A composite acoustic metamaterial sound insulation performance prediction method based on physical prior guidance

Abstract

The application relates to the technical field of acoustic metamaterial design, and discloses a composite acoustic metamaterial sound insulation performance prediction method based on physical prior guidance, which comprises the following steps: constructing a paired data set of structure parameters and sound transmission loss frequency spectrum; extracting a physical prior feature vector of a single resonance mechanism; constructing a prediction network, which comprises a multi-branch fusion neural network and a backend regression network; the multi-branch fusion neural network comprises a structure parameter branch and a physical prior branch, is used for fusing input structure parameter features and physical prior features, and generates a physical enhanced feature vector; the backend regression network realizes the prediction of a full-frequency sound transmission loss frequency spectrum according to the physical enhanced feature vector; and the prediction network is trained end to end to predict the sound insulation performance of the composite acoustic metamaterial. The application extracts physical prior guidance to provide physical navigation information for the neural network, the prediction result is more physically based, and the unexplainability of a pure data-driven model black box operation is avoided.

Description

Technical Field

[0001] This invention relates to the field of acoustic metamaterial design and artificial intelligence, and in particular to a method for predicting the sound insulation performance of composite acoustic metamaterials based on physical prior guidance. Background Technology

[0002] Composite acoustic metamaterials, by coupling multiple acoustic resonance mechanisms, can achieve excellent low-frequency sound insulation performance at the subwavelength scale, and have broad application prospects in the field of noise control. Typical composite acoustic metamaterial structures can include various configurations such as labyrinth channels, Helmholtz resonant cavities, and thin-film resonant units, achieving efficient sound insulation over a wide frequency range or in specific frequency bands through the synergistic effect of different mechanisms.

[0003] Traditionally, the acoustic performance evaluation of such complex structures has relied primarily on numerical simulation methods such as the finite element method. However, high-precision simulations are computationally expensive and struggle to meet the demands of rapid iterative optimization. In recent years, data-driven machine learning models (such as multilayer perceptrons and convolutional neural networks) have been attempted to establish a mapping between structural parameters and sound insulation performance. However, existing black-box models, when dealing with such composite structures, often suffer from amplitude passivation and position shifts in the prediction of key narrowband resonant peaks in the spectrum due to their inability to effectively distinguish and guide the model to learn the contributions of different physical mechanisms. This results in limited prediction accuracy and makes it difficult to meet practical engineering needs. Chinese invention application CN116110522A discloses a method, device, medium, and terminal for predicting the sound absorption performance of Helmholtz resonant materials. Based on a database of sound absorption performance between structural parameters and absorption coefficients of Helmholtz resonant materials in different dimensions, a machine learning-based network for predicting the sound absorption performance of Helmholtz resonant materials is trained. The network is then tested and verified using a validation set, predicting the sound absorption performance based on the structural parameters of the Helmholtz resonant material under test in different dimensions. This method solves the problem that it is difficult to predict the sound absorption performance through simulation technology due to the large number of parameters of acoustic Helmholtz resonant materials, and solves the problems of long simulation prediction time and low efficiency. However, it still has the shortcomings of not being able to effectively decouple multiple physical mechanisms in composite acoustic structures and not being able to accurately capture the coupling characteristics between multiple resonant mechanisms.

[0004] Therefore, there is an urgent need for a high-fidelity performance prediction method that can deeply integrate physical mechanism knowledge and guide machine learning models to accurately capture complex resonance features. Summary of the Invention

[0005] The purpose of this invention is to solve the technical problems of low accuracy and peak distortion in predicting resonance features caused by the coupling of multiple physical mechanisms in composite ultrasonic materials using pure data-driven models in the prior art.

[0006] The technical solution adopted by this invention to solve its technical problem is: to provide a method for predicting the sound insulation performance of composite acoustic metamaterials based on physical prior guidance, comprising the following steps:

[0007] Construct a paired dataset of structural parameters and transmission loss spectra of composite acoustic metamaterials;

[0008] Extract the physical prior feature vector of a single resonance mechanism;

[0009] A prediction network is constructed, comprising a multi-branch fusion neural network and a back-end regression network. The multi-branch fusion neural network includes a structural parameter branch and a physical prior branch, which are used to fuse the input structural parameter features and physical prior features to generate a physical enhancement feature vector. The back-end regression network is used to predict the full-band transmission loss spectrum based on the physical enhancement feature vector.

[0010] The weighted mean square error is used as the loss function to train the prediction network end-to-end.

[0011] The sound insulation performance of composite acoustic metamaterials is predicted using a trained prediction network.

[0012] Preferably, the composite acoustic material is a composite acoustic material coupled with Helmholtz resonators; the structural parameters include the thickness of the structural plate, the width of the labyrinth channel, the radius of the central channel, the curvature of the labyrinth channel, and the neck diameter of the four Helmholtz resonators HR.

[0013] The transmission loss spectrum was simulated in the frequency domain within the 200Hz to 2000Hz band with a step size of 10Hz.

[0014] Preferably, the extraction of the physical prior feature vector of a single resonance mechanism includes the following steps:

[0015] Construct a substructure dataset containing only a single resonance mechanism;

[0016] Savitzky-Golay filtering was applied to the spectrum of this subset of data for denoising.

[0017] Principal component analysis was used to extract the top 20 principal components that explained ≥99.9% of the cumulative variance, forming a 20-dimensional low-dimensional physical prior feature vector.

[0018] Preferably, the multi-branch fusion neural network includes:

[0019] The structural parameter branch receives the structural parameters of the composite acoustic metamaterial and outputs the geometric features of the linear mapping.

[0020] The physics prior branch takes a low-dimensional physics prior feature vector as input and outputs a linearly mapped physics prior feature.

[0021] The fusion layer concatenates the geometric features and physical prior features of the linear mapping to form a physically enhanced feature vector.

[0022] Preferably, the number of structural parameter branches is set according to the number of structural parameter types; each structural parameter branch inputs only one type of structural parameter.

[0023] Preferably, the multi-branch fusion neural network includes:

[0024] The maze structure parameter branch is used to receive maze structure parameters, including structural plate thickness, maze channel width, central channel radius, and maze channel curvature.

[0025] The HR structure parameter branch is used to receive HR structure parameters, including the neck diameters of the four HR parameters;

[0026] The Fano physics reconstruction branch is used to receive low-dimensional physical prior feature vectors.

[0027] Preferably, the back-end regression network adopts any one or a combination of the following:

[0028] One-dimensional convolutional neural network (1D-CNN), multilayer perceptron (MLP), bidirectional long short-term memory network (Bi-LSTM) or Transformer.

[0029] Preferably, the loss function is calculated using the following formula:

[0030] ;

[0031] in, and The first The actual and predicted transmission loss at each frequency point Total frequency points For the first The weighting coefficients for each frequency point.

[0032] Preferably, the prediction method is applicable to composite acoustic materials composed of multiple coupled acoustic resonant units.

[0033] This invention also provides a physical prior-guided system for predicting the sound insulation performance of composite acoustic metamaterials, comprising:

[0034] The dataset construction module builds a paired dataset of structural parameters and transmission loss spectra of composite acoustic metamaterials.

[0035] The prior feature extraction module extracts the physical prior feature vector of a single resonance mechanism;

[0036] A prediction network construction module is used to construct a prediction network, including a multi-branch fusion neural network and a back-end regression network. The multi-branch fusion neural network includes a structural parameter branch and a physical prior branch, which are used to fuse the input structural parameter features and physical prior features to generate a physical enhancement feature vector. The back-end regression network is used to predict the full-band transmission loss spectrum based on the physical enhancement feature vector.

[0037] The prediction network training module uses weighted mean square error as the loss function to train the prediction network end-to-end.

[0038] The performance prediction module uses a trained prediction network to predict the sound insulation performance of composite acoustic metamaterials.

[0039] The present invention has the following beneficial effects:

[0040] (1) Improved prediction accuracy through physical prior guidance: This invention provides key physical navigation information for neural networks by constructing and extracting low-dimensional physical prior vectors that represent the essential characteristics of specific resonance mechanisms. Compared with black-box models that directly use the original structural parameters, this invention can effectively decouple multiple resonance mechanisms in composite structures, achieve high-fidelity restoration of the prediction of various resonance peaks, and eliminate peak position shift and amplitude passivation;

[0041] (2) Explicit physical branches enhance interpretability: This invention enables the model to distinguish the contributions of different acoustic mechanisms through explicit physical prior branches, making the prediction results more physically based and avoiding the uninterpretability of the black box operation of pure data-driven models.

[0042] (3) Cross-configuration design ensures the universality of the method: The method of this invention is not limited to specific composite ultrasonic material configurations. Any acoustic metamaterials coupled by multiple resonance mechanisms can be constructed by extracting the physical prior features of one or more of these mechanisms to build a similar physical guidance prediction model.

[0043] (4) High-efficiency forward prediction engine accelerates design iteration: This invention provides a fast and accurate forward performance prediction engine for composite acoustic metamaterials. The trained model can complete a single performance prediction in milliseconds, which is thousands of times more efficient than traditional finite element simulation. It can significantly replace time-consuming numerical calculations and significantly accelerate the design iteration process of composite acoustic metamaterials.

[0044] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the method steps of a physical prior-guided method for predicting the sound insulation performance of composite acoustic metamaterials according to an embodiment of the present invention.

[0046] Figure 2 This is a schematic diagram of a multi-branch fusion physical prior guided neural network structure according to an embodiment of the present invention;

[0047] Figure 3 The image shows a comparison of the actual and predicted sound insulation spectra of four network models under the direct regression strategy in this embodiment of the invention, where (a) is a one-dimensional convolutional neural network 1D-CNN; (b) is a multilayer perceptron MLP; (c) is a Transformer; and (d) is a bidirectional long short-term memory network Bi-LSTM.

[0048] Figure 4 The image shows a comparison of the actual and predicted sound insulation spectra of four network models under the physical prior guidance strategy in this embodiment of the invention, where (a) is a one-dimensional convolutional neural network 1D-CNN; (b) is a multilayer perceptron MLP; (c) is a Transformer; and (d) is a bidirectional long short-term memory network Bi-LSTM.

[0049] Figure 5 This is a system structure diagram of a composite acoustic metamaterial sound insulation performance prediction system based on physical prior guidance, according to an embodiment of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0051] This embodiment provides a method for predicting the properties of composite acoustic metamaterials based on physical prior knowledge, such as... Figure 1 As shown, the specific steps include:

[0052] S101, Construct a paired dataset of structural parameters and transmission loss spectrum of composite acoustic metamaterials;

[0053] S102, extract the physical prior feature vector of a single resonance mechanism;

[0054] S103, Construct a prediction network, including a multi-branch fusion neural network and a back-end regression network;

[0055] S104, using weighted mean square error as the loss function to perform end-to-end training of the prediction network;

[0056] S105, using a trained prediction network to predict the sound insulation performance of composite acoustic metamaterials;

[0057] The multi-branch fusion neural network includes a structural parameter branch and a physical prior branch, which are used to fuse the input structural parameter features and physical prior features to generate a physical enhancement feature vector; the back-end regression network is used to predict the full-band transmission loss spectrum based on the physical enhancement feature vector.

[0058] Specifically, in S101, the composite acoustic metamaterial composed of multiple acoustic resonant units is parametrically characterized to determine multiple structural parameters and their value ranges; through an automated finite element simulation process, frequency domain scanning is performed in the target frequency band to calculate the transmission loss spectrum corresponding to each set of structural parameters, construct a paired dataset of structural parameters and transmission loss spectrum, and export and store the simulation results.

[0059] The composite acoustic metamaterial unit studied in this embodiment consists of a central labyrinthine channel coupled with an outer Helmholtz resonant cavity array. To ensure ventilation and achieve mid-to-low frequency sound insulation enhancement, eight key structural parameters were selected as design variables: structural plate thickness w, labyrinthine channel width t, central channel radius r, labyrinthine channel curvature N (discrete values, taking values ​​of 3, 4, 5, and 6), and the neck diameters d1 to d4 of the four HRs. The value ranges of each parameter were set according to engineering feasibility and geometric constraints as follows: mm, mm, mm, mm mm mm mm.

[0060] An automated simulation process was built using an automated programming platform and finite element simulation software. A multidimensional independent uniform random sampling method was employed within the parameter space to generate 20,000 parameter combinations that met geometric feasibility constraints. The scripting interface provided by the finite element simulation software was used to automatically assign parameters and perform batch scheduling. Frequency domain scanning was performed in the 200–2000 Hz frequency range with a step size of 10 Hz to calculate the transmission loss spectrum corresponding to each parameter combination. Finally, a paired dataset of 8-dimensional structural parameters and 181-dimensional transmission loss spectra containing 20,000 samples was constructed and divided into training, validation, and test sets in a 7:2:1 ratio.

[0061] Specifically, in S102, a substructure dataset containing only the target single resonance mechanism and without other coupling mechanisms is constructed; the transmission loss spectrum of the sub-dataset is smoothed and denoised; principal component analysis is used to reduce the dimensionality of the processed high-dimensional spectrum, and the top principal components whose cumulative explained variance exceeds a preset threshold are extracted as low-dimensional physical prior feature vectors.

[0062] To avoid interference from Helmholtz resonance in Fano resonance feature extraction, a pure Fano substructure dataset containing only the labyrinthine channel structure was first constructed. Within the same range of structural parameter variations as in step S1, 5000 pure Fano structure samples were generated, and their transmission loss spectra were simulated. Savitzky-Golay smoothing filtering was applied to these spectra, with a filter window length of 51 and a polynomial fitting order of 3. This processing effectively eliminated high-frequency numerical noise caused by finite element discretization errors while preserving the steep asymmetric contours of the Fano resonance peaks. Principal Component Analysis (PCA) was then performed on all smoothed pure Fano spectrum samples. Statistics showed that the cumulative explained variance of the first 20 principal components reached over 99.9%. Therefore, the scores of the first 20 principal components were taken as a 20-dimensional low-dimensional physical prior feature vector representing the essential characteristics of the Fano resonance mechanism. This vector achieved a dimensionality compression of approximately 9:1, and the original spectrum could be reconstructed with high fidelity through inverse principal component analysis, verifying its complete physical representation capability.

[0063] Specifically, in S103, a multi-branch fusion-based physical prior-guided neural network is constructed. This network includes multiple parallel feature extraction branches, at least: a structural parameter branch, which takes into account structural parameters related to various resonance units in the composite ultrasonic material; and a physical prior branch, which takes into account low-dimensional physical prior feature vectors. The output features of each branch are concatenated through a fusion layer to form a physically enhanced feature vector.

[0064] like Figure 2 As shown, the network training strategy of the present invention includes two parallel modeling paths: a physical prior strategy and a direct regression strategy (as a comparison baseline).

[0065] In the direct regression strategy, the model takes the original 8-dimensional design parameters (w,t,r,N,d1,d2,d3,d4) as input, without introducing any physical prior information, and directly maps them to the transmission loss prediction results.

[0066] In the physics prior strategy, the model adopts a multi-branch input architecture: a maze structure parameter branch (4-dimensional, including w, t, r, N), an HR structure parameter branch (4-dimensional, including d1, d2, d3, d4), and a Fano physics reconstruction branch (20-dimensional, a low-dimensional physics prior feature vector obtained by PCA dimensionality reduction). The three branches are linearly transformed and then concatenated using a feature fusion layer to form a 1536-dimensional physics-enhanced feature vector.

[0067] Specifically, in S103, the 1536-dimensional fused features are input into the back-end regression network. In this embodiment, the back-end network adopts a Transformer architecture, which utilizes a self-attention mechanism to model global frequency dependencies. The Transformer consists of 6 encoder layers, each containing an 8-head self-attention mechanism and a feedforward network, ultimately outputting a 181-dimensional vector corresponding to the predicted transmission loss values ​​for each frequency point within the 200-2000Hz frequency band. It should be noted that the back-end regression network of this invention is not limited to the Transformer architecture; in other embodiments, general regression network architectures such as one-dimensional convolutional neural networks (1D-CNN), multilayer perceptrons (MLP), and bidirectional long short-term memory networks (Bi-LSTM) can also be used.

[0068] Specifically, in S104, model training and optimization are performed. Weighted Mean Squared Error (WMSE) is used as the loss function to conduct end-to-end training of the entire network, and the training process is monitored using a validation set. The WMSE applies a higher weight coefficient to the core resonant frequency domain within the target frequency band than to the non-resonant frequency domain, guiding the model to prioritize learning key acoustic resonance features. The loss function calculation formula is as follows:

[0069] ;

[0070] in, and The first The actual and predicted transmission loss at each frequency point Total frequency points For the first The weighted mean square error applies a higher weighting coefficient to the core resonant frequency domain within the target frequency band than to the non-resonant frequency domain, guiding the model to prioritize learning key acoustic resonance features.

[0071] In this embodiment, the value is set to 5.0 in the core resonant frequency band of 600~1600Hz and to 1.0 in other frequency bands to guide the model to prioritize and accurately fit the spectral details of the resonant region.

[0072] The training hyperparameters are set as follows: the optimizer is AdamW, and the initial learning rate is 2e. -3 The training employs a cosine annealing learning rate scheduling strategy with Warmup. The batch size is 256, the total training epochs are 500, and an early stopping mechanism is implemented (training stops if the validation loss does not decrease after 50 epochs). The training process is monitored using a validation set, and the model parameters with the minimum validation loss are saved.

[0073] Specifically, to verify the effectiveness of the present invention, a comparative embodiment (direct regression strategy) was set up, which used the same training data and hyperparameters as the present invention, but only input the 8-dimensional original structural parameters and did not introduce the Fano physics prior branch. Four backend regression network architectures, namely 1D-CNN, MLP, Bi-LSTM and Transformer, were used for modeling and training.

[0074] The predictive performance of the trained model is evaluated using an independent test set. Evaluation metrics such as mean absolute error (MAE), mean squared error (MSE), and coefficient of determination (R²) are calculated, and the evaluation results are output.

[0075] Figure 3 This paper presents the comparison results of the actual and predicted transmission loss spectra of four network models under the direct regression strategy on typical test samples. Figure 3 It is evident that while the direct regression strategy can capture the overall trend of transmission loss evolution with frequency, it generally has shortcomings in handling Fano and Helmholtz resonants, manifested in the shift of peak position and significant passivation of peak amplitude, overly smooth prediction curves, and loss of the intense phase interference and local resonance characteristics in the physical simulation results.

[0076] Figure 4 This paper demonstrates the prediction results of four network models using the same test sample under the physical prior guidance strategy proposed in this invention. Figure 4 As can be seen, after introducing the Fano physical prior features, the prediction curves of all networks are highly consistent with the actual values ​​of the finite element simulation, which perfectly overcomes the amplitude passivation and peak broadening problems at the resonance peak of the direct regression strategy. It can not only accurately restore the steep phase interference characteristics unique to Fano resonance, but also accurately capture the local resonance sound insulation peak caused by the Helmholtz cavity.

[0077] To further quantitatively evaluate the performance improvement, the MAE, MSE, and R² of each model on the test set were calculated. The formulas for each metric are as follows:

[0078] Mean absolute error:

[0079] ;

[0080] Mean square error:

[0081] ;

[0082] Coefficient of determination:

[0083] ;

[0084] in, This represents the average of the actual transmission loss values.

[0085] Table 1. Comparison of model performance before and after introducing physical priors:

[0086]

[0087] The statistical results in Table 1 show that the physical prior guidance strategy significantly outperforms the direct regression strategy across all evaluation dimensions. Taking the Transformer architecture as an example, R² increased from 0.811 to 0.990, MAE decreased from 1.765 dB to 0.279 dB (a reduction of 84.2%), and MSE decreased from 3.987 dB² to 0.142 dB² (a reduction of 96.4%). These results fully demonstrate the effectiveness and superiority of the physical prior guidance method proposed in this invention.

[0088] See Figure 5 The diagram shown is a structural schematic of a composite acoustic metamaterial sound insulation performance prediction system based on physical prior guidance, including:

[0089] Dataset construction module 501 constructs a paired dataset of structural parameters and transmission loss spectra of composite acoustic metamaterials;

[0090] Prior feature extraction module 502 extracts the physical prior feature vector of a single resonance mechanism;

[0091] The prediction network construction module 503 constructs a prediction network, including a multi-branch fusion neural network and a back-end regression network. The multi-branch fusion neural network includes a structural parameter branch and a physical prior branch, which are used to fuse the input structural parameter features and physical prior features to generate a physical enhancement feature vector. The back-end regression network is used to predict the full-band transmission loss spectrum based on the physical enhancement feature vector.

[0092] The prediction network training module 504 uses weighted mean square error as the loss function to perform end-to-end training of the prediction network.

[0093] The performance prediction module 505 uses a trained prediction network to predict the sound insulation performance of composite acoustic metamaterials.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the sound insulation performance of composite acoustic metamaterials based on physical prior guidance, characterized in that, Includes the following steps: Construct a paired dataset of structural parameters and transmission loss spectra of composite acoustic metamaterials; Extract the physical prior feature vector of a single resonance mechanism; A prediction network is constructed, comprising a multi-branch fusion neural network and a back-end regression network. The multi-branch fusion neural network includes a structural parameter branch and a physical prior branch, which are used to fuse the input structural parameter features and physical prior features to generate a physical enhancement feature vector. The back-end regression network is used to predict the full-band transmission loss spectrum based on the physical enhancement feature vector. The weighted mean square error is used as the loss function to train the prediction network end-to-end. The sound insulation performance of composite acoustic metamaterials is predicted using a trained prediction network.

2. The method for predicting the sound insulation performance of composite acoustic metamaterials based on physical prior guidance according to claim 1, characterized in that, The composite acoustic material is a composite acoustic material coupled with Helmholtz resonators; the structural parameters include the thickness of the structural plate, the width of the labyrinth channel, the radius of the central channel, the curvature of the labyrinth channel, and the neck diameter of the four Helmholtz resonators HR. The transmission loss spectrum was simulated in the frequency domain within the 200Hz to 2000Hz band with a step size of 10Hz.

3. The method for predicting the sound insulation performance of composite acoustic metamaterials based on physical prior guidance according to claim 2, characterized in that, The extraction of the physical prior feature vector of a single resonance mechanism includes the following steps: Construct a substructure dataset containing only a single resonance mechanism; Savitzky-Golay filtering was applied to the spectrum of this subset of data for denoising. Principal component analysis was used to extract the top 20 principal components that explained ≥99.9% of the cumulative variance, forming a 20-dimensional low-dimensional physical prior feature vector.

4. The method for predicting the sound insulation performance of composite acoustic metamaterials based on physical prior guidance according to claim 3, characterized in that, The multi-branch fusion neural network includes: The structural parameter branch receives the structural parameters of the composite acoustic metamaterial and outputs the geometric features of the linear mapping. The physics prior branch takes a low-dimensional physics prior feature vector as input and outputs a linearly mapped physics prior feature. The fusion layer concatenates the geometric features and physical prior features of the linear mapping to form a physically enhanced feature vector.

5. The method for predicting the sound insulation performance of composite acoustic metamaterials based on physical prior guidance according to claim 4, characterized in that, The number of structural parameter branches is set according to the number of structural parameter types; each structural parameter branch can only input one type of structural parameter.

6. The method for predicting the sound insulation performance of composite acoustic metamaterials based on physical prior guidance according to claim 4, characterized in that, The multi-branch fusion neural network includes: The maze structure parameter branch is used to receive maze structure parameters, including structural plate thickness, maze channel width, central channel radius, and maze channel curvature. The HR structure parameter branch is used to receive HR structure parameters, including the neck diameters of the four HRs; The Fano physics reconstruction branch is used to receive low-dimensional physical prior feature vectors.

7. The method for predicting the sound insulation performance of composite acoustic metamaterials based on physical prior guidance according to claim 1, characterized in that, The back-end regression network adopts any one or a combination of the following: One-dimensional convolutional neural network (1D-CNN), multilayer perceptron (MLP), bidirectional long short-term memory network (Bi-LSTM) or Transformer.

8. The method for predicting the sound insulation performance of composite acoustic metamaterials based on physical prior guidance according to claim 1, characterized in that, The loss function is calculated using the following formula: ； in, and The first The actual and predicted transmission loss at each frequency point Total frequency points For the first The weighting coefficients for each frequency point.

9. The method for predicting the sound insulation performance of composite acoustic metamaterials based on physical prior guidance according to claim 1, characterized in that, The prediction method is applicable to composite acoustic materials composed of multiple coupled acoustic resonant units.

10. A system for predicting the sound insulation performance of composite acoustic metamaterials based on physical prior guidance, characterized in that, include: The dataset construction module builds a paired dataset of structural parameters and transmission loss spectra of composite acoustic metamaterials. The prior feature extraction module extracts the physical prior feature vector of a single resonance mechanism; A prediction network construction module is used to construct a prediction network, including a multi-branch fusion neural network and a back-end regression network. The multi-branch fusion neural network includes a structural parameter branch and a physical prior branch, which are used to fuse the input structural parameter features and physical prior features to generate a physical enhancement feature vector. The back-end regression network is used to predict the full-band transmission loss spectrum based on the physical enhancement feature vector. The prediction network training module uses weighted mean square error as the loss function to train the prediction network end-to-end. The performance prediction module uses a trained prediction network to predict the sound insulation performance of composite acoustic metamaterials.

Images