High-entropy alloy design method based on machine learning and nuclear high-entropy alloy

Abstract

The invention discloses a high-entropy alloy design method based on machine learning and a nuclear high-entropy alloy, and belongs to the technical field of metal material design. The design method comprises the following steps: collecting historical high-entropy alloy data, and establishing an initial data set of component-process-phase structure-mechanical property; constructing an initial feature variable composed of the element features and the empirical features, and screening to obtain a key feature variable; respectively constructing a classification model for predicting a phase structure and a regression model for predicting mechanical properties; based on the key characteristic variables, a classification model and a regression model, alloy components are screened according to a target phase structure and target mechanical properties, and design components are obtained; the design components are verified through experiments. According to the design method, on the premise that the phase structure is met, collaborative optimization of strength and toughness is achieved, and therefore the high-entropy alloy meeting various performance requirements can be designed, especially the high-entropy alloy which has an FCC phase structure and is high in toughness and resistant to high temperature and is applied to an advanced nuclear energy system can be designed.

Description

Technical Field

[0001] This application belongs to the field of metal material design technology, specifically relating to a high-entropy alloy design method based on machine learning, and high-entropy alloys obtained through this design method, particularly nuclear high-entropy alloys. Background Technology

[0002] High-entropy alloys, as an emerging design concept for metallic materials, form simple solid solution phase structures (such as face-centered cubic (FCC) or body-centered cubic (BCC)) through the mixing of multiple principal elements in equiatomic or near-equiatomic ratios. These alloys exhibit a variety of superior properties not found in traditional alloys, including high strength, high toughness, excellent high-temperature resistance, and outstanding resistance to radiation damage. Given their outstanding properties, high-entropy alloys hold great promise for applications in advanced nuclear energy systems, such as fourth-generation fission and fusion reactors. In particular, high-entropy alloys with single-phase FCC structures demonstrate significant potential as next-generation nuclear structural materials due to their good structural stability and strong plastic deformation capabilities.

[0003] Currently, materials informatics methods, represented by machine learning, have provided a novel paradigm for the design of high-entropy alloys. By constructing the relationship between "composition-structure-property", machine learning models can efficiently perform virtual screening and performance prediction in a broad composition space, thereby guiding experimental design and achieving multi-performance synergistic optimization in the design of high-entropy alloys, significantly accelerating the research and development process of new materials.

[0004] However, there is still a lack of systematic and efficient design strategies for the collaborative design of multiple objectives and constraints required by current advanced nuclear energy systems, such as "high strength and toughness" and "specific phase structure" (e.g., single-phase FCC). In particular, there is currently no relevant design method for the collaborative optimization of multiple mechanical properties such as strength (e.g., yield strength, tensile strength) and plasticity (e.g., elongation) under the constraint of specific phase structure. Summary of the Invention

[0005] In view of this, the primary objective of this application is to provide a high-entropy alloy design method based on machine learning, which achieves synergistic optimization of strength and toughness while satisfying phase structure requirements, thereby designing high-entropy alloys that meet various performance requirements, especially high-entropy alloys with FCC phase structure, high strength and toughness, and high temperature resistance for advanced nuclear energy system applications.

[0006] To achieve the above objectives, this application adopts the following technical solution: One aspect of this application discloses a high-entropy alloy design method based on machine learning, comprising the following steps: Historical high-entropy alloy data were collected to establish an initial dataset of composition, process, phase structure, and mechanical properties; and the data in the initial dataset were cleaned to obtain a clean dataset of high-entropy alloys. Construct initial feature variables composed of element features and empirical features, and filter the initial feature variables to obtain key feature variables; Based on the phase structure and mechanical property information in the cleaning dataset, a classification model for predicting phase structure and a regression model for predicting mechanical properties are constructed. A search space is defined, and based on the key feature variables and classification and regression models, alloy compositions are screened according to the target phase structure and target mechanical properties to obtain the design composition; The design components are verified experimentally. If the experimental results match the predicted results, the design ends; otherwise, the experimental results are added to the training set, and the design is repeated until a high-entropy alloy that meets the requirements for nuclear applications is designed.

[0007] Another aspect of this application discloses a high-entropy alloy design system / device based on machine learning, which executes the design method described in this application when the system / device is running.

[0008] Another aspect of this application discloses a high-entropy alloy for nuclear applications, the expression of which is: Al u Cr v Fe w Mn x Ni y Ti z In its expression, u, v, w, x, y, and z represent the atomic percentages of the corresponding elements and satisfy the following conditions: 1≤u≤5, 10≤v≤15, 45≤w≤55, 5≤x≤15, 20≤y≤30, 2≤z≤8, u+v+w+x+y+z=100.

[0009] The beneficial effects of this application are: This application first proposes a high-entropy alloy design method based on machine learning, which achieves synergistic optimization of multiple mechanical properties such as strength and plasticity under phase structure constraints. Compared with traditional parameter or experimental design, this method enables rapid screening of high-entropy alloy compositions that satisfy phase structure and rapid design based on mechanical performance targets, thereby significantly improving the efficiency of new material design and development.

[0010] Based on the aforementioned high-entropy alloy design method, this application yields a high-strength and high-toughness single-FCC phase high-entropy alloy suitable for advanced nuclear energy systems through compositional design. Compared to traditional structural materials, it not only exhibits better tensile properties at room temperature but also demonstrates significant potential for high-temperature applications, while its plasticity meets the application requirements. Furthermore, the single-phase Fe-based high-entropy alloy preparation process presented in this application is simple, low-cost, and has great potential for industrial application. Attached Figure Description

[0011] Figure 1This is a flowchart of the high-entropy alloy design method based on machine learning in Example 1.

[0012] Figure 2 The image shows the XRD pattern of the high-entropy alloy used for the Al-Cr-Fe-Mn-Ni-Ti core in Example 2.

[0013] Figure 3 The graph shows the mechanical property test results of the high-entropy alloy used for the Al-Cr-Fe-Mn-Ni-Ti core in Example 2. Detailed Implementation

[0014] The embodiments of this application will be clearly and completely described below. The technical solutions in the embodiments described below are exemplary and only possible technical implementations of this application, not all possible implementations. Those skilled in the art can combine the embodiments of this application to obtain other embodiments without creative effort, and these embodiments are also within the protection scope of this application.

[0015] The first aspect of this application discloses a high-entropy alloy design method based on machine learning, the process of which is as follows: Figure 1 As shown, the steps include: Historical high-entropy alloy data were collected to establish an initial dataset of composition, process, phase structure, and mechanical properties; and the data in the initial dataset were cleaned to obtain a clean dataset of high-entropy alloys. Construct initial feature variables composed of element features and empirical features, and filter the initial feature variables to obtain key feature variables; Based on the phase structure and mechanical property information in the cleaning dataset, a classification model for predicting phase structure and a regression model for predicting mechanical properties are constructed. A search space is defined, and based on the key feature variables and classification and regression models, alloy compositions are screened according to the target phase structure and target mechanical properties to obtain the design composition; The design components are verified experimentally. If the experimental results match the predicted results, the design ends; otherwise, the experimental results are added to the training set, and the design is repeated until a high-entropy alloy that meets the requirements for nuclear applications is designed.

[0016] In this application, the term "high-entropy alloy" refers to an alloy system composed of four or more principal elements, with each element comprising more than 5 at.% and less than 35% at.%. It is understood that high-entropy alloys, due to their multi-principal element characteristics, exhibit high thermodynamic stability, lattice distortion effect, slow diffusion effect, and "cocktail" effect, thus displaying superior comprehensive properties compared to traditional alloys, such as high strength, high hardness, good corrosion resistance, and thermal stability. In some examples, the specific systems of the high-entropy alloys include, but are not limited to: Fe-Cr-VW-Mn systems, Al-Co-Cr-Fe-Ni systems, Co-Cr-Fe-Mn-Ni systems, Mo-Nb-Ta-WV systems, etc. In this application, it specifically refers to high-entropy alloys for nuclear applications, which have high requirements for both phase structure and mechanical properties.

[0017] In this application, "historical high-entropy alloy data" refers to a collection of information about publicly disclosed or studied high-entropy alloys, gathered from existing technical literature, databases, or experimental records. This data is typically used as the basis for training and building machine learning models. In some specific examples, the historical high-entropy alloy data includes, but is not limited to, alloy composition information, preparation process parameters, microstructure, and test results of various mechanical properties (such as hardness, yield strength, tensile strength, elongation, etc.). Those skilled in the art can select appropriate historical high-entropy alloy data according to experimental objectives and research needs.

[0018] In some specific examples, for the design of high-entropy alloys for nuclear applications, the historical high-entropy alloy data selected includes composition, processing, phase structure, and mechanical properties. The mechanical properties specifically refer to yield strength, tensile strength, and total elongation. It should be understood that this data includes at least composition, phase structure, and mechanical properties; the processing details can be included in the initial dataset based on actual needs.

[0019] In this application, the term "composition" refers to the types of chemical elements that constitute a high-entropy alloy and their relative content in the alloy, usually expressed as atomic percentage (at.%) or molar percentage.

[0020] The term "process" refers to the preparation method of high-entropy alloys and its specific parameters. For example, for additive manufacturing processes, the process parameters may include laser power, scanning speed, powder feeding rate, etc.; for smelting processes, they may include the number of smelting operations, cooling method, etc.; and they may also include heat treatment, such as annealing parameters, etc.

[0021] The term "phase structure" refers to the microscopic phase composition and crystal structure type formed in high-entropy alloys under specific composition and processing conditions. Common phase structure types include body-centered cubic (BCC), face-centered cubic (FCC), hexagonal close-packed (HPC), and mixed structures between them, and may also include intermetallic compound phases or amorphous phases.

[0022] The "mechanical properties" refer to the behavior and response of high-entropy alloys under external forces, and are key indicators for evaluating their suitability as structural or functional materials. The mechanical properties of interest in this application mainly include, but are not limited to: hardness at room temperature and high temperatures, yield strength, tensile strength, and elongation.

[0023] It should be understood that the specific data selected can be chosen appropriately according to the experimental purpose and research needs, and therefore there are no special restrictions.

[0024] In this application, the term "data cleaning" refers to the steps of preprocessing collected historical high-entropy alloy data to improve data quality and ensure the reliability of subsequent modeling. In some examples, data cleaning operations include: deleting data records containing obvious outliers or missing values; calculating the relative error of the same component when there are multiple different reported values, averaging the values ​​with errors less than a preset threshold (e.g., 10%), and removing values ​​with excessive errors. As a specific example of this application, data cleaning mainly includes the following steps: (1) whether the sum of the obtained atomic percentages is 1, and if not, normalization is performed; (2) calculating the quartiles of the mechanical property data column in the initial dataset, and evaluating the data points based on the nearest neighbor samples in the composition space. If the mechanical properties of similar components differ from the point by more than 3σ, the data point is deleted; (3) manually reviewing the remaining data points, and removing sample points with differences exceeding 2σ and which cannot be explained by cited literature or whose explanations cannot be effectively supported by professional knowledge, in conjunction with phase formation criteria, mechanical properties of alloys composed of the same elements.

[0025] In this application, element features and empirical features refer to quantized descriptors derived from the original component data for the purpose of constructing a machine learning model. Together, they constitute the initial set of feature variables used for model training.

[0026] Among them, "elemental characteristics" usually refer to the atomic percentage of each element in the alloy and its inherent physicochemical parameters (such as atomic radius, electronegativity, melting point, elastic modulus, ionization energy, etc.); or they can be calculated from the aforementioned parameters using a preset mathematical formula. For example, it can be the average value of a certain physicochemical parameter of all constituent elements, the root mean square deviation of the parameter (mismatch value), or the weighted sum of the differences of the parameter between any two elements (local mismatch value), etc.

[0027] "Empirical characteristics" refer to empirical parameters in the field of materials science that have been extensively studied and verified and are closely related to the formation or properties of alloy phases. Examples include: mixing entropy, mixing enthalpy, atomic size difference, valence electron concentration, electronegativity difference, Ω parameter, Λ parameter, etc., but are not limited to these.

[0028] In this application, by screening the initial feature variables (e.g., using correlation analysis, recursive feature elimination, exhaustive search, etc.), redundant or irrelevant features can be eliminated, and a subset of key feature variables that significantly contribute to the prediction target can be obtained for subsequent construction of an efficient and highly generalizable model.

[0029] The two basic machine learning model types used in this application are classification models and regression models. Specifically: The “classification model” is used to predict the category properties of high-entropy alloys, and in this application, it is specifically used to predict the phase structure of high-entropy alloys. For example, a model can be constructed to identify whether an alloy composition tends to form a solid solution strengthening phase, or to further determine whether its main phase is a body-centered cubic structure, a face-centered cubic structure, or a mixture of both.

[0030] "Regression models" are used to predict continuous numerical performance indices of high-entropy alloys, and in this application, they are specifically used to predict the mechanical properties of high-entropy alloys. For example, models can be constructed to predict the hardness or yield strength of high-entropy alloys under specific compositions and processes.

[0031] It is understood that the machine learning algorithms used to construct the above models can employ various conventional algorithms known in the art, including but not limited to: decision trees, random forests, gradient boosting, support vector machines, K-nearest neighbors, and artificial neural networks. The specific machine algorithm can be determined based on the evaluation of the constructed model, a capability possessed by those skilled in the art.

[0032] In this application, "cross-validation" is a statistical analysis method for evaluating the generalization performance of machine learning models and preventing overfitting. Its core idea is to randomly divide the original dataset into several subsets, alternately using one subset as the test set and the remaining subsets as the training set, performing multiple training and testing cycles, and finally evaluating the model performance by the average of the multiple test results. Common cross-validation methods include k-fold cross-validation and hold-out methods, but are not limited to these.

[0033] Hyperparameters are parameters that need to be manually set before training a machine learning model, rather than being learned from training data. The choice of hyperparameters directly affects the model's architecture and learning process; examples include the maximum depth of a decision tree, the number of trees in a random forest, the kernel function and penalty coefficient in a support vector machine, and the learning rate and number of layers in a neural network. Optimizing hyperparameters is a crucial step in improving model performance, and methods such as grid search, random search, or Bayesian optimization are commonly used.

[0034] After evaluating and selecting a suitable model, alloy design is carried out. The "search space" refers to the range and rules governing the composition or process parameters to be explored when designing a new alloy composition using a trained model. In a typical example, the search space might be set as follows: at least four elements are selected from a pre-selected set of elements (such as metallic elements with low neutron activation that meet nuclear requirements in this application), with the atomic percentage of each element limited to between 0% and 100%, enumerated in 1% increments, and the sum of all element contents equal to 100%. After complete sampling, further sampling is performed, followed by uniform sampling. The constructed features are input into the constructed machine learning prediction model, and finally, experimental verification is conducted to determine the designed composition.

[0035] The second aspect of this application discloses a high-entropy alloy for nuclear applications, obtained by screening based on the high-entropy alloy design method described in the first aspect of this application. The expression for the high-entropy alloy for nuclear applications is: Al u Cr v Fe w Mn x Ni y Ti z In its expression, u, v, w, x, y, and z represent the atomic percentages of the corresponding elements and satisfy the following conditions: 1≤u≤5, 10≤v≤15, 45≤w≤55, 5≤x≤15, 20≤y≤30, 2≤z≤8, u+v+w+x+y+z=100.

[0036] The present application will be further illustrated below with reference to specific embodiments. It should be noted that the specific embodiments below are for illustrative purposes only and do not limit the scope of the present application in any way.

[0037] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0038] In addition, unless otherwise specified, methods without detailed conditions or steps are conventional methods, and the reagents and materials used are commercially available.

[0039] Example 1: High-Entropy Alloy Design Method Based on Machine Learning This embodiment takes nuclear high-entropy alloys as an example to illustrate the high-entropy alloy design method based on machine learning in this application. However, it should be understood that other high-entropy alloys with relevant design requirements can also be designed using the method in this application. The process and basic principles are the same. Depending on the experimental purpose and needs, those skilled in the art can make appropriate optimizations to the details or parameters.

[0040] The design method for nuclear high-entropy alloys in this embodiment includes the following specific steps: Step 1: Data Collection and Cleaning Step 1.1: This embodiment collected the composition and corresponding phase structure of 441 sets of high-entropy alloys in the as-cast or simple homogenized state based on literature and previous experiments, forming the initial dataset of high-entropy alloys; among them, the elemental composition includes 16 elements: Al, Co, Cr, Cu, Fe, Hf, Mn, Mo, Nb, Ni, Sn, Ta, Ti, V, W and Zr; according to whether the phase structure contains intermetallic compounds and amorphous phases, 219 sets of high-entropy alloy samples contain SS, 129 sets contain SP, and 232 sets contain FCC; in addition, 150 sets of high-entropy alloy composition and corresponding tensile mechanical properties (including YS, UTS and TE) data were collected.

[0041] Step 1.2: Perform the following checks and data cleaning steps on the high-entropy alloy data in the initial dataset obtained in Step 1.1: (1) Check whether the sum of the obtained atomic percentages is 1. If not, normalize the data. (2) Calculate the quartiles of the mechanical property data column in the initial dataset and evaluate the data points based on the nearest neighbor samples in the composition space. If the mechanical properties of similar components differ from the data point by more than 3σ, delete the data point. (3) Manually review the remaining data points and, in conjunction with the phase formation criteria, the mechanical properties and phase structure of alloys composed of the same elements, remove sample points with differences exceeding 2σ that cannot be explained by cited literature or whose explanations cannot be effectively supported by professional knowledge.

[0042] Step 2: Feature Engineering Step 2.1: In this embodiment, 44 elemental parameters corresponding to the 16 elements mentioned in Step 1 were further collected, including: (1) atomic radius R a (2) Metal radius R m (3) Covalent radius R c (4) Atomic number AN (5) clan Gr(6) Period Pe (7) Relative atomic mass Ar (8) Valence electron concentration VEC (9) Free electron concentration e / a (10) Effective nuclear charge number (Slater) N efs (11) Effective nuclear charge number (Clementi) N efc (12) Pauling electronegativity EN p (13) Allred-Rochow electronegativity EN a (14) First ionization energy I 1, (15) Second ionization energy I 2, (16) Electron work function Wf (17) Wigner-Sietz electron density N ws (18) Density ρ (19) Molar volume V m (20) Coefficient of linear expansion L t (21) Thermal conductivity K (22) Electrical conductivity EC (23) Specific heat capacity C p (24) Standard molar entropy change S 0, (25-27) elastic constant C 11 , C 12 , C 44 (28) Volume modulus B (29) Elastic modulus E (30) Shear modulus G (31) Poisson's ratio ν (32) Viscoelastic coefficient η (33) Pugh is more G / B (34) Melting point T m (35) Boiling point T b (36) Liquid phase region LR (37) Heat of fusion H fus (38) Heat of vaporization H vap (39) Sublimation feverH atm (40-41) Low temperature limit and Debye temperature at room temperature T dl , T dr (42) Polymerization energy CE, (43) Vacancy formation energy (44) Vacancy migration energy .

[0043] First, the atomic percentages and elemental parameters of the high-entropy alloy were processed according to formulas (1)-(4), generating a total of 176 elemental characteristics. Then, empirical characteristics that have been proven in the literature to be related to the target performance were collected and combined with the elemental characteristics to form characteristic variables. In this embodiment, the enthalpy of mixing Δ was collected. H mix and mixed entropy Δ S mix As empirical features, 178 feature variables were initially generated.

[0044]

[0045] In formulas (1)-(4), , , and These are the weighted average, weighted standard deviation, range, and variance value for elemental parameters and composition, respectively. Indicates the first The atomic percentage of each element; It is the first Element parameters of each element; and They are respectively the corresponding first The atomic percentage and elemental parameters of each element.

[0046] Step 2.2: Clean the two datasets, each containing component information and the 178 initial features obtained in Step 2.1. Further, randomly divide them into a 10% test set and a 90% training set using pseudo-random number generation. Using the training set data as a benchmark, normalize both the training and test sets using the training set data to ensure that no data in the test set is leaked.

[0047] The chi-square test is used to describe the correlation between phase structure information and features in the training set, or the Pearson correlation coefficient is used to describe the correlation between regression performance indicators and features. Based on these indicators, the target phase structure or performance is ranked, and the 20% of the least relevant features are removed to obtain a feature subset X. Then, based on the correlation between features, a Pearson correlation coefficient matrix is ​​constructed, and features with an absolute Pearson correlation coefficient greater than 0.9 are removed to obtain a feature subset X' composed of key feature variables, which is used for subsequent machine learning algorithm selection and model construction.

[0048] Step 3: Model Building and Evaluation Step 3.1: The algorithms used to build the machine learning models include Decision Tree (DT), Extreme Random Tree (ET), Gradient Boosting Decision Tree (GBDT), Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Random Forest (RF), and Feedforward Neural Network (FNN). In addition, for classification problems, the Logistic Regression algorithm is added; for regression problems, the Ridge Regression algorithm is added, and compared with the aforementioned algorithms. For the phase structure information obtained in Step 1, classification models are built and trained for evaluation and final model selection; for the mechanical performance information obtained in Step 1, regression models are built and trained for evaluation and final model selection.

[0049] Step 3.2: The algorithm described in Step 3.1 needs to be combined with cross-validation to optimize the hyperparameters in the training set, and the model with the highest accuracy or the smallest error that each algorithm can build should be given. The different algorithms should be compared again to obtain the applicable prediction model for each target phase structure or performance.

[0050] In this embodiment, based on the training set and test set divided in step 2.2, the hyperparameters are searched on the training set according to the above machine learning algorithm and the hyperparameters that the corresponding algorithm needs to optimize, using a Bayesian optimization strategy, and the model performance is evaluated by using 10-fold cross-validation on the training set.

[0051] For classification tasks, when using algorithms such as DT, ET, GBDT, SVM, KNN, and Logistic Regression, the F1 score is used for hyperparameter optimization; for FNN, the cross-entropy loss function (BCE Loss) is used to optimize the number of hidden layer units, optimizer, learning rate, and activation function of the neural network. The calculation of F1 score and BCE Loss is shown in formulas (5)~(7) and formula (8).

[0052]

[0053] In equations (5) to (7), TP, FP, and FN represent the number of samples predicted as positive and actually positive, predicted as positive but actually negative, and predicted as negative but actually positive, respectively. For an ideal prediction model, F1 should be 1.

[0054]

[0055] In formula (8), Indicates the first i The binary label (0 or 1) of each sample. Let N be the probability of the output label value, and N be the number of predicted samples. For an ideal prediction model, the BCE Loss should be 0. After obtaining all models, retrain the model on the complete training set and predict the results on the test set. Confusion matrix is ​​constructed based on the prediction results to comprehensively evaluate the performance of the classification model.

[0056] For regression tasks, the root mean square error is uniformly used for hyperparameter optimization of the model. The predictive power of the regression model is calculated using the normalized root mean square error (NRMSE) calculated by formula (9) and the coefficient of determination (CQD) calculated by formula (10). R 2 To conduct the evaluation.

[0057]

[0058] In formula (9)-(10), n The number of samples; and The first i Actual and predicted values ​​of each sample ( i = 1,2,..., n ); This represents the mean of the actual values. For an ideal model, NRMSE is 0. R 2 It equals 1.

[0059] The final algorithm chosen in this embodiment is a gradient boosting classification / regression algorithm. The corresponding optimized hyperparameters include the number of decision trees, learning rate, tree depth, number of leaves in subtrees, minimum number of split samples, and sampling with replacement ratio. Hyperparameter selection is based on a Bayesian optimization method using cross-validation. The optimal hyperparameter combination is determined by training / testing on multiple pseudo-random seed partitions. The model is trained on the training set and evaluated on the test set to ensure that parameter selection is not affected by a single partition.

[0060] Key features for selecting the SS prediction model include Δ H mix V K and V T mThe SP-FCC model selects key features including V. C 12 V C 44 F T b F K and The key features selected for the YS model include D V m M L t D H fus , and V C 44 Key features for UTS model selection include V R m V H fus and M e / a Key features selected for the TE model include F ν D R m V V m V T dl and D E .

[0061] Step 4: Alloy Design Step 4.1: Select at least four metallic elements with low neutron activation that meet nuclear requirements, with each element having an atomic percentage of 0-100%. Perform composition sampling across the entire composition space and construct features for the collected virtual high-entropy alloy compositions according to the requirements of Step 2. In this embodiment, Al, Cr, Fe, Mn, Ni, and Ti are selected to form the composition space. Sampling is performed in this space using a pseudo-random number method, ensuring that the compositions of the six elements are normalized. Based on the sampled data, features are constructed according to the method described in Step 2.1. Based on the phase classification model constructed in Step 3, predictions are made to screen alloys that meet FCC, SP, and SS criteria.

[0062] Step 4.2: Uniformly sample the component range given in Step 4.1, construct features based on the method in Step 2, and input them into the machine learning mechanical performance prediction model constructed in Step 3 to obtain UTS, YS, and TE results. Select the component with better UTS, YS, and TE as the design component for experimental verification.

[0063] Step 5: Experimental Verification Strictly follow the high-entropy alloy composition obtained in step 4 for melting, perform structural characterization and mechanical property testing on the as-cast sample, and compare it with the predicted results. If the experimental results match the predicted results, the design ends; otherwise, add the experimental results to the training set of the original data and redesign until a high-entropy alloy that meets the requirements is designed.

[0064] Example 2: High-entropy alloy for nuclear applications This application, referring to the design method of Example 1, obtains a high-entropy alloy Fe for nuclear applications. 50 Ni 25 Cr 10 Al5Mn5Ti5.

[0065] The components were smelted according to the designed composition: 50% Fe, 25% Ni, 10% Cr, 5% Al, 5% Mn, and 5% Ti.

[0066] The predicted values ​​for the yield strength, tensile strength, and total elongation of this high-entropy alloy are 595 MPa, 753 MPa, and 33%, respectively.

[0067] An ingot of approximately 1000g was prepared using vacuum induction melting. To ensure the uniformity of the alloy composition, the alloy was repeatedly melted and turned over at least five times during the preparation process. The XRD pattern of the as-cast alloy is shown below. Figure 2 As shown, the results indicate that the alloy is mainly composed of an FCC phase structure. Uniaxial tensile tests were conducted on this high-entropy alloy according to EN ISO 6892-1:2019 and EN ISO 6892-2:2018, and the tensile yield strength, tensile strength, and total elongation were measured to be 533 MPa, 834 MPa, and 41%, respectively. Figure 3 As shown, the error between the predicted and experimental results is small, verifying the effectiveness of the mechanical property prediction model constructed by this method.

[0068] It should be noted that this application is not limited to the above-described embodiments. The above embodiments are merely examples, and any embodiments with the same structure and effect as the technical concept within the scope of this application are included in the technical scope of this application. Furthermore, various modifications that can be conceived by those skilled in the art to the embodiments, and other ways of constructing by combining some of the constituent elements of the embodiments, without departing from the spirit of this application, are also included in the scope of this application.

Claims

1. A high-entropy alloy design method based on machine learning, characterized in that, Includes the following steps: Collect historical high-entropy alloy data to establish an initial dataset of composition, process, phase structure, and mechanical properties; The data in the initial dataset is then cleaned to obtain a cleaned dataset of high-entropy alloys. Construct initial feature variables composed of element features and empirical features, and filter the initial feature variables to obtain key feature variables; Based on the phase structure information and mechanical property information in the cleaning dataset, a classification model for predicting phase structure and a regression model for predicting mechanical properties are constructed respectively. A search space is defined, and based on the key feature variables and classification and regression models, alloy compositions are screened according to the target phase structure and target mechanical properties to obtain the design composition; The design components are verified experimentally. If the experimental results match the predicted results, the design ends; otherwise, the experimental results are added to the training set, and the design is repeated until a high-entropy alloy that meets the requirements for nuclear applications is designed.

2. The high-entropy alloy design method as described in claim 1, characterized in that, The initial dataset includes at least the following data information: Data Information 1: Elements of high-entropy alloys and the atomic percentage of each element; Data Information 2: Phase structure information of high-entropy alloys; Data 3: Mechanical property data of high-entropy alloys; Preferably, the phase structure information includes: whether it is composed of a solid solution phase, whether it is a single phase, and whether it includes a face-centered cubic phase. Preferably, the mechanical property data are tensile test data, including yield strength, tensile strength, and total elongation; Preferably, the initial dataset further includes data information 4: the preparation and / or heat treatment process of the high-entropy alloy corresponding to data information 1.

3. The high-entropy alloy design method as described in claim 1, characterized in that, The data cleaning process in the initial dataset includes the following steps: (1) Confirm whether the sum of the atomic percentages in the initial dataset is 100%. If not, perform normalization. (2) Calculate the quartiles of the mechanical property data column in the initial dataset and make a preliminary assessment of the composition, process and phase structure information of outlier data points; based on data information 1, calculate each alloy and its N nearest neighbor sample points in the composition space, where N is an integer not less than 3; at the same time, calculate the relative performance difference between it and the surrounding alloy composition and construct the difference distribution; if the relative difference between the mechanical properties of similar components and the performance of this point is greater than 3σ, it is considered that the entered alloy performance data is abnormal and the data point is deleted. (3) Manually review the remaining data points, and in conjunction with the phase formation criteria, the mechanical properties and phase structure of alloys with the same elements, remove sample points with differences greater than 2σ that cannot be explained by cited literature or whose explanations are insufficient to be supported by professional knowledge.

4. The high-entropy alloy design method as described in claim 1, characterized in that, The elemental characteristics are formed by digitizing the atomic percentage composition and elements of the high-entropy alloy in the cleaned dataset according to formulas (1)-(4): ， in, M p , S p , F p and D p These are the weighted average, weighted standard deviation, range, and variability values ​​for elemental parameters and composition, respectively; c i Indicates the first i The atomic percentage of each element; p i It is the first i The element parameters of each element; the corresponding c k and p k Then it is the first k The atomic percentage and elemental parameters of each element; The empirical features are derived from features that have been proven to be relevant to the target performance in previous literature.

5. The high-entropy alloy design method as described in claim 1, characterized in that, The initial feature variables are selected through the following steps: Based on the initial feature variables, the feature subset X is obtained by filtering the mechanical properties of the cleaned dataset based on their correlation. Then, based on the correlation between features, a correlation coefficient matrix is ​​constructed for the feature subset X, and redundant features with high correlation are removed to obtain the feature subset X' composed of key feature variables.

6. The high-entropy alloy design method as described in claim 1, characterized in that, The machine learning algorithms for the classification and regression models are selected from any one of decision trees, gradient boosting decision trees, support vector machines, k-nearest neighbors, random forests, and feedforward neural networks.

7. The high-entropy alloy design method as described in claim 1, characterized in that, The search space is set as follows: Select no fewer than four metallic elements that meet the target requirements, with each element having an atomic percentage of 0% to 100%. Preferably, the target requirement is a nuclear requirement, and the metallic elements that meet the nuclear requirement are at least four of the following: Al, Cr, Fe, Mn, Ni, Ta, Ti, V, W, and Zr.

8. The high-entropy alloy design method as described in claim 1, characterized in that, The selection process for the design components is as follows: A search space is defined, and the composition of the complete space is sampled. Features are constructed for the collected virtual alloy compositions according to the requirements of the initial feature variables. Then, a pseudo-random number method is used to sample in the composition space, ensuring that the total number of sampled compositions is 100%. Based on the sampled samples, features are constructed according to the method of constructing the initial feature variables. The phase structure is predicted based on the classification model, and high-entropy alloys that meet FCC, SP, and SS are selected. The given component range after screening is uniformly sampled, and the key feature variables are input into the regression model to obtain the predicted structure of mechanical properties. Components that meet the mechanical performance requirements are selected as design components.

9. A high-entropy alloy design system / device based on machine learning, characterized in that, When the system / device is in operation, it performs the design method according to any one of claims 1-8.

10. A high-entropy alloy for nuclear applications, characterized in that, The expression for the nuclear high-entropy alloy is: Al u Cr v Fe w Mn x Ni y Ti z In its expression, u, v, w, x, y, and z represent the atomic percentages of the corresponding elements and satisfy the following conditions: 1≤u≤5, 10≤v≤15, 45≤w≤55, 5≤x≤15, 20≤y≤30, 2≤z≤8, u+v+w+x+y+z=100.

Images