A method for constructing a strength-plasticity analysis model for composite materials

Abstract

The application discloses a kind of construction methods for strength-plasticity analysis model of composite material, comprising: high-entropy alloy / graphene composite material is generated simulation variable by multidimensional parameter space sampling, simulation variable is modeled by molecular dynamics, executes simulation, extracts feature data and label data to construct sample set;The feature data of sample set includes: atomic basic attribute, chemical property, physical property and structural feature;Label data includes: ultimate tensile strength and ductility;Define ultimate tensile strength regression prediction model set, train each model of ultimate tensile strength regression prediction model set using sample set, constitute strength-plasticity analysis model;According to strength-plasticity analysis model, obtain the most adaptive simulation variable according to hyperparameter optimization;According to the above technical solution, it can break the traditional high-entropy alloy strength-plasticity trade-off limit, provide universal design paradigm for high-entropy alloy-based composite material, and extend to other two-dimensional material reinforced system.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent design of metal matrix composites in the field of material science and engineering, in particular, to a construction method of a strength-plasticity analysis model for composites. BACKGROUND

[0002] As a new material system that breaks through the traditional alloy design paradigm, high-entropy alloys form complex solid solutions through multi-principal-element atomic ratio mixing, and exhibit excellent properties such as high strength, high temperature resistance and corrosion resistance by virtue of core advantages such as thermodynamic high-entropy effect and kinetic lag diffusion effect, and have broad application prospects in aerospace, energy equipment and other fields. However, the inherent strength-plasticity trade-off of high-entropy alloys is still the core bottleneck of engineering application, which is manifested in that high-strength materials often have insufficient plasticity, and high-plasticity materials have low strength, which is difficult to meet the stringent requirements of structural parts on comprehensive mechanical properties.

[0003] Currently, the main strategies to improve the performance of high-entropy alloys include alloy composition optimization and second phase reinforcement. Among them, graphene is widely used to construct high-entropy alloy / graphene composites due to its excellent mechanical properties and two-dimensional structure, and the strength and plasticity are synergistically improved through interface strengthening and dislocation pinning mechanisms. However, the analysis and regulation of the strength and plasticity performance of the composite material face multiple challenges: for example, due to the need for a large amount of time and resources in the traditional "melting-characterization" process, and the difficulty in observing nanoscale deformation mechanisms such as dislocation evolution and interface failure behavior, the experimental trial-and-error method for analysis and regulation has the problem of low efficiency; molecular dynamics simulation can reveal atomic-level behavior, but the exploration of parameter space (such as composition, temperature, graphene orientation angle and volume fraction) is limited by computing resources, and a single simulation takes hundreds of hours, so the pure simulation method has great limitations; the existing machine model is based on a single scale feature (such as composition parameters), but because it lacks comprehensive consideration of thermodynamics and structural characteristics, and has poor interpretability, it is difficult to analyze the coupling mechanism of multiple factors, and there is a problem of insufficient data to achieve satisfactory results.

[0004] Therefore, there is an urgent need for a technical solution that integrates multi-disciplinary technologies to break through the bottleneck of traditional empirical design, improve the inherent strength and plasticity inversion relationship of traditional high-entropy alloys, and facilitate balanced analysis of alloy material strength and plasticity during the design process. SUMMARY

[0005] To achieve the above-mentioned purpose, the present application provides a construction method of a strength-plasticity analysis model for composites, comprising the following steps:

[0006] The high-entropy alloy / graphene composite material is sampled in a multi-dimensional parameter space to generate simulation variables, molecular dynamics modeling is performed on the simulation variables, simulation is performed, feature data and label data after simulation are extracted to construct a sample set, the simulation variables include element concentration of the high-entropy alloy matrix, molecular dynamics simulation temperature, graphene volume fraction and orientation angle of graphene and the high-entropy alloy matrix, the feature data of the sample set is composed of large-class features, the large-class features include atomic basic attributes, chemical properties, physical properties and structural features, and the label data of the sample set includes ultimate tensile strength and ductility.

[0007] A set of ultimate tensile strength regression prediction models is defined, each model of the set of ultimate tensile strength regression prediction models is trained using the sample set, an optimal model is selected to constitute a strength-plasticity analysis model.

[0008] According to the strength-plasticity analysis model, hyperparameter optimization is performed to obtain the most suitable simulation variables of the strength-plasticity analysis model.

[0009] The multi-dimensional parameter space sampling is realized by Sobol quasi-random sequences.

[0010] The element concentration of the high-entropy alloy matrix includes the concentrations of Fe, Ni, Cr, Co and Cu.

[0011] The molecular dynamics simulation temperature ranges from 0 to 1100K.

[0012] The graphene volume fraction ranges from 0.4 to 3.4 vol%.

[0013] The orientation angle of graphene and the high-entropy alloy matrix ranges from 0 to 45°.

[0014] Further, the molecular dynamics modeling includes: initially constructing the atomic configuration of the composite material into a high-entropy alloy model, then inserting graphene into the high-entropy alloy matrix to form a high-entropy alloy / graphene composite model.

[0015] The high-entropy alloy / graphene composite material simulated by the high-entropy alloy / graphene composite model is composed of FCC, and the graphene is embedded in the middle of the high-entropy alloy matrix; the graphene volume fraction is represented as the volume of graphene divided by the volume of the high-entropy alloy / graphene composite material, that is: wherein, is the volume fraction of graphene, is the length of graphene, is the width of graphene, is the thickness of single-layer graphene, is the volume of the high-entropy alloy / graphene composite material.

[0016] The process of performing simulation includes: obtaining a stress-strain curve corresponding to the simulation variable after performing 0 pressure simulation and tensile simulation; wherein, during 0 pressure simulation, x and the pressure in the z-axis direction is kept at 0 bar; during tensile simulation, a 1x10 9 s -1 strain rate is applied in the positive direction of the y-axis and kept constant;

[0017] The simulated environment implementation includes: performing a minimum processing on the energy of the high-entropy alloy / graphene composite material system;

[0018] After relaxing for 50 ps in the isothermal-isobaric ensemble, further relaxing for 50 ps in the isothermal-isochoric ensemble until the system energy reaches thermodynamic equilibrium.

[0019] Further, the stress-strain curve is used to extract the ultimate tensile strength and ductility data corresponding to the simulation variable; and the ultimate tensile strength and ductility data are used to construct label data;

[0020] According to the ductility label, the ductility type includes insufficient ductility and good ductility; if the sample amount of insufficient ductility and good ductility is unbalanced, a SMOTE oversampling method is used to synthesize minority class samples to balance the data set.

[0021] The atomic basic attributes include atomic number, atomic mass, metal radius, volume, electron affinity, and valence electron concentration; the chemical properties include Pauling electronegativity, Mulliken electronegativity, ionization energy, mixing enthalpy, mixing entropy, and mismatch entropy;

[0022] The physical properties include density, melting point, specific heat capacity, thermal diffusion coefficient, elastic modulus, and Poisson's ratio;

[0023] The structural features include atomic size difference, radius difference, electronegativity difference, and theoretical molar volume;

[0024] The atomic basic attributes, chemical properties, physical properties, and structural features are integrated, and linear mixing rule, derivative mixing rule, bias mixing rule, and difference mixing rule are used to calculate and generate first feature data;

[0025] The first feature data is subjected to feature screening to form feature data; the feature data includes used for predicting the ultimate tensile strength, used for predicting the ductility; wherein, linear mixing of bulk modulus, derivative mixing of cohesive energy, linear mixing of vacancy formation energy, mixing enthalpy, TTo simulate temperature 、 To simulate the orientation angle of graphene, To simulate the volume fraction of graphene.

[0026] Further, the feature screening step includes: first step: filtering out low-contribution feature data with variance below a set threshold in the first feature data through variance threshold, and retaining data with significant variability in the first feature data;

[0027] Second step: calculate the feature importance of each element in the first feature data using the random forest model, and identify the features that have an impact on the target variable;

[0028] Third step: quantifying the linear correlation between features through Pearson correlation analysis, updating the first feature data after removing highly redundant feature data;

[0029] Fourth step: using recursive feature elimination combined with model accuracy feedback, updating the first feature data after iteratively removing secondary features in the first feature data;

[0030] Fifth step: using exhaustive search to search for the optimal combination of the first feature data to ensure the balance between prediction performance and interpretability of the feature subset, and constructing the feature data.

[0031] Further, the limit tensile strength regression prediction model set includes a limit tensile strength regression prediction model and a ductility classification prediction model;

[0032] The optimal model is selected by evaluating the prediction performance through the determination coefficient and the root mean square error, which is represented as:

[0033]

[0034] Wherein, The number of samples, The true value, The average value of The predicted value, The value of The value of And The number of values, The determination coefficient RMSE is the root mean square error;

[0035] The prediction performance is represented as: The value range of is [0, 1], The value of is closer to 1, indicating that the model fits the data better, and in the case of complete fitting, Close to 1;

[0036] The smaller the RMSE value, the smaller the prediction error of the model; The value range of the value is 0 to 1, and the value closer to 1 indicates that the classification effect of the model is more ideal, and the proportion of correct classification is higher; The value range is also 0 to 1, The value closer to 1 means that the performance of the model is better.

[0037] Further, the hyperparameter optimization refers to reducing the simulation variable space of the high-entropy alloy / graphene composite material through domain knowledge constraint and SHAP explainability method, and optimizing the simulation variable through genetic algorithm to obtain a coupling parameter combination with higher ultimate tensile strength and ductility.

[0038] Further, after obtaining the most suitable simulation variable of the strength-plasticity analysis model, a microstructure analysis is further performed to realize closed-loop verification; wherein, the closed-loop verification is used to verify whether the ultimate tensile strength and ductility corresponding to the most suitable simulation variable of the strength-plasticity analysis model are better than the original data set;

[0039] The microstructure analysis refers to further verifying the deformation mechanism through common neighbor analysis and dislocation density analysis.

[0040] According to the present application, the composition elements, temperature conditions, orientation angle and volume fraction of graphene of the high-entropy alloy / graphene material and the correlation with the mechanical properties are extracted to form a feature set, a machine learning model, domain knowledge, SHAP explainability analysis and optimization algorithm are introduced to construct an optimal strength-plasticity analysis model. Through the use of the strength-plasticity analysis model, the prediction of the ultimate tensile strength and ductility can be realized, and multi-scale collaborative optimization can also be achieved, breaking the traditional high-entropy alloy strength-plasticity trade-off limit. At the same time, a universal design paradigm is provided for high-entropy alloy-based composite materials, which can be extended to other two-dimensional material reinforced systems, and the optimized materials are suitable for extreme environment components such as aircraft engine blades, providing a new path for the development of high-performance composite materials. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is a step diagram of a method for constructing a strength-plasticity analysis model for composite materials according to an embodiment of the present application;

[0042] Figure 2 is a closed-loop verification flowchart of a strength-plasticity analysis model for composite materials according to an embodiment of the present application;

[0043] Figure 3 is a sample distribution diagram based on Sobol quasi-random sequence sampling according to an embodiment of the present application;

[0044] Figure 4is a schematic diagram of molecular dynamics modeling with different element concentrations, graphene orientation angles and graphene volume fractions according to an embodiment of the present application;

[0045] Figure 5 is a stress-strain curve obtained based on molecular dynamics simulation according to an embodiment of the present application;

[0046] Figure 6 is a distribution relationship diagram of ultimate tensile strength and ductility statistically obtained from the stress-strain curve according to an embodiment of the present application;

[0047] Figure 7 is a schematic diagram of a five-step feature screening method for reducing feature dimension according to an embodiment of the present application;

[0048] Figure 8 is a result of random forest feature importance, Pearson correlation analysis and exhaustive feature screening method in the five-step feature screening method according to an embodiment of the present application;

[0049] Figure 9 is a performance evaluation and comparison diagram of different classification models in the ultimate tensile strength regression prediction model set according to an embodiment of the present application;

[0050] Figure 10 is a prediction result schematic diagram of the optimal regression model and the classification model in the ultimate tensile strength regression prediction model set according to an embodiment of the present application;

[0051] Figure 11 is a parameter optimization strategy diagram according to an embodiment of the present application;

[0052] Figure 12 is a comparison diagram of coupled parameter combinations and parameters and performances of composite materials after parameter optimization according to an embodiment of the present application;

[0053] Figure 13 is a distribution of ultimate tensile strength and ductility in the original data set corresponding to the stress-strain curve obtained by molecular dynamics simulation according to an embodiment of the present application;

[0054] Figure 14 is a microstructure evolution analysis diagram realized by the strength-plasticity analysis model according to an embodiment of the present application;

[0055] Figure 15 is a dislocation schematic diagram in two optimized high-entropy alloys / graphenes according to an embodiment of the present application. DETAILED DESCRIPTION

[0056] The present application aims at the composition elements, temperature conditions, orientation angle and volume fraction of graphene of high-entropy alloy / graphene material, and constructs sample features related to mechanical properties, combines machine learning model, DK field knowledge, SHAP explainability analysis and features of optimization algorithm, selects the optimal machine learning model structure, constructs the optimal model that can balance strength and plasticity analysis, and determines the most suitable material parameters for strength-plasticity analysis according to the features of the optimal model, so that the comprehensive analysis of strength and plasticity can be accurately and efficiently realized in the process of designing high-entropy alloy / graphene composite materials.

[0057] The specific implementation mode of the present application will be described in detail below in combination with the drawings of the specification.

[0058] Figure 1 A construction method step diagram of a strength-plasticity analysis model for a composite material is provided, specifically comprising:

[0059] Step S100: generating simulation variables by sampling a multi-dimensional parameter space of a high-entropy alloy / graphene composite material, performing molecular dynamics modeling on the simulation variables, executing simulation, extracting feature data and label data after simulation to construct a sample set;

[0060] 1) First, determine the simulation variables for the molecular dynamics simulation of the high-entropy alloy / graphene composite material, and perform sample sampling; the simulation variables proposed by the present application include: element concentration of high-entropy alloy matrix, molecular dynamics simulation temperature, graphene volume fraction and orientation angle of graphene and high-entropy alloy matrix; for example, the element concentration of the high-entropy alloy matrix includes the concentrations of Fe, Ni, Cr, Co and Cu; the molecular dynamics simulation temperature ranges from 0 to 1100K; the graphene volume fraction ranges from 0.4 to 3.4 vol%; the orientation angle of graphene and high-entropy alloy matrix ranges from 0 to 45°.

[0061] The sample sampling method selects Sobol quasi-random sequence for multi-dimensional parameter space sampling, and in the embodiment of the present application, 1000 groups of data samples are generated based on Sobol quasi-random sequence. Figure 3 The results after sampling are shown, and according to the results, compared with the traditional Monte Carlo method, this sequence algorithm can generate a uniformly distributed sample set with low difference, which significantly improves the coverage rate and convergence speed of high-dimensional parameter space.

[0062] 2) The process of molecular dynamics modeling includes: initially constructing the atomic configuration of the composite material into a high-entropy alloy model, then inserting graphene into the high-entropy alloy matrix to form a high-entropy alloy / graphene composite model.

[0063] The present step can be realized by using open source software such as a large-scale atomic / molecular parallel simulator (such as LAMMPS) to perform molecular dynamics modeling on 1000 groups of simulation variables containing different molecular dynamics characteristics in the embodiments of the present application, and to perform molecular dynamics simulation calculation. The constructed molecular dynamics model is a periodic boundary condition with a size of 10.108 nm³, as shown in Figure 4 The size of the 1000 samples is the same, but the simulation variables of the molecular dynamics characteristics such as the element concentration and the graphene volume fraction are different.

[0064] In the simulation implementation process, first, the atomic configuration of the composite material is initially constructed as a high-entropy alloy model, and then graphene is inserted into the high-entropy alloy matrix to form a high-entropy alloy / graphene composite model; the high-entropy alloy / graphene composite material is composed of a single-phase face-centered cubic (FCC), and the FCC lattice spacing is 3.61 Å; in each model, the graphene is embedded in the middle of the high-entropy alloy matrix; different graphene volume fractions can be obtained by different side lengths; in order to achieve the required volume fraction, the length of the graphene in the stretching direction is fixed at 10.108 nm, and the width varies from 1.25 nm to 10.108 nm, and the volume fraction varies from 0.4 vol% to 3.4%; the orientation angle of the graphene varies between 0-45°; the graphene volume fraction is represented as the volume of the graphene divided by the volume of the high-entropy alloy / graphene composite material, which is: wherein, is the volume fraction of the graphene, is the length of the graphene, is the width of the graphene, is the thickness of the single-layer graphene, is the volume of the high-entropy alloy / graphene composite material.

[0065] On the other hand, a suitable potential function is selected to improve the accuracy of the simulation results: the adaptive intermolecular reaction empirical bond order (AIREBO) potential is used to describe the interaction between C-C in graphene, the embedded atom method (EAM) is used to describe the interaction between Fe, Ni, Cr, Co, Cu atoms in the HEAs matrix, and the LJ potential is used to describe the interaction between HEAs and carbon atoms.

[0066] 3) LAMMPS is used to perform all simulation processes in the present application; OVITO visualization software is used for post-processing of atomic trajectories, and combined with common neighbor analysis (CNA) and dislocation analysis (DXA) for structure analysis and visualization.

[0067] The process of performing simulation includes: after performing 0 pressure simulation and tensile simulation, the stress-strain curve corresponding to the simulation variable is obtained; wherein, in the 0 pressure simulation, xand the pressure in the z-axis direction is kept as 0 bar; during the tensile simulation, 1 x 10 9 s -1 at a constant strain rate and keeps constant;

[0068] The simulated environment implementation includes the following steps:

[0069] The energy of the high-entropy alloy / graphene composite material system is minimized;

[0070] After relaxing for 50 ps in the isothermal-isobaric ensemble, the system is further relaxed for 50 ps in the isothermal-isochoric ensemble until the system energy reaches thermodynamic equilibrium.

[0071] After the tensile simulation, the stress-strain curves of 1000 samples can be obtained, as shown in Figure 5.

[0072] The stress-strain curve can be used to extract the ultimate tensile strength and ductility data corresponding to the simulation variables (such as Figure 6 shown); the ultimate tensile strength and ductility data are used to construct the label data (i.e., the response variable);

[0073] According to the ductility label, the ductility type includes insufficient ductility and good ductility; for example, samples with a ductility greater than 13.4% are labeled as “good ductility”, and samples with a ductility less than 13.4% are labeled as “insufficient ductility”.

[0074] If the number of samples with insufficient ductility and good ductility is imbalanced, for example, the number of samples with “insufficient ductility” only occupies a small part, that is, a serious class imbalance phenomenon occurs, the SMOTE oversampling method is provided in the present application to synthesize minority class samples to balance the data set, and the problem of ductility class imbalance is solved.

[0075] SMOTE is a widely used oversampling method, which balances the data set by synthesizing minority class samples to alleviate the class imbalance problem; specifically, SMOTE generates new samples by interpolating between minority class samples, which can increase the diversity of minority classes while avoiding overfitting. By applying the SMOTE technology, more “good ductility” high-entropy alloy / graphene composite material samples are generated, thereby balancing the class distribution in the data set.

[0076] 4) After determining the label data (ultimate tensile strength and ductility) of the sample set through the simulation process, the feature data of the sample set, i.e., 1000 groups of samples, needs to be extracted:

[0077] The feature data of the sample set in the present application is composed of macro-features, including atomic basic properties, chemical properties, physical properties, and structural features.

[0078] Firstly, the atomic basic properties of the samples are determined, including atomic number, atomic mass, metallic radius, volume, electron affinity, and valence electron concentration; chemical properties include Pauling electronegativity, Mulliken electronegativity, ionization energy, mixing enthalpy, mixing entropy, and misfit entropy; physical properties include density, melting point, specific heat capacity, thermal diffusivity, elastic modulus, and Poisson's ratio; structural characteristics include atomic size difference, radius difference, electronegativity difference, and theoretical molar volume; among them, the formula of part of the parameters is:

[0079] ,

[0080] ,

[0081] ,

[0082] wherein, =8.314.

[0083] In this step, the atomic basic properties, chemical properties, physical properties, and structural characteristics are integrated, and 19 elements are selected to calculate the first feature data using linear mixing rules, derivative mixing rules, deviation mixing rules, and difference mixing rules, which are represented as:

[0084] ,

[0085] ,

[0086] ,

[0087] ,

[0088] wherein, and are the concentration of each component and the corresponding parameter value, is the linear mixing result, is the derivative mixing result, is the deviation mixing result, is the difference mixing result.

[0089] The first feature data contains 93 initial characteristic factors including thermodynamic parameters (mixing entropy) and structural characteristics (atomic size difference), combined with angular molecular dynamics simulation temperature, graphene volume fraction, and orientation angle of graphene and high-entropy alloy matrix, to construct a feature parameter space with a size of 96.

[0090] 5) Further, the first feature data is screened to form the feature data; the implementation of feature screening adopts a five-step feature screening strategy, as shown in Figure 7 .

[0091] The step of feature screening includes:

[0092] First step: filter out low-contribution feature data with variance below a set threshold in the first feature data by variance threshold, and retain data with significant variability in the first feature data. After the first step, 47 features are left for predicting ultimate tensile strength, and 47 features are left for predicting ductility.

[0093] Second step: calculate the feature importance of each element in the first feature data using the random forest model, i.e. importance score, to identify features that have an impact on the target variable, such as larger features. After the second step, 20 features are left for predicting ultimate tensile strength and ductility.

[0094] Third step: perform correlation analysis, i.e. quantify the linear correlation between features by Pearson correlation analysis. After removing highly redundant feature data (such as |r|>0.95), update the first feature data. At this time, 16 features are left for predicting ultimate tensile strength and ductility.

[0095] Fourth step: use recursive feature elimination combined with model accuracy feedback to iteratively remove secondary features in the first feature data, and update the first feature data. At this time, 8 features are left for predicting ultimate tensile strength, and 7 features are left for predicting ductility.

[0096] Fifth step, exhaustive feature elimination, i.e. search the first feature data by exhaustive method to find the optimal combination to ensure the balance between prediction performance and interpretability of the feature subset, and constitute feature data. At this time, 7 features are left for predicting ultimate tensile strength, and 6 features are left for predicting ductility.

[0097] Figure 8 The final results of calculating feature importance score by random forest model, quantifying linear correlation between features by Pearson correlation analysis, and exhaustive search are shown.

[0098] Through the feature screening of this step, 7 key features (E1, E2, E3, E4, E5, E6, E7) are effectively selected from 93 initial features (i.e. first feature data) to constitute feature data for predicting ultimate tensile strength. Similarly, 6 key features (E1, E2, E3, E4, E5, E6) are selected to constitute feature data for predicting ductility. , , , T、 , Among them, is the linear mixture of bulk modulus, is the derivative mixture of cohesive energy, is the linear mixture of vacancy formation energy, for the mixing enthalpy, T for the simulated temperature 、 for the orientation angle of graphene, for the volume fraction of graphene.

[0099] Through step S100, the key features and their corresponding labels obtained can constitute a data set to realize training of the limit tensile strength regression prediction model set. Further, 70% of the data in the data set is used as a training set, and 30% of the data is used as a test set.

[0100] Step S110: defining a limit tensile strength regression prediction model set;

[0101] The limit tensile strength regression prediction model set includes a plurality of limit tensile strength regression prediction models and a plurality of ductility classification prediction models.

[0102] The limit tensile strength regression prediction model selects 11 different machine learning models, including linear regression (LR_R), ridge regression (RR_R), Bayesian ridge regression (BR_R), support vector regression (SV_R), K-nearest neighbor (KNN_R), light gradient boosting machine regression (LightGBM_R), extreme gradient boosting tree regression (XGBoost_R), decision tree (DT_R), extreme random tree regression (ET_R), gradient boosting decision tree (GBDT_R) and random forest regression (RF_R);

[0103] The ductility classification prediction model selects 12 different machine learning models, including logistic regression (LR_C), linear discriminant analysis (LDA_C), decision tree (DT_C), random forest (RF_C), gradient boosting (GBDT_C), extreme gradient boosting tree classification (XGBoost_C), light gradient boosting machine classification (LightGBM_C), classification boosting algorithm (CatBoost_C), Gaussian naive Bayes (GNB_C), K-nearest neighbor (KNN_C), multilayer perception machine classifier (MLP_C) and support vector machine (SVM_C).

[0104] Step S120: training each model of the limit tensile strength regression prediction model set using the sample set, selecting the optimal model, and constituting a strength-plasticity analysis model;

[0105] In the present application, the optimal model is selected by the prediction performance of the regression model, which is specifically evaluated by the coefficient of determination (R 2 ) and the root mean square error (RMSE), and is expressed as:

[0106]

[0107] wherein, the number of samples, the true value, the average value of , the average value of the predicted value, is and the number of values, the determination coefficient RMSE is the root mean square error;

[0108] The prediction performance is represented as: the value of [0, 1], the value of 1 indicates that the model is better fitted with the data, and in the case of complete fitting, close to 1;

[0109] The smaller the RMSE value, the smaller the prediction error of the model; the value of 0 to 1, the closer the value to 1, the more ideal the classification effect of the model, and the higher the proportion of correct classification; the value range is also 0 to 1, the closer the value to 1, the better the performance of the model.

[0110] Figure 9 The performance evaluation results of various machine learning models are shown. Among them, Figure 9 (a)-(d) and (e)-(h) correspond to the performance of regression models and classification models. It can be found that SV_R, LGBM_R, XGBoost_R and other models have shown high prediction accuracy, R 2 >0.9, which indicates that they can well fit the data and have high accuracy in predicting the target variable. In the classification model, RF_C, XGBoost_C, LGBM_C and CatBoost_C also show an Accuracy higher than 0.9, indicating that they can effectively distinguish different categories in the classification task and have high classification reliability.

[0111] In this step, one model is selected from each of the models with Accuracy higher than 0.9 and R 2 >0.9 as the optimal model, and the trained optimal model is constructed as a strength-plasticity analysis model.

[0112] Step S130: According to the strength-plasticity analysis model, the hyperparameter optimization is performed to obtain the most suitable simulation variable of the strength-plasticity analysis model.

[0113] In this step, the size of the high-entropy alloy / graphene composite coupling parameter space is reduced by combining domain knowledge constraints and SHAP explainability analysis, and combined with genetic algorithm for parameter optimization. In order to concentrate resources, in this step, the model that has shown high performance potential is optimized, that is, the model with an evaluation score greater than 0.9 in step S120 is optimized by hyperparameter optimization to further improve the prediction performance of the model.

[0114] The hyperparameter optimization realizes the coupling search strategy by using the Bayesian optimization algorithm, and optimizes the operation based on genetic algorithm. The coupling search strategy contains four different conditions of search experiments to comprehensively evaluate the influence of different constraint conditions on the optimization effect. The four different conditions are: the whole parameter space without constraint, the constraint based on domain knowledge, the constraint based on domain knowledge and SHAP explainability analysis, and the combination of the output judgment of the classification model based on the constraint of domain knowledge and SHAP explainability analysis.

[0115] The hyperparameter optimization refers to reducing the simulation variable space of the high-entropy alloy / graphene composite by combining the domain knowledge constraint and the SHAP explainability method, optimizing the simulation variable by combining the genetic algorithm, and obtaining the coupling parameter combination with higher ultimate tensile strength and ductility. The specific steps are as shown in Figure 11

[0116] First, the original data is preprocessed, and the characteristic factors required by the model are calculated (i.e. the first characteristic data before feature selection); Then, two models XGBoost_R and GBDT_C are trained to provide a basis for subsequent optimization operations; Among them, XGBoost_R has the highest prediction accuracy compared with other regression models, so it is used to predict the tensile strength, named ML-UTS (machine learning - Ultimate tensile strength), and GBDT_C is applied to predict the ductility, named ML-EL (machine learning - elongation); Then, the parameters of the genetic algorithm are initialized, including population size, iteration number and key parameters such as crossover and mutation rate; In the optimization process, the output value of the XGBoost_R model is used as the fitness function of the genetic algorithm to evaluate the pros and cons of individuals. The algorithm continuously optimizes the population through iteration, and each iteration includes selection, crossover and mutation operations to gradually improve the overall fitness of the population. After calculating the fitness of the initial population, optimization experiments are carried out under four different constraint conditions; Each constraint condition imposes different restrictions on the optimization process, thereby affecting the final optimization result. When the preset iteration stopping condition is met, the algorithm outputs the best coupling parameter result and its corresponding fitness.

[0117] ​Among them, the domain knowledge mainly refers to reducing the parameter space through the valence electron concentration (VEC>8.225); the SHAP explainability method (SHAP) <21.5°, T <580K and <1.81vol%∪ >3vol%) refers to the range recommended by the SHAP method to reduce the parameter space.

[0118] Specifically, the population size of the genetic algorithm is 100, the crossover mutation rate is 0.1, and the iteration number is 100 times.

[0119] In the process of hyperparameter optimization, the coupling parameter space of high-entropy alloy / graphene composite materials is reduced through the domain knowledge constraint and the SHAP explainability method, and the coupling parameters are optimized by combining the genetic algorithm, so that the coupling parameter combination with higher ultimate tensile strength and ductility can be obtained.

[0120] After the hyperparameter adjustment, XGBoost_R and GBDT_C show the highest prediction performance in the regression model and the classification model, and their prediction results on the training set and the test set are shown in Figure 10 Therefore, they are used to predict the ultimate tensile strength and ductility, respectively.

[0121] Further, as shown in Figure 12 , through XGBoost_R and GBDT_C, two optimal high-entropy alloy / graphene composite materials Fe 10.7 Ni 29.6 Cr 22.8 Co 21.5 Cu 15.4 / Gr (3.26vol%, 0.04°) and Fe 24.3 Ni 12.3 Cr 23.6 Co 11.3 Cu 25.6 / Gr (0.62vol%, 1.8°) can be determined.

[0122] Overall, in the whole process of the application, data sampling and molecular dynamics simulation, feature engineering and machine learning modeling, coupling parameter space optimization and search to obtain the most suitable simulation variables of the strength-plasticity analysis model are performed; and microstructure analysis is performed, so that each link realizes closed-loop verification, as shown in Figure 2 .

[0123] The closed-loop verification refers to performing molecular dynamics simulation based on the coupling parameters of the two optimal high-entropy alloy / graphene composite materials. As shown in Figure 13As shown, the stress-strain curves of the two optimal high-entropy alloy / graphene composites are obtained by molecular dynamics simulation, and the ultimate tensile strength and ductility are obtained by statistics. As can be seen from the figure, the mechanical properties of the two composites are better than the original data set (ultimate tensile strength > 14.5 Gpa, ductility > 13.4%). This shows that by combining the DK domain knowledge and the SHAP explainability method for optimization, the mechanical properties of the high-entropy alloy / graphene composite can be significantly improved, and it has reached a new level in high strength and high ductility.

[0124] Microstructure analysis refers to further verification of deformation mechanism through common near neighbor analysis and dislocation density analysis.

[0125] The present application uses CNA to count the change of the atomic number fraction of FCC, body-centered cubic (BCC), hexagonal close-packed (HCP) and disordered structure (OTHER) in the two optimal high-entropy alloy / graphene composites. Figure 14 The evolution process of the numerical fraction of the four structures with strain is shown respectively: during the deformation process, the atomic proportion of FCC structure shows a gradually decreasing trend. This shows that with the increase of strain, the atoms originally existing in FCC structure gradually undergo structural transformation; at the same time, the atomic proportion of HCP and OTHER shows an upward trend. This is related to the stress state and energy change experienced by the material during deformation. Specifically, with the increase of strain, the stress distribution in the material changes, prompting some atoms to transform from the FCC structure with higher energy to the HCP structure with lower energy, and also producing some disordered atomic arrangement.

[0126] The present application also uses DXA to count the dislocations in the two optimized high-entropy alloy / graphene, as shown in the specific Figure 15 Dislocation, as an inherent defect, has a profound influence on the mechanical properties and deformation behavior of materials.

[0127] Specifically, the evolution of 1 / 2<110> perfect dislocation, 1 / 3<100> (Hirth dislocation), 1 / 3<111> (Frank dislocation), 1 / 6<110> (Stair-rod dislocation), 1 / 6<112> (Shockley dislocation) and OTHER dislocation in the two composites during the tensile process is shown, and the evolution of dislocation density is shown in the inset. As can be seen, in the two systems, Shockley dislocation, Stair-rod dislocation and OTHER dislocation are the dominant types, among which Shockley dislocation has the highest proportion and shows a first rising and then falling trend, and its peak value is closely related to the yield point of the material. It is worth noting that Fe 10.7 Ni 29.6 Cr 22.8 Co21.5 Cu 15.4 / Gr (3.26vol%, 0.04°) The dislocation density of the whole system is significantly lower than that of Fe 24.3 Ni 12.3 Cr 23.6 Co 11.3 Cu 25.6 / Gr (0.62vol%, 1.8°) This is in sharp contrast to its higher strength. This phenomenon can be explained by the dislocation strengthening mechanism, that is, although high dislocation density can improve the strength, excessive dislocation tangling leads to early strain localization.

[0128] Microstructure analysis shows that smaller graphite training orientation angle corresponds to better strength, and larger graphite volume fraction hinders the propagation of dislocations. When dislocations move up and down on the graphene layer, there is a high probability of collision with other dislocations to form dislocation tangling and enhance the strength of the material. In contrast, smaller graphene fails to completely occupy a plane, resulting in less interface stress concentration, thereby reducing the initiation and expansion of dislocations, making the material more prone to plastic deformation during stretching, thereby improving ductility.

[0129] The method for constructing a strength-plasticity analysis model for a composite material provided by the present application can construct an optimal model suitable for high-entropy alloy / graphene composite material design, which integrates multi-scale data and multi-technical means, significantly improves the calculation speed, and ensures high prediction accuracy. Compared with the construction of traditional network models, the method provided by the present application has lower cost and shorter cycle. Not only can it quickly and accurately predict the ultimate tensile strength and ductility before the experimental preparation of high-entropy alloy / graphene composite materials, but also can synergistically improve the strength and plasticity. At the same time, the model exhibits excellent performance on the data set, which not only verifies the effectiveness of the method, but also lays the foundation for its expansion in other different high-entropy alloy systems.

[0130] The above disclosure is only a few specific embodiments of the present application, but the present application is not limited thereto, and any changes that can be thought of by those skilled in the art should fall within the scope of the present application.

Claims

1. A method for constructing a strength-plasticity analysis model for a composite material, characterized by, The method comprises the following steps: The high-entropy alloy / graphene composite material is sampled in a multi-dimensional parameter space to generate simulation variables, molecular dynamics modeling is performed on the simulation variables, simulation is executed, feature data and label data after simulation are extracted to construct a sample set; the simulation variables include element concentration of the high-entropy alloy matrix, molecular dynamics simulation temperature, graphene volume fraction and orientation angle of graphene and the high-entropy alloy matrix; the feature data of the sample set is composed of macro features, and the macro features include atomic basic attributes, chemical properties, physical properties and structural features; the feature data includes Δx 1_BM , Fe, Ni, Co, T, θ, V for predicting the ultimate tensile strength; x R_CE , x L_VFE , S mix , T, θ, V for predicting the ductility; wherein Δx 1_BM is linear mixing of bulk modulus, x R_CE is derivative mixing of cohesive energy, x L_VFE is linear mixing of vacancy formation energy, S mix is mixing enthalpy, T is simulation temperature, θ is the orientation angle of graphene, and V is the volume fraction of graphene; the label data of the sample set includes the ultimate tensile strength and the ductility; Defining a limit tensile strength regression prediction model set, training each model of the limit tensile strength regression prediction model set using the sample set, selecting an optimal model, and constructing a strength-plasticity analysis model; According to the strength-plasticity analysis model, the super parameter optimization is performed to obtain the most adaptive simulation variables of the strength-plasticity analysis model; After obtaining the most adaptive simulation variables of the strength-plasticity analysis model, a microstructure analysis is performed to realize closed-loop verification; The super parameter optimization refers to reducing the simulation variable space of the high-entropy alloy / graphene composite material through domain knowledge constraints and SHAP interpretability methods, optimizing the simulation variables by combining a genetic algorithm, and obtaining a coupled parameter combination with higher limit tensile strength and ductility. The super parameter optimization comprises the following steps: first, pre-processing the original data to calculate the characteristic factors required by the strength-plasticity analysis model; then, training two models of ML-UTS and ML-EL, wherein ML-UTS is used to predict tensile strength, and ML-EL is applied to predict ductility; then, initializing the parameters of the genetic algorithm, including population size, iteration number, and key parameters such as crossover and mutation rate; in the optimization process, the output value of ML-UTS is used as the fitness function of the genetic algorithm to evaluate the advantages and disadvantages of individuals; when the preset iteration stopping condition is met, the genetic algorithm outputs the best coupled parameter result.

2. The method for constructing a strength-plasticity analysis model of a composite material according to claim 1, wherein The multi-dimensional parameter space sampling is realized by Sobol quasi-random sequences; The element concentration of the high-entropy alloy matrix includes the concentrations of Fe, Ni, Cr, Co, and Cu; The temperature range of the molecular dynamics simulation is 0-1100K; The volume fraction of the graphene ranges from 0.4vol% to 3.4vol%; The orientation angle between the graphene and the high-entropy alloy matrix ranges from 0° to 45°.

3. The method of claim 1, wherein, The molecular dynamics modeling comprises: initially constructing the atomic configuration of the composite material into a high-entropy alloy model, then inserting graphene into the high-entropy alloy matrix to form a high-entropy alloy / graphene composite model; The high-entropy alloy / graphene composite model simulated high-entropy alloy / graphene composite material is composed of FCC, and: graphene is embedded in the middle of the high-entropy alloy matrix; the volume fraction of the graphene is embodied as the volume of the graphene divided by the volume of the high-entropy alloy / graphene composite material, and is expressed as: Wherein, V Gr is the volume fraction of graphene, x Gr is the length of graphene, y Gr is the width of graphene, z Gr is the thickness of single-layer graphene, V HEAs / Gr is the volume of the high-entropy alloy / graphene composite material.

4. The method of claim 1, wherein, The process of performing simulation comprises: obtaining stress-strain curves corresponding to the simulation variables after performing 0 pressure simulation and tensile simulation; wherein, during the 0 pressure simulation, the pressure in the x and z axial directions is kept as 0 bar; during the tensile simulation, a strain rate of 1x10 9 s -1 is applied along the positive direction of the y axis and kept constant; The environment for performing the simulation comprises: minimizing the energy of the high-entropy alloy / graphene composite material system; After relaxing for 50ps in the isothermal-isobaric ensemble, further relaxing for 50ps in the isothermal-isochoric ensemble until the system energy reaches thermodynamic equilibrium.

5. The method for constructing a strength-plasticity analysis model of a composite material according to claim 4, wherein The stress-strain curve is used to extract the limit tensile strength and ductility data corresponding to the simulation variables; the limit tensile strength and ductility data are used to construct label data; According to the ductility label, the ductility type is determined, which includes insufficient ductility and good ductility; if the sample amounts of insufficient ductility and good ductility are imbalanced, a SMOTE oversampling method is used to synthesize minority class samples to balance the data set.

6. The method of claim 1, wherein, The atomic basic properties include atomic number, atomic mass, metal radius, volume, electron affinity, and valence electron concentration; the chemical properties include Pauling electronegativity, Mulliken electronegativity, ionization energy, mixing enthalpy, mixing entropy, and mismatch entropy. The physical properties include density, melting point, specific heat capacity, thermal diffusivity, elastic modulus and Poisson's ratio; The structural features include atomic size difference, radius difference, electronegativity difference and theoretical molar volume; The atomic basic attributes, chemical properties, physical properties and structural features are integrated, and linear mixing rule, derivative mixing rule, bias mixing rule and difference mixing rule are used to calculate and generate the first feature data; The first feature data is subjected to feature screening to form the feature data.

7. The method for constructing a strength-plasticity analysis model of a composite material according to claim 6, wherein The feature screening step includes: first step: filtering out low-contribution feature data with variance below a set threshold in the first feature data through variance threshold, and retaining data with significant variability in the first feature data; Second step: calculating the feature importance of each element of the first feature data using the random forest model to identify the features that have an impact on the target variable; Third step: quantifying the linear correlation between features through Pearson correlation analysis, updating the first feature data after removing highly redundant feature data; Fourth step: using recursive feature elimination combined with model accuracy feedback to update the first feature data after iteratively removing secondary features in the first feature data; Fifth step: searching the first feature data using the exhaustive method to find the optimal combination to ensure the balance between prediction performance and interpretability of the feature subset, and forming the feature data.

8. The method of claim 1, wherein, The limit tensile strength regression prediction model set includes a limit tensile strength regression prediction model and a ductility classification prediction model; The optimal model is selected by evaluating the prediction performance through the coefficient of determination and the root mean square error, which is represented as: wherein where n is the number of samples, y i is the true value, is the predicted value, m is the number of FPR and TPR values, R i is the average value of y is the predicted value, m is the number of FPR and TPR values, R 2 is the coefficient of determination RMSE is the root mean square error; Prediction performance is expressed as: R 2 The value range of R is [0,1]. 2 The closer the value is to 1, the better the model fits the data. In the case of a perfect fit, R0 = 1. 2 Close to 1; The smaller the RMSE value, the smaller the prediction error of the model; the value of Accuracy ranges from 0 to 1, and the closer the value is to 1, the more ideal the classification effect of the model, and the higher the proportion of correct classification; the value of AUC also ranges from 0 to 1, and the closer the value is to 1, the better the performance of the model.

9. The construction method of the strength-plasticity analysis model for composite materials according to claim 1, wherein, wherein The closed-loop verification is used to verify whether the limit tensile strength and ductility corresponding to the most suitable simulation variables of the strength-plasticity analysis model are better than the original data set; The microstructure analysis refers to further verification of the deformation mechanism through common neighbor analysis and dislocation density analysis.

Images