Application of artificial neural network algorithm in predicting solute segregation energy in aluminum grain boundaries

Abstract

The present invention discloses the application of an artificial neural network algorithm in predicting the solute segregation energy in aluminum grain boundaries, and relates to the technical field of grain boundary segregation energy prediction. The key points of the technical solution are: including obtaining characteristic data of the solute and preprocessing it; constructing an aluminum grain boundary segregation model with different solutes, and using a first-principles calculation method to obtain the segregation energy of each solute; generating a data set by extracting different features related to a target value, and using a machine learning model to analyze the segregation behavior of the solute; calculating the most important characteristic data affecting the solute segregation through the segregation behavior of the solute; substituting the main characteristic data into an artificial neural network algorithm to predict the segregation energy of the solute, and generating a prediction result of the solute segregation energy; and performing model evaluation and verification on the prediction result. The present invention generates a wider and more accurate data set by using a first-principles calculation method to ensure the accuracy of the machine learning model and a comprehensive understanding of the solute behavior.

Description

Technical Field

[0001] The present invention relates to the technical field of grain boundary segregation energy prediction, and more specifically, to the application of artificial neural network algorithm in predicting solute segregation energy in aluminum grain boundaries. Background Art

[0002] Grain refinement is considered an effective method to improve alloy properties such as strength and hardness. For example, nanocrystalline copper (Cu) with a grain size of 10 nm can achieve a strength of 1 GPa and a hardness of up to 3 GPa, while the strength of coarse-grained copper is only 50 MPa. However, one limitation of nanocrystalline metals is their poor thermal stability. Since the high density of grain boundaries provides a huge driving force for grain growth, nanocrystalline copper samples will show obvious grain coarsening after being left at room temperature for a few days. The application of stress greatly accelerates this coarsening process, even at liquid nitrogen temperature. Grain boundary segregation is not only an effective way to improve the thermal stability of nanostructured metals, but also a means to increase their strength, which has been demonstrated by many previous experimental studies.

[0003] Many studies have been conducted to improve the thermal stability and mechanical properties of aluminum alloys by grain boundary segregation. A gradient nanostructured surface layer was prepared in an Al-Mg-Cu alloy by surface sliding friction treatment, and the microhardness reached 320 HV, which was attributed to the segregation of copper and magnesium at the grain boundaries. In addition, the segregation of solute atoms at the grain boundaries can lead to the formation of second phases in the aluminum grain boundaries, which significantly affects the strength of aluminum alloys by hindering dislocation motion. The dispersion of high-density nanoscale second phases, such as η phase and Phase, which helps to improve the strength of aluminum alloy. Therefore, grain boundary segregation of solute atoms is an effective way to adjust the microstructure of aluminum alloy and improve its comprehensive mechanical properties.

[0004] However, there are still some questions about the mechanism of the effect of solute atom segregation on the properties of aluminum alloys. Grain boundary segregation is an atomic-scale phenomenon, which makes it difficult to obtain accurate and statistically significant data. These problems are related to the atomic-scale accuracy and details of grain boundaries, which hinder the quantitative study of grain boundary segregation by characterization techniques alone. With the rapid development of high-performance computing technology, computational materials science plays an important role in understanding the microscopic mechanisms of materials. For example, first-principles calculations reveal the intrinsic mechanism of solute segregation by providing information at the atomic and electronic levels. This method has also been used to study the segregation tendency of solute atoms in many alloys. Although first-principles calculations have high accuracy, they are time-consuming, which makes it difficult to perform a large number of calculations to obtain general conclusions.

[0005] Currently, machine learning models that deal with segregation energy in aluminum alloys have limitations in considering the diversity of solutes and the accuracy of data sets. Existing models mainly rely on data sets generated using embedded atom model (EAM) potential calculations, but their accuracy is still insufficient and the EAM potential is only applicable to a restricted set of elements. Therefore, the present invention aims to provide a method for predicting the segregation energy of solutes in aluminum grain boundaries using an artificial neural network algorithm. By using first-principles calculation methods, more accurate data sets and more diverse solutes can be generated to ensure the accuracy of machine learning models and a comprehensive understanding of solute behavior. Summary of the invention

[0006] The purpose of the present invention is to provide an application of an artificial neural network algorithm in predicting the segregation energy of solutes in aluminum grain boundaries. The present invention studies the segregation energy of solute atoms in aluminum through first-principles calculations, analyzes the segregation behavior of the solute, and simultaneously determines the most important characteristics affecting the segregation of the solute through calculations. Different machine learning models are used to predict the segregation energy of the solute, which can ensure the accuracy of the machine learning model and a comprehensive understanding of the solute behavior.

[0007] The above technical objectives of the present invention are achieved through the following technical solutions: Application of artificial neural network algorithm in predicting solute segregation energy in aluminum grain boundaries, comprising the following steps:

[0008] S1. Obtain characteristic data of solute and perform preprocessing;

[0009] S2. Construct a grain boundary segregation model with different solutes and calculate the segregation energy of each solute using the DFT method;

[0010] S3, extracting different features related to the target value through feature data to generate and construct a data set, and using a machine learning model to analyze the segregation behavior of the solute;

[0011] S4. Calculate the most important characteristic data affecting solute segregation through the solute segregation behavior;

[0012] S5. Substituting the most important feature data into different machine learning models to predict the solute segregation energy, and generating the prediction results of the solute segregation energy;

[0013] S6. Conduct model evaluation and verification on the prediction results.

[0014] The present invention is further configured as follows: the method for obtaining characteristic data of the solute in step S1 comprises the following steps:

[0015] S11, calculate the forces and coordinates of the extracted solute in three directions using VASP simulation software;

[0016] S12, obtain the Voronoi volume and coordination number of the solute in the grain boundary model through OVITO software analysis;

[0017] S13, use the Bader Charge software package to calculate the Bader charge to obtain the electronic interaction between the solute and the surrounding Al atoms;

[0018] S14. Collect the atomic radius, electronegativity and number of valence electrons of the solute.

[0019] The present invention is further configured as follows: the model evaluation and verification in step S6 is to evaluate and verify the prediction result using a machine learning model, and the mean absolute error and root mean square error are used as machine learning evaluation indicators, both of which reflect the deviation between the machine learning calculated value and the machine learning predicted value, that is, the smaller the value, the higher the accuracy of the machine learning model.

[0020] The present invention also provides an electronic device for predicting the solute segregation energy in aluminum grain boundaries using an artificial neural network algorithm, which is characterized by: comprising at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the method of predicting the solute segregation energy in aluminum grain boundaries using an artificial neural network algorithm.

[0021] The present invention also provides a computer-readable storage medium, characterized in that: the computer-readable storage medium stores computer instructions, and the computer instructions are used to be executed by the computer to implement a method for using an artificial neural network algorithm in predicting solute segregation energy in aluminum grain boundaries.

[0022] In summary, the present invention has the following beneficial effects:

[0023] 1. The present invention selects 44 solute atoms in aluminum and five typical grain boundary models, including Σ5(210) grain boundary, Σ5(310) grain boundary, two Σ9(221) grain boundaries and Σ11(332) grain boundary, calculates the segregation energy of solute atoms at different segregation sites on the grain boundary, and constructs a database containing 924 sets of data, describing the relationship between solute atoms, grain boundary characteristics and segregation energy. The segregation energy of solute atoms in aluminum grain boundaries was studied by first-principles calculations, the segregation behavior of solutes was analyzed, and the most important features affecting solute segregation were determined by calculation. Using different machine learning models to predict the segregation energy of solutes can ensure the accuracy of the model and a comprehensive understanding of solute behavior.

[0024] 2. The present invention uses 9 machine models to predict segregation energy. Experimental results show that the higher the complexity of the model, the better the prediction accuracy is generally. The tree-based model performs better than linear regression in the segregation energy task. Among them, the artificial neural network algorithm becomes the best performing machine model, which achieves 99% accuracy in both training and test sets, while requiring less computing time. The robustness and generalization ability of the artificial neural network model are confirmed by ten cross-validations, which can show that the combination of computing and machine learning can well predict the grain boundary segregation behavior of solute atoms, thereby accelerating the selection and design of high-performance alloys. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a workflow diagram for predicting separation energy by machine learning and DFT methods in an embodiment of the present invention;

[0026] Figure 2 It is a model diagram of Σ5(210), Σ5(310) grain boundaries, Σ11(332) and two Σ9(221) grain boundaries of Al in the embodiment of the present invention;

[0027] Figure 3 is a distribution diagram of a segregated energy dataset and a Voronoi volume dataset in an embodiment of the present invention;

[0028] Figure 4 is a functional relationship diagram of the segregation energy of five grain boundary models in the embodiment of the present invention and the (ae) Voronoi volume and (a1-e1) atomic radius;

[0029] Figure 5 is a correlation diagram of different characteristic segregation energies in an embodiment of the present invention;

[0030] Figure 6 It is a schematic diagram of the feature importance of 19 features obtained by the XGBOOST model, the Lasso regression model, and the Ridge regression model in an embodiment of the present invention;

[0031] Figure 7 It is a performance graph of the linear regression model, cubic Lasso linear regression model, cubic Ridge regression model, decision tree regression model, gradient boosting regression model, random forest regression model, XGBOOST model, support vector regression model, and artificial neural network on the training set (80%) and the test set (20%) in the embodiment of the present invention;

[0032] Figure 8 is a performance diagram of an artificial neural network plus cross validation in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The following is combined with Figure 1-8 The present invention is described in further detail.

[0034] Embodiment: A method for predicting solute segregation energy in aluminum grain boundaries using an artificial neural network algorithm comprises the following steps:

[0035] S1. In this embodiment, different grain boundary models were relaxed and the final energy was obtained by static calculation. Then, the solute segregation energy at different sites of the grain boundary was calculated, and the segregation energy data of 924 aluminum grain boundaries were generated. In order to predict the segregation energy by machine learning methods, it is necessary to properly extract the features that affect the solute segregation energy. The local environment around each segregated solute and the atomic interactions between the solute and the aluminum atoms have been shown to have an important influence on the energy characteristics around the solute atoms.

[0036] Therefore, various environmental features such as the forces (Fx, Fy, Fz) and coordinates (x, y, z) of the solute in three directions are calculated and extracted by the first-principles simulation software VASP to describe each solute. The forces and coordinates in three directions, as well as the stress conditions in six directions of the grain boundary model, are extracted from the OUTCAR file. Since the features obtained from the VASP calculations come from different files, a Python script is used to help extract these features.

[0037] In addition, the Voronoi volume and coordination number of the solute in the grain boundary model were obtained by the Voronoi analysis module of the scientific visualization and analysis software OVITO, reflecting the geometric interaction between the solute and the matrix atoms. The electronic interaction between the solute and the surrounding aluminum atoms was obtained by calculating the Bader charge using the BaderCharge software package. In addition, the physical properties of the solute atoms also affect the segregation behavior of the solute, and the radius, electronegativity and number of solute valence electrons were collected, as shown in Table 1:

[0038] Table 1 Ranges of features

[0039]

[0040] Table 1 shows 19 features related to local environment, atomic interactions and physical properties. All features are extracted as inputs to different machine models. Then, data preprocessing is performed to ensure that the features can reflect the relationship with the segregation energy in the aluminum grain boundary. For example, in this embodiment, by dividing the solute atomic radius by the aluminum atomic radius and subtracting the electronegativity of the solute from the electronegativity of the aluminum atom, the 19 input features are then standardized to speed up the machine processing and keep the order of magnitude of the features the same. Finally, a data set containing 924 data is successfully prepared.

[0041] S2. In this embodiment, the first-principles simulation software (VASP) was used for first-principles calculations, using the enhanced projected plane wave method, and the exchange-correlation potential was processed by using the generalized gradient approximation (GGA) of the Perdew-Burke-Ernzerhof form. In order to ensure the accuracy and efficiency of the calculation, a cutoff energy of 350 electron volts, a convergence energy of 10^-5 electron volts, and a 0.01 electron volt / The converged atomic forces of . Before constructing the grain boundary model, the equilibrium lattice parameters of the aluminum unit cell were calculated using a 4×4×4 k-point grid. The parameters of the fully relaxed face-centered cubic aluminum are Compared with the existing calculation results And experimental results Consistent.

[0042] Then, the coincidence symmetry lattice method was used to construct aluminum Σ5(210), Σ5(310) grain boundaries, Σ11(332) and two Σ9(221) grain boundary models. The aluminum Σ5(210) grain boundary was constructed by selecting 31 atomic layers and rotating half of the layers by 53.1°. In addition, the (210) direction includes The vacuum layer is set to ensure that there is no atomic interaction between the two (210) surfaces. Similarly, the other four grain boundary models are constructed using the same method using the VESTA software package. The detailed information of the grain boundary model is shown in Table 2.

[0043] Table 2 Grain boundary model information, including number of atoms, cell dimension and k-points

[0044]

[0045] According to the experimental results, 44 common alloying elements of Al in the periodic table were selected as segregation solutes to study the segregation behavior of grain boundaries. The information of solute atoms is shown in the table in the supplementary material, e.g. Figure 2 As shown, for each grain boundary structure, there are different numbers of segregation sites, some of which are symmetrically distributed, and the solute atoms have the same segregation tendency. Therefore, for the solute that forms segregation, the five grain boundary models have 3, 4, 4, 4 and 6 segregation sites, respectively.

[0046] S3. During structural relaxation, the atoms in the grain boundaries and bulk models underwent complete relaxation, except for the z coordinates of the outermost atoms in each model. Subsequently, a static calculation was performed in each case to obtain the final energy. During structural relaxation, the dimensions of all grain boundaries and bulk models were fixed to maintain the cross-sectional area used for the partial energy calculation.

[0047] The ability of solute to form segregation at grain boundaries can be expressed by the segregation energy E seg To represent it, as follows:

[0048]

[0049] Among them, E GB , E bulk and are the total energies of the undoped grain boundary model, the X-doped grain boundary model, the undoped Al bulk model, and the X-doped bulk model. seg Negative values ​​of indicate that solute X will segregate to the grain boundary sites at the same time, which is energetically favorable. Then, by comparing the E seg , the most preferred site of the solute can be found. Figure 3 (a) and 3(b) show the distribution of segregation energy values ​​and Voronoi volumes of different solutes in five Al grain boundaries, respectively, to explore the characteristic distribution of different solutes. It is observed that both values ​​show a Gaussian distribution with mean values ​​of -0.103 eV and In addition, more than half of E seg The values ​​are less than zero, indicating that most solutes are likely to form energetically favorable segregations. However, some high values ​​are concentrated at about 4 eV, especially for nonmetallic atoms such as carbon and boron, indicating that nonmetallic atoms have a smaller tendency to form substitutional segregations in aluminum grain boundaries.

[0050] The atomic radius significantly affects the grain boundary energy. Atoms with larger radius are more stable in grain boundaries. In addition, the site Voronoi volume of the grain boundary reflects the strain energy of the interface plane, which is another key factor affecting the grain boundary energy. In order to explore how these factors affect the segregation behavior of grain boundaries, the segregation energy distribution of atomic radius and Voronoi volume of different solutes in five grain boundaries is shown, such as Figure 4As shown in the figure, it is observed that with the increase of atomic radius, the segregation energy generally shows a downward trend, indicating that solute atoms with larger radius tend to segregate toward grain boundaries; this phenomenon is mainly attributed to the reduction of elastic strain energy at grain boundaries. Solute atoms with larger atomic radius can more effectively reduce elastic strain energy. In addition, solute segregation is often more obvious at grain boundary sites with larger Voronoi volumes, indicating that larger spaces are conducive to solute segregation. In general, the radius of solute atoms and the Voronoi volume of segregation sites have a significant effect on the segregation behavior of grain boundaries. Solutes larger than matrix atoms have a greater tendency to segregate, and solutes are more likely to segregate at sites with larger Voronoi volumes.

[0051] In addition, different grain boundary structures show different solute segregation tendencies, indicating that grain boundary structure plays a key role in the segregation behavior. For example, in the Σ11(332) grain boundary, the segregation energy of solute atoms is generally greater than -1 eV, indicating a reduced segregation tendency compared to other grain boundaries. These findings highlight the importance of considering atomic radius and site characteristics in understanding the segregation behavior in grain boundaries.

[0052] S4. In order to mitigate the risk of model overfitting and find the relationship between different features, it is crucial to calculate and check the importance of features before conducting the machine learning process as a basis for reducing the feature dimension, among which atomic radius and Voronoi volume have a significant linear relationship with the segregation behavior. In order to determine the direction (positive or negative) of the linear correlation between the input features and the segregation energy, the Pearson correlation coefficient between these features and the segregation energy was calculated. Figure 5 As shown, the Pearson correlation coefficient measures the strength and direction of the linear relationship between the feature and the partial energy. The larger the value, the stronger the linear regression. It is calculated as follows:

[0053]

[0054] Among them, cov(X,Y) and σX represent the values ​​of covariance and variance respectively, and the results of Pearson correlation coefficient are presented in Figure 5 where the target variable is the partial energy, Figure 5 It can be clearly seen that there is a strong negative correlation between atomic radius, Voronoi volume and the target variable (|ρ|>0.6), which is consistent with Figure 4 Consistent with the findings of the present study, this suggests that solutes with larger radii are more likely to form grain boundary segregation, and the linear relationship between atomic radius and Voronoi volume is stronger (ρ = 0.81), meaning that larger atoms will cause larger volumes.

[0055] In this example, charge transfer and electronegativity show relatively strong positive correlations with the target variable, with correlation coefficients of about 0.46 and 0.43, respectively, indicating that solutes with larger charge transfer are more likely to exhibit grain boundary segregation, and there is also a strong linear relationship between these two features, because larger electronegativity will lead to more charge transfer; at the same time, δy and coordination also show significant correlations with the target variable (|ρ|>0.3). On the other hand, features such as Z, Y, X, δxy, δzx, Fx, Fy, Fz, and valence electrons show weak correlations with the target variable (|ρ|<0.1), indicating that there is a negligible linear relationship between these features and the target variable.

[0056] Based on the random forest model, Lasso regression and Ridge regression, the importance of each feature for predicting partial focusing energy can be quantitatively evaluated. Figure 6 (a), (b) and (c) show the feature importance of regression partial energy. The feature importance of 19 features obtained by (a) XGBOOST model, (c) Lasso regression model, and (d) Ridge regression model. It can be observed that in the three models considering atomic radius as the most important feature, atomic radius is the most important feature because it is an intrinsic physical property of the solute. Voronoi volume was identified as the second most important feature in random forest and Lasso regression, representing a local environment metric. In contrast, features such as X, Fz, Fy, δxy and δzx have low importance in all four models, indicating that their impact on partial energy is negligible. In this embodiment, these five features are excluded from the input feature set to avoid overfitting.

[0057] S5. Before applying machine learning to the dataset, a random split was performed to create a training set (80%) and a test set (20%), resulting in 739 training set and 185 test set samples. The efficiency of different algorithms was considered by measuring the training time required for each model. In addition, there are different hyperparameters in the machine models that affect the accuracy of the algorithms. For example, the maximum depth of branches and the number of estimators in XGBOOST determine the depth and number of trees. In order to optimize the hyperparameters of the model, a grid search method was used, and the best hyperparameters were selected by the model performance in the validation set. The performance of different machine learning models is shown in Figure 2. Figure 7As shown, the performance of (a) linear regression model, (b) cubic Lasso linear regression model, (c) cubic Ridge regression model, decision tree regression model, gradient enhancement regression model, random forest regression model, XGBOOST model, support vector regression model, and artificial neural network on the training set (80%) and test set (20%) are shown. The y-axis value is the partial energy of the test data set predicted by different models, while the x-axis value is the partial energy calculated by the first principle method. If the predicted value is close to the value calculated by the first principle, the data points should be concentrated on the y=x line.

[0058] Three linear regression models were used to predict the grain boundary segregation energy, namely linear regression, cubic Lasso linear regression and cubic Ridge regression, with correlation coefficients as Figure 7 (ab) are shown. Polynomial feature expansion is used to include interaction coefficients, making it possible to capture the interactions between the original descriptors, which may be ignored in simple linear models. Figure 7 As shown in (ab), it can be found that ordinary linear regression performs poorly in these methods. 2 The R 2 is 0.635. In comparison, the cubic Lasso linear regression and cubic Ridge regression achieved better results (Lasso regression has an R 2 is 0.920). In addition, the linear model has a higher efficiency. It is worth noting that the performance of the linear model improves with the increase of feature dimension and model complexity, which suggests that other machine models may produce higher accuracy.

[0059] S6. Decision tree regression (DTR) and three decision tree-based models (gradient boosting regression model, random forest regression model, and XGBOOST model) were tested to predict the partial energy. These four machine learning algorithms are widely used in various machine learning regression tasks due to their robustness, interpretability, and ability to handle nonlinear relationships and diverse feature types. Figure 7 In (dg), it can be observed that the values ​​are more concentrated around the y=x line, indicating that the predictions of these four models are more consistent than those of machine learning. The decision tree only takes about 8 seconds to search for the best parameters. In contrast, the gradient boosting regression and random forest algorithms require more computation time (about 278 seconds and 298 seconds), but they produce higher accuracy. Among the four tree models, XGBOOST performs best, not only with a shorter computation time (about 80 seconds), but also with a higher accuracy on the training set (R 2 = 0.982, RMSE (root mean square error) = 0.14 eV) and test set (R 2=0.910, RMSE =0.36 eV) showed comparable performance.

[0060] The results of the support vector regression (SVR) model are as follows Figure 7 (h). It is more efficient and requires less time than most tree-based models, but its performance is similar to that of the four tree-based models (training R 2 =0.976, training RMSE =0.16eV). We also explored the application of an artificial neural network model for predicting grain boundary segregation energy. The neural network architecture was designed with four hidden layers containing 64 nodes, 128 nodes, 64 nodes, and 32 nodes, respectively. To ensure the best performance, the rectified linear unit (Relu) activation function was selected for the neural network as it produced the lowest root mean square error. In addition, the dropout technique was used in the neural network to mitigate the risk of overfitting, which randomly discards neurons. Figure 7 (i) shows the comparison between the predicted values ​​and the DFT calculated values. It can be observed that this ANN configuration has similar prediction accuracy to the four tree models, achieving an impressive overall R in the training set. 2 is 0.958, with a root mean square error of 0.28 eV, and R is achieved in the test set 2 The value of the proposed algorithm is 0.922, and the root mean square error is 0.022 eV. In addition, the artificial neural network algorithm only takes about 6.5 seconds to complete model training, which is much faster than the DTR-based model.

[0061] Experimental data

[0062] In order to evaluate the robustness and generalization ability of the trained model, a ten-fold cross-validation method was used. In the first cross-validation process, the dataset was divided into ten subsamples, and one sample was used to test the model in each validation iteration, while the remaining nine samples were used for training. Then, the process was repeated nine times. In each cross-validation process, a new model was generated. The robustness and generalization ability of the model were evaluated by comparing the average values ​​of these indicators. This process calculated the root mean square error and R by ten-fold cross-validation. 2 The average values ​​are shown in Table 3:

[0063] Table 3 The performance of different machine learning models

[0064]

[0065]

[0066] Table 3 collects the R values ​​of different models in the test set, training set, and cross-validation process. 2 and RMSE, we can see that the robustness of the linear regression model is relatively poor. For example, the cubic ridge regression only shows R 2 = -50.22, which is consistent with the R 2 =0.944, showing very poor generalization ability. The tree model shows better robustness than the linear regression model. The performance of SVR is slightly worse than that of the tree model, but more robust than the linear regression model. The average R of the artificial neural network model during cross validation 2 The value of the artificial neural network is 0.951, which is higher than other models. In addition, the root mean square error of the artificial neural network in the training set, test set and cross validation is similar, indicating that it has better robustness. Considering the accuracy, efficiency and robustness comprehensively, the artificial neural network algorithm performs better than other machine learning algorithms. Figure 8 Results of ten-fold cross-validation performance of the ANN model are shown, indicating that the ANN model has a strong regularization capability in predicting grain boundary segregation energy. These results highlight the effectiveness of the ANN model in capturing and modeling the complex relationship between input features and segregation energy in aluminum grain boundaries.

[0067] This specific embodiment is merely an explanation of the present invention and is not a limitation of the present invention. After reading this specification, those skilled in the art may make non-creative modifications to the present embodiment as needed. However, as long as they are within the scope of the claims of the present invention, they are protected by the patent law.

Claims

1. The method of artificial neural network algorithm in predicting the solute segregation energy in aluminum grain boundaries is characterized by: The following steps are involved: S1. Obtain characteristic data of solute and perform preprocessing; S2. Construct a grain boundary segregation model with different solutes and calculate the segregation energy of each solute using the first principle calculation method; S3, constructing a data set by extracting different features related to the target value, and using a machine learning model to analyze the solute segregation behavior; S4. Calculate the main characteristic data that affect the solute segregation through the solute segregation behavior; S5, substituting the main characteristic data into an artificial neural network algorithm to predict the solute segregation energy, and generating a prediction result of the solute segregation energy; S6. Conduct model evaluation and verification on the prediction results.

2. The method of predicting the solute segregation energy in aluminum grain boundaries using the artificial neural network algorithm according to claim 1 is characterized by: The method for obtaining characteristic data of the solute in step S1 comprises the following steps: S11, calculate and extract the forces and coordinates of the solute in three directions through VASP simulation software; S12, obtain the Voronoi volume and coordination number of the solute in the grain boundary model through OVITO software analysis; S13, use the Bader Charge software package to calculate the Bader charge to obtain the electronic interaction between the solute and the surrounding aluminum atoms; S14. Collect the atomic radius, electronegativity and number of valence electrons of the solute.

3. The method of predicting solute segregation energy in aluminum grain boundaries using an artificial neural network algorithm according to claim 1, characterized in that: In step S6, model evaluation and verification are performed, using mean absolute error and root mean square error as machine learning evaluation indicators, both of which reflect the deviation between the first principle calculation value and the machine learning prediction value, that is, the smaller the value and the shorter the time used, the higher the accuracy of the machine learning model.

4. An electronic device for predicting the solute segregation energy in aluminum grain boundaries using an artificial neural network algorithm, characterized by: It comprises at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the method of using the artificial neural network algorithm described in any one of claims 1-3 in predicting solute segregation energy in aluminum grain boundaries.

5. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to be executed by the computer to implement the method of using the artificial neural network algorithm of any one of claims 1 to 3 in predicting solute segregation energy in aluminum grain boundaries.

6. Application of artificial neural network algorithm in predicting solute segregation energy in aluminum grain boundaries.

Images