to receive target requirements, which include but are not limited to a one-dimensional vector composed of a target electromagnetic spectral response, a target incident angle, a refractive index of a target device material, and simulated random noise. This one-dimensional vector can be generated using Matlab software. It can be understood that adding simulated random noise to the inverse network to generate a full-size image can simulate the manufacturing errors caused by process conditions and environmental disturbances during the actual manufacturing process, making the device output by the network more robust. The number of layers of the fully connected layer of the inverse network is related to the complexity of the device. If the designed device structure is more complex, more parameters are required for fitting, and the corresponding number of fully connected layers also increases accordingly, generally 1 to 4 layers. After passing through the fully connected layer, the input layer is connected to the transposed convolutional layer. The number of convolutional kernels of the transposed convolution is 1, indicating that the output is a grayscale image. Specifically, the transposed convolutional layer performs upsampling based on the target requirements of the input layer to obtain reconstruction parameters. The encoding-decoding module extracts feature parameters from the reconstruction parameters, generates device parameters using the feature parameters, and converts the high-dimensional discrete data of the device structure into low-dimensional data, so that the network can learn the large-scale topological features of the device during the training process. It can be understood that the encoding-decoding module not only improves the accuracy of the network training process as a whole, but also trains at a lower spatial resolution, thus significantly reducing the computational cost. The output layer is used to output the device parameters generated based on the target requirements. The device parameters can specifically include the device structure image.

[0058] Specifically, the encoding-decoding module includes an encoder and a decoder. Among them, the encoder is a neural network, and its input and output are the data point x and the hidden representation z respectively, as Figure 3 shown. Each x is described by N features and is N-dimensional, and each z is described by K features and is K-dimensional, where K < N. The decoder is another neural network, and its input and output are the latent representation z and the reconstruction The encoding-decoding module minimizes the difference between the original x and the reconstruction through multiple training iteration optimization processes. Therefore, the encoding-decoding module creates a bottleneck for the data x, ensuring that only the main structured part of the information can pass through the encoder and be reconstructed by the decoder, thereby realizing feature extraction of the reconstruction parameters to obtain feature parameters, and then the output layer outputs the final structure image of the device.

[0059] Specifically, the inverse network can generate thousands of devices within a few seconds. Therefore, a large dataset can be made using the trained inverse network at a relatively low computational cost to generate different device parameters.

[0060] S102: Input the device parameters and real samples into the discriminant network, and let the discriminant network perform adversarial training using the device parameters and real samples to obtain the first adjustment parameter. Based on the first adjustment parameter, perform the first parameter update operation on the inverse network and the discriminant network until the device parameters generated by the inverse network according to the target electromagnetic spectral response meet the real conditions;

[0061] Specifically, also refer to Figure 2 , the discriminant network specifically includes a convolutional layer, a BN layer, two fully connected layers, and an output layer. The number of neurons in the last fully connected layer is 1, and a binary classification result is output. The discriminant network uses real samples and the device parameters generated by the inverse network as fake samples for joint training. During the training process, try to make the output result of the real samples be 1, and the output result of the fake samples generated by the inverse network be 0. After multiple rounds of training, the final generated distribution coincides with the real distribution, so that the discriminant model cannot distinguish between real images and the device images generated by the inverse network, and further makes the device parameters generated by the inverse network closer to the real device structure, improving the performance of the device structure.

[0062] As an optional embodiment, the process of making the discriminant network perform adversarial training using the device parameters and real samples to obtain the first adjustment parameter and performing the first parameter update operation on the inverse network and the discriminant network based on the first adjustment parameter includes: making the discriminant network perform adversarial training using the device parameters and real samples to obtain the probability that the device parameters are real parameters; judging whether the probability is the target value; if not, obtaining the first adjustment parameter and performing the first parameter update operation on the inverse network and the discriminant network based on the first adjustment parameter; if so, determining that the device parameters generated by the inverse network according to the target electromagnetic spectral response meet the real conditions.

[0063] Specifically, please refer to Figure 4 , Figure 4 is the flowchart for performing adversarial training on the discriminant network using device parameters and real samples, including: inputting the device parameters as fake samples and real samples into the discriminant network for training, and the discriminant network judges whether the input is a real sample or a fake sample, and judges whether the probability that the input is a real sample is the target value 0.5. If not, obtain the first adjustment parameter for backpropagation gradient update, and adjust the weight values of the inverse network and the discriminant network until the device parameters generated by the inverse network meet the real conditions, that is, the discriminant network determines that the probability that the input is a real sample is 0.5. At this time, it is determined that the device parameters output by the inverse network are close to the real device structure.

[0064] S103: Input the device parameters that meet the real conditions into the forward prediction network to obtain the actual electromagnetic spectral response. When the actual electromagnetic spectral response matches the target electromagnetic spectral response, output the device parameters generated by the inverse network.

[0065] As an alternative embodiment, after inputting device parameters that meet the real conditions into the forward prediction network to obtain the actual electromagnetic spectrum response, the device parameter acquisition method further includes: obtaining a loss function between the actual electromagnetic spectrum response and the target electromagnetic spectrum response; when the loss function converges, determining that the actual electromagnetic spectrum response and the target electromagnetic spectrum response match; when the loss function does not converge, obtaining a second adjustment parameter; performing a second parameter update operation on the inverse network and the forward prediction network based on the second adjustment parameter until the actual electromagnetic spectrum response and the target electromagnetic spectrum response match.

[0066] Specifically, after the device structure generated by the inverse network is realistic enough, then optimize the spectral performance of the generated device. The inverse network and the forward prediction network are jointly trained. Referring to Figure 2 As shown, the forward prediction network includes an input layer, a fully connected layer, a BN layer, and an output layer. Adding a BN layer to the forward prediction network can accelerate convergence, and combined with the inverse network, it can solve the one-to-many problem of the inverse problem. The joint training process of the inverse network and the forward prediction network is as Figure 5 shown. The output layer of the inverse network is used as the input layer of the forward network. Considering that there is a one-to-many problem in the inverse operation of optical device design, the gradient is updated by backpropagation by solving the loss function between the actual electromagnetic spectrum response output by the forward prediction network and the target electromagnetic spectrum response input to the inverse network. After multiple iterations, the spectral response of the device structure obtained by the inverse network will be greatly improved, and the one-to-many problem will be solved. Among them, the cross-entropy loss function can be used as the loss function.

[0067] Referring to the above, when training the discriminant network and the forward prediction network, real samples are needed. Based on this, the present application further includes an operation of pre-constructing a data set to provide the real samples required by the discriminant network and the forward prediction network.

[0068] Specifically, the COMSOL software can be used to construct a data set for training the forward prediction network and the discriminant network. The process of constructing a data set using the COMSOL software is as follows. Referring to Figure 6 shown. First, determine the device structure parameters and the value ranges of each device structure parameter. Within the value ranges, each parameter is set separately. According to the specific parameters, a two-dimensional matrix is constructed using MATLAB, and a binary image is drawn to represent the device structure. Then, the binary images representing each device structure are simulated by the COMSOL simulation software to obtain the corresponding electromagnetic spectrum response.

[0069] Taking the device as a sub-wavelength metal wire grid as an example, the device structure includes a substrate and a metal wire grid. As Figure 7 shown. Figure 8There are six reference sub - wavelength metal wire grid patterns. The necessary device structure parameters that need to be input into the COMSOL simulation software include but are not limited to parameters such as the curvature of the metal wire grid, the rotation angle, the duty cycle, the thickness, and the period. Among them, additional parameters such as the refractive indices of media including the substrate, the metal wire grid, and air also need to be set. The output parameters are the electromagnetic responses in the corresponding wavelength bands, such as optical responses like the reflectivity spectrum, the transmittance spectrum, or the polarization extinction ratio curve. Regarding the input device structure parameters: the curvature r of the metal wire grid ∈(0, 180°], with an increment of 5°, a total of 36 variables; the rotation angle R of the metal wire grid ∈(0, 180°], with an increment of 5°, a total of 36 variables; the duty cycle f of the metal wire grid ∈(0, 1), with an increment of 0.1, a total of 8 variables; the thickness h of the metal wire grid is generally in the range of [20, 200nm] according to process conditions, with an increment of 20nm, a total of 10 variables; the period needs to meet the sub - wavelength condition d < (λ / 5). According to the above parameters such as r, R, f, h, or d, use MATLAB to construct a 120×120 two - dimensional matrix to represent the metal wire grid pattern. The matrix elements are 0 or 1, representing the presence or absence of metal.

[0070] Then select germanium as the device substrate, the working wavelength λ ∈[3, 12μm], with an increment of 0.5μm, and the incident angle θ ∈[-65°, 65°], with an increment of 5°. Then, according to each pair of wavelength and angle, use different structure parameters to construct the device and generate a series of corresponding electromagnetic spectral responses. Finally, set the response threshold according to the requirements. Devices with responses higher than the threshold are high - quality devices, and devices with responses lower than the threshold are discarded.

[0071] Finally, a high - quality training set generated by COMSOL is adopted. This training set consists of 3000 images and is used for the training of the forward prediction network and the discriminant network. This initial training set is several orders of magnitude smaller than the training sets used in traditional machine vision applications.

[0072] Please refer to Figure 9 as shown in Figure 9 For the pre - training process of the forward prediction network, this pre - training process is a normal neural network training process, including constructing a network structure using pytorch or tensorflow, including fully - connected layers, BN layers, fully - connected layers, and output layers. It can be understood that adding BN layers can accelerate the network iteration rate. Fundamentally speaking, the BN layer normalizes the same feature of different samples. Let the input be B = {x1…m}, the learned hyperparameters be γ and β, and the output be y i = BN γ,β (xi), and its specific calculation process is as follows:

[0073] Mean value mean:

[0074] Variance variance:

[0075] Normalize:

[0076] Scale and shift:

[0077] The output layer represents the generated electromagnetic spectrum response. The number of nodes in the output layer represents the number of points after discretizing the spectral map of the given band. If the input band is 3 - 12 μm and discretized at intervals of 0.5 μm, the number of nodes in the output layer is 19.

[0078] After the network is constructed, hyperparameters are set. The mini - batch training method is adopted, and the batch size is set to 32. Among them, the gradient is updated by backpropagation using the loss function between the actual spectral response output by the forward prediction network and the spectral response obtained by COMSOL simulation. The loss function can adopt the cross - entropy loss function, and the specific calculation formula is:

[0079]

[0080] Among them, L is the loss function, y is the spectral response obtained by COMSOL simulation, is the actual spectral response.

[0081] In summary, a method for obtaining device parameters based on an improved neural network proposed in this application. The improved neural network includes an inverse network, a discriminant network, and a forward prediction network, which can optimize the corresponding device structure for a specific electromagnetic spectrum response. By combining the inverse network and the discriminant network, a large training set can be quickly generated, and the generated device images can be optimized to make their structures closer to real images, improving the reliability and robustness of the device. At the same time, an encoding - decoding module is added to the inverse network to obtain large - scale topological features of the device, improving the accuracy of the network training process as a whole. And training at a lower spatial resolution can significantly reduce the computational cost. Using the discriminant network and the forward prediction network with BN layers added can accelerate the network convergence speed. Then, combining the trained inverse network with the forward prediction network can solve the one - to - many problem of inverse operation while improving the performance of the network - output device. Finally, the invented neural network can output device parameters with high performance when the input is the desired electromagnetic spectrum response. Adopting the solution of this application is beneficial to promoting the solution of the inverse problem of sub - wavelength optical device design, and fine - tuning based on the trained deep - learning network to perform the inference task of optical device design can greatly improve the design efficiency of sub - wavelength devices. In addition, the method proposed in this invention is universal for the design of other sub - wavelength devices, such as two - dimensional metasurface structures, volume gratings, anti - reflection structures, etc.

[0082] Please refer to Figure 10 , Figure 10 which is a schematic structural diagram of a device parameter acquisition system based on a neural network provided by this application. The neural network includes an inverse network, a discriminant network, and a forward prediction network. The device parameter acquisition system includes:

[0083] A first processing module 11, configured to use the inverse network to generate device parameters according to target requirements, where the target requirements include a target electromagnetic spectrum response;

[0084] An adjustment module 12, configured to input the device parameters and real samples into the discriminant network, and enable the discriminant network to perform adversarial training using the device parameters and real samples to obtain a first adjustment parameter, and perform a first parameter update operation on the inverse network and the discriminant network based on the first adjustment parameter until the device parameters generated by the inverse network according to the target electromagnetic spectrum response meet the real conditions;

[0085] A second processing module 13, configured to input the device parameters that meet the real conditions into the forward prediction network to obtain an actual electromagnetic spectrum response, and when the actual electromagnetic spectrum response matches the target electromagnetic spectrum response, output the device parameters generated by the inverse network.

[0086] It can be seen that in this embodiment, the neural network composed of the inverse network, the discriminant network, and the forward prediction network optimizes the corresponding device parameters for the target electromagnetic spectrum response. In this application, the inverse network and the discriminant network are combined for adversarial training to optimize the generated device parameters to make them close to the real parameters. At the same time, the training data set required for the adversarial network is small, saving hardware resources and improving the robustness of the neural network. Then, the device parameters that meet the real conditions generated by the inverse network are input into the forward prediction network to realize the combined training of the inverse network and the forward prediction network. When the actual electromagnetic spectrum response matches the target electromagnetic spectrum response, the device parameters generated by the inverse network are output to optimize the generated device performance, and under the input of the desired target electromagnetic spectrum response, device parameters with higher performance are output.

[0087] As an alternative embodiment, the target requirements are a one-dimensional vector including a target electromagnetic spectrum response, a target incident angle, a refractive index of a target device material, and simulated random noise.

[0088] As an alternative embodiment, the process of using the inverse network to generate device parameters according to the target requirements includes:

[0089] Input the target requirements into the inverse network, enable the inverse network to obtain reconstruction parameters based on the target requirements, and perform feature extraction on the reconstruction parameters to generate device parameters.

[0090] As an alternative embodiment, the inverse network includes a transposed convolutional layer and an encoding and decoding module;

[0091] Input the target requirements into the inverse network, and the process of enabling the inverse network to obtain reconstruction parameters based on the target requirements and extracting features from the reconstruction parameters to generate device parameters includes:

[0092] Input the target requirements into the inverse network;

[0093] Perform upsampling based on the target requirements through a transposed convolutional layer to obtain reconstruction parameters;

[0094] Extract feature parameters from the reconstruction parameters through an encoding and decoding module, and generate device parameters using the feature parameters.

[0095] As an alternative embodiment, the process of enabling the discriminative network to perform adversarial training using device parameters and real samples to obtain first adjustment parameters and performing a first parameter update operation on the inverse network and the discriminative network based on the first adjustment parameters includes:

[0096] Enable the discriminative network to perform adversarial training using device parameters and real samples to obtain the probability that the device parameters are real parameters;

[0097] Determine whether the probability is the target value;

[0098] If not, obtain the first adjustment parameters, and perform a first parameter update operation on the inverse network and the discriminative network based on the first adjustment parameters;

[0099] If so, determine that the device parameters generated by the inverse network according to the target electromagnetic spectral response meet the real conditions.

[0100] As an alternative embodiment, after inputting the device parameters that meet the real conditions into the forward prediction network to obtain the actual electromagnetic spectral response, the device parameter acquisition system further includes:

[0101] An acquisition module, configured to acquire the loss function between the actual electromagnetic spectral response and the target electromagnetic spectral response;

[0102] A second processing module 13, further configured to determine that the actual electromagnetic spectral response and the target electromagnetic spectral response match when the loss function converges.

[0103] As an alternative embodiment, the second processing module 13 is further configured to:

[0104] After acquiring the loss function between the actual electromagnetic spectral response and the target electromagnetic spectral response, if the loss function does not converge, acquire second adjustment parameters; perform a second parameter update operation on the inverse network and the forward prediction network based on the second adjustment parameters until the actual electromagnetic spectral response and the target electromagnetic spectral response match.

[0105] As an alternative embodiment, both the forward prediction network and the discriminative network include BN layers.

[0106] Please refer to Figure 11 , Figure 11 which is a schematic structural diagram of a device parameter acquisition device based on a neural network provided by this application. The device parameter acquisition device includes:

[0107] A memory 21 for storing computer programs;

[0108] A processor 22 for implementing the steps of the device parameter acquisition method based on a neural network described in any one of the above embodiments when executing the computer program.

[0109] Specifically, the memory 21 includes a non-volatile storage medium and an internal memory 21. The non-volatile storage medium stores an operating system and computer-readable instructions, and the internal memory 21 provides an environment for the operation of the operating system and computer-readable instructions in the non-volatile storage medium. When the processor 22 executes the computer program saved in the memory 21, the following steps can be implemented: generating device parameters according to the target requirements by using an inverse network, where the target requirements include a target electromagnetic spectrum response; inputting the device parameters and real samples into a discriminant network, and enabling the discriminant network to perform adversarial training using the device parameters and real samples to obtain a first adjustment parameter, and performing a first parameter update operation on the inverse network and the discriminant network based on the first adjustment parameter until the device parameters generated by the inverse network according to the target electromagnetic spectrum response meet the real conditions; inputting the device parameters that meet the real conditions into a forward prediction network to obtain an actual electromagnetic spectrum response, and when the actual electromagnetic spectrum response matches the target electromagnetic spectrum response, outputting the device parameters generated by the inverse network.

[0110] It can be seen that in this embodiment, the neural network composed of an inverse network, a discriminant network, and a forward prediction network optimizes corresponding device parameters for the target electromagnetic spectrum response. In this application, the inverse network and the discriminant network are combined for adversarial training to optimize the generated device parameters to make them close to the real parameters. At the same time, the training data set required for the adversarial network is small, saving hardware resources and improving the robustness of the neural network. Then, the device parameters that meet the real conditions generated by the inverse network are input into the forward prediction network to realize the combined training of the inverse network and the forward prediction network. When the actual electromagnetic spectrum response matches the target electromagnetic spectrum response, the device parameters generated by the inverse network are output, optimizing the generated device performance, and outputting device parameters with higher performance under the input of the desired target electromagnetic spectrum response.

[0111] As an optional embodiment, when the processor 22 executes the computer subprogram saved in the memory 21, the following steps can be implemented: inputting the target requirements into the inverse network, enabling the inverse network to obtain reconstruction parameters based on the target requirements, and performing feature extraction on the reconstruction parameters to generate device parameters.

[0112] As an alternative embodiment, when the processor 22 executes the computer subroutine stored in the memory 21, the following steps may be implemented: input the target requirement into the inverse network; perform upsampling based on the target requirement through a transposed convolutional layer to obtain reconstruction parameters; extract feature parameters from the reconstruction parameters through an encoding and decoding module, and generate device parameters using the feature parameters.

[0113] As an alternative embodiment, when the processor 22 executes the computer subroutine stored in the memory 21, the following steps may be implemented: enable the discriminant network to perform adversarial training using the device parameters and real samples to obtain the probability that the device parameters are real parameters; determine whether the probability is the target value; if not, obtain the first adjustment parameter, and perform a first parameter update operation on the inverse network and the discriminant network based on the first adjustment parameter; if so, determine that the device parameters generated by the inverse network according to the target electromagnetic spectrum response meet the real conditions.

[0114] As an alternative embodiment, when the processor 22 executes the computer subroutine stored in the memory 21, the following steps may be implemented: obtain the loss function between the actual electromagnetic spectrum response and the target electromagnetic spectrum response; when the loss function converges, determine that the actual electromagnetic spectrum response and the target electromagnetic spectrum response match.

[0115] As an alternative embodiment, when the loss function does not converge, obtain the second adjustment parameter; perform a second parameter update operation on the inverse network and the forward prediction network based on the second adjustment parameter until the actual electromagnetic spectrum response and the target electromagnetic spectrum response match.

[0116] Based on the above embodiments, as a preferred implementation, refer to Figure 12 , Figure 12 which is a schematic structural diagram of another device parameter acquisition device based on a neural network provided by an embodiment of the present application. The device parameter acquisition device based on a neural network further includes:

[0117] An input interface 23, connected to the processor 22, is used to obtain externally imported computer programs, parameters, and instructions, and store them in the memory 21 under the control of the processor 22. The input interface 23 may be connected to an input device to receive parameters or instructions manually input by a user. The input device may be a touch layer covered on a display screen, or a button, a trackball, or a touchpad provided on the terminal housing.

[0118] A display unit 24, connected to the processor 22, is used to display the data sent by the processor 22. The display unit 24 may be a liquid crystal display screen or an electronic ink display screen, etc.

[0119] The network port 25, which is connected to the processor 22, is used for communication connection with various external terminal devices. The communication technology adopted for this communication connection can be a wired communication technology or a wireless communication technology, such as Mobile High-Definition Link (MHL), Universal Serial Bus (USB), High-Definition Multimedia Interface (HDMI), Wireless Fidelity (WiFi), Bluetooth communication technology, Bluetooth Low Energy communication technology, communication technology based on IEEE802.11s, etc.

[0120] On the other hand, the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the method for obtaining device parameters based on a neural network described in any one of the above embodiments.

[0121] Specifically, the present application also provides a computer-readable storage medium, which may include: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk, or an optical disc. A computer program is stored on the storage medium. When the computer program is executed by a processor, the following steps are implemented: generating device parameters according to target requirements by using an inverse network, where the target requirements include a target electromagnetic spectrum response; inputting the device parameters and real samples into a discriminant network, and enabling the discriminant network to perform adversarial training using the device parameters and real samples to obtain a first adjustment parameter, and performing a first parameter update operation on the inverse network and the discriminant network based on the first adjustment parameter until the device parameters generated by the inverse network according to the target electromagnetic spectrum response meet the real conditions; inputting the device parameters that meet the real conditions into a forward prediction network to obtain an actual electromagnetic spectrum response, and when the actual electromagnetic spectrum response matches the target electromagnetic spectrum response, outputting the device parameters generated by the inverse network.

[0122] It can be seen that in this embodiment, the neural network composed of an inverse network, a discriminant network, and a forward prediction network optimizes corresponding device parameters for the target electromagnetic spectrum response. In the present application, the inverse network and the discriminant network are combined for adversarial training to optimize the generated device parameters to make them close to the real parameters. At the same time, the training data set required for the adversarial network is small, saving hardware resources and improving the robustness of the neural network. Then, the device parameters that meet the real conditions generated by the inverse network are input into the forward prediction network to realize the combined training of the inverse network and the forward prediction network. When the actual electromagnetic spectrum response matches the target electromagnetic spectrum response, the device parameters generated by the inverse network are output, optimizing the generated device performance, and outputting device parameters with higher performance under the input of the desired target electromagnetic spectrum response.

[0123] As an alternative embodiment, when the computer subroutine stored in the computer-readable storage medium is executed by a processor, the following steps can be specifically implemented: input the target requirement into the inverse network, enable the inverse network to obtain reconstruction parameters based on the target requirement, and perform feature extraction on the reconstruction parameters to generate device parameters.

[0124] As an alternative embodiment, when the computer subroutine stored in the computer-readable storage medium is executed by a processor, the following steps can be specifically implemented: input the target requirement into the inverse network; perform upsampling based on the target requirement through a transposed convolutional layer to obtain reconstruction parameters; extract feature parameters from the reconstruction parameters through an encoding and decoding module, and generate device parameters using the feature parameters.

[0125] As an alternative embodiment, when the computer subroutine stored in the computer-readable storage medium is executed by a processor, the following steps can be specifically implemented: enable the discriminant network to perform adversarial training using device parameters and real samples to obtain the probability that the device parameters are real parameters; determine whether the probability is the target value; if not, obtain the first adjustment parameter, and perform the first parameter update operation on the inverse network and the discriminant network based on the first adjustment parameter; if so, determine that the device parameters generated by the inverse network according to the target electromagnetic spectral response meet the real conditions.

[0126] As an alternative embodiment, when the computer subroutine stored in the computer-readable storage medium is executed by a processor, the following steps can be specifically implemented: obtain the loss function between the actual electromagnetic spectral response and the target electromagnetic spectral response; when the loss function converges, determine that the actual electromagnetic spectral response and the target electromagnetic spectral response match.

[0127] As an alternative embodiment, when the computer subroutine stored in the computer-readable storage medium is executed by a processor, the following steps can be specifically implemented: when the loss function does not converge, obtain the second adjustment parameter; perform the second parameter update operation on the inverse network and the forward prediction network based on the second adjustment parameter until the actual electromagnetic spectral response and the target electromagnetic spectral response match.

[0128] It should also be noted that in this specification, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the element.

[0129] The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not intended to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for obtaining device parameters based on a neural network, characterized in that The neural network includes an inverse network, a discriminant network, and a forward prediction network. The inverse network includes an input layer, a fully connected layer, a transposed convolutional layer, an encoding and decoding module, and an output layer. The encoding and decoding module includes an encoder and a decoder. Both the forward prediction network and the discriminant network include a BN layer. The method for obtaining device parameters includes: Using the inverse network to generate multiple device parameters according to the target requirements, where the target requirements are a one-dimensional vector including a target electromagnetic spectrum response, a target incident angle, a refractive index of the target device material, and simulated random noise; the device parameters include a device structure image; Inputting the device parameters and real samples into the discriminant network, and making the discriminant network perform adversarial training using the device parameters and the real samples to obtain the probability that the device parameters are real parameters. Determine whether the probability is the target value. If not, obtain the first adjustment parameter, and perform the first parameter update operation on the inverse network and the discriminant network based on the first adjustment parameter. If so, determine that the device parameters generated by the inverse network according to the target electromagnetic spectrum response meet the real conditions; Inputting the device parameters that meet the real conditions into the forward prediction network to obtain the actual electromagnetic spectrum response. When the actual electromagnetic spectrum response matches the target electromagnetic spectrum response, output the device parameters generated by the inverse network; The process of using the inverse network to generate device parameters according to the target requirements includes: Inputting the target requirements into the inverse network; Performing upsampling based on the target requirements through the transposed convolutional layer to obtain reconstruction parameters; Extracting feature parameters from the reconstruction parameters through the encoding and decoding module, and generating device parameters using the feature parameters; After inputting the device parameters that meet the real conditions into the forward prediction network to obtain the actual electromagnetic spectrum response, the method for obtaining device parameters further includes: Obtaining the loss function between the actual electromagnetic spectrum response and the target electromagnetic spectrum response; the loss function is used to complete backpropagation to update the gradient; When the loss function converges, determine that the actual electromagnetic spectrum response matches the target electromagnetic spectrum response; When the loss function does not converge, obtain the second adjustment parameter; Performing the second parameter update operation on the inverse network and the forward prediction network based on the second adjustment parameter until the actual electromagnetic spectrum response matches the target electromagnetic spectrum response.

2. A device parameter acquisition system based on a neural network, characterized in that, The neural network includes an inverse network, a discriminant network, and a forward prediction network. The inverse network includes an input layer, a fully connected layer, a transposed convolutional layer, an encoding and decoding module, and an output layer. The encoding and decoding module includes an encoder and a decoder. Both the forward prediction network and the discriminant network include a BN layer. The system for obtaining device parameters includes: A first processing module, configured to use the inverse network to generate multiple device parameters according to the target requirements, where the target requirements are a one-dimensional vector including a target electromagnetic spectrum response, a target incident angle, a refractive index of the target device material, and simulated random noise; the device parameters include a device structure image; An adjustment module, configured to input the device parameters and the real samples into the discriminant network, and enable the discriminant network to perform adversarial training using the device parameters and the real samples, obtain the probability that the device parameters are real parameters, determine whether the probability is a target value, if not, obtain a first adjustment parameter, and perform a first parameter update operation on the inverse network and the discriminant network based on the first adjustment parameter, if so, determine that the device parameters generated by the inverse network according to the target electromagnetic spectrum response meet the real conditions; A second processing module, configured to input the device parameters that meet the real conditions into the forward prediction network to obtain an actual electromagnetic spectrum response, and output the device parameters generated by the inverse network when the actual electromagnetic spectrum response matches the target electromagnetic spectrum response; The process of generating device parameters using the inverse network according to the target requirements includes: Inputting the target requirements into the inverse network; Performing upsampling based on the target requirements through the transposed convolutional layer to obtain reconstruction parameters; Extracting feature parameters from the reconstruction parameters through the encoding and decoding module, and generating device parameters using the feature parameters; The device parameter acquisition system is further configured to: After inputting the device parameters that meet the real conditions into the forward prediction network to obtain an actual electromagnetic spectrum response, obtain a loss function between the actual electromagnetic spectrum response and the target electromagnetic spectrum response; the loss function is used to complete backpropagation to update the gradient; When the loss function converges, determine that the actual electromagnetic spectrum response matches the target electromagnetic spectrum response; When the loss function does not converge, obtain a second adjustment parameter; Perform a second parameter update operation on the inverse network and the forward prediction network based on the second adjustment parameter until the actual electromagnetic spectrum response matches the target electromagnetic spectrum response.

3. A device parameter acquisition device based on a neural network, characterized in that, Including: A memory, configured to store a computer program; A processor, configured to implement the steps of the method for acquiring device parameters based on a neural network according to claim 1 when executing the computer program.

4. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, and when the computer program is executed by the processor, the steps of the method for acquiring device parameters based on a neural network according to claim 1 are implemented.

Images