A method and system for predicting a matrix crack path of a carbon fiber composite

Abstract

The present application relates to the technical field of crack path prediction, and particularly relates to a carbon fiber composite material matrix crack path prediction method and system, the method comprising: establishing a fracture phase field model containing a total free energy of a phase field, an elastic strain energy and a dynamic evolution equation, and constructing a microstructure sample based on randomly generated fiber distribution, obtaining corresponding crack phase field data through numerical simulation to form a sample data set with physical consistency; constructing a crack path prediction model based on a convolutional neural network, introducing an attention mechanism in the skip connection between the encoder and the decoder; constructing a total loss function containing a data-driven loss term and a physical constraint loss term to constrain the predicted crack path to be more consistent with the physical reality. The present application is used to quickly and accurately predict the fracture crack path of the microstructure unit under different fiber distribution conditions, and overcomes the limitations of high calculation cost of traditional numerical simulation and the difficulty of accurately predicting physically consistent results by existing data-driven methods.

Description

Technical Field

[0001] This invention relates to the field of crack path prediction technology, specifically to a method and system for predicting crack paths in carbon fiber composite matrix. Background Technology

[0002] Carbon fiber reinforced polymer composites are widely used in aerospace, transportation, and infrastructure engineering fields due to their excellent mechanical properties and high specific strength. While the fibers significantly enhance the longitudinal load-bearing capacity of the material, the significant stiffness mismatch between the fibers and the matrix means that the material's performance under lateral stress is primarily dominated by the weaker matrix material. For example, in laminated structures subjected to high loads, matrix cracking within the 90° layer inevitably becomes the first damage mode to initiate. This microscopic damage is not only a key factor leading to subsequent delamination and fiber fracture, but also significantly weakens the overall stiffness of the structure and drastically shortens fatigue life.

[0003] In existing technologies, methods such as cohesive force models, extended finite element methods, or phase-field methods are commonly used to model representative volumetric elements containing fibers, matrix, and interfaces to simulate crack initiation and propagation. These methods can accurately reflect the stress field distribution and damage evolution process within the material. However, due to the need for detailed modeling of the microstructure and numerical solutions, they consume significant computational resources and have long computation cycles, making it difficult to meet the demand for rapid prediction. To reduce computational costs, some studies have attempted to use data-driven methods to predict crack paths by establishing a mapping relationship between microstructure and mechanical response. While these methods offer the advantage of fast prediction speed, existing methods largely rely on basic neural network models and lack effective constraints on the physical mechanisms of crack evolution. This leads to insufficient accuracy and poor physical consistency in the prediction results, especially when fiber distribution is highly random and the material exhibits strong anisotropy, causing the predicted crack path results to easily deviate from the actual crack propagation law. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for predicting crack paths in carbon fiber composite matrix in order to solve the above-mentioned technical problems. This method and system can quickly and accurately predict the fracture crack paths of microstructure units under different fiber distribution conditions, overcoming the limitations of high cost of traditional numerical simulation and the difficulty of accurately predicting physically consistent results using existing data-driven methods.

[0005] The objective of this invention can be achieved through the following technical solutions: A method for predicting crack paths in carbon fiber composite matrix, the method comprising: S1: Establish a fracture phase field model for the transverse matrix of carbon fiber composite material, including total phase field free energy, elastic strain energy, and dynamic evolution equations; S2: Generate samples that conform to the real fiber distribution for training in batches using a random generation method, perform numerical simulation based on the fracture phase field model, and output a sample dataset containing microstructure geometric features and corresponding crack phase field evolution. S3: Construct a crack path prediction model. The crack path prediction model establishes a nonlinear mapping relationship from microstructure geometric features to crack phase field distribution through a convolutional neural network architecture. The convolutional neural network model includes an encoder, a decoder, and a skip connection structure set between the encoder and the decoder. An attention mechanism is introduced into the skip connection structure. S4: Construct the total loss function, which includes a data-driven loss term and a physical constraint loss term. The data-driven loss term is used to characterize the difference between the crack phase field distribution output by the crack path prediction model and the crack phase field distribution data. The physical constraint loss term is used to penalize the non-zero probability of the prediction in the fiber region, guiding the crack to be generated and propagated in the crack preferential growth region. S5: Train the crack path prediction model. Input the sample dataset into the convolutional neural network architecture, optimize the total loss function through the backpropagation algorithm, and output the trained crack path prediction model. S6: Obtain the microstructure image to be predicted, input the microstructure image into the crack path prediction model, and output the corresponding crack phase field distribution.

[0006] Furthermore, methods for establishing fracture phase-field models include: S11: Set the order parameter d to characterize the damage state of the material and the order parameter c to describe the phase distribution state of the fibrous phase and the matrix phase; construct the total free energy function of the phase field that includes the elastic strain energy term and the crack surface dissipation energy term. S12: Calculate the density of elastic strain energy. Elastic strain energy is decomposed into elastic energy caused by tension and compression of the material, and the degradation of strain energy during crack evolution is characterized by a monotonically decreasing damage degradation function. S13: The crack surface is regularized and the crack surface energy is approximated as the density function distributed in the solid domain. The gradient of the order parameter d is used to describe the rate of change of the crack field in space. At the same time, a normalization coefficient is introduced to control the crack dispersion width. S14: Establish the field dynamic evolution equation of the order parameter d, introduce historical variables to avoid crack closure, and control the crack propagation rate through damage migration rate.

[0007] Furthermore, the random generation method is a fiber random generation method based on spatial exclusion, which includes: S21: Determine the material's dimensional parameters and randomly generate fiber center positions within a preset fiber region; set the fiber volume fraction, which determines the number of fibers per unit area. S22: Set the minimum fiber spacing. The minimum fiber spacing is set based on the gap between fibers under the actual fiber distribution of the material. Based on the minimum fiber spacing, the generated fibers and the surrounding area are marked as areas where fibers cannot be generated. S23: Continue to generate fibers in the remaining areas of the fiber region until the fiber volume fraction is reached.

[0008] Furthermore, the process of constructing the sample dataset includes: applying preset loading conditions and boundary conditions to each microstructure unit, and performing numerical simulation based on the fracture phase field model to output the corresponding crack phase field distribution data.

[0009] Furthermore, the encoder is used to extract multi-scale features from the microstructure image, and the decoder is used to reconstruct the multi-scale features step by step to output the crack phase field distribution.

[0010] Furthermore, the process of introducing an attention mechanism into the skip connection structure includes: Calculate attention coefficient ,set up This is a skip connection feature. For gating signals from coarse-grained levels, the attention coefficient The calculation formula is as follows: in, It is the ReLU activation function. For the Sigmoid function, , , This is a linear transformation achieved through 1×1 convolution. The bias term corresponds to the gated signal; the final skip connection feature is obtained through element-wise multiplication. get; Furthermore, the data-driven loss term includes binary cross-entropy loss. Mean square error loss and focus Tversky loss ; Total loss function Represented as: in, This is the physical constraint loss term, used to constrain the physical consistency of the prediction results; , , , Represents network weight parameters; binary cross-entropy loss. Used to enhance focus on cracked pixels; mean square error loss Used to ensure the smoothness of the predicted probability field; Focus Tversky loss Represented as: in, , , These represent true positive, false negative, and false positive, respectively. , , This is represented as a hyperparameter.

[0011] Furthermore, the physical constraint loss term Training methods include: S41. Extract all fiber coordinate information within the fiber region and calculate the center-to-center distance between any two adjacent fibers. And according to the distance from the center Calculate the fiber gap with fiber radius r ; Calculate the angle between the line connecting the centers of two adjacent fibers and the loading direction. When the fiber gaps The angle is not greater than the preset gap threshold and the included angle is not greater than the preset gap threshold. When the angle is not greater than a preset threshold, the matrix region between two adjacent fibers is marked as a crack-preferential growth region. In a preferred embodiment, the preset gap threshold is determined based on the fiber radius r. When the fiber radius r is 4 micrometers, the preset gap threshold ranges from 0.4 micrometers to 2.0 micrometers. The preset angle threshold is the maximum allowable deviation between the line connecting the centers of two adjacent fibers and the loading direction, and its value ranges from 10° to 30°.

[0012] S42. Penalize the non-zero crack probability of pixels located in the fiber region and matrix pixels not located in the crack preferential growth region in the prediction results, so as to guide the crack to be generated and propagated in the crack preferential growth region.

[0013] Furthermore, the training methods for the crack path prediction model include: S51. Divide the sample dataset into a training set and a validation set; S52. Input the training set into the crack path prediction model for iterative training; during the training process, update the network weight parameters of the crack path prediction model through the Adam optimizer; minimize the total loss function through the backpropagation algorithm; S53. Monitor the model performance on the validation set. When the loss on the validation set no longer decreases, adaptively adjust the network weight parameters to reduce the learning rate until the model converges and the optimal network weight parameters are saved.

[0014] Furthermore, a crack path prediction system for carbon fiber composite matrix, employing the aforementioned method, includes: The fracture phase field module is used to establish a fracture phase field model of the transverse matrix of carbon fiber composite materials, which includes the total phase field free energy, elastic strain energy, and dynamic evolution equations. The dataset construction module is used to generate samples that conform to the real fiber distribution in batches through random generation methods, and to perform numerical simulation based on the fracture phase field model, outputting a sample dataset containing microstructure geometric features and corresponding crack phase field evolution. The crack path prediction module is used to acquire the microstructure image to be predicted, input the microstructure image into the crack path prediction model, and output the corresponding crack phase field distribution. processor; A storage unit is used to store one or more programs that, when executed by one or more processors, enable the one or more processors to implement the aforementioned carbon fiber composite matrix crack path prediction method.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention generates microstructure input maps and corresponding fracture crack fields in batches based on a fracture phase-field model, forming a dataset construction method for deep learning. By analyzing the fracture mechanism and generating corresponding data on microstructure and crack phase-field distributions based on the fracture phase-field model, the training data achieves physical consistency, thereby improving the accuracy of crack path prediction results.

[0016] 2. This invention improves the accuracy of crack path prediction by introducing an attention mechanism into the skip connection structure of the convolutional neural network architecture, thereby enhancing crack-related features and suppressing interference from irrelevant regions.

[0017] 3. This invention improves the physical consistency of prediction results by setting a physical constraint loss term and penalizing the prediction of non-zero probability cracks within the fiber region, so that the prediction results meet the physical constraint that cracks do not occur in the fiber region and are more likely to initiate in the dangerous region.

[0018] 4. This invention trains a crack path prediction model and makes predictions based on the trained model, thereby achieving rapid prediction from microstructure images to crack phase field distribution, thus reducing computational costs and improving prediction efficiency. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the carbon fiber composite matrix crack path prediction method of the present invention; Figure 2 This is a schematic diagram illustrating the selection of sequence parameters for the fracture phase field model in the carbon fiber composite matrix crack path prediction method of the present invention; Figure 3 This is a schematic diagram of batch modeling of microstructures in the carbon fiber composite matrix crack path prediction method of the present invention; Figure 4 This is a schematic diagram of the data processing process in the carbon fiber composite matrix crack path prediction method of the present invention; Figure 5 This is a schematic diagram of the deep learning framework constructed in the carbon fiber composite matrix crack path prediction method of the present invention; Figure 6 This diagram illustrates the physical driving loss guidance mechanism in the carbon fiber composite matrix crack path prediction method of this invention. Figure 7 This is a comparison chart of the prediction results between the carbon fiber composite matrix crack path prediction method of the present invention and existing methods. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0021] like Figure 1 The method shown is a matrix crack path prediction method for carbon fiber composites, the method includes: S1: Establish a fracture phase field model for the transverse matrix of carbon fiber composite material, including total phase field free energy, elastic strain energy and dynamic evolution equation; by introducing fracture mechanics and energy evolution mechanism, achieve physical consistency modeling of crack evolution process.

[0022] S2: Samples that conform to the real fiber distribution are generated in batches for training using a random generation method. Numerical simulation is performed based on the fracture phase field model, and the output is a sample dataset containing microstructure geometric features and corresponding crack phase field evolution. While ensuring the randomness of fiber distribution, crack evolution data consistent with real materials is obtained, so that the dataset has both diversity and physical authenticity, thereby improving the generalization ability and prediction accuracy of the crack path prediction model for different microstructures.

[0023] S3: Construct a crack path prediction model. The crack path prediction model establishes a nonlinear mapping relationship from microstructure geometric features to crack phase field distribution through a convolutional neural network architecture. The convolutional neural network model includes an encoder, a decoder, and a skip connection structure set between the encoder and decoder. An attention mechanism is introduced into the skip connection structure to realize the nonlinear mapping from microstructure geometric features to crack phase field distribution. The attention mechanism enhances crack-related features and suppresses interference from irrelevant regions, thereby improving the accuracy and detail representation ability of crack path prediction.

[0024] S4: Construct the total loss function, which includes a data-driven loss term and a physical constraint loss term. The data-driven loss term is used to characterize the difference between the crack phase field distribution output by the crack path prediction model and the crack phase field distribution data. The physical constraint loss term is used to penalize the non-zero probability of the prediction within the fiber region. On the basis of ensuring the consistency between the prediction results and the training data, a physical constraint mechanism is introduced to make the model more inclined to fit the crack in the marked region, and the crack path predicted by the model is more in line with physical reality.

[0025] S5: Train the crack path prediction model by inputting the sample dataset into the convolutional neural network architecture, optimizing the total loss function through the backpropagation algorithm, and outputting the trained crack path prediction model; enabling the crack path prediction model to learn the mapping relationship between microstructure and crack evolution.

[0026] S6: Obtain the microstructure image to be predicted, input the microstructure image into the crack path prediction model, and output the corresponding crack phase field distribution.

[0027] This invention generates microstructure input images and corresponding fracture crack fields in batches based on a fracture phase-field model, forming a dataset construction method for deep learning. By analyzing the fracture mechanism and generating corresponding data of microstructure and crack phase-field distributions based on the fracture phase-field model, the training data achieves physical consistency, thereby improving the accuracy of crack path prediction. An attention mechanism is introduced into the skip connection structure of the convolutional neural network architecture to enhance crack-related features and suppress interference from irrelevant regions, thus improving the accuracy of crack path prediction. By setting a physical constraint loss term and penalizing the prediction of non-zero cracks within the fiber region, the prediction results satisfy the physical constraint that cracks do not occur in the fiber region, thereby improving the physical consistency of the prediction results. By training the crack path prediction model and performing predictions based on the trained model, rapid prediction from microstructure images to crack phase-field distributions is achieved, thereby reducing computational costs and improving prediction efficiency.

[0028] Specifically, the methods for establishing the fracture phase-field model include: S11: Set sequence parameters Sum of parameters Describes the phase distribution state of the fracture phase field model, order parameters Order parameters are used to characterize the damage state of materials. Used to describe the phase distribution state of the fibrous phase and the matrix phase; where, Takes a value between 0 and 1, when When, it indicates that the material is in the crack phase, when When, it indicates that the material is in the solid phase; Indicates fibrous phase, Representing the matrix phase; the sequence parameters of the fracture phase field model are selected as follows: Figure 2 As shown.

[0029] The total phase field free energy F comprises two major terms: elastic strain energy and dissipated energy, as shown in the following formula: in, This represents the elastic strain energy within the solid domain Ω. This represents the dissipated energy Γ acting on the surface region of the crack. Γ represents the elastic strain energy density, d represents the order parameter used to characterize the damage state of the material, Ω represents the solid domain, and Γ represents the crack surface region. Represented as strain tensor; It is the displacement field vector; Fracture toughness is related to the material's inherent properties; S12: Calculate the density of elastic strain energy using the following formula: Elastic strain energy is decomposed into elastic energy resulting from tension on the material. Elastic properties resulting from material compression w(d) is a monotonically decreasing damage degradation function that reflects the strain energy degradation caused by crack evolution during the tensile process. When d=1, w(d)=0, and when d=0, w(d)=1.

[0030] S13: A regularized description of the crack surface, where the crack surface is approximately represented as a density function γ distributed within the solid domain Ω. The energy dissipated at the crack surface is then calculated. The density function γ is: in, Let be the gradient of the order parameter d, used to describe the rate of change of the crack field d in space. is the length dimension parameter used to determine the crack dispersion width of the order parameter d; 3 / 8 is the normalization coefficient used to ensure that the numerical model is consistent with the energy conservation of classical fracture mechanics. S14: The field dynamic evolution equation is: in, Historical variables are used to avoid crack closure; It represents the damage migration rate, which determines how fast the crack propagates; This is the time derivative operator, used to express the damage evolution rate.

[0031] Specifically, such as Figure 3 and Figure 4As shown, the random generation method is a fiber random generation method based on spatial exclusion, and the method includes: S21: Determine the material's dimensional parameters. Here, the fiber radius is set to 4 micrometers. Randomly generate the fiber center positions within a preset fiber region. Considering computation time (larger unit sizes result in slower computation and higher dataset construction costs) and model training difficulty (excessively large regions increase the difficulty of model fitting), a square region with a boundary range of 30x30 micrometers is determined while ensuring the randomness of fiber distribution. Set the fiber volume fraction, which determines the number of fibers per unit area; in this embodiment, the fiber volume fraction is 60%.

[0032] S22: Randomly generate the center coordinates of the first fiber within a square space. Set the minimum fiber spacing, which is based on the actual fiber distribution of the material and the gap between fibers. Preferably, the minimum gap between fibers under ideal material manufacturing conditions is determined based on the actual fiber distribution found in the SEM electron microscopy test results of the real material. Mark the generated fibers and surrounding areas as areas where fibers cannot be generated based on the minimum fiber spacing. S23: Continue generating fibers in the remaining area of ​​the fiber region until the preset fiber volume fraction is reached, then stop generating fibers to complete one microstructure unit image construction; the microstructure unit image is a high-dimensional tensor image with a resolution of 512×512, reflecting the geometric features of the microstructure, that is, the distribution of 30×30 micrometer fibers that may appear in the actual material structure.

[0033] Specifically, such as Figure 4 As shown, the process of constructing the sample dataset includes: applying preset loading and boundary conditions to each microstructure unit, performing numerical simulation based on the fracture phase field model, and outputting the corresponding crack phase field distribution data. Using finite element analysis software, high-fidelity numerical simulations are performed on 1000 microstructure units in the dataset based on the fracture phase field model in S1 to obtain crack field images. Samples conforming to the fiber distribution requirements of actual materials are submitted to open-source multiphysics simulation software for numerical analysis, and finally, a script program is used to obtain the corresponding crack field images. Based on the above generation method, 1000 microstructure units are generated for constructing the sample dataset. In subsequent steps, 800 microstructure units are used for model training, and 100 microstructure units are used as the validation set and test set of the model, respectively.

[0034] Specifically, the convolutional neural network architecture is a fully convolutional network structure, including an encoder, decoder, attention module, and physical constraint loss function. The network architecture uses a fully convolutional network structure; the encoder extracts multi-scale features from the microstructure image, and the decoder reconstructs these multi-scale features step-by-step to output the crack phase field distribution. Considering that the presence of fibers in the predicted crack field does not change, only the phase field order parameter d is predicted to simplify the model prediction task. Correspondingly, the input and target images also need to be processed, and the predicted order parameter d is post-processed to restore detailed information about the crack field. The image processing process is as follows: Figure 4 As shown. The framework of a convolutional neural network is as follows. Figure 5 As shown, the microstructure images in the dataset are used as input, and the encoder uses ResNet34 as the backbone network to extract hierarchical features. The initial convolutional layers are adjusted to have 7×7 kernels and a stride of 2 to adapt to the 5-channel input data. The encoder contains four stages of residual blocks, with the number of feature channels increasing progressively to {64, 128, 256, 512}, while the spatial resolution is downsampled to 1 / 32 of the original image. To capture multi-scale contextual information without sacrificing resolution, the bottleneck layer integrates a dilated spatial pyramid pooling module; this module uses parallel dilated convolutions with dilation rates of {1, 6, 12, 18}, enabling the network to simultaneously perceive local crack details and global structural constraints.

[0035] Specifically, the process of introducing an attention mechanism into the skip connection structure includes: Calculate attention coefficient ,set up This is a skip connection feature. For gating signals from coarse-grained levels, the attention coefficient The calculation formula is as follows: in, It is the ReLU activation function. For the Sigmoid function, , , This is a linear transformation achieved through 1×1 convolution. The bias term corresponds to the gated signal; the final skip connection feature is obtained through element-wise multiplication. This effectively filters out background noise. The decoder progressively upsamples through 2×2 transposed convolutions until the original resolution is restored. Skip connections adaptively weight the high-gradient feature responses at crack tips and microstructure interfaces, suppressing background noise interference. Specifically, the total loss function addresses the class imbalance between cracks and background and imposes physical constraints. The data-driven loss term measures the mean square error between the network's predicted values ​​and the finite element label values, and includes the binary cross-entropy loss. Mean square error loss and focus Tversky loss ; Total loss function Represented as: in, The physical constraint loss term is used to constrain the physical consistency of the prediction results. The physical driving loss penalizes the non-zero probability of the prediction in the fiber region, forcing the model to learn the physical rule of "fiber crack prohibition". At the same time, it combines the physical information that the initial crack is more likely to be generated in the transverse matrix region with small fiber gaps. , , , Represents network weight parameters; binary cross-entropy loss. Used to enhance focus on cracked pixels; mean square error loss Used to ensure the smoothness of the predicted probability field; Focus Tversky loss Represented as: in, , , These represent true positive, false negative, and false positive, respectively. , , This is represented as a hyperparameter.

[0036] Specifically, such as Figure 6 As shown, the physical constraint loss term Training methods include: S41. Extract all fiber coordinate information within the fiber region. Combining the distance between the centers of adjacent fibers and the angle between the line connecting the centers and the loading direction, mark the matrix region with smaller fiber gaps and a line connecting the fiber centers that is closer to the loading direction as the crack preferential growth region; specifically, calculate the distance between the centers of any two adjacent fibers. And according to the distance between the centers Calculate the fiber gap with fiber radius r ; Calculate the angle between the line connecting the centers of two adjacent fibers and the loading direction. When the fiber gap The included angle is not greater than a preset gap threshold. When the angle is not greater than the preset threshold, the matrix region between the two adjacent fibers is marked as the crack preferential growth region. S42. A non-zero probability penalty is applied to fibers predicted not to be in the crack-preferred growth region, guiding crack generation and propagation within this region. The weight coefficients of the physical constraint loss term employ a dynamic adjustment strategy, focusing on data fitting in the early stages of training and gradually increasing the physical constraint weights in the later stages to correct the physical consistency of the prediction results.

[0037] Specifically, the training methods for crack path prediction models include: S51. Divide the sample dataset into a training set and a validation set; S52. Input the training set into the crack path prediction model for iterative training; during training, update the network weight parameters of the crack path prediction model using the Adam optimizer; minimize the total loss function using the backpropagation algorithm; in this embodiment, the entire convolutional neural network is implemented based on the PyTorch framework and trained on an NVIDIA 5070 GPU. The initial learning rate is set to... The weight decays to .

[0038] S53. Monitor model performance on the validation set. When the validation set loss no longer decreases, adaptively adjust the network weight parameters to reduce the learning rate until the model converges and the optimal network weight parameters are saved. The training strategy uses the ReduceLROnPlateau scheduler. If the validation set loss does not decrease within 4 epochs, the learning rate is halved. The model training consists of 30 epochs with a batch size of 4. The first 10 epochs are for data fitting, the 11th to 20th epochs are for introducing physical constraints, and the 21st to 30th epochs are for joint optimization to ensure training stability and convergence.

[0039] The process of predicting crack phase field distribution after deploying the crack path prediction model is as follows: Obtain the microstructure image to be predicted, input it into the trained crack path prediction model, and output the corresponding crack phase field distribution prediction result through a single forward propagation. This allows for rapid assessment of the fracture path and damage degree of the material under specific loads, eliminating the need for time-consuming finite element iterative calculations. The predicted results are then compared with the results calculated using the corresponding finite element analysis method in S2. Figure 7 As shown, it can be found that the crack image predicted by the model proposed in this invention for the matrix fracture task is closer to the crack image result of the fracture phase field model than the prediction result of the U-Net model, and the crack path is more consistent with the real situation.

[0040] Specifically, a crack path prediction system for carbon fiber composite matrix, the system employing the aforementioned method, includes: The fracture phase field module is used to establish a fracture phase field model of the transverse matrix of carbon fiber composite materials, which includes the total phase field free energy, elastic strain energy, and dynamic evolution equations. The dataset construction module is used to generate samples that conform to the real fiber distribution in batches through random generation methods, and to perform numerical simulation based on the fracture phase field model, outputting a sample dataset containing microstructure geometric features and corresponding crack phase field evolution. The crack path prediction module is used to acquire the microstructure image to be predicted, input the microstructure image into the crack path prediction model, and output the corresponding crack phase field distribution. processor; A storage unit is used to store one or more programs that, when executed by one or more processors, enable the one or more processors to implement the aforementioned carbon fiber composite matrix crack path prediction method.

[0041] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0042] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in sequences other than those illustrated or described herein.

[0043] The present invention has been further described above with reference to specific embodiments. However, it should be understood that the specific description herein should not be construed as limiting the nature and scope of the present invention. Various modifications made to the above embodiments by those skilled in the art after reading this specification are all within the scope of protection of the present invention.

Claims

1. A method for predicting crack paths in a carbon fiber composite matrix, characterized in that, The method includes: S1: Establish a fracture phase field model for the transverse matrix of carbon fiber composite material, including total phase field free energy, elastic strain energy, and dynamic evolution equations; S2: Generate samples that conform to the real fiber distribution for training in batches using a random generation method, perform numerical simulation based on the fracture phase field model, and output a sample dataset containing microstructure geometric features and corresponding crack phase field evolution. S3: Construct a crack path prediction model, which establishes a nonlinear mapping relationship from microstructure geometric features to crack phase field distribution through a convolutional neural network architecture; the convolutional neural network model includes an encoder, a decoder, and a skip connection structure set between the encoder and the decoder; an attention mechanism is introduced into the skip connection structure; S4: Construct the total loss function, which includes a data-driven loss term and a physical constraint loss term; the data-driven loss term is used to characterize the difference between the crack phase field distribution output by the crack path prediction model and the crack phase field distribution data; The training method for the physical constraint loss term includes: S41. Extract all fiber coordinate information within the fiber region and calculate the center-to-center distance between any two adjacent fibers. And according to the distance between the centers Calculate the fiber gap with fiber radius r ; Calculate the angle between the line connecting the centers of two adjacent fibers and the loading direction. When the fiber gap The included angle is not greater than a preset gap threshold. When the angle is not greater than the preset threshold, the matrix region between the two adjacent fibers is marked as the crack preferential growth region. S42. Penalize the non-zero crack probability of pixels located within the fiber region and matrix pixels not located within the crack preferential growth region in the prediction results, and guide cracks to generate and propagate within the crack preferential growth region. S5: Train the crack path prediction model by inputting the sample dataset into the convolutional neural network architecture, optimizing the total loss function through the backpropagation algorithm, and outputting the trained crack path prediction model. S6: Obtain the microstructure image to be predicted, input the microstructure image into the crack path prediction model, and output the corresponding crack phase field distribution.

2. The method for predicting crack paths in a carbon fiber composite matrix according to claim 1, characterized in that, The method for establishing the fracture phase field model includes: S11: Set the order parameter d to characterize the damage state of the material and the order parameter c to describe the phase distribution state of the fibrous phase and the matrix phase; construct the total free energy function of the phase field that includes the elastic strain energy term and the crack surface dissipation energy term. S12: Calculate the density of elastic strain energy, which is decomposed into elastic energy caused by tension and compression of the material, and the degradation of strain energy during crack evolution is characterized by a monotonically decreasing damage degradation function. S13: The crack surface is regularized and the crack surface energy is approximately represented as a density function distributed in the solid domain. The gradient of the order parameter d is used to describe the rate of change of the crack field in space. At the same time, a normalization coefficient is introduced to control the crack dispersion width. S14: Establish the field dynamic evolution equation of the order parameter d, introduce historical variables to avoid crack closure, and control the crack propagation rate through damage migration rate.

3. The method for predicting crack paths in a carbon fiber composite matrix according to claim 1, characterized in that, The random generation method is a fiber random generation method based on spatial exclusion, and the method includes: S21: Determine the material's dimensional parameters, and randomly generate fiber center positions within the preset fiber region; set the fiber volume fraction, which is used to determine the number of fibers per unit area; S22: Set a minimum fiber spacing, which is set based on the gap between fibers under the actual fiber distribution of the material; mark the generated fibers and the surrounding area as areas where fibers cannot be generated based on the minimum fiber spacing; S23: Continue generating fibers in the remaining area of ​​the fiber region until the fiber volume fraction is reached.

4. The method for predicting crack paths in a carbon fiber composite matrix according to claim 1, characterized in that, The process of constructing the sample dataset includes: applying preset loading conditions and boundary conditions to each microstructure unit, performing numerical simulation based on the fracture phase field model, and outputting the corresponding crack phase field distribution data.

5. The method for predicting crack paths in a carbon fiber composite matrix according to claim 1, characterized in that, The encoder is used to extract multi-scale features from the microstructure image, and the decoder is used to reconstruct the multi-scale features step by step to output the crack phase field distribution.

6. The method for predicting crack paths in a carbon fiber composite matrix according to claim 1, characterized in that, The process of introducing an attention mechanism into the skip connection structure includes: Calculate attention coefficient ,set up This is a skip connection feature. The attention coefficients are gating signals from a coarse-grained level. The calculation formula is as follows: in, It is the ReLU activation function. For the Sigmoid function, , , This is a linear transformation achieved through 1×1 convolution. This is the bias term corresponding to the gated signal; the final skip connection feature is obtained through element-wise multiplication. get.

7. The method for predicting crack paths in a carbon fiber composite matrix according to claim 1, characterized in that, The data-driven loss term includes binary cross-entropy loss. Mean square error loss and focus Tversky loss ; The total loss function Represented as: in, This is the physical constraint loss term, used to constrain the physical consistency of the prediction results; , , , Represents network weight parameters; the binary cross-entropy loss Used to enhance focus on cracked pixels; the mean square error loss Used to ensure the smoothness of the predicted probability field; the focal Tversky loss Represented as: in, , , These represent true positive, false negative, and false positive, respectively. , , This is represented as a hyperparameter.

8. The method for predicting crack paths in a carbon fiber composite matrix according to claim 7, characterized in that, The training method for the crack path prediction model includes: S51. Divide the sample dataset into a training set and a validation set; S52. Input the training set into the crack path prediction model for iterative training; during the training process, update the network weight parameters of the crack path prediction model through the Adam optimizer; minimize the total loss function through the backpropagation algorithm; S53. Monitor the model performance on the validation set. When the loss on the validation set no longer decreases, adaptively adjust the network weight parameters to reduce the learning rate until the model converges and the optimal network weight parameters are saved.

9. A crack path prediction system for carbon fiber composite matrix, characterized in that, The system employs the method according to any one of claims 1-8, and the system comprises: The fracture phase field module is used to establish a fracture phase field model of the transverse matrix of carbon fiber composite materials, which includes the total phase field free energy, elastic strain energy, and dynamic evolution equations. The dataset construction module is used to generate samples that conform to the real fiber distribution in batches through a random generation method, and to perform numerical simulation based on the fracture phase field model, outputting a sample dataset containing microstructure geometric features and corresponding crack phase field evolution. The crack path prediction module is used to acquire the microstructure image to be predicted, input the microstructure image into the crack path prediction model, and output the corresponding crack phase field distribution. processor; A storage unit is used to store one or more programs that, when executed by one or more processors, enable the one or more processors to implement the carbon fiber composite matrix crack path prediction method according to any one of claims 1 to 8.

Images