Construction method of high-entropy alloy mechanical property prediction model

Abstract

The invention discloses a construction method of a high-entropy alloy mechanical property prediction model, which comprises the following steps: constructing a high-entropy alloy model set, namely defining proportional distribution of high-entropy alloy components, determining component structures, and enabling one group of component structures to correspond to one high-entropy alloy model; performing force field parameter setting stable configuration on the high-entropy alloy model; simulating and acquiring mechanical characteristics of the high-entropy alloy model through a stretching process; obtaining a multi-scale feature data set; the multiple scales comprise environment features, reference features and analog calculation features; performing feature screening on the multi-scale feature data set to obtain a sample data set; defining a plurality of regression machine learning models; and training and verifying the regression machine learning model by adopting the sample data set, selecting an optimal regression machine learning model, and constructing a mechanical property prediction model. According to the technical scheme, the limitation of a traditional scheme can be broken through, the calculation efficiency and the prediction credibility are improved, the new material research and development period is remarkably shortened, and the experiment cost is reduced.

Description

Technical Field

[0001] The present invention relates to the field of combining artificial intelligence technology with molecular dynamics simulation, and in particular to a method for constructing a high entropy alloy mechanical property prediction model. Background Art

[0002] High-entropy alloys (HEAs), an emerging class of metallic materials, demonstrate exceptional comprehensive properties unmatched by conventional alloys through their unique design concept of multiple principal components in equimolar or near-equimolar ratios. These alloys not only possess excellent strength-ductility matching and significant work-hardening capabilities, but also exhibit exceptional high-temperature resistance and excellent corrosion resistance. These outstanding performance advantages hold great promise for their application in aerospace engine hot-end components, nuclear power system structural materials, and chemical equipment for extreme environments.

[0003] However, the research and development of high-entropy alloys (HEAs) faces several challenges: their vast design space and complex multi-component interactions lead to a high degree of diversity and uncertainty in their microstructural characteristics. Traditional alloy research and development relies primarily on empirical rules and trial-and-error, an approach that is both time-consuming and costly.

[0004] In recent years, with the rapid development of high-throughput computational simulation, it is possible to quickly simulate and generate performance characteristic data related to high-entropy alloy structures. However, the mechanical properties extracted from simulated high-entropy alloy structures often have limitations in certain specific environments: for example, in high-temperature environments, the single FCC phase of the high-entropy alloy structure will lead to poor creep resistance, reduced hardness, and insufficient wear resistance and oxidation resistance. However, if a large amount of performance characteristic data is extracted from the simulation structure of a wide range of design application scenarios, the physical mechanisms of the cross-scenario models will vary significantly, resulting in redundancy in the performance characteristic data and affecting computational efficiency and accuracy. Therefore, it is necessary to start from the optimization link of the high-entropy alloy structure simulation to establish a reliable composition-structure-performance correlation model, thereby achieving the design of high-entropy alloys with excellent mechanical properties. Summary of the Invention

[0005] To achieve the above objectives, the present application provides a method for constructing a high entropy alloy mechanical property prediction model, characterized in that the high entropy alloy is AlCoCrFeNi, and the construction method comprises the following steps: Constructing a high-entropy alloy model set, including: defining a proportional distribution of the high-entropy alloy components, determining a component structure of the high-entropy alloy model based on the proportional distribution, wherein a set of the component structures corresponds to a high-entropy alloy model; setting force field parameters for the high-entropy alloy model to stabilize the configuration; and obtaining mechanical characteristics of the high-entropy alloy model through stretching process simulation; Obtaining a multi-scale feature data set through the high entropy alloy model set; the multi-scale includes: environmental features, reference features, and simulation calculation features; Performing feature screening on the multi-scale feature dataset to obtain a sample dataset; wherein the target variables in the sample dataset are simulation calculation features, including Young's modulus features and ultimate tensile strength features; the feature screening refers to performing redundant feature screening and feature importance selection on reference features, and eliminating non-critical features; Define multiple regression machine learning models; use the sample data set to train and verify the regression machine learning models respectively, select the optimal regression machine learning model, and construct a mechanical property prediction model.

[0006] The proportion distribution of high entropy alloy components is defined by using a basic proportion increasing and remaining proportion equal distribution strategy, including: Define the base ratio, maximum ratio and growth step value; One component is extracted from the five components of the high-entropy alloy model in turn as a basic component, a single proportion value of the basic component is cyclically defined, and the remaining proportion values are evenly distributed among the other four components; wherein the remaining proportion value = 100% - single proportion value.

[0007] The loop definition includes the following steps: Step 1: Define a single ratio value as the basic ratio value; Step 2: defining the ratio of the basic components as a single ratio value; Step 3: Increment the single ratio value by the growth step value, and determine whether the single ratio value is greater than the maximum ratio value. If it is greater than the maximum ratio value, end the loop definition; otherwise, execute steps 2 to 3.

[0008] Furthermore, the environmental characteristics are determined based on temperature, strain rate and element ratio, and the reference characteristics are determined based on the component element characteristics of the high entropy alloy model; the simulated calculation characteristics are generated through a process simulating the stretching process.

[0009] The reference characteristics include the basic characteristics of the five components, the enhanced model mismatch value, the weighted standard difference, the local mismatch value and the thermodynamic parameters; Wherein, the basic characteristics provide the basic physical and chemical properties of the components; The enhanced model mismatch value is used to quantify the macroscopic differences in mechanical and electrical properties between components; The weighted standard difference reflects the degree of dispersion of properties in the multi-component system; The local misfit value is used to reveal atomic-scale strain; The thermodynamic parameters are used to predict phase separation, solid solution or compound formation.

[0010] Furthermore, feature screening includes: Calculate the linear correlation between the reference features and the target variable, eliminate redundant features, and establish the optimal feature subset; quantifying the contribution of the reference features in the optimal feature subset to the model prediction ability, and removing the reference features with the lowest contribution from the optimal feature subset; The Lasso regression model is used to eliminate features that have little impact on the target variable.

[0011] Furthermore, when removing redundant features, the PCC value calculated on the original dataset was used for screening; If the two reference features are closely correlated, the correlation between the two reference features and the target feature is further determined, and only the reference features with the largest correlation are retained.

[0012] Among them, multiple regression machine learning models include: random forest, extreme gradient boosting, lightweight gradient boosting and category feature gradient boosting algorithms.

[0013] Choosing the best regression machine learning model involves: Find the parameter combination that performs best on the dataset; The 10-fold cross validation was used to extract the mean of 10 validation results as the evaluation model indicator; The mean square error and goodness of fit are used to select the optimal model from the multiple regression machine learning models.

[0014] Furthermore, after constructing the mechanical property prediction model, a SHAP global interpretation is performed on the Young's modulus and ultimate tensile strength predicted by the mechanical property prediction model, and a SHAP local interpretation is performed on the local dependence of important features of the Young's modulus and ultimate tensile strength predicted by the mechanical property prediction model.

[0015] The method for constructing a high-entropy alloy mechanical property prediction model, provided by this invention, ensures the integrity of features from the perspective of sample composition and optimizes them to ensure effective model training. From the perspective of the model itself, it also employs multiple verification and interpretation methods to ensure efficient and accurate predictions. From the perspective of prediction results, it analyzes the contribution of features using SHAP values, providing interpretable theoretical guidance for the compositional design and performance optimization of novel high-entropy alloys. Compared with traditional trial-and-error methods, this method overcomes the limitations of traditional approaches, significantly improving computational efficiency and prediction reliability, significantly accelerating the new material R&D cycle, and reducing experimental costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 1 is a schematic diagram of the steps of a method for constructing a high entropy alloy mechanical property prediction model provided by an embodiment of the present invention; Figure 2 A molecular dynamics simulation model for a specific simulated process provided according to an embodiment of the present invention; Figure 3 A schematic diagram of a process flow of a three-step feature screening method according to an embodiment of the present invention; Figure 4 Results of random forest and Lasso regression feature screening based on Young's modulus and ultimate tensile strength provided in an embodiment of the present invention; Figure 5 Schematic diagram of the fitting effect of predicting Young's modulus and ultimate tensile strength using four machine learning models provided in an embodiment of the present invention; Figure 6 A schematic diagram of a model interpretation process according to an embodiment of the present invention; Figure 7 A SHAP global interpretation diagram of Young's modulus and ultimate tensile strength predicted by the CatBoost model provided in an embodiment of the present invention; Figure 8 This is a local dependency graph of the top three important features when predicting Young's modulus according to an embodiment of the present invention. DETAILED DESCRIPTION

[0017] This study focuses on the adjustability of element ratios in a high-entropy alloy (AlCoCrFeNi) by designing an element ratio gradient simulation experiment. This study constructs a feature dataset based on multiple sets of high-entropy alloy models with varying element ratios. A multi-scale feature system is established based on the relationships between these multiple feature sets. A targeted feature screening method is then used to optimize the feature dimensions and construct a sample dataset. Multiple models are selected within this sample dataset, and during the model training phase, a nested hyperparameter optimization framework combining 10-fold cross-validation and grid search is employed to systematically optimize model parameters. Finally, an interpretable machine learning framework based on game-theoretic Shapley values is constructed to accurately predict the mechanical properties of high-entropy alloys.

[0018] The specific implementation of the present invention is described in detail below with reference to the accompanying drawings.

[0019] The method for constructing the high entropy alloy mechanical property prediction model provided by the present invention is as follows Figure 1 As shown, the following steps are included: Step S100: constructing a high-entropy alloy model set, including: defining a proportional distribution of high-entropy alloy components, determining a component structure of the high-entropy alloy model based on the proportional distribution, wherein a set of component structures corresponds to one high-entropy alloy model; setting force field parameters for the high-entropy alloy model to stabilize the configuration; and obtaining the mechanical characteristics of the high-entropy alloy model through stretching process simulation; This step can be achieved by using a systematic component design method using software such as LAMMPS: 1) First determine the model size, for example, 15nm×5nm×5nm.

[0020] 2) Next, define the proportional distribution of the high-entropy alloy components, further distributing the five components, Al, Co, Cr, Fe, and Ni. Currently, the commonly used high-entropy alloy component distribution method uses an isocratic composition method. However, the tensile strength of an isocratic AlCoCrFeNi high-entropy alloy in most environments is approximately 0.8-1.2 GPa, and the elongation is 10%-15%. However, there are mechanical property limitations in certain specific environments. For example, at high temperatures, the single FCC phase leads to poor creep resistance, hardness drops to 120-200 HV, corrosion resistance is weak, and wear resistance and oxidation resistance also decrease.

[0021] To address this shortcoming, the present invention adopts a distribution strategy of increasing the basic proportion and evenly dividing the remaining proportion. The high entropy alloy designed by this distribution strategy has a component structure with non-equal proportion distribution. The implementation method of the distribution strategy includes: 1) Define the base ratio, maximum ratio, and growth step value: for example, define the base ratio as 10%, the maximum ratio as 30%, and the growth step as 2%; 2) Taking turns to extract one component from the five components of the high-entropy alloy model as the basic component, using the basic proportion value as the initial value and the growth step value as the growth value, cyclically define the single proportion value of the basic component, and evenly distribute the remaining proportion values to the other four components; wherein the remaining proportion value = 100% - the single proportion value; that is, take turns to distribute the proportion to each element of the high-entropy alloy model, and the proportion of each element starts from 10% (basic proportion value), and uses an incremental step of 2% for proportion distribution. In each round of distribution, the proportions of the remaining four elements are evenly distributed as the remaining atomic percentages.

[0022] The loop definition is implemented through the following steps: Step 1: Define a single ratio value as the base ratio value; for example, start allocation from 10%; Step 2: defining the ratio of the basic components as a single ratio value; In the first allocation, the proportion of the specified component (such as Al) is 10%, and the remaining atomic percentage is 90%; Step 3: Increment the single ratio value by the increment step value, and determine whether the single ratio value is greater than the maximum ratio value. If it is greater than the maximum ratio value, end the loop definition; otherwise, execute steps 2 to 3; In step 3, 90% is evenly distributed to other components (such as Co, Cr, Fe, and Ni); the next round of loop definition is continued. In the next round of processing, the proportion of the specified component is 12%, and the remaining atomic percentage is 88%, and so on. During the loop definition process, the proportion of each element is limited to 10% to 30%, and the model building process ensures that the total atomic percentage maintains 100% stoichiometric balance.

[0023] Through this method, by controlling the changes in a single variable (the proportion of a certain element), we can accurately analyze the influence of specific element content on material properties, achieve systematic component design, and try to ensure the integrity of the structural composition to achieve complete characteristics. In this way, a component-performance relationship database can be established to provide sufficient and balanced training data samples for machine learning modeling.

[0024] The present invention provides an embodiment. Through this method, a total of 51 groups of high-entropy alloy models with different element ratios are constructed, including single adjustment models for each component (such as adjusting only the Al content). The high-entropy alloy models designed through this step can be used for special application scenarios.

[0025] After the high entropy alloy model is constructed, the potential function is set and the force field parameters are set to minimize its energy and molecular dynamics balance so that the constructed model has a stable configuration.

[0026] Furthermore, the mechanical characteristics of the high entropy alloy model are obtained by simulating the stretching process: in the specific implementation, relaxation is first performed under the NPT ensemble with a relaxation time of 10 ps; then these model materials are uniaxially stretched at a constant rate along the y direction under the NVT ensemble to obtain the corresponding stress-strain curve (mechanical characteristics can be calculated). The simulated model and the stretching direction are as follows Figure 2 shown.

[0027] Step S110: obtaining a multi-scale feature data set through the high entropy alloy model set; The multi-scale here includes three dimensions: environmental characteristics, reference characteristics and simulation calculation characteristics; 1) Environmental features are manually set based on environmental conditions, including temperature, strain rate, and element ratio information. When manually setting environmental features, they are usually determined and set based on experience.

[0028] The present invention provides an embodiment to determine seven types of environmental characteristics (temperature, strain rate and ratio of five elements); the temperature includes 10K, 300K, 600K and 900K, covering low temperature, room temperature and high temperature; the strain rate includes 10 8 s -1 , 10 9 s -1 and 10 10 s-1 ; The proportions of the five elements refer to the atomic proportions set when the model is constructed.

[0029] 2) Reference features are features determined based on the characteristics of the component elements of the high-entropy alloy model. They can also be found through literature search, such as by searching for the values of the relevant characteristics of these five elements in published research or some related websites, and then calculating the characteristics of this alloy material. The present invention provides an embodiment in which 48 reference features are obtained.

[0030] Specifically, the reference characteristics include the basic characteristics composed of these five groups, the enhanced model misfit value, the weighted standard difference, the local misfit value and the thermodynamic parameters; Basic features provide basic physical and chemical properties of components, including: shear modulus ( G ), bulk modulus ( B ), Poisson's ratio( ν ), atomic radius ( R ),density( D )、Boiling point( TB ), atomic mass ( m ), Pauling electronegativity ( X )、Melting point( T m ), valence electron concentration ( VEC ), free electron concentration ( e / a ), cohesive energy ( E c ), vacancy formation energy ( E v ), first ionization energy ( E I-1st ), the second ionization energy ( E I-2nd )wait; The strengthening model mismatch value is used to quantify the macroscopic differences in mechanical and electrical properties between components, including: the shear modulus mismatch value in the strengthening model ( η G ), the bulk modulus mismatch value in the strengthening model ( η B ); Weighted standard difference is used to reflect the degree of dispersion of properties in a multi-component system, including: shear modulus difference ( δ G ), bulk modulus difference ( δ B ), Poisson's ratio difference ( δ ν ), atomic radius difference ( δ R ), density difference ( δ D ), boiling point difference ( δTB ), atomic mass difference ( δ m )wait; Local mismatch values are used to reveal atomic-scale strains, including: local shear modulus mismatch values ( D G ), local bulk modulus mismatch value ( D B ), local Poisson's ratio mismatch value ( D ν ), local atomic radius mismatch value ( D R ); Thermodynamic parameters used to predict phase separation, solid solution, or compound formation include the entropy of mixing ( ), mixing enthalpy ( ) and phase stability descriptors ( ).

[0031] Some reference features are calculated as follows: , , , , , , , in, As the basic feature, To strengthen the mismatch value in the model, is the weighted standard difference of the basic characteristic parameters, is the local mismatch value of the basic characteristic parameters, Indicates the number of elements in the alloy (the high entropy alloy in the present invention contains five elements), Indicates the ratio of elements, Represents the basic parameters of the element, represents the Boltzmann constant, represents the mixing enthalpy between two different elements, represents the mixing entropy, represents the mixing enthalpy, is the phase stability descriptor, is the average melting temperature of the alloy system.

[0032] 3) The simulation calculation feature is to simulate the tensile process of high entropy alloys with different components under different conditions through molecular dynamics simulation, obtain stress-strain curves, and calculate the two mechanical characteristics of Young's modulus and ultimate tensile strength through stress-strain curves; The calculation of Young's modulus uses the least squares method to find the slope of the curve from 0 to 0.02. The value of the slope is defined as Young's modulus. The calculation formula of the least squares slope is: ,in, represents the strain value of the i-th data point, represents the stress value of the i-th data point, represents the average value of all strain values, represents the average stress value of all data points, represents Young's modulus.

[0033] The ultimate tensile strength is the maximum stress value in the stress-strain curve.

[0034] This step, combined with step S110, comprehensively extracts feature data based on a rich and complete structural structure across a wide range of application scenarios. For example, the high-entropy alloy model constructed in step S110 supports various component ratios. The combination of different component ratios (such as Al concentration) and temperature significantly affects the phase structure and, consequently, the mechanical properties of the high-entropy alloy. As Al concentration increases, the alloy's crystal structure gradually transforms from the FCC phase to the BCC phase. Increasing temperature accelerates this transformation by increasing atomic diffusion rates, often accompanied by an increase in plasticity. Temperature's regulation of phase transformation exhibits a phased pattern: at low temperatures, due to limited atomic mobility, short-range ordered structures or nano-precipitates are more likely to form, enhancing matrix strength through solid solution strengthening and precipitation strengthening. From room temperature to high temperatures, thermodynamically unstable phases decompose or transform into disordered structures, ultimately leading to a significant decrease in strength due to the failure of strengthening mechanisms. In this step, based on the high-entropy alloy model structure at different concentrations, environmental characteristics are defined for various temperature environments, and simulated computational features that change during the FCC / BCC phase transformation are calculated.

[0035] The present invention provides an embodiment, and the multi-scale feature data set constructed includes 57 features and 612 data. Among them, the alloy structure (such as Al) with significant advantages in mechanical properties under special environments is realized based on the allocation strategy of increasing the basic proportion and evenly dividing the remaining proportion. 18 Co 20.5 Cr 20.5 Fe 20.5 Ni 20.5 ), the tensile strength of the structural alloy under low temperature conditions corresponds to the characteristic data of tensile simulation, which reaches about 6.4GPa. 30 Co 17.5 Cr 17.5 Fe 17.5 Ni 17.5 During the stretching process, the structure undergoes a significant transformation from FCC to BCC, forming a two-phase coexistence structure of FCC and BCC.

[0036] Step S120: performing feature screening on the multi-scale feature data set to obtain a sample data set; wherein the target variables in the sample data set are simulation calculation features, including Young's modulus features and ultimate tensile strength features; Too many feature dimensions in the feature pool of a multi-scale feature dataset will significantly increase the computational complexity of model training and prediction, reduce overall efficiency, generate noise features, and affect the generalization performance of the model. In order to build an efficient and robust prediction model, the present invention uses a three-step feature screening method to optimize the features. The environmental features in the multi-scale feature dataset can facilitate the observation of the impact of manually input features on mechanical properties. The simulated calculation features are the target parameters for model training. Therefore, these two types of features are retained and not screened. The reference features are optimized using the feature screening method. The reference features are screened for redundant features and feature importance selection through feature screening, and non-critical features are eliminated.

[0037] The process of feature selection is as follows Figure 3 As shown, it includes the following three steps: Step S310: Calculate the linear correlation between the reference features and the target variable, eliminate redundant features, and establish the optimal feature subset.

[0038] When removing redundant features, the PCC value calculated on the original data set was used for screening to obtain the first step data set first_data.

[0039] The calculation formula of PCC value is: ,in, E is the expected value, X and Y are two features, and are the mean and standard deviation of X, and are the mean and standard deviation of Y, respectively. PCC values range from -1 to 1: a PCC value greater than 0 indicates a positive correlation between X and Y; a value less than zero indicates a negative correlation between X and Y; and a value equal to zero indicates no linear correlation between X and Y. A larger absolute value of the PCC value indicates a closer correlation between the two features.

[0040] As shown in the figure, if there is a close correlation between two reference features (the absolute value of PCC is greater than the specified value), it means that the two features are redundant. We further judge the correlation between the two reference features and the target feature, and only retain the reference features with high correlation.

[0041] Step S320: quantifying the contribution of the reference features in the optimal feature subset to the model prediction capability, and deleting the reference features with the lowest contribution from the optimal feature subset.

[0042] This step uses the random forest algorithm to calculate feature importance, selects the top 50% features with the highest importance score, and obtains the dataset second_data; Random forest feature selection is based on its ensemble learning mechanism and the characteristics of decision trees. It implements feature importance assessment by quantifying the contribution of features to the model's predictive ability. It mainly looks at how much each feature contributes to each tree in the random forest, then takes the average, and finally compares the contribution of different features.

[0043] When implementing the evaluation, build a decision model, define a test set based on the feature set, calculate the initial performance of the decision model on the test set, randomly disrupt the value of a certain feature, keep other features unchanged, and re-evaluate the performance of the decision model. The greater the performance, the more important the feature is to the decision model. The importance score calculation formula is: in, Representation characteristics The importance score of reflects the decrease in model performance after the feature is destroyed. represents the performance index of the model on the original test set, Indicates that the feature The performance indicators of the model on the same test set after the values of are randomly shuffled.

[0044] In the embodiment of the present invention, the top 50% features are retained according to the feature importance ranking to update the optimal feature subset. After using this method to perform feature screening on first_data, both datasets still retain 6 features, such as Figure 4 As shown, it is defined as E_second_data and UTS_second_data.

[0045] Step S330: Use the Lasso regression model to eliminate features that have little impact on the target variable; through the sparsity constraint of the Lasso regression, eliminate non-critical features with coefficients close to zero to obtain the data set data.

[0046] Lasso regression feature selection is achieved by minimizing the objective function, which includes squared loss and L1 regularization. During the optimization process, Lasso regression reduces the coefficients of some features to 0, thereby achieving feature sparsity and eliminating features with little impact on the target variable.

[0047] The objective function of Lasso regression is: in, n is the sample size, is the number of features, For thei The true value of the sample, For the i The first sample j eigenvalues, For the j The coefficient of the feature ( is the intercept term), represents the regularization strength parameter.

[0048] The objective function of Lasso regression can be transformed into an equivalent form: , in, is the target variable, is the feature matrix, is the L2 norm (squared loss), is the L1 norm (L1 regularization term).

[0049] The objective function can be transformed into a constrained optimization problem: , λ and t Inversely proportional, that is λ The bigger, t The smaller it is, the stricter the constraint is and the sparser the coefficients are.

[0050] Lasso is mainly through the coefficient The absolute value of determines the importance of the feature. The larger it is, the more important the feature is. features are completely removed.

[0051] After screening, Young's modulus still has 6 features, and the ultimate tensile strength dataset still retains 5 features, such as Figure 4 As shown in the figure, together with the 7 manually set features and the target features calculated by simulation, the Young's modulus dataset contains 14 features, defined as E_data, and the ultimate tensile strength dataset contains 13 features, defined as UTS_data.

[0052] In subsequent steps, the E_data and UTS_data datasets are used to form a sample dataset to train the machine learning model.

[0053] Step S130: define multiple regression machine learning models, use sample data sets to train and verify the multiple regression machine learning models respectively, select the optimal regression machine learning model from the multiple regression machine learning models, and construct a mechanical property prediction model.

[0054] Multiple regression machine learning models include: Random Forest (RF), Extreme Gradient Boosting (XGBoost), Lightweight Gradient Boosting (LightGBM), and Category Feature Gradient Boosting (CatBoost).

[0055] In this step, four different machine learning models and algorithms are used to predict the mechanical properties of AlCoCrFeNi high-entropy alloy. In terms of data partitioning, 80% of the dataset is used as a training set and 20% as a test set.

[0056] The process of selecting the optimal regression machine learning model involves the following steps: 1) Set hyperparameter ranges: In this step, we first set hyperparameter ranges and use grid search techniques to fine-tune the parameters of each model. Typical model parameters include the number of weak learners in the ensemble (n_estimators), the maximum depth of a single decision tree (max_depth), and the learning rate (learning_rate) in the gradient boosting algorithm. This method exhaustively searches through a predefined space of parameter combinations to find the parameter combination that yields the best performance on the dataset.

[0057] 2) Verification method: Generally speaking, if an unoptimized model is used directly for learning and training, overfitting or underfitting will occur, and the prediction effect will be unstable. This invention adopts a 10-fold cross-validation method to reduce the problem of overfitting and underfitting. 10-fold cross-validation refers to: The dataset is randomly divided into 10 subsets (folds) of equal size to maintain the consistency of category distribution; Perform 10 rounds of training / validation, using 9 folds (90% of the data) for training and the remaining 1 fold (10% of the data) for validation in each round; The mean of 10 validations is set as the evaluation model indicator.

[0058] 3) Use mean squared error and goodness of fit to select the best model from multiple regression machine learning models: A good model usually has a high R 2 Value and lower MSE value.

[0059] The calculation formulas for mean square error and goodness of fit as evaluation indicators are: in,n Indicates the amount of data, represents the true value, Represents the predicted value of the model Indicates the average value.

[0060] During the evaluation process, the correlation coefficient was used As an indicator to measure the performance of the model, the range is 0 to 1. The closer the value is to 1, the better the variables in the equation are. The stronger the explanatory power, the better the model fits the data.

[0061] Figure 5 The fitting regression graphs of four machine learning models for predicting Young's modulus and ultimate tensile strength are provided: Among them, the CatBoost model has the best prediction effect on the test sets of both data sets. On the test set of Young's modulus, The value reached 0.997, on the tensile strength testing machine, The value reached 0.9896.

[0062] Step S140: Introduce model explanation: Because the machine learning model is a black box model, after determining the optimal regression machine learning model, the present invention also introduces the SHAP method to explain the prediction results of the model with excellent prediction performance on the test set. The model interpretation process is as follows: Figure 6 This method is based on the Shapley value in cooperative game theory and constructs an additive explanatory model. It considers all features as "contributors" and calculates the marginal contribution of each feature to the model output. It then explains the mechanical properties prediction model from both global and local levels.

[0063] The CatBoost model performs well in predicting Young's modulus and ultimate tensile strength. In this step, the SHAP technology is used to interpret the model, analyze the importance of each feature to the model, and perform a SHAP global interpretation of the Young's modulus and ultimate tensile strength predicted by the mechanical property prediction model: Figure 7 The following is a SHAP global interpretation diagram of the Young's modulus and ultimate tensile strength predicted by the CatBoost model. Figure 7 As shown in the figure, the horizontal axis is the shap value. Each point in the figure represents a sample. The redder the color, the larger the value of the feature itself, and the bluer the color. If the sample point is red and the shap value is less than 0, it means that the feature has a negative impact. If the sample point is red and the shap value is greater than 0, it means that the feature has a positive impact. and It has a negative effect on Young's modulus. Ev It has a positive effect on Young's modulus. T and It has a negative impact on the ultimate tensile strength. Ev It has a positive effect on Young's modulus.

[0064] Although the SHAP feature globally maps the relationship trends among environmental features, reference features, and simulated features, it does not provide detailed information about the relationships between individual features. In this step, a SHAP local interpretation is performed on the local dependencies of the important features used in the prediction of Young's modulus and ultimate tensile strength by the mechanical property prediction model: Figure 8 This chart shows the top three most important features for CatBoost Young's modulus prediction. The horizontal axis represents the feature value, and the vertical axis represents the SHAP value of each feature. Each feature has a critical value that divides the SHAP value into positive and negative regions. A SHAP value greater than zero indicates that improving that feature will increase the ultimate tensile strength of the AlCoCrFeNi high-entropy alloy. For example: T <600K, <0.0579, Ev >1.3967, which provides a reference for micro variable settings.

[0065] The present invention provides a method for constructing a high-entropy alloy mechanical property prediction model. The high-entropy alloy mechanical property prediction model constructed by the present invention realizes the prediction of high-entropy alloy mechanical properties by combining machine learning and multi-scale feature engineering. When constructing a high-entropy alloy mechanical property prediction model, first define a high-entropy alloy model with as rich a ratio as possible, fully obtain multiple types of feature information corresponding to each model to construct a multi-dimensional feature data set to ensure the integrity of the feature information; to ensure the accuracy of the feature information, a systematic strategy is adopted to perform feature screening and retain sample data sets to ensure the efficiency and accuracy of subsequent model training. On the other hand, multiple machine learning models are provided to ensure the selection of high-precision predictions of the model.

[0066] The above disclosures are only a few specific embodiments of the present invention. However, the present invention is not limited thereto. Any changes that can be conceived by those skilled in the art should fall within the scope of protection of the present invention.

Claims

1. A method for constructing a high entropy alloy mechanical property prediction model, characterized in that: The high entropy alloy is AlCoCrFeNi, and the construction method comprises the following steps: Constructing a high-entropy alloy model set, including: defining a proportional distribution of the high-entropy alloy components, determining a component structure of the high-entropy alloy model based on the proportional distribution, wherein a set of the component structures corresponds to a high-entropy alloy model; setting force field parameters for the high-entropy alloy model to stabilize the configuration; and obtaining mechanical characteristics of the high-entropy alloy model through stretching process simulation; Obtaining a multi-scale feature data set through the high entropy alloy model set; the multi-scale includes: environmental features, reference features, and simulation calculation features; Performing feature screening on the multi-scale feature dataset to obtain a sample dataset; wherein the target variables in the sample dataset are simulation calculation features, including Young's modulus features and ultimate tensile strength features; the feature screening refers to performing redundant feature screening and feature importance selection on reference features, and eliminating non-critical features; Define multiple regression machine learning models; use the sample data set to train and verify the regression machine learning models respectively, select the optimal regression machine learning model, and construct a mechanical property prediction model.

2. The method for constructing a high entropy alloy mechanical property prediction model according to claim 1, wherein: The method of defining the proportion distribution of the high entropy alloy components adopts a strategy of increasing the basic proportion and evenly dividing the remaining proportions, including: Define the base ratio, maximum ratio and growth step value; One component is extracted from the five components of the high-entropy alloy model in turn as a basic component, a single proportion value of the basic component is cyclically defined, and the remaining proportion values are evenly distributed among the other four components; wherein the remaining proportion value = 100% - single proportion value.

3. The method for constructing a high entropy alloy mechanical property prediction model according to claim 2, wherein: The loop definition includes the following steps: Step 1: Define a single ratio value as the basic ratio value; Step 2: defining the ratio of the basic components as a single ratio value; Step 3: Increment the single ratio value by the growth step value, and determine whether the single ratio value is greater than the maximum ratio value. If it is greater than the maximum ratio value, end the loop definition; otherwise, execute steps 2 to 3.

4. The method for constructing a high entropy alloy mechanical property prediction model according to claim 1, wherein: The environmental characteristics are determined based on temperature, strain rate and element ratio, the reference characteristics are determined based on the component element characteristics of the high entropy alloy model; and the simulation calculation characteristics are generated through a process of simulating a stretching process.

5. The method for constructing a high entropy alloy mechanical property prediction model according to claim 1, wherein: The reference characteristics include basic characteristics of five components, enhanced model mismatch values, weighted standard differences, local mismatch values and thermodynamic parameters; Wherein, the basic characteristics provide the basic physical and chemical properties of the components; The enhanced model mismatch value is used to quantify the macroscopic differences in mechanical and electrical properties between components; The weighted standard difference reflects the degree of dispersion of properties in the multi-component system; The local misfit value is used to reveal atomic-scale strain; The thermodynamic parameters are used to predict phase separation, solid solution or compound formation.

6. The method for constructing a high entropy alloy mechanical property prediction model according to claim 1, wherein: The feature screening includes: Calculate the linear correlation between the reference features and the target variable, eliminate redundant features, and establish the optimal feature subset; quantifying the contribution of the reference features in the optimal feature subset to the model prediction ability, and removing the reference features with the lowest contribution from the optimal feature subset; The Lasso regression model is used to eliminate features that have little impact on the target variable.

7. The method for constructing a high entropy alloy mechanical property prediction model according to claim 6, characterized in that: When removing redundant features, the PCC value calculated on the original data set is used for screening; If the two reference features are closely correlated, the correlation between the two reference features and the target feature is further determined, and only the reference features with the largest correlation are retained.

8. The method for constructing a high entropy alloy mechanical property prediction model according to claim 1, wherein: The multiple regression machine learning models include: random forest, extreme gradient boosting, lightweight gradient boosting and category feature gradient boosting algorithms.

9. The method for constructing a high entropy alloy mechanical property prediction model according to claim 1, wherein: The selecting of the optimal regression machine learning model includes: Find the parameter combination that performs best on the dataset; The 10-fold cross validation was used to extract the mean of 10 validation results as the evaluation model indicator; The mean square error and goodness of fit are used to select the optimal model from the multiple regression machine learning models.

10. The method for constructing a high entropy alloy mechanical property prediction model according to claim 1, wherein: After the mechanical property prediction model is constructed, a SHAP global interpretation is performed on the Young's modulus and ultimate tensile strength predicted by the mechanical property prediction model, and a SHAP local interpretation is performed on the local dependence of important features of the Young's modulus and ultimate tensile strength predicted by the mechanical property prediction model.

Images