Method for predicting mechanical properties of silicon carbide particle reinforced aluminum matrix composites

Abstract

The application provides a method for predicting the mechanical properties of a silicon carbide particle reinforced aluminum matrix composite material, comprising: constructing a structure-property correlation data set of the silicon carbide particle reinforced aluminum matrix composite material; constructing a deep learning model based on an encoder-decoder structure, the deep learning model taking a material structure picture as input, the encoder being used to obtain a feature encoding vector of the structure, the decoder being used to obtain a stress-strain evolution relationship according to the feature encoding vector, and the deep learning model outputting a stress-strain curve corresponding to the structure; dividing the structure-property correlation data set into a training set, a test set and a validation set, and training the deep learning model; inputting a newly generated to-be-predicted silicon carbide particle reinforced aluminum matrix composite material configuration into the trained deep learning model to obtain a corresponding stress-strain curve, and extracting mechanical properties therefrom. The application can realize accurate prediction of the stress-strain curve of the material during the stretching process.

Description

Technical Field

[0001] This invention relates to the field of composite materials technology, and more specifically, to a method for predicting the mechanical properties of silicon carbide particle-reinforced aluminum matrix composites. Background Technology

[0002] Silicon carbide particle-reinforced aluminum matrix composites are widely used in aerospace, transportation, and electronic packaging fields due to their excellent comprehensive properties, including high modulus, high strength, high wear resistance, and low coefficient of thermal expansion. The mechanical properties of silicon carbide particle-reinforced metal matrix composites are closely related to their composite configuration. Rapid reasoning and prediction of their structure-property relationships can effectively guide the design and optimization of composite configurations, thereby further leveraging their performance advantages and application value.

[0003] However, in silicon carbide particle-reinforced aluminum matrix composites, the content, shape, size, and distribution of particles directly determine the microstructure of the composite. Due to the random combination of numerous configurational features, the composite configuration space exhibits extremely vast complexity. Simultaneously, the addition of particles indirectly affects the matrix's own microstructure, thus triggering complex strengthening and toughening effects. Traditional experimental methods rely on repeated trial and error, while single numerical calculation methods depend on model and parameter selection, resulting in expensive and lengthy research processes with limited flexibility. These limitations become increasingly apparent in the face of complex composite systems, making it difficult to conduct in-depth and systematic research on the coupling relationship between numerous composite configurations and mechanical properties. Therefore, developing a new method capable of rapidly predicting the corresponding mechanical properties from a given microstructure of silicon carbide particle-reinforced aluminum matrix composites is crucial for achieving rapid derivation from new material structures to mechanical properties.

[0004] A search revealed Chinese invention patent application CN116959647A, which discloses a method for predicting the mechanical properties of micro / nano composite materials based on finite element analysis. The method includes: S1: Based on microscopic images of the composite material observed under a microscope, 5000 virtual images of various structures composed of spherical particles are first generated. These virtual images are automatically generated based on distribution patterns by setting basic particle parameters; S2: The virtual images of the microstructures of different composite materials are imported into finite element simulation analysis software, and then the equivalent mechanical property parameters of the material are extracted from the simulation results using the program; S3: The virtual images of the composite material's microstructure and the corresponding extracted equivalent mechanical property parameters are used as basic input data to train a neural network and develop a prediction method based on deep learning. Finally, the validated neural network, i.e., a CNN network, is used to predict the mechanical properties of the composite material in the prediction loop. In this patent, the input structure used is regular circular particles, which limits the range of generated data; this patent only uses a simple CNN network with a network structure of several simple convolutional layers stacked together, and its learning ability needs to be improved; this patent only predicts a single mechanical property and cannot reflect the entire process of material response during use. Summary of the Invention

[0005] In view of the deficiencies in the prior art, the purpose of this invention is to provide a method for predicting the mechanical properties of silicon carbide particle-reinforced aluminum matrix composites.

[0006] According to one aspect of the present invention, a method for predicting the mechanical properties of silicon carbide particle-reinforced aluminum matrix composites is provided, comprising:

[0007] Construct a structure-property correlation dataset for silicon carbide particle-reinforced aluminum matrix composites;

[0008] A deep learning model based on an encoder-decoder structure is constructed. The deep learning model takes an image of the material's microstructure as input. The encoder of the deep learning model is used to obtain the feature encoding vector of the microstructure. The decoder of the deep learning model obtains the stress-strain evolution relationship based on the feature encoding vector. The deep learning model outputs the stress-strain curve corresponding to the microstructure.

[0009] The structure-performance correlation dataset is divided into a training set, a test set, and a validation set, and the deep learning model is trained on the dataset.

[0010] The newly generated configuration of silicon carbide particle-reinforced aluminum matrix composite material to be predicted is input into the trained deep learning model to obtain the corresponding stress-strain curve, and the mechanical properties are extracted from it.

[0011] Furthermore, the structure-property correlation dataset for constructing silicon carbide particle-reinforced aluminum matrix composites includes:

[0012] Geometric modeling of silicon carbide particles;

[0013] During the assembly process of silicon carbide particles and aluminum matrix, the silicon carbide particles are spatially distributed without cross-intersection in the aluminum matrix, thus obtaining the RVE model of silicon carbide particle reinforced aluminum matrix composite material.

[0014] High-throughput finite element analysis was performed on the RVE model, and a constitutive model of silicon carbide particle-reinforced aluminum matrix composite was introduced, including an elastic-brittle fracture model of silicon carbide particles and an elastic-plastic deformation / ductile fracture model of aluminum matrix, as well as a cohesive model of the interface behavior between silicon carbide particles and aluminum matrix. Based on the RVE model and the constitutive model, uniaxial tensile deformation simulation was performed using the Abaqus explicit solver, thereby obtaining the stress-strain data corresponding to the RVE model of silicon carbide particle-reinforced aluminum matrix composite.

[0015] The digital images and stress-strain data corresponding to the RVE model are standardized to obtain a structure-property correlation dataset of silicon carbide particle-reinforced aluminum matrix composites.

[0016] Furthermore, the process of geometrically modeling the silicon carbide particles involves: using randomly generated convex polygons to geometrically model the silicon carbide particles, and using a unit circle boundary point offset method to generate the vertices of the convex polygons.

[0017] Furthermore, in the assembly process of silicon carbide particles and aluminum substrate, the silicon carbide particles are arranged in a spatially non-intersecting manner within the aluminum substrate, including: using a method of pre-generating proposed positions, using Poisson disk sampling to ensure that the distance between points does not exceed a set threshold, and using incremental rotation and slight scaling to adjust the particle position and size.

[0018] Furthermore, the digital images and stress-strain data corresponding to the RVE model are standardized, including: the reconstructed microstructure is represented as a 224*224 binarized image, with 0 representing the aluminum matrix and 1 representing silicon carbide particles; the stress sequence is processed using Min-Max normalization to scale the stress data to the range [0,10].

[0019] Furthermore, the deep learning model constructed based on the encoder-decoder structure is wherein: the encoder part adopts a ResNet-based convolutional neural network as the backbone, and the decoder part adopts a long short-term memory neural network LSTM as the backbone.

[0020] Furthermore, the encoder portion includes:

[0021] The preconvolution module consists of three 3×3 convolutional layers with kernel sizes of 3, and no pooling operation is performed after each convolution layer;

[0022] The residual blocks adopt the BootleNeck structure, with the number of stacks in the depth direction being 3, 3, 9, and 3 respectively, and each BootleNeck structure contains 3 convolutional layers;

[0023] The location fusion module consists of a depth-separable convolutional layer with the kernel size being the same as the input feature map size.

[0024] Furthermore, the deep learning model constructed based on the encoder-decoder structure is as follows: the initial external state h0 and initial internal state c0 of the Long Short-Term Memory (LSTM) neural network are each initialized by a fully connected layer, and the dimensions of the external and internal states are consistent with the dimension of the feature encoding; in the LSTM, the feature encoding combined with the position encoding at each step is used as the input for each step; the LSTM output vector h at each step is... t The strain is converted into the target strain value through a fully connected layer, with each fully connected layer sharing parameters.

[0025] Furthermore, the structure-performance correlation dataset is divided into a training set, a test set, and a validation set to train the deep learning model. Different optimizers and learning rates are used for the encoder and decoder parts, respectively. After multiple rounds of training iterations, a trained deep learning model is obtained.

[0026] Further, the structure-performance correlation dataset is divided into a training set, a test set, and a validation set, and the deep learning model is trained, including:

[0027] The structure-performance correlation dataset is divided into three parts: training set, validation set and test set in a ratio of 8:1:1.

[0028] The loss function used is the mean squared error loss function, which, for sequence data, is defined as:

[0029]

[0030] Among them, L i d is the loss value of the i-th sample. seq This refers to the length of the stress-strain sequence. and Let represent the values ​​at step j in the predicted stress sequence and the actual stress sequence, respectively;

[0031] The encoder part of the deep learning model uses the AdamW optimizer, and the decoder part uses the SGD optimizer.

[0032] A smooth transition during the initial training process of the network is achieved by using a linear learning rate warm-up method;

[0033] A fixed-step learning rate decay strategy is adopted, in which the learning rate is decayed at regular intervals.

[0034] Train the deep learning model for at least 100 iterations and save the model with the smallest error on the validation set during the training process.

[0035] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0036] This invention constructs a high-fidelity dataset of complex multi-dimensional material structure-property correlations to provide sufficient data that closely matches real experimental data, thereby ensuring the robustness and generalization of the deep learning model's predictive ability for different feature configurations. Furthermore, based on the data characteristics of silicon carbide particle-reinforced aluminum matrix composites, this invention constructs a deep learning model with an encoder-decoder structure, which can effectively predict the stress-strain curves of the material during tensile service, thus improving the speed of predicting the mechanical properties of silicon carbide particle-reinforced aluminum matrix composites. This invention can achieve accurate prediction of the stress-strain curves of materials during tensile processes, which is beneficial to improving the efficiency of material development. Attached Figure Description

[0037] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0038] Figure 1 This is a flowchart illustrating the prediction method in one embodiment of the present invention;

[0039] Figure 2 The experimental and simulation results of the stress-strain curve of the SiCp / Al composite material in one embodiment of the present invention are shown.

[0040] Figure 3 This is an overall architecture diagram of a deep learning model based on an encoder-decoder structure in one embodiment of the present invention;

[0041] Figure 4 This is a schematic diagram of the encoder portion in one embodiment of the present invention;

[0042] Figure 5 This is a graph showing the predicted stress-strain curves of silicon carbide particle-reinforced aluminum matrix composites in the embodiments of the present invention, wherein: (a) is the random structure, (b) is the strip structure, (c) is the cluster structure, and (d) is the network structure.

[0043] Figure 6This is a quantitative graph of the prediction results of the prediction model for mechanical properties in the embodiments of the present invention, wherein: (a) is Young's modulus, (b) is yield strength, (c) is tensile strength, and (d) is fracture toughness. Detailed Implementation

[0044] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0045] To address the problems of low efficiency and limited applicability of existing prediction methods, one embodiment of the present invention provides a method for predicting the mechanical properties of silicon carbide particle-reinforced aluminum matrix composites. Specifically, it is a method for predicting the stress-strain curve of silicon carbide particle-reinforced aluminum matrix composites during tensile testing based on a spatiotemporal network model, referring to... Figure 1 The method includes:

[0046] S1. Construct a structure-property correlation dataset for silicon carbide particle-reinforced aluminum matrix composites;

[0047] S2. Construct a deep learning model based on an encoder-decoder structure. The deep learning model takes the image of the material's microstructure as input. The encoder of the deep learning model is used to extract the deep semantic information of the two-dimensional reconstructed material microstructure to obtain the feature encoding vector of the microstructure. Then, the feature encoding vector is input into the recurrent neural network. The decoder of the deep learning model obtains the stress-strain evolution relationship based on the feature encoding vector. The deep learning model outputs the stress-strain curve corresponding to the microstructure.

[0048] S3. Divide the structure-performance correlation dataset into training set, test set and validation set, and train the deep learning model;

[0049] S4. Input the newly generated configuration of the silicon carbide particle-reinforced aluminum matrix composite material to be predicted into the trained deep learning model to obtain the corresponding stress-strain curve, and extract mechanical properties such as modulus, yield strength, tensile strength, and toughness from it.

[0050] The method provided in this invention constructs a structure-property correlation dataset for silicon carbide particle-reinforced aluminum matrix composites. By building a deep learning model based on an encoder-decoder structure, using images of the material's microstructure as input and stress-strain curves as output, the trained deep learning model accurately predicts the stress-strain curves of the material during tensile testing. Based on this, physical information such as modulus, strength, and toughness of the material can be obtained. The method in this invention will help reduce material development costs and accelerate the configuration screening process for silicon carbide particle-reinforced aluminum matrix composites, thereby meeting the urgent need for high-performance structural materials in industrial applications.

[0051] In step S1, the structure involved is a two-dimensional reconstructed material microstructure, representing the model established in the finite element method, used to reflect the core topological information of the material. To construct a structure-property correlation dataset for silicon carbide particle-reinforced aluminum matrix composites, in some embodiments, step S1 specifically includes:

[0052] S11. Perform geometric modeling on silicon carbide particles;

[0053] S12. During the assembly process of silicon carbide particles and aluminum matrix, the silicon carbide particles are distributed without spatial overlap in the aluminum matrix, thus obtaining the representative volume element (RVE) model of silicon carbide particle reinforced aluminum matrix composite material.

[0054] S13. High-throughput finite element analysis was performed on the RVE model of silicon carbide particle-reinforced aluminum matrix composites. A constitutive model of silicon carbide particle-reinforced aluminum matrix composites was introduced, including the elasto-brittle fracture model of silicon carbide particles and the elasto-plastic deformation / ductile fracture model of aluminum matrix, as well as the cohesive force model of the interface behavior between silicon carbide particles and aluminum matrix. Based on the above RVE model and constitutive model, uniaxial tensile deformation simulation was performed using the Abaqus explicit solver to obtain the stress-strain data corresponding to the RVE model of silicon carbide particle-reinforced aluminum matrix composites.

[0055] S14. Standardize the digital images and stress-strain data corresponding to the RVE model to obtain the structure-property correlation dataset of silicon carbide particle-reinforced aluminum matrix composites.

[0056] In some implementations, using polygons to geometrically abstract and model silicon carbide particles during the geometric modeling process can better reflect the shape characteristics of the silicon carbide particles. The silicon carbide particles are geometrically modeled using randomly generated convex polygons, and the vertices of the convex polygons are generated using a unit circle boundary point offset method, thereby achieving flexible control over the polygon shape. The specific steps are as follows:

[0057] S111, Generate angle sequence A 1:E={A1,A2,...,A e ,...,A E},in:

[0058]

[0059] S112, Generate a length sequence L 1:E ={L1,L2,...,L e ,...,L E},in:

[0060] L e ~N(1,σ2)

[0061] S113. Generate the vertex sequence (x, y). 1:E ={(x1,y1),(x2,y2),...,(x e ,y e ),...,(x E ,y E )},in:

[0062]

[0063] S114. Connect the vertex sequence sequentially to obtain polygonal particles;

[0064] The size of the particles is controlled by the diameter of the boundary circle, and the shape of the particles is controlled by shape control factors σ1 and σ2. That is, statistically speaking, the mean size of the particles is equal to the diameter of the boundary circle. The closer σ1 and σ2 are to 0, the closer the particles are to regular polygons. The larger σ1 and σ2 are, the more the particles deviate from regular polygons, that is, the higher the degree of shape variation.

[0065] In some implementations, during the assembly of silicon carbide particles with the aluminum substrate, the silicon carbide particles are spatially distributed without cross-intersection within the aluminum substrate. This includes: using a method of pre-generating proposed positions, first generating a set of proposed points, and using Poisson disk sampling to ensure that the distance between points does not exceed a set threshold, thereby reducing the probability of particle cross-intersection. If particle cross-intersection occurs, the particle position and size are adjusted by incremental rotation and slight scaling, effectively improving the generation efficiency of large-scale configurations.

[0066] In the above embodiments, a structure-property correlation dataset for silicon carbide particle-reinforced aluminum matrix composites is constructed through microstructure modeling and high-throughput finite element analysis. By performing random polygon modeling on the silicon carbide particles and generating RVE models using pre-proposed points, the realism and diversity of the reconstructed microstructure are ensured. The generated RVE models include different particle content, size, spatial distribution, and other configurational features. Compared to existing technologies that use regular circles as input microstructures, this invention comprehensively considers the shape, size, distribution, and content of particles, resulting in a wider range of generated data.

[0067] In step S13 above, models are generated in batches, and high-throughput finite element analysis (tensile mechanics simulation) is performed on the established RVE model of silicon carbide particle-reinforced aluminum matrix composite to extract reconstructed microstructure images and corresponding stress-strain curves as raw data. Specifically, a constitutive model including the elasto-brittle fracture of silicon carbide particles and the elasto-plastic deformation / ductile fracture of the aluminum matrix is ​​introduced into Abaqus, and a cohesion model is applied to describe the interfacial behavior between silicon carbide particles and the aluminum matrix. Based on the created 2D RVE model and the constitutive model of silicon carbide particle-reinforced aluminum matrix composite, uniaxial tensile mechanical deformation tests are conducted using the Abaqus explicit solver. Figure 2 The verification between the stress-strain experimental and simulation results of the SiCp / Al composite material shown indicates that the two are in good agreement, demonstrating the reliability of the data obtained through simulation.

[0068] Since the images exported by Abaqus are three-channel RGB images, the input images need to be standardized (preprocessed). In some implementations, the digitized images and stress-strain data corresponding to the RVE model are standardized, including: the reconstructed microstructure is represented as a 224*224 binarized image, with 0 representing the aluminum matrix and 1 representing silicon carbide particles; the stress sequence is normalized using Min-Max, scaling the stress data to the range [0, 10].

[0069] For example, Abaqus is used to extract the reconstructed microstructure model of silicon carbide particle-reinforced aluminum matrix composite material into a digital image. The image is then processed into grayscale and normalized. The image size is then scaled to 224*224 pixels, and binarized, with a pixel value of 1 representing the silicon carbide particle phase and a pixel value of 0 representing the aluminum matrix phase. Regarding the preprocessing of the stress sequence (stress-strain curve), since the time step is fixed during the Abaqus solution process, the generated stress-strain curves actually share strain data. Therefore, only the stress sequence needs to be predicted during training. Because the stress value range is large, for the micron-sized silicon carbide particle-reinforced 7A04Al composite material studied in this embodiment, the stress value is typically between 500-700 MPa. To improve the lateral comparability of the sequence data and avoid gradient problems during model training, the following formula is used for sequence standardization, and Min-Max standardization is used for normalization, scaling the stress data to the range [0, 10]. Let X∈R be the stress sequence in the global sample space. 10000×60 Where 10000 is the total sample size and 100 is the sequence length, denoted as X. i,j Let X be the j-th value in the i-th sample sequence. i,j The standardized value is:

[0070]

[0071] By standardizing the configuration images and stress sequences, more standardized and easily processed input data is provided for the subsequent training of the neural network, thereby improving the training efficiency and generalization ability of the model.

[0072] It should be noted that the same method can be used to process stress sequences for other types of silicon carbide particle-reinforced aluminum matrix composites to improve the training efficiency and generalization ability of the model.

[0073] In some implementations, a deep learning model based on an encoder-decoder structure is constructed. The encoder extracts configuration space feature information, and the decoder performs sequence modeling of stress-strain curves. The overall structure of the model is as follows: Figure 3As shown, the encoder uses a ResNet-based convolutional neural network as its backbone, while the decoder uses a Long Short-Term Memory (LSTM) neural network as its backbone. A deep convolutional neural network is used as the encoder to extract deep semantic information of the two-dimensional reconstructed material microstructure to obtain feature encoding vectors of the microstructure. These feature vectors are then input into a recurrent neural network to capture the stress-strain evolution relationship, ultimately outputting the stress-strain curve corresponding to the material microstructure. In other embodiments, the encoder and decoder can be adjusted and improved to better adapt to the prediction task.

[0074] In step S2, an encoder for extracting configurational spatial feature information is first constructed, followed by a decoder for sequential modeling of stress-strain curves. The encoder uses a convolutional neural network based on residual connections as the backbone, comprising a pre-convolutional module, residual blocks, and a location fusion module. The pre-convolutional module consists of 3×3 convolutional layers with three kernels, and no pooling operation is performed after each convolution, used to extract shallow features. Then, four stages of residual blocks are applied to extract high-level features, using a BootleNeck structure. The number of stacked residual blocks in the depth direction is 3, 3, 9, and 3 respectively, and each residual block is a BootleNet structure containing three convolutional layers, for a total of 54 convolutional layers. Finally, the location fusion module consists of a depth-separable convolutional layer with a kernel size the same as the input feature map, used to assign location weights to pixels in each channel, thus aggregating absolute location information. In addition, in the encoder, the activation function is LeakyReLU with a slope of 0.05 in the negative region, so as to allow negative values ​​to flow in the neural network, and all pooling layers used are average pooling layers to avoid the loss of effective information in the regression task.

[0075] The detailed structure of the encoder section is as follows: Figure 4 As shown, the input image first undergoes shallow feature extraction through the PreConv block, which contains 3 convolutional layers. Then, it goes through four stages of residual blocks (ResBlock) for high-level feature extraction. The stacking number of the four residual blocks along the depth direction is (3, 3, 9, 3) respectively. Each residual block contains 3 convolutional layers, for a total of 54 layers, which is the main part of the network. Finally, the feature maps obtained above are aggregated through the PositionFusion block to obtain the final feature encoding vector. This process can be regarded as 1 convolutional layer. The final overall network depth is 3 + 54 + 1 = 58 layers.

[0076] PreConv corresponds to the stem stage of the original ResNet, i.e., the pre-convolution stage, which is mainly used for shallow feature extraction and expanding the network width. It mainly contains three 3x3 convolutions, where the stride of the second 3x3 convolution is 2, in order to reduce the number of parameters and adapt to small-resolution input (small resolution here corresponds to a small RVE model size). Secondly, since this task is a regression task, the excessive downsampling factor in the first stage will cause some loss of small-scale information, so no pooling layer is added after the convolutional layer. In addition, the activation function is replaced by LeakyReLU, and the slope of the negative region is set to 0.05.

[0077] ResBlock is the core module of ResNet, and its main idea is residual learning. The ResBlock structure used in this embodiment of the invention adopts a bottleneck layer structure. This structure can effectively compress sparse information and improve computational efficiency. The number of channels is adjusted by 1x1 convolutions at both ends, and 3x3 convolutions with stride=1 or 2 are used in the middle. Batch normalization (BN) layers are added after the first two convolution layers, and an activation function is followed by the last convolution layer. D(x) is a "short-circuit connection" layer, which is used to adjust the shape of the input tensor x and establish residual connections between x and F(x). When the tensor shape of F(x) is consistent with that of x, D(x) = x. After F(x) has been adjusted in terms of channels and size, D(x) consists of conv1x1 + BN + AvgPool. conv1x1 is used to align the number of channels, and AvgPool is used to align the feature map size. In this embodiment of the invention, the combination of residual blocks is divided into four stages, and the number of residual blocks is allocated as (3,3,9,3). For each stage, channel adjustment and output size adjustment are only performed on the last residual block. That is, after each stage, the number of output channels is doubled and the size is halved.

[0078] The PositionFusion block is primarily used to fuse absolute positional information from tissue images. In some classification tasks, Global Average Pooling (GAP) is often used to compress the feature map output from the last residual block into a feature vector. The calculation process is as follows: Given an input feature map X∈R... M×N×C The output vector y∈R 1×C C represents the number of channels, X d Let X be all slices along the channel direction, then:

[0079]

[0080] That is, an average value is calculated for all pixels of the feature map of each channel of the feature map to obtain a feature vector of length C equal to the number of channels. GAP can effectively integrate global spatial information and is more robust to changes in target position in classification tasks. In other words, GAP increases the robustness of the classification model by ignoring absolute position information to a certain extent. However, considering the actual background of this task, the absolute position of SiC particle clusters with fixed relative information in the RVE model will affect the calculation results of the finite element method. Therefore, it is necessary to consider the absolute position information. Thus, this embodiment of the invention aggregates the absolute position information by assigning position weights to pixels of each channel. The specific calculation process is as follows: Given an input feature map X∈R M ×N×C Learnable parameters W∈R M×N×C ,b∈R 1×C The output vector y∈R 1×C ,but:

[0081]

[0082] In the specific implementation, a 7x7 depthwise separable convolution was used.

[0083] In some implementations, a deep learning model based on an encoder-decoder structure is constructed, wherein: the main part of the decoder is a Long Short-Term Memory (LSTM) neural network, and the initial external state h0 and initial internal state c0 of the LSTM are initialized by a fully connected layer, with the dimensions of the external and internal states being consistent with the dimension of the feature encoding; in the LSTM, the output y from the previous step is typically used. t-1 As the next input x t In this embodiment of the invention, since the output y at each step is a specific strain value, which is a scalar and contains too little information, feature encoding combined with the position encoding at each step is used as the input for each step; finally, the LSTM output vector h at each step t The strain is converted into the target strain value through a fully connected layer, with each fully connected layer sharing parameters.

[0084] In some implementations, the structure-performance correlation dataset is divided into a training set, a test set, and a validation set to train the deep learning model. Different optimizers and learning rates are used for the encoder and decoder parts, and after multiple rounds of training iterations, a trained deep learning model is obtained.

[0085] In some implementations, the structure-performance correlation dataset is divided into training, testing, and validation sets for training the deep learning model, including:

[0086] S31. Based on the amount of available data and the complexity of the dataset, the structure-performance related dataset is divided into three parts—training set, validation set, and test set—in an 8:1:1 ratio to evaluate the generalization ability of the deep learning model and prevent overfitting.

[0087] S32. The loss function adopted is the mean squared error loss function, which is defined as follows for sequence data:

[0088]

[0089] Among them, L i d is the loss value of the i-th sample. seq This refers to the length of the stress-strain sequence. and Let represent the values ​​at step j in the predicted stress sequence and the actual stress sequence, respectively;

[0090] S33. For the encoder part of the deep learning model, select the AdamW optimizer, and for the decoder part, select the SGD optimizer.

[0091] S34. Use the linear learning rate preheating method to achieve a smooth transition during the initial training process of the network;

[0092] S35. Use a fixed step learning rate decay strategy (StepLR) to decay the learning rate at regular intervals.

[0093] S36. Train the deep learning model for no less than 100 training iterations, and save the model with the smallest error on the validation set during the training process, i.e., the trained deep learning model.

[0094] The embodiments of this invention employ a residual network model and have made adjustments to methods such as pooling and activation functions for regression tasks, resulting in stronger learning capabilities compared to using a simple CNN network. Unlike existing technologies that only predict single mechanical properties of materials, the embodiments of this invention use a method combining convolutional neural networks and long short-term memory neural networks to output the stress-strain curves of materials during service, thereby reflecting a series of comprehensive mechanical properties of the materials.

[0095] In some implementations, the newly generated image of the silicon carbide particle-reinforced aluminum matrix composite structure to be predicted is input into a trained deep learning model, and the corresponding stress-strain curve is obtained through prediction, such as... Figure 5 The image shows the model's prediction performance on the stress-strain curves of several typical configurations (free-form, strip, cluster, and network configurations). Then, mechanical properties such as modulus, yield strength, tensile strength, and toughness are extracted from these curves and analyzed using R... 2The predictive performance of the model is measured by R and the mean absolute percentage error (MAPE), where R... 2 The calculation formula is:

[0096]

[0097] Where n is the number of samples in the test set, y i It is the actual mechanical property. It predicts mechanical properties. It is the mean of the true mechanical properties in the sample, and a higher R... 2 This indicates that the model can better explain the variation in the dependent variable, while a lower R-squared value... 2 This indicates that the model fits poorly. The formula for calculating MAPE is:

[0098]

[0099] MAPE can intuitively represent the degree to which predicted values ​​deviate from true values; the closer the value is to 0, the better the model's performance. For example... Figure 6 The figure shows the quantitative graph of the model's prediction effect on several specified mechanical properties. It can be seen that the actual value and the predicted value are basically distributed around y=x, indicating that the model has a good prediction effect on several mechanical properties.

[0100] The embodiments of the present invention generate a large number of high-fidelity composite configuration-mechanical property correlation datasets through microstructure reconstruction modeling and high-throughput finite element simulation. Then, a deep learning model based on an encoder-decoder structure is constructed, which can directly output the stress-strain curve of the material under tensile process. Thus, it is possible to accurately and efficiently predict the comprehensive mechanical properties of silicon carbide particle-reinforced aluminum matrix composites, providing new research methods and ideas for the design and manufacture of high-performance metal matrix composites, and has broad application prospects.

[0101] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention. The above preferred features can be used in any combination without conflict.

Claims

1. A method for predicting the mechanical properties of silicon carbide particle-reinforced aluminum matrix composites, characterized in that, include: Construct a structure-property correlation dataset for silicon carbide particle-reinforced aluminum matrix composites; A deep learning model based on an encoder-decoder structure is constructed. The deep learning model takes an image of the material's microstructure as input. The encoder of the deep learning model is used to obtain the feature encoding vector of the microstructure. The decoder of the deep learning model obtains the stress-strain evolution relationship based on the feature encoding vector. The deep learning model outputs the stress-strain curve corresponding to the microstructure. The structure-performance correlation dataset is divided into a training set, a test set, and a validation set, and the deep learning model is trained on the dataset. The newly generated configuration of the silicon carbide particle-reinforced aluminum matrix composite material to be predicted is input into the trained deep learning model to obtain the corresponding stress-strain curves, and the mechanical properties are extracted from them; where: The structure-property correlation dataset for constructing silicon carbide particle-reinforced aluminum matrix composites includes: Geometric modeling of silicon carbide particles; During the assembly process of silicon carbide particles and aluminum matrix, the silicon carbide particles are spatially distributed without cross-intersection in the aluminum matrix, thus obtaining the RVE model of silicon carbide particle reinforced aluminum matrix composite material. High-throughput finite element analysis was performed on the RVE model, and a constitutive model of silicon carbide particle-reinforced aluminum matrix composite was introduced, including an elastic-brittle fracture model of silicon carbide particles and an elastic-plastic deformation / ductile fracture model of aluminum matrix, as well as a cohesive model of the interface behavior between silicon carbide particles and aluminum matrix. Based on the RVE model and the constitutive model, uniaxial tensile deformation simulation was performed using the Abaqus explicit solver, thereby obtaining the stress-strain data corresponding to the RVE model of silicon carbide particle-reinforced aluminum matrix composite. The digital images and stress-strain data corresponding to the RVE model are standardized to obtain a structure-property correlation dataset of silicon carbide particle-reinforced aluminum matrix composites. The deep learning model constructed is based on an encoder-decoder structure, wherein the encoder part adopts a ResNet-based convolutional neural network as the backbone, and the decoder part adopts a Long Short-Term Memory (LSTM) neural network as the backbone.

2. The method according to claim 1, characterized in that, The process of geometrically modeling silicon carbide particles involves: using randomly generated convex polygons to model the silicon carbide particles, and generating the vertices of the convex polygons using a unit circle boundary point offset method.

3. The method according to claim 1, characterized in that, During the assembly process of silicon carbide particles and aluminum substrate, the silicon carbide particles are arranged in a spatially non-intersecting manner within the aluminum substrate. This includes: using a method of pre-generating proposed positions; using Poisson disk sampling to ensure that the distance between points does not exceed a set threshold; and using incremental rotation and slight scaling to adjust the particle position and size.

4. The method according to claim 1, characterized in that, The digital images and stress-strain data corresponding to the RVE model are standardized, including: the reconstructed microstructure is represented as 224. A 224-bit binary image was generated, with 0 representing the aluminum matrix and 1 representing the silicon carbide particles. The stress sequence was processed using Min-Max normalization to scale the stress data to the range [0, 10].

5. The method according to claim 1, characterized in that, The encoder portion includes: The preconvolution module consists of three 3×3 convolutional layers with kernel sizes of 3, and no pooling operation is performed after each convolution layer; The residual blocks adopt the BootleNeck structure, with the number of stacks in the depth direction being 3, 3, 9, and 3 respectively, and each BootleNeck structure contains 3 convolutional layers; The location fusion module consists of a depth-separable convolutional layer with the kernel size being the same as the input feature map size.

6. The method according to claim 1, characterized in that, The construction of the deep learning model based on the encoder-decoder structure includes: the initial external state of the Long Short-Term Memory (LSTM) neural network. h 0 and initial internal state c 0 Each state is initialized by a fully connected layer, with the dimensions of the external and internal states matching the dimension of the feature encoding. In the LSTM, the feature encoding combined with the positional encoding at each step serves as the input for each step. The LSTM output vector at each step... h t The strain is converted into the target strain value through a fully connected layer, with each fully connected layer sharing parameters.

7. The method according to claim 1, characterized in that, The structure-performance correlation dataset is divided into a training set, a test set, and a validation set. The deep learning model is trained by using different optimizers and learning rates for the encoder and decoder parts, and after multiple rounds of training iterations, a trained deep learning model is obtained.

8. The method according to claim 1, characterized in that, The structure-performance correlation dataset is divided into a training set, a test set, and a validation set. The deep learning model is then trained, including: The structure-performance correlation dataset is divided into three parts: training set, validation set and test set in a ratio of 8:1:

1. The loss function used is the mean squared error loss function, which, for sequence data, is defined as: ； in, It is the first i The loss value for each sample. This refers to the length of the stress-strain sequence. and Representing the predicted stress sequence and the actual stress sequence respectively. j The value of the step; The encoder part of the deep learning model uses the AdamW optimizer, and the decoder part uses the SGD optimizer. A smooth transition during the initial training process of the network is achieved by using a linear learning rate warm-up method; A fixed-step learning rate decay strategy is adopted, in which the learning rate is decayed at regular intervals. Train the deep learning model for at least 100 iterations and save the model with the smallest error on the validation set during the training process.

Images