Rapid forecasting method for thermal-mechanical-electrical response of quartz / epoxy resin composite material under laser irradiation based on VQ-VAE

Abstract

The invention relates to the technical field of composite material damage evaluation and performance prediction, and discloses a method and a system for rapidly predicting thermal-mechanical-electrical response of a quartz / epoxy resin composite material under laser irradiation based on a vector quantization variational auto-encoder (VQ-VAE). According to the method, a cascade deep learning neural network architecture of a DNN prediction network and a VQ-VAE decoder is constructed, laser working condition parameters are mapped to a discrete potential space, and complete temperature field, stress field and dielectric constant parameters are rapidly reconstructed. Wherein the VQ-VAE adopts a vector quantization mechanism to extract material ablation phase change and damage features, the visual field resolution is adjusted by improving the size of a convolutional layer in an encoder, and extraction of small-scale damage features is achieved. According to the method, second-level forecasting from parameter input to full-field output is achieved, the calculation efficiency and the forecasting precision are remarkably improved, and the method is suitable for rapid performance evaluation and design optimization of the composite material under laser irradiation.

Description

Technical Field

[0001] This invention relates to the field of composite material damage assessment and performance prediction technology, and in particular to a rapid prediction method for the thermo-mechanical-electric response of quartz / epoxy resin composite materials under laser irradiation based on VQ-VAE. Background Technology

[0002] Quartz fiber reinforced epoxy resin composites are widely used in high-end fields such as aerospace, military defense, high-speed aircraft, and radar systems due to their excellent specific strength, heat resistance, and electromagnetic wave transmission properties. Especially in key functional components such as radomes, antenna radomes, and electromagnetic windows, these materials not only need to withstand complex aerodynamic loads and thermal environments, but also must maintain stable electromagnetic wave transmission performance to ensure the realization of core functions such as radar detection and communication transmission.

[0003] In real-world applications, especially in military confrontations or high-energy weapon testing scenarios, these composite material components may be subjected to irradiation attacks from high-energy laser weapons. Laser beams are characterized by concentrated energy and rapid action, capable of causing drastic temperature rises, thermal stress concentration, matrix pyrolysis, fiber damage, and even ablation on the material's surface and interior within a very short time. This damage not only weakens the material's structural integrity but also significantly alters its dielectric properties, leading to decreased or even complete failure of its wave transmission performance, thereby severely impacting the functionality of the entire radar or communication system.

[0004] Therefore, accurate and rapid prediction of the thermo-mechanical-electric multiphysics coupled response of composite materials under laser irradiation has become an indispensable technical link in materials design, damage assessment, protection strategy formulation, and system reliability assurance. Traditionally, this prediction relies on numerical simulation, experimental testing, or simplified models, but these methods have significant shortcomings in terms of efficiency, accuracy, and applicability.

[0005] Currently, in the prediction of composite material properties under laser irradiation, there are mainly three technical approaches and their corresponding technical problems:

[0006] 1. Multiphysics coupled numerical simulation (MPS) methods, which establish a detailed finite element model and solve the equations of heat conduction, thermoelasticity, and electromagnetic fields in a coupled manner, can obtain relatively accurate physical field distributions. However, to capture high-gradient phenomena such as ablation fronts and stress concentration zones, extremely fine meshes and extremely small time steps are required, resulting in a single complete simulation taking hours or even days. This severe inefficiency makes it unsuitable for practical engineering needs such as engineering design optimization, real-time damage assessment, or rapid parameter scanning.

[0007] 2. Traditional machine learning proxy model methods, which utilize neural networks and other models to establish a mapping from laser parameters to key responses, can improve computational speed, but generally suffer from the following problems:

[0008] It can only predict limited scalar or low-dimensional vector results (such as maximum temperature and maximum stress), and cannot reconstruct the complete high-resolution physical field distribution, losing a large number of local damage details;

[0009] Traditional autoencoders (VAEs) use a continuous latent variable space, which makes it difficult to effectively characterize physical processes with abrupt and discrete characteristics, such as material phase transitions and ablation pit formation, resulting in unclear prediction results in key areas.

[0010] Standard convolutional neural network (CNN) structures are not optimized for micron- to millimeter-scale features of laser damage to composite materials, making it difficult to effectively extract and retain small-scale details such as ablation pit edges and microcracks, thus affecting the accuracy of electromagnetic performance prediction.

[0011] 3. Experimental testing method: This method relies on physical samples and specialized testing equipment, involving laser irradiation experiments and measuring the response using equipment such as temperature recorders and vector network analyzers. Its limitations include:

[0012] The experiment has a long cycle, high cost, and is significantly affected by factors such as the consistency of sample preparation and interference from the testing environment;

[0013] It cannot flexibly support parametric research and design iteration, and it is difficult to effectively predict extreme or dangerous working conditions;

[0014] It has poor repeatability and lacks feasibility for rapid forecasting and engineering application.

[0015] In summary, existing technologies struggle to balance high accuracy, full-physics field output, and rapid computation. There is an urgent need for a rapid forecasting method for the thermo-mechanical-electric response of quartz / epoxy resin composites under laser irradiation, based on a vector quantization variational autoencoder (VQ-VAE), that can achieve efficient, high-accuracy, and full-field forecasting. Summary of the Invention

[0016] The core objective of this invention is to solve the problem of achieving second-level rapid prediction of the thermo-mechanical-electric multi-physics coupled response of quartz / epoxy resin composite materials under laser irradiation while maintaining high accuracy. This is mainly reflected in the following aspects:

[0017] First, how to overcome the shortcomings of existing numerical simulation methods that take a very long time to calculate, and while ensuring the accuracy of physical field prediction, to significantly improve the prediction speed, so as to meet the urgent needs of engineering practice for rapid evaluation, parameter optimization and real-time feedback;

[0018] Second, how to address the limitations of existing data-driven methods based on traditional machine learning in prediction, overcome their inability to generate complete spatial physical field distributions with high precision, overcome their insufficient ability to characterize discrete and abrupt physical processes such as material phase transitions and damage evolution, and overcome the weakness of their standard network architecture in capturing small-scale details of ablation damage; ultimately achieving accurate predictions with complete, high-resolution, and clear physical features.

[0019] Third, how to overcome the limitations of traditional experimental methods, such as high cost, long cycle and poor flexibility, and provide a numerical alternative that can quickly, cost-effectively and repeatably predict various laser parameters and complex working conditions to support efficient material design and performance evaluation; the above-mentioned problems.

[0020] Therefore, this invention designs a rapid prediction method for the thermo-mechanical-electric response of quartz / epoxy resin composite materials under laser irradiation based on VQ-VAE. The core technology of this method lies in:

[0021] VQ-VAE is introduced for discretized physical field characterization: continuous features are mapped to discrete codes through a vector quantization encoding table, which effectively characterizes abrupt physical processes such as material phase transition and damage evolution, and overcomes the ambiguity problem at the interface in traditional continuous latent variable models.

[0022] Construct a cascaded forecast architecture of "parameter-latent variable-physical field": use a deep neural network (DNN) to establish a mapping from laser parameters to the discrete latent space of VQ-VAE, and then use a pre-trained VQ-VAE decoder to quickly reconstruct the complete physical field, so as to achieve end-to-end second-level forecast from parameter input to full field output;

[0023] Design a feature extraction network for small-scale damage: Because the damage radius caused by laser ablation is relatively small compared to the geometric dimensions of the thin plate in the dataset and differs significantly from the surrounding feature values, ablation hole damage appears as a small-scale data feature in the dataset, such as... Figure 4 As shown, the receptive field of multiple convolutional layers is crucial for the correct generation of small local features. The receptive field sizes of convolutional layers with different structures are compared to... Figure 5 As shown, a multi-layer small-size convolutional kernel stacked structure is adopted in the encoder of the VQ-VAE reconstruction network to enhance the characterization ability of minute features such as ablation pit edges and microcracks, and to ensure the accurate correlation between damage details and electromagnetic performance;

[0024] A systematic process from data generation to model deployment is achieved: integrating Latin hypercube sampling, high-precision coupled simulation, data preprocessing, and two-stage model training and verification to form a reproducible and scalable rapid forecasting technology system.

[0025] The specific technical solution of the present invention to achieve the above objectives is a rapid prediction method for the thermo-mechanical-electric response of quartz / epoxy resin composite materials under laser irradiation based on VQ-VAE, comprising the following steps:

[0026] Step 1: Construct a multiphysics dataset covering the target working conditions based on high-precision numerical simulation;

[0027] Step 2: Construct and train a cascaded prediction model consisting of a VQ-VAE reconstruction network and a DNN prediction network;

[0028] Step 3: Use the trained cascaded prediction model to quickly predict and verify new laser operating conditions.

[0029] The process of simulating and constructing a multiphysics dataset in step 1 includes:

[0030] The Latin hypercube sampling method is used to extract sample points in the parameter space of laser loading conditions.

[0031] For each sample point, a thermo-mechanical-electric coupled finite element simulation is performed to extract data such as temperature field, stress field, and dielectric constant parameters.

[0032] The extracted data is normalized and its size is standardized to construct a multi-physics dataset, which includes a training set and a test set.

[0033] Step 2, the VQ-VAE network reconstruction includes:

[0034] The encoder employs a multi-layer, small-sized convolution kernel stacking structure;

[0035] Vector quantization layer, containing a learnable encoding table, is used to map continuous features to discrete codes;

[0036] A decoder is used to reconstruct physical field data from discrete codes.

[0037] In step 2, the encoder has a convolution kernel size of 3×3 and a network depth of 4 layers.

[0038] In step 2, the DNN prediction network is a fully connected neural network, and its output dimension is consistent with the dimension of the continuous feature vector output by the encoder in the VQ-VAE reconstruction network.

[0039] The training of the DNN prediction network employs a joint loss function, which includes: feature space mean squared error loss and reconstruction mean squared error loss.

[0040] The rapid forecasting process in step 3 is as follows:

[0041] Input the laser loading condition parameters into the DNN prediction network and output a continuous feature vector;

[0042] It is converted into discrete code through a vector quantization layer;

[0043] Input the decoder in the frozen VQ-VAE reconstruction network and output the multiphysics prediction results.

[0044] The composite material is a quartz fiber reinforced epoxy resin composite material.

[0045] A rapid forecasting system, the system for implementing the steps of the method, includes the following modules:

[0046] The data generation module is used to perform high-precision coupled simulation and data extraction;

[0047] The model training module is used to build and train the VQ-VAE reconstruction network and the DNN prediction network;

[0048] The forecast execution module is used to receive laser parameters and output physical field forecast results.

[0049] A computer-readable storage medium having a computer program stored thereon, which, when executed, can perform the steps of the method.

[0050] Compared with the prior art, the technical solution disclosed in this application has the following non-obvious technical features:

[0051] The first distinguishing technical feature is that this application uses a vector quantization variational autoencoder (VQ-VAE) to discretize, compress, and reconstruct multiphysics data. It maps continuous features to discrete indices through a learnable encoding table, specifically targeting abrupt physical processes such as material phase transitions and ablation damage, thus solving the ambiguity problem at physical interfaces in traditional continuous latent variable models.

[0052] The second distinguishing technical feature is that this application constructs a cascaded prediction architecture of "DNN prediction network + VQ-VAE reconstruction network (especially: decoder)", which transforms the time-consuming iterative physical solution process into a single feedforward neural network calculation, realizing a second-level fast mapping from laser parameters to the complete physical field distribution;

[0053] The third distinguishing technical feature is that this application designs a multi-layer small-sized convolution kernel stacked structure in the encoder of the VQ-VAE reconstruction network, which is optimized for the micron to millimeter-level features of laser damage to composite materials, and improves the feature extraction capability of key details such as ablation front and microcracks.

[0054] The fourth distinguishing technical feature is that this application proposes a systematic technical solution covering the entire process of intelligent working condition design, high-precision simulation data generation, standardized preprocessing, two-stage model training and integrated verification, to ensure the reliability, generalization ability and engineering practicality of the model.

[0055] The fifth distinguishing technical feature is that this application uses a joint loss function to train the DNN prediction network, which not only constrains its output to align with the discrete features of the VQ-VAE reconstruction network, but also ensures the accuracy of the final physical field output through reconstruction loss, thus achieving the end-to-end optimization goal.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] 1. This invention can significantly improve forecasting speed by replacing traditional iterative solutions with a cascaded neural network architecture, reducing the time for a single forecast from several hours to seconds, thus meeting the engineering requirements for rapid evaluation and real-time feedback.

[0058] 2. The physical field reconstruction of this invention has high accuracy. By utilizing the discretized latent space of the VQ-VAE reconstruction network, it clearly characterizes abrupt changes such as material phase transitions and damage boundaries, significantly improving the clarity and accuracy of the full-field reconstruction.

[0059] 3. This invention can effectively preserve small-scale damage features. The encoder structure is optimized for micro-damage features, ensuring that details such as the edge of ablation pits and micro-cracks are captured and mapped into electromagnetic performance prediction, thereby improving the physical reliability of the prediction.

[0060] 4. This invention utilizes a systematic process to ensure the reliability of forecasts. The entire process design, from data generation to model deployment, ensures that the model has strong generalization ability and engineering applicability, reducing the reliance on repeated experiments or simulations.

[0061] 5. This invention adopts integrated multi-physics field output, which can simultaneously output high-resolution distributions of temperature field, stress field, dielectric constant, etc., providing complete data support for the collaborative evaluation of multiple properties of composite materials under laser irradiation. Attached Figure Description

[0062] Figure 1 This is a flowchart of the method described in this invention;

[0063] Figure 2 This is a schematic diagram of the structure and principle of the VQ-VAE reconstruction network described in this invention;

[0064] Figure 3 This is a structural block diagram of the system described in this invention;

[0065] Figure 4 This is the receptive field of small-sized data in the downsampling layer as described in this invention;

[0066] Figure 5 This is a comparison table of the receptive fields of different convolutional layer structures described in this invention. Detailed Implementation

[0067] The present invention will now be described in detail with reference to the accompanying drawings:

[0068] A rapid prediction method for the thermo-mechanical-electric response of quartz / epoxy resin composites under laser irradiation based on VQ-VAE is proposed, and its specific steps are as follows:

[0069] Step 1: Construct a high-precision and diverse physical field dataset that covers the target working conditions and is used to train deep learning models based on high-precision numerical simulation.

[0070] Step 2: Construct and train a cascaded prediction model consisting of a VQ-VAE reconstruction network and a DNN prediction network; establish a generative neural network model that can directly and quickly generate a complete physical field distribution from laser operating parameters;

[0071] Step 3: Perform rapid forecasting and verification based on the fully trained model. Use the trained model to quickly forecast new working conditions and verify the forecasting accuracy and efficiency by comparing with high-precision simulation results.

[0072] The specific process of step 1 includes:

[0073] Step 1.1: Design and sampling of the parameter space for laser loading conditions. Determine the key laser parameters affecting the material response and their variation ranges, including laser power density P and the coordinates (X,Y) of the application position. Use the Latin hypercube sampling (LHS) method to extract N (e.g., 500) sample points with good space-filling properties and no correlation within the above multidimensional parameter space. LHS sampling divides the value range of each parameter into N equally probable intervals and randomly selects a value within each interval, ensuring that the sample points uniformly cover the entire parameter space and that the correlation between variables is low.

[0074] Step 1.2: High-precision multiphysics coupled numerical calculation and data extraction. For each laser loading sample point in Step 1.1, high-precision finite element analysis is performed to generate physical field response data. Model establishment and solution: A three-dimensional fine geometric model of the quartz / epoxy resin composite material is established in finite element software (such as ABAQUS), defining the thermophysical parameters (thermal conductivity, specific heat capacity), mechanical parameters (elastic modulus, Poisson's ratio, coefficient of thermal expansion), and dielectric property evolution model of the material as a function of temperature. A user subroutine is used to define a laser heat source (such as Gaussian distributed surface heat flow) as a transient load, and a thermo-mechanical-electric coupled analysis is performed to obtain the transient temperature field and stress / strain field. Based on the temperature field and material damage state (such as resin pyrolysis, interlaminar cracking) obtained from the thermo-mechanical analysis, and based on the electromagnetic parameters of each component material (original quartz / epoxy resin, pyrolytic carbon), corresponding complex permittivity and other electromagnetic parameters are assigned to each region according to the logarithmic mixing law, thereby constructing a material property field for electromagnetic simulation. Subsequently, Maxwell's equations were solved using the finite element method in electromagnetic simulation software (such as CST and HFSS) to calculate the scattering parameters (S-parameters) and spatial dielectric constant distribution under this damage state. Structured data extraction: A data extraction algorithm was developed to rearrange the numbering of all elements in the finite element model according to spatial location (column W, row H, depth C dimension). After the solution was completed, the physical quantity results (temperature, equivalent stress, z-displacement, dielectric constant, etc.) of each element or node were extracted in this numbered order and recorded according to time steps (e.g., at 0.1-second intervals). The extracted data was written into a structured text file or array file by layer and coordinate order, and the corresponding operating parameters (P, X, Y) were embedded in the filename for association.

[0075] Step 1.3: Data Preprocessing and Dataset Construction; Normalization and Resizing: All extracted physics data are normalized based on their maximum values, mapping the values ​​of each physical quantity to the [0,1] interval. Simultaneously, to accommodate the input requirements of the subsequent convolutional neural network, bilinear interpolation and other methods are used to uniformly scale the data to a standard size (e.g., 256×256). Dataset Encapsulation and Partitioning: The processed multiphysics data (temperature field, stress field, dielectric constant) are stacked along the channel dimension to form a multi-channel three-dimensional tensor, which is then encapsulated into a specific data format file (e.g., .npy). Finally, all data samples are randomly divided into training and validation / test sets according to a preset ratio (e.g., 9:1).

[0076] The specific process of step 2 includes:

[0077] Step 2.1: Construct and train an optimized generative VQ-VAE reconstruction network capable of efficient discretization compression and high-precision reconstruction of multiphysics data. The specific implementation is as follows: First, construct a complete VQ-VAE network, consisting of an encoder, a vector quantization layer, and a decoder. The encoder comprises four cascaded convolutional downsampling modules, used to progressively compress the input high-dimensional physics data (size 256×256) into low-dimensional continuous feature vectors (e.g., 16×16×128). This encoder structure is specifically optimized for the small spatial scale features of ablation damage: a multi-layer, small-sized convolutional kernel (e.g., 3×3) stacked design is adopted to enhance the feature extraction capability of key details such as the ablation front and microcracks, ensuring their effective capture and encoding. The vector quantization layer contains a learnable encoding table, which consists of K C-dimensional embedding vectors. Nearest neighbor search is used to replace the C-dimensional vector corresponding to the thickness in each continuous feature vector output by the encoder with the nearest embedding vector in the encoding table, thus discretizing the continuous features into a spatial index matrix and corresponding quantized continuous feature vectors. This discretization mechanism is the core of this method's ability to clearly reconstruct abrupt physical processes such as material phase transitions and damage evolution. The decoder structure is symmetrical to the encoder, and it reconstructs the quantized feature vectors into multiphysics data of the original size through a series of upsampling operations. The loss function used during network training consists of three parts: reconstruction loss (mean square error between output and input data), encoding table loss (pushing the encoding table vector closer to the encoder output), and commitment loss (constraining the encoder output to be closer to the selected encoding table vector). To overcome the non-differentiability of vector quantization operations, gradient pass-through estimation is used during backpropagation, directly copying the gradient of the loss function calculated by the decoder to the encoder output, thus providing effective update gradients for the encoder and encoding table. Finally, gradient descent is used to jointly optimize the weights of the encoder and decoder, as well as the encoding table vectors. Through this process, the VQ-VAE network not only learns data compression and reconstruction, but also establishes an intrinsic mapping relationship from fine damage features to macroscopic electromagnetic parameter distribution in the discrete latent space.

[0078] Step 2.2: Construct and train a DNN feedforward neural prediction network to establish a fast mapping from laser operating parameters to the VQ-VAE discrete latent space. The specific implementation is as follows: First, construct a deep fully connected neural network. The input layer dimension is consistent with the number of laser operating parameters (e.g., 3D), and the output layer dimension strictly matches the dimension of the flattened continuous feature vectors output by the VQ-VAE encoder (e.g., 16×16×128=32768). The network contains multiple hidden layers with batch normalization layers and ReLU activation functions. During the training phase, all parameters of the previously trained VQ-VAE reconstruction network are kept fixed and frozen. The training objective of the DNN prediction network is defined as: to make its output continuous vector approximate the discrete feature vectors of the VQ-VAE encoding table as closely as possible. To this end, the designed loss function consists of two parts: one is the feature space loss, which calculates the mean square error between the DNN output and the discrete feature vectors of the VQ-VAE encoding table; the other is the reconstruction loss, which calculates the mean square error between the reconstructed physical field and the real physical field after the DNN output is reshaped and processed by the frozen VQ-VAE quantization layer and decoder. By optimizing this joint loss function, the DNN is trained to directly generate a continuous feature vector based on the input laser parameters. After the vector is processed by the subsequent fixed VQ-VAE quantization and decoding process, the complete multiphysics distribution can be reconstructed quickly and accurately.

[0079] The specific process of step 3 includes:

[0080] Step 3.1: Model Integration and Fast Forecasting. The trained DNN prediction network and VQ-VAE decoder network are integrated to form an end-to-end real-time forecasting network. For any new set of laser condition parameters, they are directly input into this integrated network. The DNN prediction network first maps them to a continuous feature vector; then, this vector is processed by the frozen VQ-VAE vector quantization layer (the same as during training), and is converted into discrete quantized features by finding the nearest neighbor embedding vector in the fixed encoding table; finally, this quantized feature is input into the frozen VQ-VAE decoder, and through one forward propagation, the normalized multiphysics prediction result is output.

[0081] Step 3.2: Post-processing and output of results. The normalized physical field data output by the forecast network is denormalized according to the normalization parameters in Step 1.3 to restore the temperature field, stress field and dielectric constant parameters with actual physical units; finally, the complete forecast results are output in the form of two-dimensional cloud map, three-dimensional distribution or data file, etc.

[0082] Step 3.3: Model Validation. The prediction model is quantitatively evaluated using the test set reserved in Step 1. Accuracy verification: The model prediction results are compared with the high-precision finite element benchmark solution; the relative errors of key physical quantities (such as maximum temperature, maximum stress, and transmittance at specific locations) are calculated, as well as quantitative indicators such as root mean square error (RMSE) and mean absolute error (MAE) of the whole field data; in addition, the structural similarity index (SSIM) can be used to evaluate the similarity between the predicted field and the actual field in the overall structure.

[0083] Example 1:

[0084] A rapid prediction method for the thermo-mechanical-electric response of quartz / epoxy resin composites under laser irradiation based on VQ-VAE is presented, with the overall process as follows: Figure 1 As shown, it includes three stages: data preparation, model training, and forecast validation.

[0085] Data preparation stage:

[0086] Laser parameter sampling: Determine key parameters and their ranges such as laser power density P and action position (X,Y), and extract 500 sample points using Latin hypercube sampling;

[0087] High-precision coupled simulation: A three-dimensional model of the composite material is established in ABAQUS, temperature-dependent thermophysical and mechanical parameters are defined, a Gaussian distributed laser heat source is applied, and transient thermo-mechanical coupling analysis is performed. Based on the damage state, the dielectric constant and S-parameters are calculated in CST.

[0088] Data extraction and preprocessing: Extract the physical field data according to the spatial grid, normalize to the [0,1] interval, uniformly scale to 256×256 size, and divide the training set and test set at a ratio of 9:1;

[0089] Model training phase:

[0090] VQ-VAE Reconstructs Network Training:

[0091] Encoder: 4 convolutional layers, 3×3 kernels, stride 2, output feature map size 16×16×128;

[0092] Vector quantization layer: The encoding table contains 512 128-dimensional vectors, and discretization is achieved through nearest neighbor search;

[0093] Decoder: Symmetric upsampling structure;

[0094] Loss function: reconstruction loss + encoding table loss + commitment loss, with gradient pass-through estimation used for backpropagation;

[0095] DNN prediction network training:

[0096] Network structure: 3-dimensional input layer, 32768-dimensional output layer, containing 3 fully connected layers (including batch normalization and ReLU activation).

[0097] Training method: Fix the VQ-VAE parameters and train the DNN to make the output approximate the discrete features of VQ-VAE;

[0098] Loss function: Feature space MSE + Reconstructed MSE;

[0099] Forecasting and Verification Phase:

[0100] Model ensemble: Integrating the trained DNN with the VQ-VAE decoder into an end-to-end network.

[0101] Rapid forecast: Input new laser parameters, and output the physical field forecast result after one forward propagation;

[0102] Post-processing: Inverse normalization processing, outputting contour plots or data files of temperature field, stress field, and dielectric constant parameter field;

[0103] Accuracy verification: RMSE, MAE, SSIM and other indicators are calculated using a test set and compared with the high-precision simulation results.

[0104] Example 2:

[0105] A rapid prediction method for the thermo-mechanical-electric response of quartz / epoxy resin composites under laser irradiation based on VQ-VAE is presented, with the overall process as follows: Figure 1 As shown, taking a quartz / epoxy resin laminate as an example, the laser power density is 500W / cm², the spot radius is 2mm, and the irradiation time is 5s. Using this method, the prediction takes about 0.5 seconds, outputting the temperature, stress, and dielectric constant distribution across the entire field. Compared with the finite element reference solution, the overall RMSE is less than 5%, and the SSIM is greater than 0.92, meeting the engineering accuracy requirements. Other aspects are the same as in Example 1.

[0106] Example 3:

[0107] A rapid forecasting system, such as Figure 3 As shown, the system for implementing the method includes the following modules:

[0108] The data generation module is used to perform high-precision coupled simulation and data extraction;

[0109] The model training module is used to build and train the VQ-VAE reconstruction network and the DNN prediction network;

[0110] The forecast execution module is used to receive laser parameters and output physical field forecast results.

[0111] Example 4:

[0112] A computer-readable storage medium having a computer program stored thereon, which, when executed, can perform the steps of the method.

[0113] The above technical solutions only embody the preferred technical solutions of the present invention. Any modifications that may be made by those skilled in the art to certain parts thereof embody the principles of the present invention and fall within the protection scope of the present invention.

Claims

1. A rapid prediction method for the thermo-mechanical-electric response of quartz / epoxy resin composite materials under laser irradiation based on VQ-VAE, characterized in that, Includes the following steps: Step 1: Construct a multiphysics dataset covering the target working conditions based on high-precision numerical simulation; Step 2: Construct and train a cascaded prediction model consisting of a VQ-VAE reconstruction network and a DNN prediction network; Step 3: Use the trained cascaded prediction model to quickly predict and verify new laser operating conditions.

2. The method according to claim 1, characterized in that, The process of simulating and constructing a multiphysics dataset in step 1 includes: The Latin hypercube sampling method is used to extract sample points in the parameter space of laser loading conditions. For each sample point, a thermo-mechanical-electric coupled finite element simulation is performed to extract data on temperature field, stress field, and dielectric constant parameters. The extracted data is normalized and its size is standardized to construct a multi-physics dataset, which includes a training set and a test set.

3. The method according to claim 1, characterized in that, Step 2, the VQ-VAE network reconstruction includes: The encoder employs a multi-layer, small-sized convolution kernel stacking structure; Vector quantization layer, containing a learnable encoding table, is used to map continuous features to discrete codes; A decoder is used to reconstruct physical field data from discrete codes.

4. The method according to claim 3, characterized in that, In step 2, the encoder has a convolution kernel size of 3×3 and a network depth of 4 layers.

5. The method according to claim 4, characterized in that, In step 2, the DNN prediction network is a fully connected neural network, and its output dimension is consistent with the dimension of the continuous feature vector output by the encoder in the VQ-VAE reconstruction network.

6. The method according to claim 5, characterized in that, The training of the DNN prediction network employs a joint loss function, which includes: feature space mean squared error loss and reconstruction mean squared error loss.

7. The method according to claim 1, characterized in that, The rapid forecasting process in step 3 is as follows: Input the laser loading condition parameters into the DNN prediction network and output a continuous feature vector; It is converted into discrete code through a vector quantization layer; Input the decoder in the frozen VQ-VAE reconstruction network and output the multiphysics prediction results.

8. The method according to claim 1, characterized in that, The composite material is a quartz fiber reinforced epoxy resin composite material.

9. A rapid forecasting system, said system being used to implement the steps of the method according to any one of claims 1 to 8, characterized in that, Includes the following modules: The data generation module is used to perform high-precision coupled simulation and data extraction; The model training module is used to build and train the VQ-VAE reconstruction network and the DNN prediction network; The forecast execution module is used to receive laser parameters and output physical field forecast results.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed, the computer program is capable of performing the steps of the method according to any one of claims 1 to 8.

Images