Configuration generation method of adjustable photonic crystal based on deep learning

Abstract

The invention discloses a deep learning-based adjustable photonic crystal configuration generation method, a computing device, a computer program product and a storage medium, and the deep learning-based adjustable photonic crystal configuration generation method is executed in the computing device. The method comprises the following steps: inputting a first characteristic frequency vector indicating a dispersion relationship of a to-be-generated photonic crystal configuration into a configuration generation model for processing to obtain at least one candidate configuration image; performing image preprocessing on each candidate configuration image to obtain a plurality of similar images; respectively inputting the similar images into a dispersion relation prediction model for processing to obtain a second characteristic frequency vector representing the dispersion relation of the similar images; and respectively calculating an average absolute percentage error value of the first characteristic frequency vector and each second characteristic frequency vector, and selecting a similar image indicated by the second characteristic frequency vector corresponding to the minimum average absolute percentage error value as a configuration image of the photonic crystal to be generated.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a configuration generation method of an adjustable phononic crystal based on deep learning, a computing device, a computer program product and a storage medium. Background Art

[0002] With the acceleration of urbanization, noise pollution has become one of the important sources of environmental pollution in modern cities. Noise pollution refers to unnecessary sounds or sound signals in the environment, which come from a variety of sources, such as transportation, industry, and social activities. Noise not only comes from a wide range of sources, but also may have a negative impact on human health. Specifically, it may cause hearing loss, tinnitus, sleep disorders, mental health problems, etc. Noise pollution is very likely to occur in residential areas, workplaces, or learning environments, reducing people's work efficiency and quality of life. Noise pollution may also cause social problems and cause economic losses. At the community level, noise may cause neighborhood disputes, social unrest, and other problems; for the commercial and tourism industries, noise may affect the business environment and tourist experience, leading to economic losses. Therefore, noise reduction technology is needed to control the propagation of noise. Phononic crystals have attracted many researchers to conduct in-depth research on their sound absorption properties because they can reduce the propagation of sound within a specific frequency range.

[0003] As a new type of artificial composite structure, phononic crystals are composed of different material phases and are periodically distributed in space. If elastic waves cannot propagate inside phononic crystals within a certain characteristic frequency range, this property is called a band gap. Using the band gap characteristics, the propagation of sound or elastic waves can be artificially controlled, such as acoustic damping, noise control, waveguide filters, acoustic focusing and acoustic energy collection.

[0004] In recent years, phononic crystals with tunable band gaps have received extensive attention in multiple disciplines. Early studies mainly focused on the tunable materials or geometric parameters of phononic crystals. However, most of the proposed structures are difficult to manufacture, which limits the application of phononic crystals.

[0005] Topological optimization is a method for calculating material distribution. In terms of structural design, topological optimization has shown great potential. Currently, many studies have used it to design the configuration of phononic crystals. However, the topological design of phononic crystals with tunable band gaps is difficult. In addition, many existing methods make it difficult to design phononic crystals with tunable band gaps under given dispersion relations. Summary of the invention

[0006] In view of the above problems, the present invention is proposed to provide a method, a computing device and a storage medium that overcome the above problems or at least partially solve the above problems.

[0007] According to one aspect of the present invention, a configuration generation method for an adjustable phononic crystal based on deep learning is provided, which is executed in a computing device, and comprises: inputting a first characteristic frequency vector indicating the dispersion relation of the configuration of the phononic crystal to be generated into a configuration generation model for processing to obtain at least one candidate configuration image; performing image preprocessing on each candidate configuration image to obtain a plurality of similar images; inputting each similar image into a dispersion relation prediction model for processing to obtain a second characteristic frequency vector representing the dispersion relation of each similar image; respectively calculating the average absolute percentage error value of the first characteristic frequency vector and each second characteristic frequency vector, and selecting the similar image indicated by the second characteristic frequency vector corresponding to the smallest average absolute percentage error value as the configuration image of the phononic crystal to be generated.

[0008] Optionally, in the configuration generation method according to the present invention, for each candidate configuration image, image preprocessing is performed on it to obtain multiple similar images, including: performing image symmetry transformation processing on the candidate configuration image to obtain multiple initial similar images, the image symmetry transformation includes horizontal flipping, vertical flipping and diagonal transformation; using a median filter to filter discrete pixels of each initial similar image; binarizing the initial similar image with the discrete pixels filtered to obtain a similar image.

[0009] Optionally, the configuration generation method according to the present invention also includes: generating a second training sample set based on the constructed configuration image database, the second training sample set including configuration images of multiple phononic crystals and their corresponding dispersion relations, wherein the dispersion relations are represented by characteristic frequency vectors; based on the second training sample set, generating a dispersion relation prediction model by training using a regression loss function.

[0010] Optionally, in the configuration generation method according to the present invention, generating a second training sample set based on the constructed configuration image database includes:

[0011] The phononic crystal configuration images were generated by MATLAB, and the generated phononic crystal configuration images were used to build a configuration image database. The phononic crystal configuration images had various shapes and were single-material phononic crystal configurations. Equibiaxial stretching was used to perform stretching deformation to generate band gaps. MATLAB interface is used to input the phononic crystal configuration image into COMSOL to establish a geometric model; according to the defined relevant parameters and boundary conditions, the geometric model is simulated and solved to obtain the dispersion relation atlas corresponding to different phononic crystal configuration images, the relevant parameters include the hyperelastic material properties and shape control parameters of the geometric model, the boundary conditions include specified displacement conditions and periodic conditions, the dispersion relation atlas includes the dispersion relation diagram corresponding to the phononic crystal under different tensile strain states, the dispersion relation diagram consists of 11 curves, each curve consists of 31 characteristic frequency points, among which the band gap characteristics corresponding to different tensile strain states are different, and any dispersion relation diagram indicates the corresponding relationship between the wave vector and the frequency under the target tensile strain state; for each dispersion relation diagram, its corresponding fourth characteristic frequency vector is extracted; the second training sample set is constructed using the fourth characteristic frequency vector and its corresponding dispersion relation diagram.

[0012] Optionally, the configuration generation method according to the present invention also includes the step of training a configuration generation model: constructing a generative adversarial model based on the Wasserstein distance, the generative adversarial model including a generator and a discriminator; generating a first training sample set, wherein any training sample of the first training sample set includes a real image of the sample phononic crystal and a third characteristic frequency vector representing the dispersion relation of the configuration of the sample phononic crystal; inputting each third characteristic frequency vector into the generator, and outputting a generated image after processing; inputting the generated image and the corresponding real image into the discriminator together, and outputting a metric value indicating whether the generated image is consistent with the corresponding real image after processing; based on the metric value, iteratively training the generator and the discriminator using a loss function with a gradient penalty until the number of training rounds reaches a preset number of rounds and the training is completed, and the generator generated by the training is used as the configuration generation model.

[0013] Optionally, in the configuration generation method according to the present invention, the discriminator includes: a plurality of coupled convolution modules and an output module; the convolution module includes a convolution layer, an instance normalization layer, an activation function layer and a random inactivation layer coupled in sequence; the output module includes a fully connected module.

[0014] Optionally, in the configuration generation method according to the present invention, the dispersion relationship prediction model includes: a basic network component, a multi-layer perception component and a prediction output component coupled in sequence; wherein the basic network component includes multiple residual modules, and each residual module includes an attention mechanism module, and the attention mechanism module includes a channel attention sub-module and a spatial attention sub-module; the multi-layer perception component includes a coupled fully connected layer, an activation function layer and a random inactivation layer.

[0015] Optionally, in the configuration generation method according to the present invention, the regression loss function is:

[0016]

[0017] Where y represents the true value of each training sample in the second training sample set, f(x) represents the predicted value of the training sample, δ is the hyperparameter of the regression loss function, δ=1.

[0018] Optionally, in the configuration generation method according to the present invention, the geometric model is represented by the following formula:

[0019] x1=r(θ)cosθ,x2=r(θ)sinθ

[0020] Among them, r(θ)=r0[1+c1cos(n1θ)+c2cos(n2θ)+c3cos(n3θ)+c4cos(n4θ)]

[0021] (x1,x2) are the position coordinates of the intersection points of the scatterer and the matrix in the unit cell model, θ is the filling rate control parameter of the scatterer, c1~c4, n1~n4 and r0 are shape control parameters, the range of c1~c4 is [-0.35,0.35], the range of n1~n4 is [1,20], and the range of r0 is [0.65,2].

[0022] Optionally, in the configuration generation method according to the present invention, the specified displacement condition is to adopt an equibiaxial stretching method.

[0023] Optionally, in the configuration generation method according to the present invention, the first characteristic frequency vector indicating the dispersion relationship of the phononic crystal configuration to be generated is input into the configuration generation model for processing to obtain at least one candidate configuration image, including: reducing the dimension of the first characteristic frequency vector; fusing the first characteristic frequency vector after the dimension reduction processing with a random vector by dot product to obtain a fused vector; inputting the fused vector into the configuration generation model for processing to obtain at least one candidate configuration image.

[0024] Optionally, in the configuration generation method according to the present invention, the first characteristic frequency vector is subjected to dimensionality reduction processing, including: sequentially inputting the first characteristic frequency vector into two groups of fully connected layers coupled to each other for processing, to obtain a first characteristic frequency vector having the same dimension as the random vector.

[0025] Optionally, in the configuration generation method according to the present invention, it also includes: the discriminator is iteratively trained 5 times per training, and the generator is iteratively trained once.

[0026] Optionally, in the configuration generation method according to the present invention, the average absolute percentage error value of the first characteristic frequency vector and each second characteristic frequency vector is calculated respectively, including: scaling each characteristic frequency value contained in the first characteristic frequency vector and the second characteristic frequency vector; and calculating the average absolute percentage error value using each characteristic frequency value after scaling.

[0027] Optionally, in the configuration generation method according to the present invention, the characteristic frequency value is scaled using the following formula:

[0028] p t =a+k*(pa)

[0029] Among them, p t represents the characteristic frequency value after scaling, p represents the characteristic frequency value before scaling, a is a constant, and a=0.001, k is the scaling factor, and b=1,c max is the maximum value among the characteristic frequency values, c min is the minimum value of the characteristic frequency.

[0030] Optionally, in the configuration generation method according to the present invention, the mean absolute percentage error value is calculated by the following formula:

[0031]

[0032] in, and p ij are respectively the j-th predicted characteristic frequency value of the i-th second characteristic frequency vector and the j-th actual characteristic frequency value of the corresponding first characteristic frequency vector, M is the number of second characteristic frequency vectors, and N is the number of characteristic frequency values ​​contained in the second characteristic frequency vector.

[0033] According to another aspect of the present invention, there is provided a computing device, comprising: at least one processor; and a memory storing program instructions, wherein the program instructions are configured to be suitable for being executed by the at least one processor, and the program instructions include instructions for executing the above method.

[0034] According to yet another aspect of the present invention, there is provided a computer program product, comprising a computer program / instruction, wherein the computer program / instruction implements the above method when executed by a processor.

[0035] According to another aspect of the present invention, a readable storage medium storing program instructions is provided. When the program instructions are read and executed by a computing device, the computing device executes the above method.

[0036] According to the solution of the present invention, a deep learning framework that combines a configuration generation model and a dispersion relation prediction model is proposed. By using the deep learning framework to generate adjustable phononic crystal configurations, candidate configurations close to the optimal can be selected, thereby improving the accuracy of generating the adjustable phononic crystal configurations. At the same time, the time cost of generating the adjustable phononic crystals can be greatly reduced, thereby improving the configuration generation speed of the adjustable phononic crystals.

[0037] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0039] Figure 1 A block diagram of the physical components (ie, hardware) of a computing device 100 is shown;

[0040] Figure 2 A flowchart of a method 200 for generating a configuration of an adjustable phononic crystal based on deep learning according to an embodiment of the present invention is shown;

[0041] Figure 3 A phononic crystal configuration diagram (3a) according to an embodiment of the present invention and a corresponding dispersion relationship diagram (3b) thereof when no tensile deformation occurs are shown;

[0042] Figure 4a A schematic diagram of a network structure of a generative adversarial model generator according to an embodiment of the present invention is shown;

[0043] Figure 4b A schematic diagram of a network structure of a generative adversarial model discriminator according to an embodiment of the present invention is shown;

[0044] Figure 5 A schematic diagram of an image preprocessing process according to an embodiment of the present invention is shown;

[0045] Figure 6 A schematic diagram showing a candidate configuration image and a corresponding similar image according to an embodiment of the present invention;

[0046] Figure 7 A schematic diagram of a ResNet18 network residual block according to an embodiment of the present invention is shown;

[0047] Figure 8 A schematic diagram of a multi-layer perception component according to an embodiment of the present invention is shown;

[0048] Fig. 9 A schematic diagram of a construction process of a second training sample set according to an embodiment of the present invention is shown;

[0049] Fig.10 A schematic structural diagram of a phononic crystal unit cell configuration according to an embodiment of the present invention is shown;

[0050] Fig.11 A schematic diagram of the boundary conditions of a phononic crystal unit cell configuration according to an embodiment of the present invention is shown;

[0051] Fig.12a and Figure 12b Schematic diagrams of a phononic crystal configuration image and its corresponding dispersion relation diagram according to an embodiment of the present invention are respectively shown;

[0052] Fig.13 A schematic diagram showing a comparison between a first target dispersion relation and a corresponding first simulated dispersion relation;

[0053] Fig.14 A schematic diagram showing a comparison between a second target dispersion relation and a corresponding second simulated dispersion relation;

[0054] Fig.15 A schematic flow chart of a method for generating a configuration of an adjustable phononic crystal based on deep learning according to another embodiment of the present invention is shown. DETAILED DESCRIPTION

[0055] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0056] Due to their special properties, phononic crystals have received widespread attention in many disciplines, especially their band gap characteristics. The band gap characteristics of phononic crystals provide new solutions to practical engineering problems such as sound insulation and noise reduction. Tunable soft phononic crystals made of soft materials can perform real-time and reversible control of the band gap, so the band gap regulation technology has been further developed.

[0057] Due to its powerful learning ability, deep learning has also been applied to the field of phononic crystal design, mainly focusing on two aspects: forward prediction of properties and reverse design of configurations. However, there is currently little research on the application of deep learning in the design of tunable phononic crystals.

[0058] In order to solve the problems existing in the above-mentioned prior art, the solution of the present invention is proposed. This application proposes a deep learning framework that combines a configuration generation model and a dispersion relation prediction model. The deep learning framework is used to generate an adjustable phononic crystal configuration, and a candidate configuration close to the optimal can be selected, thereby improving the accuracy of generating the adjustable phononic crystal configuration. At the same time, the time cost of generating the adjustable phononic crystal can be greatly reduced, and the configuration generation speed of the adjustable phononic crystal can be improved.

[0059] The configuration generation method of the tunable phononic crystal based on deep learning can be executed in the computing device 100. The computing device 100 can be implemented as a server, such as an application server, a Web server, etc.; it can also be implemented as a desktop computer, a laptop computer, a processor chip, a tablet computer, etc., but is not limited thereto. Figure 1 1 shows a block diagram of the physical components (i.e., hardware) of a computing device 100. In a basic configuration, the computing device 100 includes at least one processing unit 102 and a system memory 104. According to one aspect, depending on the configuration and type of the computing device, the system memory 104 includes, but is not limited to, volatile storage (e.g., random access memory), non-volatile storage (e.g., read-only memory), flash memory, or any combination of such memories.

[0060] According to one aspect, system memory 104 includes operating system 105. System memory 104 also includes program modules 150. According to one aspect, operating system 105, for example, is suitable for controlling the operation of computing device 100. Furthermore, examples are practiced in conjunction with graphics libraries, other operating systems, or any other application programs, and are not limited to any particular application or system. Figure 1 This basic configuration is illustrated in FIG. 1 by those components within dashed line 108. According to one aspect, computing device 100 has additional features or functionality. For example, according to one aspect, computing device 100 includes additional data storage devices (removable and / or non-removable), such as magnetic disks, optical disks, or tapes. Such additional storage Figure 1 1 is illustrated by a removable storage device 109 and a non-removable storage device 110.

[0061] As stated above, according to one aspect, a program module 150 is stored in the system memory 104. According to one aspect, the program module 150 can be implemented as one or more computer program products. The present application does not limit the type of computer program products, and may include, for example, email, word processing application, spreadsheet application, database application, slide show application, painting or computer-aided application, web browser, etc. In some embodiments according to the present application, the computer program / instructions related to the configuration generation method 200 of the tunable phononic crystal based on deep learning are packaged as a computer program product. When these computer programs / instructions are executed by the processor (i.e., the processing unit 102), the configuration generation method 200 of the tunable phononic crystal based on deep learning according to the present application is implemented.

[0062] According to one aspect, the examples may be practiced on a circuit comprising discrete electronic components, a packaged or integrated electronic chip containing logic gates, a circuit utilizing a microprocessor, or a single chip containing electronic components or a microprocessor. Figure 1 Each or many components shown in can be integrated into a system on a chip (SOC) on a single integrated circuit to practice examples. According to one aspect, such a SOC device may include one or more processing units, a graphics unit, a communication unit, a system virtualization unit, and various application functions, all of which are integrated (or "burned") into a chip substrate as a single integrated circuit. When operated via SOC, the functions described in this article can be operated via a dedicated logic integrated with other components of the computing device 100 on a single integrated circuit (chip). Embodiments of the present invention can also be practiced using other technologies capable of performing logical operations (such as AND, OR, and NOT), including but not limited to mechanical, optical, fluid, and quantum technologies. In addition, embodiments of the present invention can be practiced in a general-purpose computer or in any other circuit or system.

[0063] According to one aspect, the computing device 100 may also have one or more input devices 112, such as a keyboard, a mouse, a pen, a voice input device, a touch input device, etc. Output devices 114 may also be included, such as a display, a speaker, a printer, etc. The aforementioned devices are examples and other devices may also be used. The computing device 100 may include one or more communication connections 116 that allow communication with other computing devices 118, which may be printing devices, such as printers. Examples of suitable communication connections 116 include, but are not limited to: RF transmitters, receivers and / or transceiver circuits; Universal Serial Bus (USB), parallel and / or serial ports.

[0064] The term computer-readable medium as used herein includes computer storage media. Computer storage media may include volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (e.g., computer-readable instructions, data structures, or program modules). System memory 104, removable storage device 109, and non-removable storage device 110 are all examples of computer storage media (i.e., memory storage). Computer storage media may include random access memory (RAM), read-only memory (ROM), electrically erasable read-only memory (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disk (DVD) or other optical storage, cassettes, tapes, disk storage or other magnetic storage devices, or any other products that can be used to store information and can be accessed by computer device 100. According to one aspect, any such computer storage medium can be a part of computing device 100. Computer storage media does not include carrier waves or other propagated data signals.

[0065] According to one aspect, communication media is implemented by computer readable instructions, data structures, program modules, or other data in a modulated data signal (e.g., a carrier wave or other transport mechanism), and includes any information delivery media. According to one aspect, the term "modulated data signal" describes a signal that has one or more characteristics set or changed in a manner that encodes information in the signal. By way of example and not limitation, communication media includes wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, radio frequency (RF), infrared, and other wireless media.

[0066] In an embodiment according to the present invention, the computing device 100 is configured to execute the configuration generation method 200 of the tunable phononic crystal based on deep learning according to the present invention. The computing device 100 includes one or more processors, and one or more readable storage media storing program instructions. When the program instructions are configured to be executed by one or more processors, the computing device executes the configuration generation method 200 of the tunable phononic crystal based on deep learning in the embodiment of the present invention.

[0067] Figure 2 A flowchart of a method 200 for generating a configuration of an adjustable phononic crystal based on deep learning according to an embodiment of the present invention is shown. The method 200 is suitable for being executed in a computing device (such as the aforementioned computing device 100).

[0068] like Figure 2As shown, the purpose of method 200 is to realize a deep learning framework that combines a configuration generation model and a dispersion relation prediction model to generate an adjustable phononic crystal configuration, which can select a candidate configuration close to the optimal one, thereby improving the accuracy of generating the adjustable phononic crystal configuration, and at the same time can greatly reduce the time cost of generating the adjustable phononic crystal, thereby improving the configuration generation speed of the adjustable phononic crystal.

[0069] The method 200 begins at step 202 , in which a first characteristic frequency vector indicating a dispersion relation of a to-be-generated phononic crystal configuration is input into a configuration generation model for processing to obtain at least one candidate configuration image.

[0070] The dispersion relation of the phononic crystal configuration describes the relationship between the frequency and wave vector of the phonon, that is, the energy-momentum (Ek) relationship of the phonon. The band gap characteristic in the dispersion relation is a key property of the phononic crystal. By utilizing the band gap characteristic, the propagation of sound waves or elastic waves can be artificially controlled, such as acoustic damping, noise control, waveguide filters, acoustic focusing and acoustic energy collection.

[0071] The dispersion relation of the phononic crystal configuration is generally characterized by a dispersion relation diagram, such as Figure 3 As shown, Figure 3 A phononic crystal configuration diagram (3a) according to an embodiment of the present invention and its corresponding dispersion relationship diagram (3b) when no tensile deformation occurs are shown. According to some embodiments of the present application, each phononic crystal configuration diagram corresponds to 3 dispersion relationship diagrams under different tensile deformations, for example, a first dispersion relationship diagram with a tensile displacement of 0 mm, a second dispersion relationship diagram with a tensile displacement of 0.5 mm, and a third dispersion relationship diagram with a tensile displacement of 1 mm. The three dispersion relationship diagrams are each composed of 11 curves, and each curve is composed of 31 characteristic frequency points, that is, each phononic crystal configuration corresponds to 3*11*31=1023 characteristic frequency points, and the characteristic frequency values ​​of these 1023 characteristic frequency points constitute the characteristic frequency vector indicating the dispersion relationship.

[0072] It is worth noting that in order to avoid the insufficient fusion of the first eigenfrequency vector and the noise information (random vector) in this embodiment, the configuration generation model is difficult to capture the label features corresponding to the phononic crystal configuration, resulting in a poor diversity of the generated image. To this end, it is necessary to reduce the dimension of the first eigenfrequency vector before inputting it into the network. Specifically, the first eigenfrequency vector is subjected to dimensionality reduction. The first eigenfrequency vector after dimensionality reduction is fused with the random vector by dot multiplication to obtain a fused vector. The fused vector is used as the input vector of the configuration generation model.

[0073] In some embodiments, the first characteristic frequency vector is subjected to dimensionality reduction processing, including: sequentially inputting the first characteristic frequency vector into two groups of fully connected layers coupled to each other for processing, to obtain a first characteristic frequency vector having the same dimension as the random vector.

[0074] In some embodiments, the step of training a configuration generation model is also included:

[0075] Firstly, a generative adversarial model (WGAN) based on Wasserstein distance is constructed. The generative adversarial model includes a generator and a discriminator.

[0076] Referring to FIG. 4 , FIG. 4 shows a schematic diagram of a network structure of a generative adversarial model generator according to an embodiment of the present invention ( Figure 4a ) and the discriminator network structure diagram ( Figure 4b ), wherein the generator includes: an input module, a fully connected module, a batch normalization processing module, a plurality of groups of convolution modules and an output module coupled in sequence. The discriminator includes: a plurality of coupled convolution modules and an output module, the convolution module includes a convolution layer, an instance normalization layer, an activation function layer and a random dropout layer coupled in sequence, and the output module includes a fully connected module.

[0077] Then, a first training sample set is generated, and any training sample of the first training sample set includes a real image of the sample phononic crystal and a third characteristic frequency vector representing the dispersion relation of the sample phononic crystal configuration.

[0078] Afterwards, each third characteristic frequency vector is input into the generator, and the generated image is output after processing.

[0079] Subsequently, the generated image and the corresponding real image are input into the discriminator together, and after processing, a metric value indicating whether the generated image is consistent with the corresponding real image is output.

[0080] Finally, based on the metric value, the generator and the discriminator are iteratively trained using a loss function with a gradient penalty until the number of training rounds reaches a preset number of rounds, the training is completed, and the generator generated by the training is used as the configuration generation model.

[0081] The WGAN model is prone to problems such as gradient explosion or gradient disappearance. In order to solve these problems, this embodiment introduces the gradient penalty technology to limit the gradient norm of WGAN. In fact, the gradient penalty (GP) is used to constrain the gradient norm of the WGAN network discriminator. Compared with the weight clipping technology, the gradient penalty can better meet the Lipschitz constraint and make the weight distribution of the WGAN network discriminator more uniform. In general, adding a gradient penalty term to the discriminator loss function can make the training process of the network model more stable and can better solve problems such as mode collapse. The WGAN with the gradient penalty term is called WGAN-GP, and its objective function is shown in the following formula.

[0082]

[0083] Among them, the hyperparameter γ is the gradient penalty coefficient, Satisfying equation 4.11, the range of ε is [0,1].

[0084]

[0085] Preferably, the preset number of rounds can be set to 500 rounds. In addition, it is worth noting that the discriminator is used to distinguish the true or false of the configuration, and it takes a longer time to train. The generator only needs to generate images, and the training time is less. Therefore, in order to ensure the effectiveness of the training, this embodiment trains the generator once after training the discriminator for 5 iterations, that is, the discriminator is trained 5 times for each iteration, and the generator is trained once for each iteration.

[0086] After the candidate configuration images are obtained, the process proceeds to step 204 , where each candidate configuration image is subjected to image preprocessing to obtain a plurality of similar images.

[0087] In some embodiments, Figure 5 As shown, the candidate configuration image is preprocessed, including:

[0088] Firstly, the candidate configuration images are subjected to image symmetry transformation to obtain multiple initial similar images. The image symmetry transformation includes horizontal flipping, vertical flipping and diagonal transformation.

[0089] Then, the discrete pixels of each initial similar image are filtered using a median filter.

[0090] Finally, the initial similarity image after filtering the discrete pixels is binarized to obtain a similarity image. Figure 6 As shown, Figure 6 Schematic diagram of two candidate configuration images and corresponding similar images after image preprocessing. It is worth noting that Figure 6Only five similar images of each candidate configuration image are shown, and not all the corresponding similar images are shown. The present application does not limit the specific number of similar images.

[0091] Afterwards, in step 206, each similar image is input into a dispersion relation prediction model for processing to obtain a second characteristic frequency vector representing the dispersion relation of each similar image.

[0092] The dispersion relation prediction model includes: a basic network component, a multi-layer perception component and a prediction output component coupled in sequence. In this embodiment, ResNet18 is selected as the basic network component, which can further improve the accuracy of dispersion relation prediction.

[0093] In this embodiment, if Figure 7 As shown, in order to further reduce the error, a CBAM attention mechanism module is added to all residual blocks of the ResNet18 network, that is, the basic network component includes multiple residual modules, and each residual module contains an attention mechanism module, and the attention mechanism module includes a channel attention submodule and a spatial attention submodule.

[0094] In a specific example, the ResNet18 with the CBAM attention mechanism module added is compared with the original ResNet18, and the experimental results are shown in Table 1. As can be seen from the table, after adding the CBAM attention mechanism, the error of ResNet18 has decreased from 2.26% to 1.29%.

[0095] Table 1 Comparison of models with and without CBAM

[0096]

[0097]

[0098] like Figure 8 As shown, the multi-layer perception component (MLP component) includes a coupled fully connected layer, an activation function layer, and a random dropout layer.

[0099] The fully connected layer is one of the basic layers in the deep learning model. In the fully connected layer, the outputs of all input neurons are connected to all output neurons, forming a fully connected network structure. Generally, a fully connected layer is added at the end of the neural network to convert the feature maps extracted by the previous layer into the final output results. The fully connected layer can be used for tasks such as classification and regression. By learning the complex relationship between the input data, it can effectively represent and predict the input data. It should be noted that the fully connected layer has a large number of parameters and is prone to overfitting. In order to avoid overfitting, some regularization methods can be used, such as Dropout, L1 / L2 regularization, etc.

[0100] ReLU (Rectified Linear Unit) is a commonly used activation function, which is widely used in deep learning. Its mathematical expression is f(x) = max(0,x), that is, when the input is greater than 0, the output is x, otherwise the output is 0. The ReLU activation function has a simple calculation and derivative calculation method, and can alleviate the gradient disappearance problem and accelerate the network convergence speed. Compared with traditional activation functions (such as Sigmoid and Tanh), ReLU can usually provide better performance and faster training speed when training deep neural networks.

[0101] Dropout is a widely used regularization technique that aims to reduce the risk of overfitting in deep neural networks. During the training process, Dropout randomly sets the output of some neurons to zero with a specified probability, that is, temporarily shuts down these neurons, thereby reducing the dependencies between neurons and preventing the model from over-relying on specific neurons, which can make the network more robust and improve generalization capabilities. In the testing phase, Dropout does not work, but instead maintains the desired output by scaling the output of neurons that were shut down during training. Dropout layers are usually added between fully connected layers, and of course can also be used in convolutional layers. Appropriate Dropout probability can help improve the performance and generalization ability of the model, and is often used in the training of deep neural networks.

[0102] In some embodiments, the MLP component is set at the end of ResNet18, and the number is set to 1.

[0103] In some embodiments, the method 200 further includes the step of training and generating a dispersion relation prediction model:

[0104] Based on the constructed configuration image database, a second training sample set is generated, wherein the second training sample set includes configuration images of multiple phononic crystals and their corresponding dispersion relations, wherein the dispersion relations are represented by characteristic frequency vectors. And

[0105] Based on the second training sample set, a dispersion relation prediction model is generated by training using a regression loss function.

[0106] Among them, Fig.10 As shown, based on the constructed configuration image database, generating the second training sample set specifically includes:

[0107] First, MATLAB is used to generate a phononic crystal configuration image, and the generated phononic crystal configuration image is used to construct the configuration image database, wherein the phononic crystal configuration image has a variety of shapes and is a single-material phononic crystal configuration, and an equibiaxial stretching method is used to perform stretching deformation to generate a band gap.

[0108] Then, through the COMSOL with MATLAB interface, the phononic crystal configuration image is input into COMSOL to establish a geometric model.

[0109] Then, according to the defined relevant parameters and boundary conditions, the geometric model is simulated and solved to obtain a dispersion relation atlas corresponding to different phononic crystal configuration images, wherein the relevant parameters include the hyperelastic material properties and shape control parameters of the geometric model, and the boundary conditions include specified displacement conditions and periodic conditions. The dispersion relation atlas includes dispersion relation diagrams corresponding to phononic crystals under different tensile strain states, and the dispersion relation diagram consists of 11 curves, each of which consists of 31 characteristic frequency points, wherein the band gap characteristics corresponding to different tensile strain states are different, and any dispersion relation diagram indicates the corresponding relationship between the wave vector and the frequency under the target tensile strain state.

[0110] Afterwards, for each dispersion relation diagram, the corresponding fourth eigenfrequency vector is extracted.

[0111] Finally, the second training sample set is constructed using the fourth eigenfrequency vector and its corresponding dispersion relation diagram.

[0112] In some embodiments, the regression loss function is:

[0113]

[0114] Where y represents the true value of each training sample in the second training sample set, f(x) represents the predicted value of the training sample, δ is the hyperparameter of the regression loss function, δ=1.

[0115] In a specific example, the choice of loss function is also crucial, which can make the model converge better. Commonly used loss functions in regression problems include MAE (mean absolute error), MSE (mean square error), Huber Loss (regression loss function), etc. The MAE curve is V-shaped, continuous but not differentiable at yf(x) = 0, and the gradient is equal everywhere else. This means that even for small loss values, its gradient is relatively large. This property of MAE makes the function convergence and model learning ability relatively poor. Since MAE calculates the absolute error, there is no square term, and the penalty intensity is the same, it is not sensitive to outliers. The MSE curve is characterized by smoothness, continuity, and differentiability. When the difference between y and f(x) is greater than 1, MSE will impose a relatively large penalty on the model, while a relatively small penalty will be imposed when the difference is less than 1. Therefore, it is more sensitive to outliers. However, as the error continues to decrease, the gradient will continue to decrease, which is actually conducive to the convergence of the function. Even if the model fixes the learning rate, the function can converge to the minimum value relatively quickly. The Huber loss function has the advantages of both MSE and MAE. Compared with MAE, it can be differentiated everywhere and can converge faster. Compared with MSE, it is not sensitive to outliers and abnormal values. The formulas of the three loss functions of MAE, MSE and Huber Loss are as follows.

[0116]

[0117] In the above formula, y i and y represent the true value of the training sample, f(x i ) and f(x) represent the predicted values ​​of the training samples. δ is the hyperparameter of Huber Loss. The size of δ determines the emphasis of Huber Loss on MSE and MAE. The default value is 1.

[0118] In a specific example, Fig.10 FIG. 2 shows a schematic diagram of a phononic crystal unit cell configuration according to an embodiment of the present invention. Fig.10 As shown, a0 is the lattice constant, which is 10 mm, and the image pixel size is 128 × 128. In this embodiment, the unit cell model is composed of only one material, that is, the black part (matrix) in the figure, and the white part is air (scatterer), which belongs to a two-dimensional solid-fluid type tunable phononic crystal.

[0119] 1. Related parameters

[0120] (1) Material parameters.

[0121] This embodiment uses the neo-Hookean hyperelastic model to describe the properties of the material, and the corresponding strain energy density function is shown in the following formula: Wherein, I1 is the first strain deviator invariant, and J is the local volume change in the theoretical model.

[0122]

[0123] Since the stretching deformation simulation of the phononic crystal is required, this embodiment considers using a hyperelastic material, which is assumed to be silicone rubber. Its parameters are shown in Table 2, where ρ0 represents the initial density, K is the initial bulk modulus, and C 10 is a parameter in the strain energy density function. In this case, the initial Poisson's ratio is about 0.49, indicating that the material is almost incompressible and only stretchable.

[0124] Table 1

[0125]

[0126] (2) Shape control parameters

[0127] Fig.11 The white part in the figure is air, and its outline shape is represented by the (x1,x2) coordinates of the equation, which is shown below. Among them, r0 is a parameter that controls the length of r(θ).

[0128]

[0129] The area of ​​the white part and the area of ​​the entire phononic crystal configuration image are:

[0130]

[0131] Therefore, the fill rate of the white part can be expressed as:

[0132]

[0133] In this embodiment, a phononic crystal unit cell model is generated by using a fourth-order equation, as shown in the following equation.

[0134] x1=r(θ)cosθ,x2=r(θ)sinθ

[0135] r(θ)=r0[1+c1cos(n1θ)+c2cos(n2θ)+c3cos(n3θ)+c4cos(n4θ)]

[0136] The above formula has a total of 9 shape control parameters: c1~c4, n1~n4 and r0. This paper sets the range of c1~c4 to [-0.35,0.35], the range of n1~n4 to [1,20], and the range of r0 to [0.65,2]. These 9 control parameters can generate complex phononic crystal samples with rich and diverse shapes. Most of them are uniaxially symmetrical. In order to facilitate subsequent analysis, it is considered to appropriately reduce the sample space.

[0137] This embodiment applies symmetry conditions to ensure that the generated image is symmetrical in the vertical, horizontal, and diagonal directions. In addition, the increase in the filling rate helps to enrich the initial band gap of the undeformed phononic crystal configuration. In order to make the evolution of the band gap opening or closing as obvious as possible, the filling rate should be appropriately increased. Preferably, the filling rate φ is set in the range of [0.5, 0.75].

[0138] 2. Boundary conditions

[0139] In this embodiment, the specified displacement condition determines that the deformation analysis is performed by equibiaxial stretching, and fixed constraints and specified displacements are set at the same time, such as Fig.11 As shown, fixed constraints are added to the two edges on the coordinate axis, and the displacements are set for the other two edges, which are Δl x and Δl y (Δl x =Δl y ), and then add periodic conditions to the geometry in both directions.

[0140] Fig.12a and Figure 12b Schematic diagrams of a phononic crystal configuration image and its corresponding dispersion relation diagram according to an embodiment of the present invention are shown respectively. During simulation calculation, this embodiment scans 31 wave vectors along the path GXMG to calculate the first 11 eigenfrequencies, and the stretching displacements are set to 0mm, 0.5mm and 1mm respectively, so one phononic crystal configuration image can obtain three dispersion relation diagrams. Fig.14 As shown, two phononic crystal configuration images are shown respectively ( Fig.12a and Figure 12b ) and its corresponding three dispersion relation graphs. The three dispersion relation graphs corresponding to the same phononic crystal configuration image in the figure respectively represent the dispersion relation conditions when the stretching displacement is 0 mm, 0.5 mm, and 1 mm.

[0141] from Fig.12a It can be seen that the phononic crystal configuration has rich band gap characteristics. As the stretching displacement increases, the band gap width between the fifth and sixth energy bands gradually narrows until it closes. At the same time, the band gap between the ninth and tenth energy bands opens and the width gradually widens.

[0142] from Figure 12b It can be seen that the phononic crystal configuration also has rich band gap characteristics. As the tensile displacement increases, the band gap width between the fifth and sixth energy bands gradually narrows until it is closed; at the same time, the band gap between the third and fourth energy bands is opened. This further illustrates that the change in configuration geometry caused by external mechanical tensile loads can effectively change the band gap and regulate the opening and closing of the band gap.

[0143] Finally, in step 208, the mean absolute percentage error (MAPE) of the first eigenfrequency vector and each second eigenfrequency vector is calculated respectively, and the similarity image indicated by the second eigenfrequency vector corresponding to the smallest mean absolute percentage error is selected as the configuration image of the phononic crystal to be generated.

[0144] The step of respectively calculating the average absolute percentage error value of the first characteristic frequency vector and each second characteristic frequency vector comprises: firstly scaling each characteristic frequency value included in the first characteristic frequency vector and the second characteristic frequency vector, and then calculating the average absolute percentage error value using each characteristic frequency value after scaling.

[0145] In some embodiments, the characteristic frequency value is scaled using the following formula:

[0146] p t =a+k*(pa)

[0147] Among them, p t represents the characteristic frequency value after scaling, p represents the characteristic frequency value before scaling, a is a constant, and a=0.001, k is the scaling factor, and b=1,c max is the maximum value of the characteristic frequency value (since the characteristic frequency value is a decimal, it can be rounded to 7005 in this embodiment), c min is the minimum value of the characteristic frequency value (0).

[0148] In some embodiments, the mean absolute percentage error value may be calculated by the following formula:

[0149]

[0150] in, and p ij are respectively the j-th predicted characteristic frequency value of the i-th second characteristic frequency vector and the j-th actual characteristic frequency value of the corresponding first characteristic frequency vector, M is the number of second characteristic frequency vectors, and N is the number of characteristic frequency values ​​contained in the second characteristic frequency vector.

[0151] In addition, in order to verify the accuracy of the phononic crystal configuration generated by the above method 200, in some embodiments, the method 200 further includes verifying the configuration image of the selected phononic crystal to be generated by using a finite element simulation method.

[0152] Specifically, the target configuration image and target dispersion relation diagram of the phononic crystal generated by method 200 are obtained, and the target configuration image is used as the real configuration, and the target dispersion relation is used as the real dispersion relation. The real configuration is input into the COMSOL finite element for simulation to obtain the generated configuration and the simulated dispersion relation diagram, and the real dispersion relation diagram and the simulated dispersion relation diagram are compared to verify the accuracy of the generated phononic crystal configuration.

[0153] In some embodiments, Fig.13 and 14 As shown, Fig.13 A schematic diagram showing a comparison between a first target dispersion relation and a corresponding first simulated dispersion relation is shown. Fig.14 A schematic diagram showing a comparison between the second target dispersion relation and the corresponding second simulated dispersion relation is shown.

[0154] Fig.13 The MAPE between the first target dispersion relation (corresponding to the real configuration on the upper left) and the first simulated dispersion relation (corresponding to the generated configuration on the lower left) is 2.95%. Fig.14 The MAPE between the second target dispersion relation and the second simulated dispersion relation is shown to be 1.91%.

[0155] As can be seen from the figure, the energy bands in the low frequency range can be matched more accurately, while the accuracy in the higher frequency range decreases. However, overall, the simulated dispersion relation of the generated configuration is consistent with the real dispersion relation in terms of characteristic frequency values ​​and overall trends, indicating that the phononic crystal configuration generated by method 200 is more accurate.

[0156] In a specific example, combining Fig.15 The configuration generation method of the present application is further explained.

[0157] like Fig.15 As shown, when a dispersion relation (target dispersion relation) is given, the characteristic frequency indicating the target dispersion relation is used as input data, and after adding random noise to the input data, it is input into the ACWGAN-GP generator (configuration generation model).

[0158] After being processed by the ACWGAN-GP generator, candidate configuration images are generated, and the candidate configuration images are post-processed (pre-processed) to obtain multiple similar images.

[0159] Each similar image is input into ReNet (dispersion relation prediction model) for processing to obtain the predicted dispersion relation corresponding to each similar image.

[0160] The mean absolute percentage error (MAPE) between the characteristic frequency of the target dispersion relation and the characteristic frequency of the predicted dispersion relation of each similar image is calculated respectively.

[0161] The similarity image with the smallest MAPE value is taken as the final configuration image of the target dispersion relation.

[0162] The final configuration image is simulated using finite element simulation (COMSOL) to obtain a simulated dispersion relation. The error between the simulated dispersion relation and the target dispersion relation is used to detect the accuracy of the generated configuration.

[0163] The method provided by the present invention proposes a deep learning framework that combines a configuration generation model with a dispersion relation prediction model. The deep learning framework is used to generate an adjustable phononic crystal configuration, and a candidate configuration close to the optimal one can be selected, thereby improving the accuracy of generating the adjustable phononic crystal configuration. At the same time, the time cost of generating the adjustable phononic crystal can be greatly reduced, thereby improving the configuration generation speed of the adjustable phononic crystal.

[0164] A8. The method as described in A3, wherein the regression loss function is: Wherein, y represents the true value of each training sample in the second training sample set, f(x) represents the predicted value of the training sample, δ is the hyperparameter of the regression loss function, δ = 1. A9. The method as described in A7, wherein the geometric model is represented by the following formula: x1 = r(θ)cosθ, x2 = r(θ)sinθ wherein r(θ) = r0[1+c1cos(n1θ)+c2cos(n2θ)+c3cos(n3θ)+c4cos(n4θ)](x1,x2) are the position coordinates of each intersection point of the scatterer and the matrix in the unit cell model, θ is the filling rate control parameter of the scatterer, c1~c4, n1~n4 and r0 are shape control parameters, the range of c1~c4 is [-0.35,0.35], the range of n1~n4 is [1,20], and the range of r0 is [0.65,2]. A10. The method as described in A7, wherein the specified displacement condition is to adopt an equibiaxial stretching method. A11. The method as described in A1, wherein the first characteristic frequency vector indicating the dispersion relation of the phononic crystal configuration to be generated is input into the configuration generation model for processing to obtain at least one candidate configuration image, including: reducing the dimension of the first characteristic frequency vector; fusing the first characteristic frequency vector after the dimension reduction processing with the random vector by dot product to obtain a fused vector; inputting the fused vector into the configuration generation model for processing to obtain at least one candidate configuration image. A12. The method as described in A11, wherein the first characteristic frequency vector is reduced in dimension, including: inputting the first characteristic frequency vector into two groups of fully connected layers coupled to each other in sequence for processing to obtain a first characteristic frequency vector having the same dimension as the random vector. A13. The method as described in 5, wherein it also includes: the discriminator is trained 5 times per iteration, and the generator is trained once per iteration. A14. The method as described in A1, wherein the average absolute percentage error value of the first characteristic frequency vector and each second characteristic frequency vector is calculated respectively, including: scaling each characteristic frequency value contained in the first characteristic frequency vector and the second characteristic frequency vector; and calculating the average absolute percentage error value using each characteristic frequency value after scaling. A15. The method as described in A14, wherein the characteristic frequency value is scaled using the following formula: p t =a+k*(pa) where p t represents the characteristic frequency value after scaling, p represents the characteristic frequency value before scaling, a is a constant, and a=0.001, k is the scaling factor, and b=1,c max is the maximum value among the characteristic frequency values, c min is the minimum value of the characteristic frequency value. A16. The method as described in A14 or A15, wherein the mean absolute percentage error value is calculated by the following formula: in, and p ij are respectively the j-th predicted characteristic frequency value of the i-th second characteristic frequency vector and the j-th actual characteristic frequency value of the corresponding first characteristic frequency vector, M is the number of second characteristic frequency vectors, and N is the number of characteristic frequency values ​​contained in the second characteristic frequency vector.

[0165] Although the present invention has been described in terms of a limited number of embodiments, it will be apparent to those skilled in the art, having benefit of the above description, that other embodiments are contemplated within the scope of the invention thus described. Furthermore, it should be noted that the language used in this specification is selected primarily for readability and didactic purposes, rather than for the purpose of explaining or defining the subject matter of the present invention. Therefore, many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the appended claims.

Claims

1. A method for generating a configuration of an adjustable phononic crystal based on deep learning, executed in a computing device, the method comprising: Inputting a first characteristic frequency vector indicating a dispersion relation of a phononic crystal configuration to be generated into a configuration generation model for processing to obtain at least one candidate configuration image; For each candidate configuration image, image preprocessing is performed to obtain multiple similar images; Inputting each similar image into a dispersion relation prediction model for processing, respectively, to obtain a second characteristic frequency vector representing the dispersion relation of each similar image; The average absolute percentage error values ​​of the first eigenfrequency vector and each second eigenfrequency vector are calculated respectively, and the similar image indicated by the second eigenfrequency vector corresponding to the smallest average absolute percentage error value is selected as the configuration image of the phononic crystal to be generated.

2. The method of claim 1, wherein: For each candidate configuration image, image preprocessing is performed to obtain multiple similar images, including: Performing image symmetry transformation processing on the candidate configuration image to obtain multiple initial similar images, wherein the image symmetry transformation includes horizontal flipping, vertical flipping and diagonal transformation; Filter the discrete pixels of each initial similar image using a median filter; The initial similar image after filtering the discrete pixels is binarized to obtain the similar image.

3. The method of claim 1, wherein: Also includes: Based on the constructed configuration image database, a second training sample set is generated, wherein the second training sample set includes configuration images of a plurality of phononic crystals and their corresponding dispersion relations, wherein the dispersion relations are represented by characteristic frequency vectors; Based on the second training sample set, a dispersion relation prediction model is generated by training using a regression loss function.

4. The method of claim 3, wherein: Based on the constructed configuration image database, a second training sample set is generated, including: Generate a phononic crystal configuration image using MATLAB, and construct the configuration image database using the generated phononic crystal configuration image, wherein the phononic crystal configuration image has a variety of shapes and is a single-material phononic crystal configuration, and is stretched and deformed by equibiaxial stretching to generate a band gap; Through the COMSOL with MATLAB interface, the phononic crystal configuration image is input into COMSOL to establish a geometric model; According to the defined relevant parameters and boundary conditions, the geometric model is simulated and solved to obtain dispersion relation atlases corresponding to different phononic crystal configuration images, wherein the relevant parameters include the hyperelastic material properties and shape control parameters of the geometric model, and the boundary conditions include specified displacement conditions and periodic conditions. The dispersion relation atlases include dispersion relation diagrams corresponding to the phononic crystal under different tensile strain states, and the dispersion relation diagrams are composed of 11 curves, each of which is composed of 31 characteristic frequency points, wherein the band gap characteristics corresponding to different tensile strain states are different, and any dispersion relation diagram indicates the corresponding relationship between the wave vector and the frequency under the target tensile strain state; For each dispersion relation diagram, extract the corresponding fourth eigenfrequency vector; The second training sample set is constructed by using the fourth characteristic frequency vector and its corresponding dispersion relation diagram.

5. The method of claim 1, wherein: It also includes the steps of training the configuration generation model: Constructing a generative adversarial model based on Wasserstein distance, wherein the generative adversarial model includes a generator and a discriminator; Generate a first training sample set, wherein any training sample of the first training sample set includes a real image of a sample phononic crystal and a third characteristic frequency vector representing a dispersion relation of a configuration of the sample phononic crystal; Input each third characteristic frequency vector into the generator, and output a generated image after processing; The generated image and the corresponding real image are input to the discriminator together, and after processing, a metric value indicating whether the generated image is consistent with the corresponding real image is output; Based on the metric value, the generator and the discriminator are iteratively trained using a loss function with a gradient penalty until the number of training rounds reaches a preset number of rounds, and the training is completed, and the generator generated by the training is used as the configuration generation model.

6. The method of claim 3, wherein: The discriminator comprises: A plurality of coupled convolution modules and output modules; The convolution module includes a convolution layer, an instance normalization layer, an activation function layer and a random dropout layer coupled in sequence; The output module includes a fully connected module.

7. The method of claim 1, wherein: The dispersion relation prediction model comprises: Sequentially coupled basic network components, multi-layer perception components, and prediction output components; The basic network component includes a plurality of residual modules, and each residual module includes an attention mechanism module, and the attention mechanism module includes a channel attention submodule and a spatial attention submodule; The multi-layer perception component includes coupled fully connected layers, activation function layers, and random dropout layers.

8. A computing device comprising: at least one processor; and A memory storing program instructions, wherein the program instructions are configured to be executed by the at least one processor, and the program instructions include instructions for executing the method according to any one of claims 1 to 7.

9. A computer program product comprising a computer program / instructions, wherein: When the computer program / instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

10. A readable storage medium storing program instructions, when the program instructions are read and executed by a computing device, the computing device executes the method according to any one of claims 1 to 7.

Images