GCN-LSTM-based dual-phase titanium alloy grain level stress prediction method

Abstract

The invention relates to the field of computer algorithms and material engineering, in particular to a two-phase titanium alloy grain level stress prediction method based on GCN-LSTM. Firstly, a GCN module is used for extracting high-dimensional spatial features from crystal grain nodes and connection relations thereof; then, an LSTM module is adopted to carry out time sequence modeling on grain stress data which are preprocessed and spliced into a time sequence, and the hidden state of the last time step of the sequence serves as a dynamic feature; and then fusing the features of the two parts element by element, and mapping the fused features into a grain-level stress prediction value through a full connection layer. The stress data predicted for microstructures of different texture types are reversely reconstructed into a three-dimensional visual image which is highly consistent with a crystal plasticity finite element calculation result, the method has a remarkable advantage in calculation speed, and more accurate local information is provided at the same time. The method provides efficient and accurate technical support for microstructure design, alloy performance and preparation process optimization and structural safety evaluation of materials such as titanium alloy.

Description

Technical Field

[0001] The present invention relates to the fields of computer algorithms and materials engineering, and specifically to a method for predicting grain-level stress of duplex titanium alloy based on GCN-LSTM, which is applicable to quickly and accurately extracting information on microscopic stress and strain distribution, assisting in the design of material microstructure and properties, and analyzing the structural safety. Background Art

[0002] Duplex titanium alloys are widely used in the fields of aviation, aerospace, medical, and automotive due to their light weight, high strength, and excellent corrosion resistance. However, the internal grain distribution is complex, and the stress changes dynamically with time during the loading process. When using crystal plasticity finite element methods and experimental techniques for grain-level stress prediction, due to the large amount of crystal plasticity simulation calculations, the cumbersome data preprocessing process, insufficient accuracy, and incomplete capture of dynamic responses, it is often difficult to balance spatial structure and time dynamics. In addition, the extremely large amount of calculations requires extremely high computing resources, and the long calculation cycle also limits the practical application of traditional methods. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for predicting grain-level stress of duplex titanium alloy based on GCN-LSTM, which can efficiently and automatically extract microscopic structure and dynamic information, and achieve fast and accurate grain-level stress prediction.

[0004] The technical solution of the present invention is as follows: A method for predicting grain-level stress of duplex titanium alloy based on GCN-LSTM, which combines the tissue information obtained by computational simulation and experiments for prediction, and includes the following steps: (1) Data acquisition: Read the microscopic structure data of duplex titanium alloy, including the node features and edge information of the grain map and the incremental data related to the grains. This part of the data is obtained through modeling with DREAM.3D software or through the EBSD experimental method; (2) Graph feature extraction: Use the graph convolutional neural network (GCN) module to perform forward propagation on the graph structure data to extract high-dimensional spatial features, including an initial convolutional layer and at least one hidden convolutional layer; (3) Temporal feature extraction: Use the long short-term memory network (LSTM) module to perform forward propagation on the temporal tensor composed of the incremental data to extract the dynamic features of the last time step of the sequence; (4) Feature fusion and prediction: Add the features obtained in steps (2) and (3) element by element, and map them to the grain-level stress prediction value through a fully connected layer; (5) Reverse reconstruction: Reverse reconstruct the prediction value into a three-dimensional stress distribution map, which is basically consistent with the grid stress distribution map based on finite elements; (6) Model training and evaluation: The mean squared error (MSE) is used as the loss function. The model parameters are updated through backpropagation using the Adam optimizer, and the model performance is evaluated using the MAE, MARE, and R² metrics. The results are also presented through scatter plots and training loss curves.

[0005] In the described method for predicting the grain-level stress of a duplex titanium alloy based on GCN-LSTM, the GCN module further includes: An initial convolutional layer for mapping the input features to a preset hidden dimension; At least one set of hidden convolutional layers for gradually enhancing the node feature representation; A final convolutional layer for outputting graph features with the same dimension as the output of the LSTM module.

[0006] In the described method for predicting the grain-level stress of a duplex titanium alloy based on GCN-LSTM, the LSTM module uses fixed-time-step sampling to construct an input sequence, and the output at the last time step of the sequence is used as the temporal feature.

[0007] In the described method for predicting the grain-level stress of a duplex titanium alloy based on GCN-LSTM, the model training uses the Adam optimizer, and the stable convergence of the model parameters is achieved by adjusting the learning rate and the number of training epochs.

[0008] The described method for predicting the grain-level stress of a duplex titanium alloy based on GCN-LSTM also includes preprocessing and normalization of the microstructure diagram data and incremental data, as well as data storage and loading processing using the pickle, dictionary, or CSV file format.

[0009] The described method for predicting the grain-level stress of a duplex titanium alloy based on GCN-LSTM is used for predicting the mechanical properties of the microstructural alloy and evaluating the evolution of the stress distribution in the microstructure with deformation, including: Quantitatively describing the prediction error using the MAE, MARE, and R² metrics; Intuitively showing the relationship between the true stress and the predicted stress and the loss change during the training process using scatter plots, kernel density estimation plots, and line plots.

[0010] The design concept of the present invention is: The present invention adopts a method based on a GCN-LSTM fusion model, combines the advantages of a graph convolutional neural network (GCN) and a long short-term memory network (LSTM), and automatically extracts the grain spatial structure and deformation dynamic characteristics through the joint processing of the microstructure diagram data and incremental time series data of different microstructures of the alloy. First, grid-level stress data of 26 time slices are obtained through multi-scale finite element simulation, and then these data are converted into grain-level time series data through automated preprocessing to form the model input. The GCN module extracts high-dimensional spatial features from grain nodes and their edge information, and the LSTM module performs time series modeling on the pre-selected time series data to capture the dynamic evolution law during the loading process. Then, the two parts of the features are fused element by element and the grain-level stress prediction value is output through a fully connected layer. At the same time, the present method uses the reverse reconstruction technology to restore the prediction result to a three-dimensional visual stress distribution map, which is highly consistent with the grid-based calculation result of the traditional finite element method.

[0011] The advantages and beneficial effects of the present invention are as follows: 1. The stress data predicted by the present invention for different texture types of microstructures can not only be reversely reconstructed into a three-dimensional visualization image, which is highly consistent with the calculation result of crystal plasticity finite element method, but also has a significant advantage in calculation speed. For the same set of microstructures, the traditional open-source crystal plasticity finite element calculation platform DAMASK (used to study the mechanical response of materials at multiple scales, can simulate multi-physical phenomena (such as crystal plasticity, heat, damage, etc.) in the range from single crystal scale to structural component scale, and realize multi-scale and multi-field coupled crystal plasticity analysis) takes about 100 minutes to calculate, while the present method only takes about 5 minutes to train 90 microstructures and less than 30 seconds to test 10 microstructures. Compared with the traditional DAMASK calculation, the present method has a significant acceleration advantage in data training and prediction, and can obtain finer and more accurate local stress information.

[0012] 2. The present invention provides efficient and accurate technical support for the microstructure design, alloy performance and preparation process optimization of materials such as titanium alloys, as well as the structural safety assessment.

[0013] (1) For the microstructure design and optimization, the present invention can play a role in the following three aspects: High-resolution stress field reconstruction: The three-dimensional stress distribution map obtained through reverse reconstruction can truly reflect the stress concentration areas and distribution laws among grains inside the material. Design engineers can use this information to analyze the influence of grain size, shape, grain boundary distribution and phase ratio on the local stress distribution, so as to optimize the acute angle or contact mode of grain boundaries in the microstructure design and avoid the formation of stress concentration areas.

[0014] Optimize crystal orientation and phase ratio: By statistically analyzing the simulation data of different microstructures, the local mechanical responses under different crystal orientations and different α / β phase ratios can be explored. Using this method, a quantitative relationship between grain orientation and stress distribution can be quickly established to assist in determining the optimal heat treatment process and processing method, aiming to optimize the microstructure and improve the overall mechanical properties.

[0015] Rapid design iteration: Since the time required for training and prediction of this model is extremely short (only about 5 minutes for training 90 microstructures and less than 30 seconds for testing 10 microstructures), designers can conduct simulation tests on multiple microstructure schemes in a short time, obtain local stress distribution data in real time, and achieve rapid iteration and optimization of microstructure design.

[0016] (2) For the optimization of alloy properties and preparation processes, the present invention can also play a role in the following three aspects: Enhance the prediction of material mechanical properties: High-precision prediction of grain-level mechanical responses can not only identify local high-stress regions but also reveal the overall stress evolution law. By comparing the differences in microstress distributions under different preparation processes (such as casting, forging, and heat treatment processes), the alloy composition ratio and process parameters can be optimized, thereby improving fatigue resistance, impact resistance, and fracture resistance.

[0017] Real-time feedback and process control: The rapid prediction ability enables this method to be embedded in the production process monitoring system to achieve online or quasi-online evaluation. When deviations occur in the actual production process, the local stress data quickly feedback by the prediction model can be used to adjust process parameters (such as temperature, cooling rate, strain rate, etc.), thereby ensuring that all important indicators in production meet the design requirements.

[0018] Multi-scale coupling verification: Combining the prediction results of traditional finite element calculations and this method can verify the effectiveness of process parameter optimization. It not only improves the overall deformation performance macroscopically but also accurately restores the local stress state at the micro level, providing a scientific basis for process adjustment.

[0019] (3) In terms of structural safety assessment, the present invention can play a role in the following three aspects: Predict local failure potential: Through high-precision local stress prediction and three-dimensional stress reconstruction diagrams, regions where stress concentration, fatigue crack initiation, or local yielding may occur can be identified in a timely manner, and preventive design can be carried out on the weak links in the structure.

[0020] Optimized Design and Safety Margin: After adopting this method, engineers can predict the stress distribution of structural components under service conditions at the early stage, providing improvement suggestions for structural design. By optimizing the microstructure and distribution of materials, the safety margin of the overall structure can be improved, and virtual fatigue tests and life predictions can be carried out through digital means, reducing the actual test cost.

[0021] Integrated into a Complete Digital Twin System: The highly restored three-dimensional visual stress map obtained by this method can serve as important input data for the digital twin model, realizing the comparison and correction with sensor measurement data, thereby monitoring and warning the structural health status in real time, and improving the accuracy and timeliness of safety assessment. Description of the Drawings

[0022] Figure 1 is the structural schematic diagram of the present invention.

[0023] Figure 2 is the scatter plot of the true value - predicted value of the test set of the present invention. Among them, (a) Training set: lnc_1~3, Test set: lnc_26, Coefficient of determination R²: 0.839, Mean Absolute Error MAE: 0.32, Mean Absolute Relative Error MeanARE: 109%; (b) Training set: lnc_1~10, Test set: lnc_26, Coefficient of determination R²: 0.930, Mean Absolute Error MAE: 0.21, Mean Absolute Relative Error MeanARE: 45%; (c) Training set: lnc_1~25, Test set: lnc_26, Coefficient of determination R²: 0.976, Mean Absolute Error MAE: 0.13, Mean Absolute Relative Error MeanARE: 30%; (d) Training set: lnc_1~3, Test set: lnc_4, Coefficient of determination R²: 0.988, Mean Absolute Error MAE: 0.06, Mean Absolute Relative Error MeanARE: 14%; (e) Training set: lnc_1~10, Test set: lnc_11, Coefficient of determination R²: 0.976, Mean Absolute Error MAE: 0.12, Mean Absolute Relative Error MeanARE: 30%; (f) Training set: lnc_1~25, Test set: lnc_26, Coefficient of determination R²: 0.992, Mean Absolute Error MAE: 0.08, Mean Absolute Relative Error MeanARE: 18%; In the figure, the abscissa True stress represents the true stress (MPa), and the ordinate Predicted stress represents the predicted stress (MPa).

[0024] Figure 3 is the reverse reconstruction diagram of the grain stress distribution of the present invention. Among them, (a) is the two-phase distribution diagram, (b) is the finite element simulation stress distribution diagram, and (c) is the reverse reconstruction stress distribution diagram of the GCN-LSTM model.

[0025] Figure 4 It is a comparison chart of the time used by the method of the present invention and the crystal plasticity finite element method for calculating the same data set. Detailed implementation manners

[0026] In the specific implementation process, the present invention first uses the GCN module to extract high-dimensional spatial features from the grain nodes and their connection relationships; then uses the LSTM module to perform time series modeling on the grain stress data that has been preprocessed and spliced into a time series, and uses the hidden state of the last time step of the sequence as the dynamic feature; subsequently, the two parts of the features are fused element by element, and mapped to the grain-level stress prediction value through a fully connected layer.

[0027] Next, the present invention will be further elaborated in detail through examples and drawings. Example

[0028] In this example, a method for predicting the grain-level stress of a duplex titanium alloy based on GCN-LSTM includes the following steps: Step 1. Data acquisition and preprocessing 1. Generation of finite element simulation data Use DAMASK to perform multi-scale finite element simulation on the duplex titanium alloy under dynamic loading. The specific parameters are as follows: (1) Elastic stage: 40 time steps, each step t = 10; (2) Plastic deformation stage: 60 time steps, each step t = 60, total deformation amount 7%; (3) Set the output frequency to 4 to obtain 26 slice data, named Inc_1 to Inc_26.

[0029] 2. Conversion of mesh data to grain data To meet the input requirements of the graph neural network, the mesh-level data is converted to grain-level data through the following steps: (1) Automatic acquisition and reading: Deploy a batch processing program to automatically traverse the folder storing multiple groups of microstructure data, and read the text file containing the grain ID (each grain is assigned a unique ID) and phase information.

[0030] Read the microstructure data of the duplex titanium alloy, including the node features and edge information constituting the grain graph and the incremental data related to the grains. This part of the data can be obtained by modeling through DREAM.3D software or by experimental methods such as electron backscatter diffraction (EBSD); (2) Information extraction and screening: Use data parsing and segmentation algorithms to extract the grain ID and phase data from each text file to form a complete grain information data set.

[0031] (3) Stress matching and calculation: Match the stress data of the grids within each time slice, and calculate the average stress value of each grain based on the grain ID; (4) Data storage: Organize the average stress values of each grain under 26 slices into a one-dimensional array, preprocess and normalize the microstructure diagram data and incremental data, and store them as a dictionary or CSV file for subsequent model calls.

[0032] Step 2: Model construction and data transfer 1. GCN module As Figure 1 shown, load the initial grain features (such as a 5D vector: Euler angles (α), Euler angles (β), Euler angles (Γ), grain size, and phase type, etc.) and the graph structure (edge information) from the preprocessed data; Map the input to 64 dimensions through the first convolutional layer (conv1) and process it through the activation function (ReLU); Subsequently, enhance the spatial interaction features between grains through the second convolutional layer (conv2) to ensure that the output dimension is consistent with the temporal features.

[0033] Use the graph convolutional neural network (GCN) module to perform forward propagation on the graph structure data and extract high-dimensional spatial features, including an initial convolutional layer and at least one hidden convolutional layer; the initial convolutional layer is used to map the input features to a preset hidden dimension, and at least one group of hidden convolutional layers gradually enhance the node feature expression. The final convolutional layer outputs spatial features (output feature 1) with the same dimension as the output of the LSTM module.

[0034] 2. LSTM module Construct a time series tensor according to the sampling strategy (such as sampling every 4 slices and selecting 9 slices); Feed this time series into the LSTM network, adopt the batch-first mode of the LSTM module, and obtain the hidden state of the last time step of the sequence as the dynamic feature.

[0035] Use the long short-term memory network (LSTM) module to perform forward propagation on the time series tensor composed of incremental data and extract the dynamic feature of the last time step of the sequence (output feature 2); the LSTM module uses fixed-time step sampling to construct the input sequence and takes the output of the last time step of the sequence as the temporal feature.

[0036] 3. Feature fusion and output Add the spatial features output by the GCN and the dynamic features output by the LSTM element-wise for fusion; The fused features are mapped to the grain-level stress prediction values through a fully connected layer, and the prediction results are output for a specified slice (such as Inc_26).

[0037] 4. Data Transfer Details The model input includes data.x (grain feature matrix), data.edge_index (graph structure information), and a field containing stress data for each time slice (e.g., increment_n); During the forward propagation process, the data flows through the GCN module, LSTM module, feature fusion layer, and fully connected layer in sequence, thus outputting the final prediction value; The predicted value is reversely reconstructed into a three-dimensional stress distribution map, which is basically consistent with the grid stress distribution map based on finite elements.

[0038] Step Three: Model Training and Evaluation 1. Training Process As Figure 2 shown, different training modes are set according to different experimental cases (e.g., Figure 2 (a) Use Inc_1 - 3 for training and Inc_26 for testing; Figure 2 (b) Use Inc_1 - 10 for training, Figure 2 (c) Use Inc_1 - 25, etc.), perform forward propagation on each graph data sample, and calculate the MSE between the prediction result and the true stress value; Use the mean square error (MSE) as the loss function, perform backpropagation and parameter update through the Adam optimizer, and record the training loss in each training cycle; The model training uses the Adam optimizer, and by adjusting the learning rate and training cycle, stable convergence of the model parameters is achieved.

[0039] 2. Model Evaluation Use the mean absolute error (MAE), mean absolute relative error (MARE), and coefficient of determination (R²) metrics to quantify the prediction effect and evaluate the model performance; At the same time, through the form of scatter plots and training loss curves, visually display the corresponding relationship between the true and predicted stress data; Step Four: Reverse Reconstruction and Visualization of Three-Dimensional Stress Distribution The predicted grain-level stress data is reassembled into a three-dimensional array through a specific reverse reconstruction algorithm, as Figure 3 shown, and a three-dimensional stress distribution map is generated using visualization software (such as ParaView); By comparing with the grid stress distribution image obtained from crystal plasticity finite element simulation, it can be observed that the reverse reconstruction image is highly consistent in both local details and overall distribution, verifying the accuracy and high restoration degree of this method.

[0040] Step Five: Computational Efficiency and Precision Advantages In the crystal plasticity finite element method, for the same set of 10 microstructure data, the DAMASK calculation time is about 100 minutes; while this method only takes about 5 minutes when using 90 microstructures for model training and less than 30 seconds when testing the same set of 10 microstructures. As Figure 4 shown in the bar chart of the time used by different models in the same dataset, as Figure 4 can be seen, the efficiency of training the crystal plasticity finite element dataset using this method is nearly 1000 times higher than the direct simulation efficiency. Using the trained model, the prediction efficiency of the performance of new microstructures is increased by 10000 times, which will greatly shorten the time for predicting alloy performance and avoid waste of resources.

[0041] In addition to significantly shortening the calculation time, this method can also provide more refined local stress information and highly restored three-dimensional stress visualization images, providing real-time and accurate prediction support for practical engineering applications.

[0042] Step 6. Physical meaning and interpretability of the model The experiment uses the stress distributions of microstructures at different deformation stages of the material corresponding to different test slice data (Inc_4, Inc_11, and Inc_26), and the prediction results correspond to Figure 2 (d), Figure 2 (e), Figure 2 (f). Through comparative analysis, it can be seen that this change process from local stress concentration to global equilibrium and then to zonal concentration is closely related to the deformation mechanism of the dual-phase titanium alloy. Typically, in the initial elastic stage, the stress is mainly concentrated at the grain boundaries. As the deformation progresses, in the yield stage, the interaction between grains gradually balances and the local stress diffuses evenly; while in the plastic stage, due to the phase ratio difference, different degrees of plastic deformation occur within the grains respectively, resulting in the segmentation phenomenon of local high-stress regions. This also verifies the model's ability to depict real physical phenomena and its interpretability.

[0043] The experimental results show that: (1) As the training set contains more historical slice data (such as Figure 2 (c)), the model captures the stress evolution law more comprehensively, thus achieving high-precision prediction; (2) The local and global stress characteristics at different deformation stages (elastic, yield, plastic) are effectively reflected, which is consistent with physical phenomena such as the initial stress concentration at the grain boundaries and the formation of local high-stress regions; (3) The generalization test results show that this model has good robustness and applicability to unknown data, providing a reliable basis for the safety of engineering structures and the optimization of material properties.

[0044] The present invention realizes a complete process from the preprocessing of crystal plasticity finite element simulation data, the extraction of grain stress, the joint learning of spatio-temporal features, to stress prediction and three-dimensional inverse reconstruction by adopting a grain-level stress prediction method based on a GCN-LSTM fusion model. This method not only greatly improves the prediction accuracy and the ability to restore local details, but also significantly shortens the calculation time. It provides efficient, accurate and visual technical support for material property analysis, optimization design and structural safety assessment.

Claims

1. A method for predicting the stress at the grain level of a duplex titanium alloy based on GCN-LSTM, characterized in that, This method combines computational simulation and experimentally obtained tissue information for prediction, including the following steps: (1) Data acquisition: Read the microstructure data of the dual-phase titanium alloy, including the node features and edge information constituting the grain map and the incremental data related to the grains. This part of the data is obtained through modeling using DREAM.3D software or through the EBSD experimental method; (2) Graph feature extraction: Use a graph convolutional neural network (GCN) module to perform forward propagation on the graph structure data and extract high-dimensional spatial features, including an initial convolutional layer and at least one hidden convolutional layer; (3) Temporal feature extraction: Use a long short-term memory network (LSTM) module to perform forward propagation on the temporal tensor composed of the incremental data and extract the dynamic features at the last time step of the sequence; (4) Feature fusion and prediction: Add the features obtained in steps (2) and (3) element by element and map them to the grain-level stress prediction value through a fully connected layer; (5) Inverse reconstruction: Inversely reconstruct the predicted value into a three-dimensional stress distribution map, which is basically consistent with the mesh stress distribution map based on finite elements; (6) Model training and evaluation: Use the mean square error (MSE) as the loss function, update the model parameters through backpropagation using the Adam optimizer, and evaluate the model performance using the MAE, MARE, and R² metrics. At the same time, display the results through scatter plots and training loss curves.

2. The method for predicting the stress at the grain level of a duplex titanium alloy based on GCN-LSTM according to claim 1, wherein The GCN module further includes: An initial convolutional layer for mapping the input features to a preset hidden dimension; At least one group of hidden convolutional layers for gradually enhancing the node feature representation; A final convolutional layer that outputs graph features with the same dimension as the output of the LSTM module.

3. The method for predicting the stress at the grain level of a duplex titanium alloy based on GCN-LSTM according to claim 1, wherein The LSTM module uses fixed-time-step sampling to construct the input sequence and uses the output at the last time step of the sequence as the temporal feature.

4. The stress prediction method for duplex titanium alloy grain level based on GCN-LSTM according to claim 1, characterized in that The model training uses the Adam optimizer, and by adjusting the learning rate and the number of training epochs, stable convergence of the model parameters is achieved.

5. The method for predicting the grain-level stress of a duplex titanium alloy based on GCN-LSTM according to claim 1, wherein It also includes the preprocessing and normalization of the microstructure graph data and the incremental data, as well as the data storage and loading processing using the pickle, dictionary, or CSV file format.

6. The method for predicting the grain-level stress of a duplex titanium alloy based on GCN-LSTM according to claim 1, wherein, For predicting the mechanical properties of the microstructural alloy and evaluating the evolution of the stress distribution in the tissue with deformation, including: Quantitatively describing the prediction error using the MAE, MARE, and R² metrics; Intuitively displaying the relationship between the true stress and the predicted stress and the loss change during the training process using scatter plots, kernel density estimation plots, and line plots.

Images