Material microstructure evolution prediction method and system based on autoencoder

Abstract

The application provides a material microstructure evolution prediction method and system based on an autoencoder, and comprises the following steps: obtaining material microstructure image information; processing the material microstructure image information to obtain corresponding graph representation; constructing an autoencoder and performing unsupervised learning on the graph representation to obtain a reduced dimension feature vector in a material microstructure evolution process; inputting the reduced dimension feature vector into a recurrent neural network constitutive model for training to obtain an updated feature vector, and reconstructing the updated feature vector by using the autoencoder to obtain evolved material microstructure image information; taking the evolved material microstructure image information as input, and repeatedly executing the above steps to predict the evolution of a material microstructure strain field in a deformation process. The material microstructure image information is reduced to a low-dimensional feature vector by using the autoencoder, and the unsupervised learning reduces the dimension, thereby reducing the consumption of computing resources.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence material modeling and prediction, in particular, to a material microstructure evolution prediction method and system based on an autoencoder. In particular, it relates to a method for learning material microstructure evolution by an autoencoder to achieve efficient cross-scale modeling and prediction. BACKGROUND

[0002] Establishing a reasonable material constitutive model to accurately describe the deformation behavior of the material is the basis for ensuring the effectiveness and accuracy of numerical simulation. For a long time, the main problem of material constitutive modeling is that the material deformation behavior at the current time is affected by the history deformation path of the material. The main reason for this problem is that the microstructure of the material gradually evolves during the deformation process, and different deformation paths will lead to different material microstructure information at the current time, which will further affect the current material deformation behavior.

[0003] The classical phenomenological constitutive model is difficult to describe the complex evolution process of the material microstructure and the mechanical property behavior at the same time because it does not consider or cannot perfectly consider the influence of the material microstructure. Cross-scale methods such as the representative volume element method can effectively couple the material microstructure and the macroscopic deformation behavior, but the computational resource consumption is too large, and even high-performance computing centers are difficult to complete the calculation for the deformation process of the macroscopic component, making it difficult to be applied to engineering practice efficiently and at low cost.

[0004] Neural networks are an excellent data mining method in the field of machine learning, among which recurrent neural networks can effectively consider the influence of historical information on the model at the current time step, and have been used to replace traditional constitutive models in many studies. The constitutive model based on recurrent neural networks can use a large amount of stress-strain data of materials under different loading paths as a driving force to learn and describe the influence of the history deformation path on the material deformation behavior. For example, Professor Cao Jian of Northwestern Polytechnical University applied a gate recurrent neural network (GRU) to describe the deformation behavior of a representative volume element of an elastic-plastic material.

[0005] However, the current material constitutive model based on recurrent neural networks is still a phenomenological model, and the good performance of the model depends on the excellent fitting ability of the neural network "black box" and a large amount of training data, without truly considering the evolution of the material microstructure information during the material deformation process, and its interpretability and robustness are low. SUMMARY

[0006] In view of the defects in the prior art, the purpose of the present application is to provide a material microstructure evolution prediction method and system based on an autoencoder.

[0007] According to the material microstructure evolution prediction method and system based on the autoencoder provided by the present application, the method comprises:

[0008] Step S1: obtaining material microstructure image information;

[0009] Step S2: preprocessing the material microstructure image information to obtain corresponding graph representation;

[0010] Step S3: constructing an autoencoder and performing unsupervised learning on the graph representation to obtain a reduced dimension feature vector in the material microstructure evolution process;

[0011] Step S4: inputting the reduced dimension feature vector into a recurrent neural network constitutive model for training to obtain an updated feature vector, and using the autoencoder to reconstruct to obtain evolved material microstructure image information;

[0012] Step S5: taking the evolved material microstructure image information as input, repeating steps S3 to S4, thereby predicting the evolution of the material microstructure strain field in the deformation process.

[0013] Preferably, step S2 comprises:

[0014] Step S2.1: extracting integral points i in each grid cell in the material microstructure image information as nodes v of the graph i , to form a node set V;

[0015] Step S2.2: extracting the adjacency relationship of the element integral points of two common finite element grid nodes as edges e of the graph ij , to form an edge set E, wherein i and j represent integral point i and integral point j, respectively;

[0016] Step S2.3: extracting information on integral point i as weights w of the graph i , to form a weight set W;

[0017] Step S2.4: combining the node set V, the edge set E, and the weight set W to form an undirected weighted graph G(V, E, W).

[0018] Preferably, the autoencoder comprises an encoder and a decoder;

[0019] The encoder comprises an information passing layer, a global pooling layer, and a fully connected layer, for reducing the material microstructure image information to a reduced dimension feature vector;

[0020] The decoder comprises a fully connected layer and an information passing layer, for upscaling the reduced dimension feature vector and restoring the material microstructure image information.

[0021] Preferably, the information on the integration point i includes the position coordinates of the integration point i and any one or more of the phase component information, the stress tensor information and the strain tensor information of the integration point i.

[0022] Preferably, the recurrent neural network constitutive model includes two sub neural networks, and both are fully connected neural networks.

[0023] According to the material microstructure evolution prediction system based on an autoencoder provided by the application, the system comprises:

[0024] Module M1: obtaining material microstructure image information;

[0025] Module M2: pre-processing the material microstructure image information to obtain corresponding graph representation;

[0026] Module M3: constructing an autoencoder and performing unsupervised learning on the graph representation to obtain a reduced dimension feature vector in a material microstructure evolution process;

[0027] Module M4: inputting the reduced dimension feature vector into a recurrent neural network constitutive model for training to obtain an updated feature vector, and reconstructing the updated feature vector by using the autoencoder to obtain evolved material microstructure image information;

[0028] Module M5: taking the evolved material microstructure image information as input, repeatedly triggering modules M3 to M4, so as to predict the evolution of a material microstructure strain field in a deformation process.

[0029] Preferably, module M2 comprises:

[0030] Module M2.1: extracting integration points i in each grid cell in the material microstructure image information as nodes v of a graph i to form a node set V;

[0031] Module M2.2: extracting the adjacent relationship of the element integration points of two common finite element grid nodes as edges e of a graph ij to form an edge set E, wherein i and j represent the integration point i and the integration point j, respectively;

[0032] Module M2.3: extracting the information on the integration point i as weights w of a graph i to form a weight set W;

[0033] Module M2.4: combining the node set V, the edge set E and the weight set W to form an undirected weighted graph G(V, E, W).

[0034] Preferably, the autoencoder comprises an encoder and a decoder.

[0035] The encoder includes an information transmission layer, a global pooling layer, and a fully connected layer, which are used to reduce the dimensionality of the material microstructure image information into a feature vector.

[0036] The decoder includes a fully connected layer and an information transmission layer, which are used to increase the dimensionality of the reduced feature vector and restore the microstructure image information of the raw material.

[0037] Preferably, the information at integration point i includes the position coordinates of integration point i and any one or more of the phase composition information, stress tensor information and strain tensor information of integration point i.

[0038] Preferably, the recurrent neural network constitutive model includes two sub-neural networks, both of which are fully connected neural networks.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. This invention reduces the dimensionality of material microstructure image information into low-dimensional feature vectors through an autoencoder and applies them. This unsupervised learning dimensionality reduction reduces the consumption of computing resources.

[0041] 2. This invention embeds material microstructure image information into the material constitutive model, enabling the final recurrent neural network constitutive model to consider the coupling effect of material microstructure evolution and macroscopic deformation behavior, thereby improving the interpretability and robustness of the neural network constitutive model.

[0042] 3. The present invention uses an autoencoder and a decoder to reconstruct the microstructure image information of the material after updating the material feature vector, which can predict and observe the evolution of the material microstructure during the deformation process. Attached Figure Description

[0043] 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:

[0044] Figure 1 This is a schematic diagram of the workflow of the present invention.

[0045] Figure 2 This is a schematic diagram illustrating the process of constructing a graphical representation of the microstructure image information of materials in this invention.

[0046] Figure 3 This is a schematic diagram of the autoencoder architecture in this invention.

[0047] Figure 4 This is a schematic diagram of the constitutive model framework of the recurrent neural network in this invention.

[0048] Figure 5 This is a schematic diagram comparing the strain field predicted by the finite element model and the strain field predicted by the model of this invention. Detailed Implementation

[0049] 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 several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0050] This invention utilizes an autoencoder to perform unsupervised learning on the microstructure image information of materials, which can extract dimensionality-reduced feature vectors in the evolution process of material microstructure. Then, the corresponding feature vectors are used as internal variables to introduce a neural network material mechanics model, which can achieve efficient cross-scale mapping of material microstructure and mechanical properties, and solve the problems of poor reliability of traditional mechanical models and difficulty in efficiently connecting material microstructure.

[0051] According to the present invention, a method for predicting the microscopic evolution of materials based on an autoencoder is provided, such as... Figure 1 As shown, it includes:

[0052] Step S1: Obtain image information of the material's microstructure.

[0053] Step S2: Preprocess the microstructure image information of the material to obtain the corresponding graphical representation. That is, divide the material's microstructure into a finite element mesh and process it into a graphical representation usable by an autoencoder. Specifically, as follows... Figure 2 As shown, step S2 includes:

[0054] Step S2.1: Extract the integration point i from each grid cell in the material microstructure image information and use it as the node v of the graph. i This forms the node set V.

[0055] Step S2.2: Extract the adjacency relationship of the element integration points of two shared finite element mesh nodes as the edge e of the graph. ij , forming the edge set E, where i and j represent integration points i and j, respectively.

[0056] Step S2.3: Extract the information at integration point i as the weight w of the graph. i This constitutes the weight set W. The information at the integration point i includes the position coordinates of the integration point i and any one or more of the phase composition information, stress tensor information, and strain tensor information of the integration point i.

[0057] Step S2.4: Combine the node set V, edge set E, and weight set W to form an undirected weighted graph G(V,E,W).

[0058] Step S3: Construct an autoencoder and perform unsupervised learning on the graph representation to obtain a dimensionality-reduced feature vector in the material's microstructure evolution process. Specifically, the autoencoder includes an encoder and a decoder. The encoder includes an information transfer layer, a global pooling layer, and a fully connected layer, used to reduce the dimensionality of the material's microstructure image information into a feature vector. The decoder includes a fully connected layer and an information transfer layer, used to increase the dimensionality of the reduced feature vector and restore the material's microstructure image information. Wherein, as... Figure 3 As shown, the encoder comprises an information delivery layer, a global pooling layer, and two fully connected layers, while the decoder comprises a fully connected layer and an information delivery layer. The information delivery layer is a type of graph neural network layer, such as a graph isomorphic neural network layer (GIN). This dimensionality reduction of the feature vectors solves the problem of high computational resource consumption caused by directly using material microstructure images in some current neural network constitutive models and multi-scale simulation models, which results in high-dimensional data input.

[0059] One variation of this invention is that the finite element meshing method used is pixel meshing, and when pixel meshing is used, the information transmission layer is selected as an image convolutional neural network layer.

[0060] Step S4: Input the reduced feature vector into the recurrent neural network constitutive model for training to obtain the updated feature vector, and reconstruct it using an autoencoder to obtain the evolved material microstructure image information. The recurrent neural network constitutive model includes two sub-neural networks, both of which are fully connected neural networks. Specifically, as shown... Figure 4 As shown, the input to the recurrent neural network is the feature vector ζ corresponding to the microstructure image information of the material at the current time step. i The stress tensor σ at the current time step i and the strain tensor ε at the next time step i The output of the first sub-neural network is the feature vector ζ for the next time step. i+1 The output of the second sub-neural network is the stress tensor α for the next time step. i+1 By embedding material microstructure image information into the material constitutive model, the final recurrent neural network constitutive model can consider the coupling effect between the evolution of material microstructure and macroscopic deformation behavior. This solves the problem that current recurrent neural network constitutive models cannot consider material microstructure, and brings better interpretability and robustness to neural network constitutive models.

[0061] Step S5: Using the evolved material microstructure image information as input, repeat steps S3 to S4 to predict the evolution of the material microstructure strain field during deformation. The feature vector output from the recurrent neural network at the next time step is used to reconstruct the material microstructure image information for the next time step using the decoder in the autoencoder.

[0062] This invention utilizes an autoencoder to reduce the dimensionality of material microstructure image information into corresponding feature vectors, and embeds these vectors as internal variables into a recurrent neural network material mechanics model to fit the material constitutive model. It then observes and predicts the evolution of image information during deformation. Machine learning methods can be used to discover, mine, and learn the hidden mapping relationship between material microstructure image information and material deformation behavior, and can be combined with general-purpose finite element software to achieve efficient cross-scale prediction.

[0063] Furthermore, a material constitutive model based on an autoencoder is used to specifically describe the hypothetical representative volume element of the material. Finite element simulation is then performed in the commercial finite element software ABAQUS by applying periodic boundary conditions to the constructed representative volume element model. The invention is described in detail below:

[0064] First, the microstructure image information of the material is extracted and processed. Based on its finite element mesh, a graphical representation of this microstructure information is constructed. The weight information of the graph includes the position coordinates and strain tensor of each integration point. 100 frames of images during the material deformation process are extracted as the training dataset.

[0065] Then, an autoencoder framework suitable for image information of material microstructure is constructed and unsupervised learning is performed. The autoencoder structure is as follows: Figure 3 As shown in the figure. The information transmission layer uses a graph isomorphic neural network layer.

[0066] Finally, the recurrent neural network constitutive model of the material is trained, and its structure is as follows: Figure 4 As shown in the diagram, at each loop time step, the encoder part of the autoencoder is first used to reduce the dimensionality of the material microstructure image information, and the corresponding feature vectors are embedded as internal variables into the neural network for training. The decoder part of the autoencoder then reconstructs the updated feature vectors into the material microstructure image information for the next frame.

[0067] The feature vectors updated by the recurrent neural network constitutive model were extracted, and the strain field along the principal strain directions of the material's microstructure was reconstructed using a decoder. The prediction results of the autoencoder-based model in this invention were compared with those of the finite element simulation model, and the results are as follows: Figure 5 As shown, it can be seen that the present invention can effectively predict the evolution of the strain field of the material microstructure during the deformation process.

[0068] The present invention also provides a material micro-evolution prediction system based on an autoencoder. Those skilled in the art can implement the material micro-evolution prediction system based on an autoencoder by executing the steps of the material micro-evolution prediction method based on an autoencoder. That is, the material micro-evolution prediction method based on an autoencoder can be understood as a preferred embodiment of the material micro-evolution prediction system based on an autoencoder.

[0069] According to the present invention, a material micro-evolution prediction system based on an autoencoder includes:

[0070] Module M1: Acquires image information of the material's microstructure.

[0071] Module M2: Preprocesses the microstructure image information of the material to obtain the corresponding graph representation. Module M2 includes: Module M2.1: Extracts the integration point i from each grid cell in the microstructure image information of the material as a node v of the graph. i This forms a node set V. Module M2.2: Extracts the adjacency relationship of the element integration points of two shared finite element mesh nodes as edges e of the graph. ij This forms the edge set E, where i and j represent integration points i and j, respectively. Module M2.3: Extracts the information at integration point i as the weight w of the graph. i The weight set W is formed by combining the node set V, edge set E, and weight set W. The information at the integration point i includes the position coordinates of the integration point i and any one or more of the phase composition information, stress tensor information, and strain tensor information of the integration point i. Module M2.4: Combines the node set V, edge set E, and weight set W to form an undirected weighted graph G(V,E,W).

[0072] Module M3: Constructs an autoencoder and performs unsupervised learning on the graph representation to obtain a dimensionality-reduced feature vector during the material's microstructure evolution. The autoencoder includes an encoder and a decoder. The encoder includes an information transfer layer, a global pooling layer, and a fully connected layer, used to reduce the dimensionality of the material's microstructure image information into a feature vector. The decoder includes a fully connected layer and an information transfer layer, used to increase the dimensionality of the reduced feature vector and restore the material's microstructure image information.

[0073] Module M4: Inputs the reduced feature vector into the recurrent neural network constitutive model for training, obtains the updated feature vector, and uses an autoencoder to reconstruct the evolved material microstructure image information. The recurrent neural network constitutive model includes two sub-neural networks, both of which are fully connected neural networks.

[0074] Module M5: Takes the evolved material microstructure image information as input and repeatedly triggers modules M3 to M4 to predict the evolution of the material microstructure strain field during deformation.

[0075] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0076] 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 changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for predicting the microscopic evolution of materials based on an autoencoder, characterized in that, include: Step S1: Obtain image information of the material's microstructure; Step S2: Preprocess the microstructure image information of the material to obtain the corresponding image representation; Step S3: Construct an autoencoder and perform unsupervised learning on the graph representation to obtain the dimensionality-reduced feature vector in the material microstructure evolution process; Step S4: Input the reduced feature vector into the recurrent neural network constitutive model for training to obtain the updated feature vector, and use an autoencoder to reconstruct the evolved material microstructure image information. Step S5: Using the evolved material microstructure image information as input, repeat steps S3 to S4 to predict the evolution of the material microstructure strain field during deformation. Step S2 includes: Step S2.1: Integrate the points in each grid cell of the material microstructure image information. Extracted as nodes in the graph , forming a node set ; Step S2.2: Extract the adjacency relationship of the element integration points of two shared finite element mesh nodes as edges of the graph. , forming an edge set ,in , They represent the integration points respectively. Integral points ; Step S2.3: Integrate the points The information is extracted and used as the weights of the graph. , forming a weight set ; Step S2.4: Combine the node set Edge set and weight set This forms an undirected weighted graph. .

2. The method for predicting the microscopic evolution of materials based on an autoencoder according to claim 1, characterized in that, The autoencoder includes an encoder and a decoder; The encoder includes an information transmission layer, a global pooling layer, and a fully connected layer, which are used to reduce the dimensionality of the material microstructure image information into a feature vector. The decoder includes a fully connected layer and an information transmission layer, which are used to increase the dimensionality of the reduced feature vector and restore the microstructure image information of the raw material.

3. The material micro-evolution prediction method based on autoencoders according to claim 1, characterized in that, The integration point The information includes integration points. Position coordinates and integration points The information includes any one or more of the phase composition information, stress tensor information, and strain tensor information.

4. The method for predicting the microscopic evolution of materials based on an autoencoder according to claim 1, characterized in that, The recurrent neural network constitutive model includes two sub-neural networks, both of which are fully connected neural networks.

5. A material micro-evolution prediction system based on an autoencoder, characterized in that, include: Module M1: Acquires image information of the material's microstructure; Module M2: Preprocesses the microstructure image information of the material to obtain the corresponding graphic representation; Module M3: Constructs an autoencoder and performs unsupervised learning on the graph representation to obtain a dimension-reduced feature vector in the material microstructure evolution process; Module M4: Input the reduced feature vector into the recurrent neural network constitutive model for training to obtain the updated feature vector, and use an autoencoder to reconstruct the evolved material microstructure image information; Module M5: Takes the evolved material microstructure image information as input and repeatedly triggers modules M3 to M4 to predict the evolution of the material microstructure strain field during deformation. Module M2 includes: Module M2.1: Integrates the points in each grid cell of the material's microstructure image information. Extracted as nodes in the graph , forming a node set ; Module M2.2: Extracts the adjacency relationship of element integration points of two shared finite element mesh nodes as edges of the graph. , forming an edge set ,in , They represent the integration points respectively. Integral points ; Module M2.3: Integral points The information is extracted and used as the weights of the graph. , forming a weight set ; Module M2.4: Combines the aforementioned node set Edge set and weight set This forms an undirected weighted graph. .

6. The material micro-evolution prediction system based on an autoencoder according to claim 5, characterized in that, The autoencoder includes an encoder and a decoder; The encoder includes an information transmission layer, a global pooling layer, and a fully connected layer, which are used to reduce the dimensionality of the material microstructure image information into a feature vector. The decoder includes a fully connected layer and an information transmission layer, which are used to increase the dimensionality of the reduced feature vector and restore the microstructure image information of the raw material.

7. The material micro-evolution prediction system based on an autoencoder according to claim 6, characterized in that, The integration point The information includes integration points. Position coordinates and integration points The information includes any one or more of the phase composition information, stress tensor information, and strain tensor information.

8. The material micro-evolution prediction system based on an autoencoder according to claim 5, characterized in that, The recurrent neural network constitutive model includes two sub-neural networks, both of which are fully connected neural networks.

Images