Acoustic vector multiplexing structure parameter generation method and system based on physical coupling constraint

Abstract

The application discloses a kind of acoustic vector multiplexing structure parameter generation method and system based on physical coupling constraint, first, obtain the target tensor set containing the first and second vector components of sound velocity field, after formatting, input physical driving neural network, output the binary structure parameter matrix of metasurface unit;Differentiable acoustic physics layer is embedded into network, based on wave equation and linear Euler equation, the physical mapping of parameter and vector field is established, to generate the prediction field data containing coupling characteristics;Construct a multi-channel joint loss function, while constraining the component difference of prediction field and target field and the physical coupling relationship;Gradient is calculated using automatic differentiation and backpropagation, iteratively update network weights to convergence, output structure parameters that meet physical coupling constraints;The application solves the difficulty of vector field decoupling and multiplexing crosstalk caused by physical strong coupling in the prior art, and realizes efficient generation of high-fidelity vector holographic multiplexing acoustic structure parameters.

Description

Technical Field

[0001] This invention belongs to the field of computer data processing and artificial intelligence technology, and particularly relates to a method and system for generating acoustic vector multiplexing structural parameters based on physical coupling constraints. Background Technology

[0002] Sound field manipulation is a key technology in acoustic engineering and information processing, with broad application prospects in non-destructive testing, medical imaging, virtual reality, and communications. In recent years, holographic multiplexing technology has attracted significant attention due to its ability to significantly increase the capacity of information transmission or storage. In the field of optics, utilizing the polarization degree of freedom of light to achieve orthogonal state multiplexing has become a mature solution. However, sound waves, as longitudinal waves, do not inherently possess polarization characteristics. Traditional holographic sound field manipulation has long relied on a single scalar sound pressure field, limiting the information dimension and making efficient multiplexing difficult.

[0003] To overcome the limitations of scalar fields, existing technologies have begun to explore the use of the vector characteristics of acoustic particle velocity fields for multidimensional information encoding. Typical methods are usually based on linear acoustic theory (such as the Euler equations), coupling the particle velocity vector with the sound pressure gradient in a model, and employing numerical optimization algorithms (such as the adjoint method, topology optimization, or iterative inversion algorithms) to inversely design the structural parameters of metasurfaces or metamaterials, aiming to generate the desired vector field distribution in the target region. These methods attempt to achieve independent or joint control of the individual components of the acoustic vector field by solving complex inverse problems of physical constraints.

[0004] However, existing technologies have significant limitations: due to the strong coupling constraint between the sound pressure field and the particle velocity field through physical laws, traditional numerical optimization algorithms face challenges such as extremely high computational complexity, difficulty in convergence, and heavy reliance on initial guesses in such problems, making it difficult to efficiently obtain the global optimal solution. While directly applying general data-driven deep learning models can improve computational speed, the lack of strict physical equation constraints in the learning process often results in physically infeasible structural parameters, leading to severe crosstalk between the components of the final sound field. This fails to meet the multi-channel independence and functional decoupling requirements necessary for high-fidelity vector holographic multiplexing. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a method and system for generating acoustic vector multiplexing structural parameters based on physical coupling constraints. By constructing a deep learning network with embedded physical equations, this invention solves the problems of difficulty in decoupling acoustic vector fields and severe multiplexing crosstalk caused by strong physical coupling in existing technologies, and efficiently generates acoustic structural parameters that can achieve high-fidelity vector holographic multiplexing.

[0006] Technical solution: The acoustic vector multiplexing structure parameter generation method based on physical coupling constraints described in this invention includes the following steps:

[0007] S1. Obtain a vector holographic multiplexing target data tensor set, wherein the target data tensor set includes at least a first channel data representing the first vector component of the acoustic particle velocity field and a second channel data representing the second vector component of the acoustic particle velocity field, and perform data formatting processing on the target data tensor set to construct a network input tensor;

[0008] S2. Construct and initialize a physical driving neural network model, which includes a feature extraction module and a nonlinear mapping module. The physical driving neural network model is used to perform multi-channel feature extraction and nonlinear mapping on the network input tensor and output a binary structural parameter matrix that corresponds one-to-one with the acoustic metasurface structural unit.

[0009] S3. Construct a differentiable acoustic physical layer and embed the differentiable acoustic physical layer into the physical driving neural network model. The differentiable acoustic physical layer establishes a physical mapping relationship between the structural parameter matrix and the acoustic vector field based on the acoustic wave equation and the linear Euler equation. Through forward propagation operation, the binarized structural parameter matrix is ​​mapped into predicted vector field data containing vector component coupling characteristics.

[0010] S4. Construct a multi-channel joint loss function, which is used to characterize the difference between the predicted vector field data and the target data tensor set in each vector component, and introduces the physical coupling relationship embodied by the differentiable acoustic physical layer as a constraint term into the multi-channel joint loss function;

[0011] S5. Calculate the gradient of the multi-channel joint loss function with respect to the weight parameters of the physical driving neural network model using an automatic differentiation mechanism, and iteratively update the weight parameters through a backpropagation algorithm until the preset convergence condition is met, and output the acoustic vector multiplexing structure parameter matrix that satisfies the physical coupling constraint.

[0012] This invention constructs a differentiable physical layer embedding acoustic wave equations and linear Euler equations, directly integrating the inherent physical coupling relationships between particle velocity components in a vector acoustic field as hard constraints into a deep learning network. This fundamentally solves the core bottlenecks of existing technologies, such as the difficulty in decoupling vector fields and severe crosstalk between multiplexed channels due to neglecting strong physical coupling characteristics. Specifically, S1 formats the multiplexed target containing multiple vector components of the velocity field, providing a data foundation for multi-channel joint constraints; S2's feature extraction and nonlinear mapping module enables efficient reverse design from high-dimensional target fields to binary structural parameters. S3 establishes a strictly differentiable physical mapping between structural parameters and predicted vector fields through forward simulation of differentiable physical layers, enabling the network to "perceive" the real coupling fluctuation process during training; S4 constructs a multi-channel joint loss function to uniformly optimize the reproduction accuracy of each vector component and explicitly quantize the physical coupling relationship as a constraint term, forcing the network to output a cooperative solution that satisfies the physical laws; finally, S5's automatic differentiation and end-to-end iterative optimization ensure that the network autonomously converges to the optimal solution under physical constraints, thereby efficiently generating acoustic metasurface structural parameters that can accurately reproduce complex vector holographic targets and are naturally decoupled between multiplexed channels.

[0013] Preferably, step S1 includes:

[0014] The system receives two independent target images input by the user. The first image serves as the target image of the first channel, corresponding to the horizontal component of the acoustic particle velocity field, and the second image serves as the target image of the second channel, corresponding to the vertical component of the acoustic particle velocity field.

[0015] The first channel target image and the second channel target image are respectively converted into normalized floating-point data tensors to form the first channel data representing the horizontal component of the acoustic particle velocity field and the second channel data representing the vertical component of the acoustic particle velocity field, which serve as the target data tensor set.

[0016] By mapping two independent target images to the horizontal and vertical components of the acoustic particle velocity field, and constructing a multi-channel target data tensor set in the form of normalized floating-point data tensors, this invention achieves a seamless transformation from arbitrarily customized vector holographic patterns to physically quantifiable sound field distributions. This provides a clear and physically aligned supervisory benchmark for subsequent multi-channel joint optimization under physical coupling constraints, ensuring that the reverse design process can accurately correspond to the preset multiplexing function, and fundamentally guaranteeing the independent controllability and design flexibility of the vector holographic multiplexing channel.

[0017] Preferably, the construction and initialization of the physical-driven neural network model in step S2 includes:

[0018] Convolutional neural networks based on the U-Net architecture are constructed, including an encoder module, a decoder module, and a binarization output layer;

[0019] The encoder module is used to receive a dual-channel target image composed of the first and second vector components of the acoustic particle velocity field, and to extract features and downsample them step by step through multi-level convolutional layers, batch normalization layers and Leaky ReLU activation function to obtain high-dimensional semantic features.

[0020] The decoder module is used to restore spatial resolution step by step through transposed convolution operations, and to fuse shallow features of the corresponding layer of the encoder with deep features in the decoder through skip connections.

[0021] The binarized output layer is located at the end of the network. It uses the Gumbel-Sigmoid activation function to perform binarization approximation on the output features. During forward propagation, it generates structural parameters with values ​​close to 0 or 1. During backward propagation, it maintains gradient continuity and outputs a binarized acoustic transmission amplitude coefficient distribution matrix.

[0022] By constructing a physical-driven neural network based on the U-Net architecture, the high-dimensional semantic features in the dual-channel vector sound field target are efficiently extracted using multi-level convolution and downsampling of the encoder module. Then, multi-scale spatial details are fused through transposed convolution and skip connections of the decoder module, achieving high-resolution end-to-end mapping from the target sound field to structural parameters. In particular, the Gumbel-Sigmoid activation function is introduced as a binary output layer to generate strictly binary structural parameters in forward propagation to adapt to the discrete manufacturing constraints of acoustic metasurfaces, while maintaining gradient continuity in backward propagation. This effectively solves the gradient blocking problem caused by discrete binary variables, thus achieving accurate differentiable generation and end-to-end optimization of acoustic structural parameters while ensuring process compatibility.

[0023] Preferably, the construction of the differentiable acoustic physical layer in step S3 includes modulation operations, propagation operations, and coupling constraint operations, and the differentiable acoustic physical layer does not contain trainable parameters; the execution process of the differentiable acoustic physical layer is as follows:

[0024] structural parameter matrix The transmission amplitude coefficient distribution of the acoustic metasurface is compared with the complex sound pressure distribution of the pre-defined incident sound field. By multiplying point by point, the complex sound pressure distribution at the metasurface exit surface is obtained:

[0025]

[0026] in, For spatial coordinate variables, The complex sound pressure distribution at the exit surface is shown; the complex sound pressure distribution at the exit surface is analyzed using the angular spectrum method. Propagation calculations are performed by transforming the complex sound pressure distribution to the frequency domain using a Fast Fourier Transform (FFT), multiplying it by the propagation transfer function corresponding to the propagation distance z, and then obtaining the scalar sound pressure field distribution at the target plane using an Inverse Fast Fourier Transform (IFFT). Where z is the propagation distance of the sound wave in free space; for the scalar sound pressure field distribution Spatial gradient difference calculations are performed to establish the coupling relationship between sound pressure and particle velocity based on the linear Euler equation.

[0027] By constructing a pure physical mapping layer without trainable parameters, the structural parameter matrix is ​​sequentially transformed into outgoing complex sound pressure through modulation operations, spatial diffraction propagation is simulated through the angular spectrum method, and the particle velocity field is inverted through gradient difference coupling with the linear Euler equation. Thus, the physical laws of the entire process of sound wave modulation from metasurface to vector field formation are completely reproduced within the neural network. This process solidifies the inherent coupling relationship between sound pressure and velocity vector components into a deterministic forward propagation path in a completely differentiable form. This not only ensures the high fidelity of the physical simulation, but also achieves a balance between computational efficiency and gradient smoothness through numerical methods based on FFT and differential operators. This enables the network to have strict physical interpretability and end-to-end trainability without additional parameter burden.

[0028] Preferably, in establishing the coupling relationship between sound pressure and particle velocity based on the linear Euler equation, the vector components of the acoustic particle velocity field in the x and y directions are calculated according to the following formulas:

[0029]

[0030]

[0031] in, and These represent the complex vector components of the particle velocity in the x and y directions, respectively. The complex acoustic pressure field distribution at the target plane. and These represent the spatial partial derivatives of the complex acoustic pressure field with respect to the x-direction and the y-direction, respectively. For the density of the propagation medium, ω is the angular frequency of the sound wave, and j is the imaginary unit;

[0032] The structural parameter matrix of the transmission amplitude coefficient of a single acoustic metasurface is established through the above calculations. With multiple particle velocity vector components , The physical coupling mapping relationship between them.

[0033] By explicitly mapping the spatial gradient of the scalar acoustic pressure field to two orthogonal vector components of particle velocity using the linear Euler equation, a one-to-many physical coupling relationship between the parameter matrix of a single acoustic metasurface structure and the multi-channel vector output field is precisely constructed within a differentiable physical layer. This mapping mechanism reveals the inherent consistency between the acoustic pressure gradient and the velocity vector from the bottom layer of the wave equation, ensuring that the network strictly follows the law of conservation of momentum during training. This ensures that the generated structural parameters can simultaneously modulate the horizontal and vertical velocity components and satisfy the interlocking constraints between them. This fundamentally eliminates the multiplexing crosstalk caused by neglecting physical coupling when designing each channel independently in traditional methods, and realizes the coordinated control and high-fidelity reproduction of multiple vector holographic channels by a single metasurface structure.

[0034] Preferably, the multi-channel joint loss function constructed in step S4 includes a data error term, which is measured by the mean square error between the predicted vector field data and the target data tensor set, and is expressed as:

[0035]

[0036] in, This represents the data error term in the multi-channel joint loss function. This represents the total number of sampling points involved in the error calculation, where i represents the vector component index. ,in This represents the first vector component of the acoustic particle velocity field. This represents the second vector component of the acoustic particle velocity field. This represents the predicted vector field data of the i-th vector component obtained through forward propagation via a differentiable acoustic physical layer. This represents the target data tensor of the i-th vector component obtained in step S1. This represents the square of the difference between the predicted value and the target value at the corresponding sampling point.

[0037] By constructing a data error term based on multi-channel mean square error, the prediction accuracy of the horizontal and vertical components of the particle velocity field is uniformly quantified into a joint optimization objective, enabling the network to simultaneously minimize the reproduction deviation of the two vector channels during training. This loss function directly aligns the multi-channel prediction field output by the physical layer with the independent target image input by the user. The explicit constraint of energy error drives the network to simultaneously consider the amplitude fidelity and spatial distribution accuracy of each vector component. Thus, based on the hard constraint of physical coupling, the independent control accuracy of the multiplexed channels is further enhanced, ensuring that the generated structural parameters can balance and optimize the reconstruction quality of the two holographic targets, effectively avoiding the single-channel performance degradation problem caused by inter-channel competition.

[0038] Preferably, step S5 further includes a process for materializing structural parameters:

[0039] The binary structural parameter matrix obtained after training convergence is mapped to the design drawing of acoustic metasurface hologram, where the unit in the matrix is ​​used to represent the acoustic transparency or blocking characteristics of the corresponding region. The transparent region allows sound waves to pass through, while the blocking region prevents sound waves from passing through.

[0040] According to the design drawings, the rigid substrate material is processed using CNC machining equipment, including laser cutting equipment or computer numerical control (CNC) machining equipment, and the rigid substrate material includes stainless steel plate or other acoustic rigid materials, thereby preparing a solid acoustic metasurface device.

[0041] The solid acoustic metasurface device was placed in the experimental testing system, and the first vector component and the second vector component were measured and verified at the target plane to verify the independent reconstruction effect of each vector component in space and the non-interference characteristic between the components.

[0042] By mapping the converged binary structural parameter matrix of the training to a metasurface design drawing that clearly identifies the transparent and blocked regions, and using CNC machining equipment to precisely manufacture a rigid substrate, a non-destructive process transformation from virtual structural parameters to physical acoustic devices was achieved. This materialization process strictly maintains the binary spatial distribution characteristics optimized by the network, ensuring that the sound transmission / insulation state of each structural unit completely corresponds to the actual physical modulation characteristics. Finally, through independent measurement of the dual vector components by the experimental testing system, the designed device was verified in a real sound field to accurately reconstruct horizontal and vertical velocity targets without interference between channels. This completed the entire chain of technology from physically constrained digital design to high-fidelity vector holographic functional objects.

[0043] Secondly, the acoustic vector multiplexing structure parameter generation system based on physical coupling constraints described in this invention includes:

[0044] The target data acquisition and formatting module is used to acquire a vector holographic multiplexing target data tensor set, wherein the target data tensor set includes at least a first channel data representing a first vector component of the acoustic particle velocity field and a second channel data representing a second vector component of the acoustic particle velocity field, and performs data formatting processing on the target data tensor set to construct a network input tensor;

[0045] The physical-driven neural network module, connected to the target data acquisition and formatting module, includes a feature extraction submodule and a nonlinear mapping submodule, used to perform multi-channel feature extraction and nonlinear mapping on the network input tensor, and output a binary structural parameter matrix that corresponds one-to-one with the acoustic metasurface structural unit.

[0046] A differentiable acoustic physics layer module is embedded in the physical driving neural network module. It establishes a physical mapping relationship between the structural parameter matrix and the acoustic vector field based on the acoustic wave equation and the linear Euler equation. Through forward propagation operation, the binarized structural parameter matrix is ​​mapped into predicted vector field data containing vector component coupling characteristics.

[0047] A multi-channel joint loss construction module is connected to the target data acquisition and formatting module and the differentiable acoustic physics layer module, respectively, and is used to construct a multi-channel joint loss function. The multi-channel joint loss function is used to characterize the difference between the predicted vector field data and the target data tensor set in each vector component, and the physical coupling relationship embodied by the differentiable acoustic physics layer module is introduced into the multi-channel joint loss function as a constraint term.

[0048] The parameter iteration optimization module, connected to the multi-channel joint loss construction module and the physical driving neural network module, is used to calculate the gradient of the multi-channel joint loss function with respect to the weight parameters of the physical driving neural network module using an automatic differentiation mechanism, and iteratively update the weight parameters through a backpropagation algorithm until a preset convergence condition is met, and outputs an acoustic vector multiplexing structure parameter matrix that satisfies the physical coupling constraint.

[0049] Thirdly, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed by the described method for generating acoustic vector multiplexing structure parameters based on physical coupling constraints.

[0050] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for generating acoustic vector multiplexing structure parameters based on physical coupling constraints.

[0051] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: 1. By constructing a physical-driven deep learning network with embedded acoustic wave equations and linear Euler equations, this invention explicitly establishes the physical coupling mapping relationship between structural parameters and multiple vector components during model forward propagation, and uses this coupling relationship as a constraint term in the loss function. This solves the problems of difficult vector field decoupling and severe crosstalk caused by strong physical coupling in existing technologies, and achieves efficient generation of acoustic structural parameters for high-fidelity vector holographic reuse; 2. The physical-driven neural network model constructed in this invention can directly map from a dual-channel vector holographic target to a binary structural parameter matrix without iterative optimization or heuristic search, significantly shortening the time from the target field to the processing drawing. The design cycle is shortened, providing a real-time intelligent design framework for complex sound field control; 3. By embedding a differentiable acoustic physical layer and adopting an automatic differentiation mechanism, this invention achieves seamless joint training of the physical model and the neural network. While ensuring that the output structure strictly meets the laws of sound field propagation and coupling, it maintains end-to-end gradient flow, significantly improving the physical authenticity and optimization efficiency of the design results; 4. By adopting Gumbel-Sigmoid differentiable binarization technology, the network directly outputs structural parameter matrices that correspond one-to-one with the binary metasurface units such as perforated plates. These matrices can be mapped to processing drawings without post-processing and realized as physical devices through CNC machining, ensuring a high degree of consistency between simulation design, physical processing and experimental verification. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0053] Figure 2 This is a schematic diagram illustrating the principle of dual-channel vector acoustic holographic multiplexing of the present invention;

[0054] Figure 3 This is a flowchart of the reverse design process of the present invention based on physics-driven deep learning;

[0055] Figure 4 This is a schematic diagram illustrating the generation and preparation of the binary acoustic metasurface of the present invention;

[0056] Figure 5 The image shows the optimized structural parameter matrix and the corresponding dual-channel vector holographic reconstruction effect of the present invention, where (a) is the structural parameter matrix and (b) is the reconstructed horizontal and vertical component holographic image. Detailed Implementation

[0057] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0058] Example 1:

[0059] like Figure 1As shown, this embodiment provides a method for generating acoustic vector multiplexing structure parameters based on physical coupling constraints, used to generate metasurface structure parameters capable of realizing acoustic vector holographic multiplexing. The method mainly includes the following steps:

[0060] Step S1: Construct vector multiplexed target data. This invention aims to achieve dual-channel vector acoustic holography. The system first receives two independent target images input by the user; Figure 2 This paper demonstrates how incident sound waves are modulated using a metasurface into independent target images (“X” and “Y”) on two orthogonal particle velocity components. In this embodiment, the input image “X” serves as the first channel target, corresponding to the horizontal component of the acoustic particle velocity field; the input image “Y” serves as the second channel target, corresponding to the vertical component of the acoustic particle velocity field. The system converts these images into normalized floating-point data tensors, which are then used as training targets for subsequent neural networks.

[0061] Step S2: Construct a deep neural network as follows Figure 3 As shown, this embodiment constructs a convolutional neural network based on the U-net architecture as the core engine for reverse engineering. (1) Encoder: The left side of the network receives stacked dual-channel target images ("X" and "Y"), and extracts high-dimensional semantic features of the image step by step through a series of convolutional layers, batch normalization layers, and Leaky ReLU activation functions, and performs downsampling. (2) Decoder: The spatial resolution is restored through transposed convolution operation, and shallow texture features and deep semantic features are fused using skip connections. (3) Binarization output layer: as shown Figure 3 As shown in the network output section, the network's end incorporates a Gumbel-Sigmoid activation function. This function allows for simulated binary sampling during forward propagation, outputting values ​​close to 0 or 1, while maintaining gradient continuity during backward propagation. This enables the network to directly learn and generate a binarized acoustic transmission amplitude coefficient distribution matrix M (such as...). Figure 3 (As shown in "Output: Structure Parameter Matrix").

[0062] Step S3: Perform forward computation of physical constraints. The network outputs a parameter matrix. Instead of direct comparison with a target, the input is fed into a pre-defined "differentiable acoustic physics layer." This layer does not contain trainable parameters but strictly executes mathematical operations based on the laws of acoustic physics.

[0063] Step 1: Modulation calculation, calculating the complex sound pressure distribution of the metasurface emission field. :

[0064]

[0065] Step 2: Propagation Calculation

[0066] The sound pressure distribution of a sound wave propagating in free space to a target plane at a distance z is calculated using the angular spectrum method. This process performs convolution operations in the frequency domain using Fast Fourier Transform (FFT) and Inverse Fourier Transform (IFFT), which is not only fast but also differentiable throughout the entire process.

[0067] Step 3: Euler Coupling Constraint Calculation

[0068] Based on the linear Euler equations, the spatial gradient of the sound pressure field is calculated using the finite difference operator, thereby deriving the particle velocity vector components:

[0069]

[0070]

[0071] in For the density of the medium, This represents the angular frequency. This step simulates the strong coupling between sound pressure and particle velocity in the physical world. Through this computational layer, the system forces the establishment of a single acoustic transmission amplitude coefficient distribution matrix. With multiple vector field components , The physical mapping between them.

[0072] Step S4: Model training and parameter update, such as Figure 3 As shown at the bottom, the difference between the predicted vector field data and the input target data (Loss function) is calculated. The gradient is calculated using an automatic differentiation mechanism, and the weights of the U-net network are updated via backpropagation (the dashed arrow points to "update parameters"). After multiple iterations, the network converges, outputting an optimized acoustic transmission amplitude coefficient distribution matrix. The Loss function is defined as the difference between the predicted vector field data (…). ) and the target data in step S1 ( The mean square error between )

[0073] Step S5: Solidification and fabrication of structural parameters. Figure 4 This demonstrates the mapping process from discrete structural parameter matrices to holographic design drawings, and then to solid fabricated samples (such as perforated stainless steel plates). The optimized binarized structural parameter matrix ( Figure 4 The left-hand 0 / 1 matrix is ​​mapped to the holographic design blueprint ( Figure 4 (Middle section). The value "1" corresponds to the sound-transmitting area (white circular hole), and the value "0" corresponds to the sound-blocking area (gray background). This design drawing can be directly used to guide CNC machining. For example... Figure 4As shown in the "Processed Sample" on the right, solid acoustic metasurface devices can be fabricated by drilling holes in hard materials (such as stainless steel plates) using laser cutting or CNC technology.

[0074] Verification results: The structure parameter matrix generated using this method is as follows: Figure 5 As shown in (a) above, physical verification was performed. A clear dual-channel vector image was successfully reconstructed at the target plane, as shown below. Figure 5 As shown in (b): Channel The "X" pattern is clearly presented, and the channel... The "Y" pattern is clearly presented, and the two do not interfere with each other, proving the effectiveness of this method in handling vector field coupling problems.

[0075] Example 2:

[0076] This invention also provides a system for generating acoustic vector multiplexing structure parameters based on physical coupling constraints, comprising:

[0077] The target data acquisition and formatting module is used to acquire a vector holographic multiplexing target data tensor set, wherein the target data tensor set includes at least a first channel data representing a first vector component of the acoustic particle velocity field and a second channel data representing a second vector component of the acoustic particle velocity field, and performs data formatting processing on the target data tensor set to construct a network input tensor;

[0078] The physical-driven neural network module, connected to the target data acquisition and formatting module, includes a feature extraction submodule and a nonlinear mapping submodule, used to perform multi-channel feature extraction and nonlinear mapping on the network input tensor, and output a binary structural parameter matrix that corresponds one-to-one with the acoustic metasurface structural unit.

[0079] A differentiable acoustic physics layer module is embedded in the physical driving neural network module. It establishes a physical mapping relationship between the structural parameter matrix and the acoustic vector field based on the acoustic wave equation and the linear Euler equation. Through forward propagation operation, the binarized structural parameter matrix is ​​mapped into predicted vector field data containing vector component coupling characteristics.

[0080] A multi-channel joint loss construction module is connected to the target data acquisition and formatting module and the differentiable acoustic physics layer module, respectively, and is used to construct a multi-channel joint loss function. The multi-channel joint loss function is used to characterize the difference between the predicted vector field data and the target data tensor set in each vector component, and the physical coupling relationship embodied by the differentiable acoustic physics layer module is introduced into the multi-channel joint loss function as a constraint term.

[0081] The parameter iteration optimization module, connected to the multi-channel joint loss construction module and the physical driving neural network module, is used to calculate the gradient of the multi-channel joint loss function with respect to the weight parameters of the physical driving neural network module using an automatic differentiation mechanism, and iteratively update the weight parameters through a backpropagation algorithm until a preset convergence condition is met, and outputs an acoustic vector multiplexing structure parameter matrix that satisfies the physical coupling constraint.

[0082] Example 3:

[0083] The present invention also discloses an electronic device.

[0084] Specifically, the electronic device can be a desktop computer, laptop computer, handheld computer, or cloud server, etc. This computer device may include, but is not limited to, a processor and memory. The processor and memory can be connected via a bus or other means. The processor can be a Central Processing Unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, graphics processing units (GPUs), embedded neural network processing units (NPUs) or other dedicated deep learning coprocessors, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0085] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules. The processor executes various functional applications and data processing by running non-transitory software programs, instructions, and modules stored in memory. Memory may include a program storage area and a data storage area. The program storage area may store the control unit and the application program required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, memory may include high-speed random access memory and non-transitory memory. In some embodiments, memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0086] Example 4:

[0087] The present invention also discloses a computer-readable storage medium.

[0088] Specifically, the computer-readable storage medium is used to store a computer program, which, when executed by a processor, implements the methods described in the above method implementation.

[0089] Those skilled in the art will understand that all or part of the processes in the methods described above can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium can also include combinations of the above types of memory.

Claims

1. A method for generating acoustic vector multiplexing structure parameters based on physical coupling constraints, characterized in that, Includes the following steps: S1. Obtain a vector holographic multiplexing target data tensor set, wherein the target data tensor set includes at least a first channel data representing the first vector component of the acoustic particle velocity field and a second channel data representing the second vector component of the acoustic particle velocity field, and perform data formatting processing on the target data tensor set to construct a network input tensor; S2. Construct and initialize a physical driving neural network model, which includes a feature extraction module and a nonlinear mapping module. The physical driving neural network model is used to perform multi-channel feature extraction and nonlinear mapping on the network input tensor and output a binary structural parameter matrix that corresponds one-to-one with the acoustic metasurface structural unit. S3. Construct a differentiable acoustic physical layer and embed the differentiable acoustic physical layer into the physical driving neural network model. The differentiable acoustic physical layer establishes a physical mapping relationship between the structural parameter matrix and the acoustic vector field based on the acoustic wave equation and the linear Euler equation. Through forward propagation operation, the binarized structural parameter matrix is ​​mapped into predicted vector field data containing vector component coupling characteristics. S4. Construct a multi-channel joint loss function, which is used to characterize the difference between the predicted vector field data and the target data tensor set in each vector component, and introduces the physical coupling relationship embodied by the differentiable acoustic physical layer as a constraint term into the multi-channel joint loss function; S5. Calculate the gradient of the multi-channel joint loss function with respect to the weight parameters of the physical driving neural network model using an automatic differentiation mechanism, and iteratively update the weight parameters through a backpropagation algorithm until the preset convergence condition is met, and output the acoustic vector multiplexing structure parameter matrix that satisfies the physical coupling constraint.

2. The method according to claim 1, characterized in that, Step S1 includes: The system receives two independent target images input by the user. The first image serves as the target image of the first channel, corresponding to the horizontal component of the acoustic particle velocity field, and the second image serves as the target image of the second channel, corresponding to the vertical component of the acoustic particle velocity field. The first channel target image and the second channel target image are respectively converted into normalized floating-point data tensors to form the first channel data representing the horizontal component of the acoustic particle velocity field and the second channel data representing the vertical component of the acoustic particle velocity field, which serve as the target data tensor set.

3. The method according to claim 1, characterized in that, Step S2, which involves constructing and initializing the physics-driven neural network model, includes: Convolutional neural networks based on the U-Net architecture are constructed, including an encoder module, a decoder module, and a binarization output layer; The encoder module is used to receive a dual-channel target image composed of the first and second vector components of the acoustic particle velocity field, and to extract features and downsample them step by step through multi-level convolutional layers, batch normalization layers and Leaky ReLU activation function to obtain high-dimensional semantic features. The decoder module is used to restore spatial resolution step by step through transposed convolution operations, and to fuse shallow features of the corresponding layer of the encoder with deep features in the decoder through skip connections. The binarized output layer is located at the end of the network. It uses the Gumbel-Sigmoid activation function to perform binarization approximation on the output features. During forward propagation, it generates structural parameters with values ​​close to 0 or 1. During backward propagation, it maintains gradient continuity and outputs a binarized acoustic transmission amplitude coefficient distribution matrix.

4. The method according to claim 1, characterized in that, Step S3 describes constructing a differentiable acoustic physical layer, which includes modulation operations, propagation operations, and coupling constraint operations. The differentiable acoustic physical layer does not contain trainable parameters. The execution process of the differentiable acoustic physical layer is as follows: structural parameter matrix The transmission amplitude coefficient distribution of the acoustic metasurface is compared with the complex sound pressure distribution of the pre-defined incident sound field. By multiplying point by point, the complex sound pressure distribution at the metasurface exit surface is obtained: ;in, For spatial coordinate variables, The complex sound pressure distribution at the exit surface is shown; the complex sound pressure distribution at the exit surface is analyzed using the angular spectrum method. Propagation calculations are performed by transforming the complex sound pressure distribution to the frequency domain using a Fast Fourier Transform (FFT), multiplying it by the propagation transfer function corresponding to the propagation distance z, and then obtaining the scalar sound pressure field distribution at the target plane using an Inverse Fast Fourier Transform (IFFT). Where z is the propagation distance of the sound wave in free space; for the scalar sound pressure field distribution Spatial gradient difference calculations are performed to establish the coupling relationship between sound pressure and particle velocity based on the linear Euler equation.

5. The method according to claim 4, characterized in that, In establishing the coupling relationship between sound pressure and particle velocity based on the linear Euler equation, the vector components of the acoustic particle velocity field in the x and y directions are calculated according to the following formulas: ; ;in, and These represent the complex vector components of the particle velocity in the x and y directions, respectively. The complex acoustic pressure field distribution at the target plane. and These represent the spatial partial derivatives of the complex acoustic pressure field with respect to the x-direction and the y-direction, respectively. For the density of the propagation medium, ω is the angular frequency of the sound wave, and j is the imaginary unit; The structural parameter matrix of the transmission amplitude coefficient of a single acoustic metasurface is established through the above calculations. With multiple particle velocity vector components , The physical coupling mapping relationship between them.

6. The method according to claim 1, characterized in that, The multi-channel joint loss function constructed in step S4 includes a data error term, which is measured by the mean square error between the predicted vector field data and the target data tensor set. The expression for the data error term is: ;in, This represents the data error term in the multi-channel joint loss function. This represents the total number of sampling points involved in the error calculation, where i represents the vector component index. ,in This represents the first vector component of the acoustic particle velocity field. This represents the second vector component of the acoustic particle velocity field. This represents the predicted vector field data of the i-th vector component obtained through forward propagation via a differentiable acoustic physical layer. This represents the target data tensor of the i-th vector component obtained in step S1. This represents the square of the difference between the predicted value and the target value at the corresponding sampling point.

7. The method according to claim 1, characterized in that, Step S5 also includes the materialization process of structural parameters: The binary structural parameter matrix obtained after training convergence is mapped to the design drawing of acoustic metasurface hologram, where the unit in the matrix is ​​used to represent the acoustic transparency or blocking characteristics of the corresponding region. The transparent region allows sound waves to pass through, while the blocking region prevents sound waves from passing through. According to the design drawings, the rigid substrate material is processed using CNC machining equipment, including laser cutting equipment or computer numerical control (CNC) machining equipment, and the rigid substrate material includes stainless steel plate or other acoustic rigid materials, thereby preparing a solid acoustic metasurface device. The solid acoustic metasurface device was placed in the experimental testing system, and the first vector component and the second vector component were measured and verified at the target plane to verify the independent reconstruction effect of each vector component in space and the non-interference characteristic between the components.

8. A system for generating structural parameters for acoustic vector multiplexing based on physical coupling constraints, characterized in that, include: The target data acquisition and formatting module is used to acquire a vector holographic multiplexing target data tensor set, wherein the target data tensor set includes at least a first channel data representing a first vector component of the acoustic particle velocity field and a second channel data representing a second vector component of the acoustic particle velocity field, and performs data formatting processing on the target data tensor set to construct a network input tensor; The physical-driven neural network module is connected to the target data acquisition and formatting module. It includes a feature extraction submodule and a nonlinear mapping submodule, which are used to perform multi-channel feature extraction and nonlinear mapping on the network input tensor, and output a binary structure parameter matrix that corresponds one-to-one with the acoustic metasurface structure unit. A differentiable acoustic physics layer module is embedded in the physical driving neural network module. It establishes a physical mapping relationship between the structural parameter matrix and the acoustic vector field based on the acoustic wave equation and the linear Euler equation. Through forward propagation operation, the binarized structural parameter matrix is ​​mapped into predicted vector field data containing vector component coupling characteristics. A multi-channel joint loss construction module is connected to the target data acquisition and formatting module and the differentiable acoustic physics layer module, respectively, and is used to construct a multi-channel joint loss function. The multi-channel joint loss function is used to characterize the difference between the predicted vector field data and the target data tensor set in each vector component, and the physical coupling relationship embodied by the differentiable acoustic physics layer module is introduced into the multi-channel joint loss function as a constraint term. The parameter iteration optimization module, connected to the multi-channel joint loss construction module and the physical driving neural network module, is used to calculate the gradient of the multi-channel joint loss function with respect to the weight parameters of the physical driving neural network module using an automatic differentiation mechanism, and iteratively update the weight parameters through a backpropagation algorithm until a preset convergence condition is met, and outputs an acoustic vector multiplexing structure parameter matrix that satisfies the physical coupling constraint.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the acoustic vector multiplexing structure parameter generation method based on physical coupling constraints as described in any one of claims 1 to 7.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method for generating acoustic vector multiplexing structure parameters based on physical coupling constraints according to any one of claims 1 to 7.

Images