Thin film type acoustic metamaterial reverse design method based on deep learning

Abstract

The invention discloses a film type acoustic metamaterial reverse design method based on deep learning, and the method comprises the following steps: generating a sound transmission loss STL curve data set according to a predefined film type acoustic metamaterial structure parameter range; preprocessing and dividing the sound transmission loss STL curve data set to obtain processed data; based on the processed data, constructing and training a convolutional neural network (CNN) model, and obtaining a trained CNN model through hyper-parameter optimization; and predicting the new STL curve image according to the trained CNN model, and outputting structural parameters of the thin film type acoustic metamaterial. According to the method, the structural parameters of the thin-film acoustic metamaterial meeting the target STL curve can be quickly predicted and reversely designed, and the design efficiency is remarkably improved; according to the method, the STL curve features are automatically extracted through the deep learning model, the dependence on traditional simulation and artificial experience is reduced, and the design threshold is lowered.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of acoustic metamaterial design and deep learning technology, and particularly relates to a thin film type acoustic metamaterial reverse design method based on deep learning. BACKGROUND

[0002] Thin film acoustic metamaterials have become a research hotspot in the field of acoustic control due to their low-frequency sound insulation and high-efficiency sound absorption, as well as their flexible structural design. However, the traditional design method of thin film acoustic metamaterials has the following limitations:

[0003] Firstly, the design process is highly dependent on empirical formulas and numerical simulations (such as COMSOL and MATLAB), which have high computational costs, long design cycles, and are difficult to meet the rapid iteration and optimization requirements. Secondly, the performance of thin film acoustic metamaterials and the structural parameters (such as radius, thickness, and density) have a complex nonlinear relationship, which is difficult for traditional methods to accurately capture, resulting in deviations between the design results and actual requirements.

[0004] Deep learning technology provides a new approach to the design of thin film acoustic metamaterials. By converting the complex relationship between STL curves and structural parameters into an end-to-end prediction problem, the design efficiency and accuracy can be significantly improved. However, existing deep learning models still have the following problems in predicting thin film acoustic metamaterial parameters: first, the model structure is simple and difficult to fully extract the acoustic features of STL curves; second, there is a lack of enhancement mechanism for key frequency characteristics, resulting in insufficient prediction accuracy; third, the model training process is unstable and easily affected by noise and outliers, making it difficult to ensure the reliability of the prediction results. SUMMARY

[0005] To address the problems of long design cycle, low parameter optimization efficiency, and difficulty in balancing multiple performance objectives in existing thin film acoustic metamaterial design methods, the present application provides a thin film type acoustic metamaterial reverse design method based on deep learning, including the following steps:

[0006] Generating an acoustic transmission loss (STL) curve dataset according to a predefined range of structural parameters of thin film acoustic metamaterials;

[0007] Preprocessing and dividing the acoustic transmission loss (STL) curve dataset to obtain processed data;

[0008] Based on the processed data, a convolutional neural network (CNN) model is constructed and trained, and a trained CNN model is obtained through hyperparameter optimization;

[0009] According to the trained CNN model, a new STL curve image is predicted, and the structural parameters of the thin film acoustic metamaterial are output.

[0010] Preferably, the method for generating the STL curve dataset comprises:

[0011] A physical model of the thin-film acoustic metamaterial is established using COMSOL software;

[0012] The physical model is called by MATLAB programming to define the range of structure parameters;

[0013] An STL curve image is generated using the mphplot command to form the STL curve dataset.

[0014] Preferably, the constructed CNN model comprises:

[0015] EfficientNet is used as a backbone network for feature extraction;

[0016] A feature pyramid network (FPN) is used for multi-scale feature fusion;

[0017] A dual attention mechanism is used to optimize the weights of channel and spatial features.

[0018] Preferably, the EfficientNet backbone network comprises a plurality of MBConv blocks, each MBConv block comprising, in sequence: an expansion point convolution layer for expanding feature channels; a depth separable convolution layer for extracting spatial features; an SE attention module for dynamically adjusting channel weights; a projection layer for compressing feature dimensions; wherein each MBConv block is connected by a residual connection.

[0019] Preferably, the training method of the CNN model comprises: using a custom loss function, the custom loss function being a weighted combination of mean square error (MSE) loss, smooth L1 loss, and relative error loss; dynamically adjusting the weights of each loss in the custom loss function according to the training process.

[0020] Preferably, the final loss function comprises:

[0021]

[0022] wherein α, β, and δ represent weight coefficients; L MSE represents the mean square error loss, L smoothL1 represents the L1 smooth loss; and L rel represents the relative error loss.

[0023] Preferably, the method for training the CNN model further comprises: dynamically adjusting the learning rate using the OneCycleLR learning rate scheduling strategy; optimizing the model parameters using the exponential moving average (EMA) mechanism; setting a gradient clipping threshold to prevent gradient explosion; and setting an early stopping mechanism to stop training when the loss on the validation set no longer decreases.

[0024] Preferably, the structural parameters include: mass block radius, film thickness, mass block density and film density.

[0025] Compared with the prior art, the application has the following advantages:

[0026] The application can quickly predict and inversely design film acoustic metamaterial structure parameters meeting target STL curves, significantly improving design efficiency; the application automatically extracts STL curve features through a deep learning model, reduces dependence on traditional simulation and artificial experience, reduces design threshold, and can be extended to acoustic metamaterial design in different types and application scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the present application, the following briefly introduces the drawings needed to be used in the embodiments, obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0028] Figure 1 The method flowchart of the embodiment of the present application is shown in the figure;

[0029] Figure 2 The acoustic physical model structure diagram of the embodiment of the present application is shown in the figure;

[0030] Figure 3 The CNN convolution structure diagram of the embodiment of the present application is shown in the figure;

[0031] Figure 4 The fully connected neural network structure diagram of the embodiment of the present application is shown in the figure;

[0032] Figure 5 The loss curve diagram used in the embodiment of the present application is shown in the figure;

[0033] Figure 6 The comparison diagram of acoustic prediction parameter STL curve and real parameter STL curve of the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0034] The technical solutions in the embodiments of the present application will be described clearly and completely below with the drawings in the embodiments of the present application, obviously, the described embodiments are only some embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0035] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with the drawings and specific embodiments.

[0036] Embodiment one

[0037] As Figure 1 shown, the method flowchart of the present embodiment, the steps include:

[0038] S1. According to the pre-defined thin film type acoustic metamaterial structure parameter range, the sound transmission loss STL curve data set is generated.

[0039] The model of thin film type acoustic metamaterial is built by COMSOL, the impedance tube length is 100, the mass block shape is selected as circular, the model is saved as.mph format after building, COMSOL Multiphysics 6.2 with MATLAB is opened to show that it starts to listen on port 2306, programming in MATLAB, calling the model generated in COMSOL, defining the parameter range, using mphlopt command to get STL sound insulation picture and get data set. The COMSOL model is shown in Figure 2 , and the range of hyperparameter optimization is shown in Table 1.

[0040] Table 1

[0041] .

[0042] S2. The sound transmission loss STL curve data set is preprocessed and divided to obtain the processed data.

[0043] The generated data set is divided into training set, validation set and test set, and the data preparation and processing of neural network are carried out; the specific steps are as follows:

[0044] S2.1: The data set is randomly divided into training set, validation set and test set according to the ratio of 6:2:2.

[0045] S2.2: The STL curve is normalized, and the formula is as follows:

[0046]

[0047] Among them, indicates the data after standardization; x indicates the original data; μ indicates the mean (average) of the data set where the original data is located, reflecting the central tendency of the data; indicates the standard deviation of the data set where the original data is located, which reflects the dispersion degree of the data. The larger the standard deviation, the more dispersed the data.

[0048] S2.3: The training set data is enhanced, including Gaussian noise, curve translation and scaling.

[0049] The Gaussian noise formula is:

[0050]

[0051] wherein, is the data after adding noise (enhanced data), which is the result of noise injection on the original data x, obeys a Gaussian distribution with mean 0 and variance , The greater the noise, the more obvious the noise effect.

[0052] wherein the curve translation (x shift ) and scaling (x scale ) formula is:

[0053]

[0054]

[0055] wherein, is the translation amount, which controls the translation direction and distance; α0 is the scaling factor (constant), which controls the scaling ratio.

[0056] S2.4: Compare the data and the enhanced data, display the parameter distribution, help to adjust the data enhancement strategy, and the correlation coefficient matrix between parameters:

[0057]

[0058] wherein, represents the covariance (Covariance) of the feature maps X and Y, reflecting the joint change trend of the two variables; , and are the standard deviations (Standard Deviation) of the feature maps X and Y, respectively, used to standardize the covariance.

[0059] S3. Based on the processed data, build and train a convolutional neural network CNN model, and obtain a trained CNN model through hyperparameter optimization.

[0060] Step 3.1: Based on the convolutional neural network model, build an innovative acoustic CNN model that integrates EfficientNet, feature pyramid network, and dual attention mechanism, such as Figure 3As shown, EfficientNet backbone network optimizes depth, width and resolution through compound scaling, and realizes performance maximization combined with efficient module design. The core module MBConv adopts an inverted residual structure: first, expand the channel number through 1x1 convolution, then extract spatial features through depth separable convolution, and finally compress the channel through 1x1 convolution, and optionally add SE attention mechanism to dynamically calibrate channel weights. Each module works together: compound scaling ensures the balance between model size and computing resources, MBConv greatly reduces the parameter amount by separating space and channel calculation, and SE module enhances key features. Depth separable convolution and inverted residual design together improve feature reuse efficiency, and compound scaling makes the expansion of these modules more systematic. The specific structure is as follows:

[0061] (1) EfficientNet backbone network, containing 5 stage MBConv block sequence, structure as follows:

[0062] First stage: input channel 3, output channel 32, step 2, downsample to 112x112

[0063] Second stage: input channel 32, output channel 64, step 2, downsample to 56x56

[0064] Third stage: input channel 64, output channel 128, step 2, downsample to 28x28

[0065] Fourth stage: input channel 128, output channel 256, step 2, downsample to 14x14

[0066] Fifth stage: input channel 256, output channel 512, step 2, downsample to 7x7

[0067] Each MBConv block is composed of expansion point convolution, depth separable convolution, SE attention module and projection layer, and uses residual connection to optimize gradient flow.

[0068] Using two-dimensional convolution, local feature extraction and weight sharing are used to efficiently extract spatial features, as follows:

[0069]

[0070] Where, represents the value of the output feature map at position (i, j) and channel o after convolution operation; C in is the number of channels of input feature map X; K h , K w represent the height and width of the convolution kernel respectively; is the convolution kernel parameter, the dimension is (C outis the output channel number), represents the weight of the convolution kernel at the (h, w) position, input channel c, and output channel o; b o is the bias item corresponding to the output channel o.

[0071] Using deep separable convolution, reducing the risk of overfitting, maintaining the ability to extract features, and improving the efficiency of the network model, the formula is as follows:

[0072]

[0073]

[0074] wherein,, X is the input feature map; Y is the output feature map of the deep convolution, the dimension is consistent with X; k is the size of the convolution kernel; is the kernel parameter of the convolution, which represents the weight of the (m, n) position of the cth channel; is the bias of the cth channel; Z is the final output feature map of the pointwise convolution; is the kernel parameter of the pointwise convolution, is the bias of the pointwise convolution; represents the value of the input feature map at position (i, j) of the cth channel.

[0075] Using SE attention module, the model performance is significantly improved at a very small calculation cost, the intensity of the feature map channel is dynamically adjusted through compression, excitation, scaling, redundant information is suppressed, the feature dimension is compressed by projection layer, the dimension of residual connection is matched, the gradient vanishing is relieved through residual connection, and the training stability is improved.

[0076] (2) Feature Pyramid Network (FPN), realizing multi-scale feature fusion, integrating features at each level through horizontal connection and top-down path, the specific workflow is as follows:

[0077] First, use 5 horizontal connection convolutions to map the features of each scale to 256 channels.

[0078] Second, use 5 3x3 convolutions to extract and refine the fused feature representation.

[0079] Third, use the final 1x1 fusion convolution to fuse the spliced multi-scale features (256x5 channels) into 512 channel output.

[0080] The specific method of feature fusion is to first construct the vertical transmission path of the pyramid, and then guide the low-level features to enhance through the high-level semantics:

[0081]

[0082] Then, through the horizontal aggregation of the pyramid, the complementary information of each layer is integrated into a unified representation:

[0083]

[0084] wherein: represents the output of the i-th layer of the feature pyramid; represents the i-th layer feature connected transversely; represents the up-sampling operation; represents a 3x3 convolution for feature smoothing and channel adjustment; represents the concatenation of multi-level features along the channel dimension; represents a 1x1 convolution for adjusting the number of channels after fusion; represents the final fused feature map.

[0085] (3) Dual attention mechanism, simultaneously optimizing channel and spatial feature weights, enhancing key information capture ability, and improving model precision and generalization.

[0086] Channel attention, based on channel statistical information to generate channel weights. The formula is as follows:

[0087]

[0088] wherein F represents the input feature map; represents global average pooling; represents global maximum pooling; represents a multi-layer perceptron; represents activation function; represents the channel attention weight map.

[0089] Spatial attention: based on spatial statistical information to generate spatial weights. The formula is as follows:

[0090]

[0091] wherein, represents the spatial attention weight map; represents the output feature map of the channel attention module; , represents the average pooling and maximum pooling operation in the channel dimension; f represents the convolution operation.

[0092] Normalization, activation function and dropout operation are applied after each component to improve training stability and prevent overfitting.

[0093] S3.2: After convolution feature extraction and multi-module enhancement, the feature is compressed into a vector by global average pooling, and input into a 4-layer fully connected network for parameter prediction. The fully connected neural network diagram is shown in Figure 4 .

[0094] First layer: input dimension 512, output dimension 256, batch normalization, ReLU activation, 40% dropout rate;

[0095] Second layer: input dimension 256, output dimension 128, batch normalization, ReLU activation, 40% dropout rate;

[0096] Third layer: input dimension 128, output dimension 32, batch normalization, ReLU activation, 40% dropout rate;

[0097] Fourth layer: input dimension 32, output dimension 4, corresponding to the prediction of 4 acoustic parameters.

[0098] The mathematical expression of the fully connected layer is:

[0099]

[0100] Wherein: and are the weights and biases of each layer; BN is the batch normalization operation standardizing the intermediate features, accelerating the training and improving the stability; ReLU is the activation function, introducing nonlinearity to enable the network to fit complex functions.

[0101] S3.3: Train the main program, call the dataset and deep neural network model, and get the trained deep neural network model.

[0102] The specific steps are as follows:

[0103] In the training network part, a custom loss function is used, which is a weighted combination of multiple items, including mean square error (MSE), smooth L1 loss and relative error, and different types of parameters are optimized with weights. The formula is:

[0104] Mean square error loss:

[0105] Smooth L1 loss:

[0106] Relative error loss:

[0107]

[0108] In the formula, N represents the number of samples; represents the predicted value of the i-th sample; represents the true value of the i-th sample; represents the prediction error of a single sample; represents a constant to prevent the denominator from being zero; the operation of smoothL1 is as follows:

[0109] .

[0110] In the above formula, Always equal to 1, in the early stage of training, α is 0.35, β is 0.63, δ is 0.02, with the increase of training rounds, the loss ratio l of training set and validation set is used to dynamically adjust α, β, δ, to the later stage of training, L smoothL1 The weight will dominate and enhance the robustness of the later stage to avoid overfitting.

[0111] Then run the Bayesian optimization code to get the optimal set of hyperparameters through the loss ratio, the value of about equal to 1 is considered as the ideal state, which represents that the training loss and validation loss are declining synchronously, and the optimal hyperparameters obtained are shown in Table 2, and the loss curve graph obtained after the end is shown in Figure 5 .

[0112] Step 3.4: Set the training round to 300, use AdamW optimizer, configure weight decay, and randomly initialize the parameters using Kaiming method.

[0113] Step 3.5: Refer to OneCycleLR learning rate scheduling strategy to dynamically adjust the learning rate, the formula is as follows:

[0114]

[0115] Where, is the learning rate of the t-th step of training; is the initial learning rate; is the total number of steps in the warm-up phase; T is the total number of steps in the entire training.

[0116] Step 3.6: Introduce EMA mechanism to improve model stability, the formula is as follows:

[0117]

[0118] Where, is the sliding average parameter maintained by EMA after the t-th step of training; is the original parameter of the model after the t-th step of training; λ is the smoothing coefficient.

[0119] Step 3.7: Set the gradient clipping threshold to 1.0 to prevent gradient explosion, the formula is as follows:

[0120]

[0121] Where, g is the original gradient; is the clipped gradient; is the threshold of gradient clipping; is the norm of the gradient g.

[0122] Step 3.8: Set early stopping parameters: patience value is 50, i.e. stop training when the validation set loss does not decrease for 50 consecutive training rounds.

[0123] Step 3.9: Run Bayesian optimization, define 10 key hyperparameters, including learning rate, regularization strength, EMA decay rate, etc. in multiple dimensions, as shown in Table 2. Finally, get the loss curve graph.

[0124] Table 2

[0125] .

[0126] S4. According to the trained CNN model, predict the new STL curve image and output the structure parameters of the film type acoustic metamaterial.

[0127] The new STL curve image is first loaded into the AcousticCNN2D deep learning model enhanced by the EfficientNet backbone network and the feature pyramid network, and then the input 224x224 image is preprocessed. The image preprocessing process includes adaptive histogram equalization, Gaussian filter denoising, median filter outlier removal, and edge enhancement, etc. to ensure the quality of the input image. The preprocessed grayscale image is converted to 3-channel RGB format and normalized to meet the model input requirements.

[0128] The new STL sound insulation curve image is predicted, and running the prediction code will get 4 different predicted parameter values. The mass radius error is between 3.4%-6.6%, the film thickness error is between 8.3%-11.35%, the mass density error is between 0.01%-2.5%, and the film density error is between 3.6%-4.5%. As shown in Table 3.

[0129] Table 3

[0130] .

[0131] According to the 4 different predicted parameter values, the STL sound insulation curve is simulated in COMSOL, and then compared with the STL sound insulation curve obtained by the original parameters. As shown in Table 3. Figure 6

[0132] The above-described embodiments are only descriptions of the preferred modes of the present application and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.​

Claims

1.A deep learning-based inverse design method for a thin-film acoustic metamaterial, characterized in that, The method comprises the following steps: According to the pre-defined thin film type acoustic metamaterial structure parameter range, the sound transmission loss STL curve data set is generated; The sound transmission loss STL curve data set is pre-processed and divided to obtain the processed data; Based on the processed data, a convolutional neural network CNN model is constructed and trained, and the trained CNN model is obtained through hyperparameter optimization; According to the trained CNN model, a new STL curve image is predicted, and the structure parameters of the thin film type acoustic metamaterial are output. 2.The deep learning-based inverse design method of a thin film acoustic metamaterial, as claimed in claim 1, wherein The method for generating the sound transmission loss STL curve data set comprises: A physical model of the thin film type acoustic metamaterial is established using COMSOL software; The physical model is called through MATLAB programming, and the structure parameter range is defined; The STL curve image is generated using the mphplot command to form the STL curve data set. 3.The deep learning-based inverse design method of thin film acoustic metamaterials according to claim 1, wherein, The constructed CNN model comprises: EfficientNet is used as the backbone network for feature extraction; The feature pyramid network FPN is used for multi-scale feature fusion; The dual attention mechanism is used to optimize the channel and spatial feature weights. 4.The deep learning-based inverse design method of thin film acoustic metamaterials according to claim 3, wherein, The EfficientNet backbone network comprises a plurality of MBConv blocks, each of which comprises, in sequence: an expansion point convolution layer for expanding feature channels; a depth separable convolution layer for extracting spatial features; an SE attention module for dynamically adjusting channel weights; a projection layer for compressing feature dimensions; wherein each MBConv block is connected in a residual manner. 5.The deep learning-based inverse design method of thin film type acoustic metamaterials according to claim 1, wherein, The training method of the CNN model comprises: using a custom loss function, wherein the custom loss function is a weighted combination of mean square error MSE loss, smooth L1 loss and relative error loss; and dynamically adjusting the weights of each loss in the custom loss function according to the training process. 6.The deep learning-based inverse design method of thin film acoustic metamaterials according to claim 5, wherein, The final loss function comprises: Wherein, α, β, δ represent weight coefficients; L MSE represents the mean square error loss, L smoothL1 represents the L1 smoothing loss; L rel represents the relative error loss. 7.The deep learning-based inverse design method of thin film type acoustic metamaterials according to claim 1, wherein, The method for training the CNN model further comprises: dynamically adjusting the learning rate using the OneCycleLR learning rate scheduling strategy; optimizing the model parameters using the exponential moving average EMA mechanism; setting a gradient clipping threshold to prevent gradient explosion; and setting an early stopping mechanism to stop training when the loss of the validation set no longer decreases. 8.The deep learning-based inverse design method of thin film type acoustic metamaterials according to claim 1, wherein, The structure parameters comprise: mass block radius, film thickness, mass block density and film density.

Images