Design method of multi-performance collaborative optimization high-entropy alloy components based on machine learning

Abstract

The present invention discloses a method for designing high entropy alloy components with multi-performance collaborative optimization based on machine learning. In this method, a high entropy alloy component-performance initial data set is created, data cleaning is performed, the cleaned data is standardized, a high entropy alloy performance training set is obtained, a base model for predicting each performance of the high entropy alloy is determined by the high entropy alloy performance training set, an expected improvement (EI) value of each performance of the high entropy alloy in the component search space is determined, an alloy performance EI value Pareto front is determined in the component search space, a cluster analysis is performed to obtain the cluster center of the EI value Pareto front, a high entropy alloy and its related component information corresponding to the EI value are determined, the high entropy alloy is manufactured, and a characteristic evaluation test is performed for each target performance of the manufactured high entropy alloy, and if the test data does not meet the requirements, the high entropy alloy component-performance initial data set is updated according to the test data. The present invention can realize an efficient component design of the multi-performance collaborative optimization of high entropy alloy.

Description

Technical Field

[0001] The present invention belongs to the technical field of metal material design, and specifically relates to a method for designing multi-performance collaborative optimization high-entropy alloy components based on machine learning.

Background Art

[0002] With the rapid development of the economy and society, the requirements for the performance of metal materials have been continuously improved. Traditional metal alloy systems based on single-element design have been continuously improved, and performance development tends to saturate, and there are bottlenecks in the development, design, and application of high-performance metal materials. Taiwanese scholar Ye Junwei proposed the concept of multi-element alloy composition design in 2004. The high-entropy alloy thus defined provides a rich exploration space for the design of metal materials and greatly improves the freedom of alloy composition selection and performance adjustment. Recent research has shown that high-entropy alloys are superior to conventional alloy materials in many aspects such as high-temperature strength, low-temperature toughness, thermal stability, and corrosion resistance, and have broad application prospects.

[0003] However, the composition system of high-entropy alloys is complex and depends on traditional trial-and-error material design methods, making it difficult to meet its requirements for rapid and accurate composition design. On the other hand, in the research and development of high-entropy alloys for engineering applications, it is usually necessary to synergistically optimize multiple properties. Alloy design faces the problem that all performances cannot be comprehensively optimized, resulting in low efficiency, long cycle, and high cost in the research and development of new high-performance alloy materials. Therefore, how to accelerate the exploration of target materials in a wide composition space and develop an efficient design method for high-performance alloy materials has become a technical issue that urgently needs to break through in the field of metal alloys and even the entire field of material design.

[0004] In recent years, with the rise of materials genetic engineering, a new paradigm has emerged in materials research and development, and data technologies represented by machine learning are playing an increasingly important role in materials design. Machine learning methods can establish implicit relationships between material composition, process, etc. and target performance based on existing material data information, and can quickly and accurately predict material properties by excavating hidden material laws and knowledge based on existing material data information, thus accelerating the design of material composition. Currently, it has been reported that machine learning methods are used to guide the component and performance design of high-entropy alloys.

[0005] For example, Chinese Patent CN113870957A discloses a design method and device for eutectic high-entropy alloy components based on machine learning. By training a machine learning model, important elements that have a great influence on eutectic formation and elements strongly related to them are predicted, and based on this, the element content is adjusted to predict a eutectic high-entropy alloy and obtain the components of the eutectic high-entropy alloy.

[0006] Chinese Patent CN114678086A discloses a machine learning method for low-activation high-entropy alloy design. By training a classification model and a regression model, Fe-Cr-V-W-Mn-based high-entropy alloy components with low-activation characteristics having a BCC phase structure are predicted.

[0007] In order to solve the problems of time-consuming, labor-intensive and inaccurate existing in the conventional methods for predicting and searching high-entropy alloys with excellent hardness characteristics, Chinese Patent CN112216356A discloses a hardness prediction method for high-entropy alloys based on machine learning, and Chinese Patent CN115061435A discloses a machine learning method for quickly predicting the hardness of high-entropy alloys and its manufacturing process.

[0008] Chinese Patent CN114464274A discloses a hardness prediction method for high-entropy alloys that improves genetic algorithm feature screening based on machine learning. By training a machine learning model, improving the genetic algorithm, and screening for feature combinations highly correlated with the hardness of high-entropy alloys, the prediction accuracy of hardness is improved, guiding the design of high-entropy alloy components according to the required hardness.

[0009] However, the machine learning methods reported so far are usually designed for specific high-entropy alloy systems within a limited alloy space, and it is difficult to quickly search for the target alloy in a high-dimensional and complex composition space. On the other hand, the reported methods mainly aim at the material composition design for the optimization of single-target performance, and cannot solve the problem of high-entropy alloy composition design under the requirement of co-optimization design of multi-target performance.

[0010] From the above, for the problems of the huge space of high-entropy alloy components and the material design requirements for multi-performance collaborative optimization, based on the concept of materials gene engineering, it is necessary to design a new design method for high-entropy alloy components that is data-driven and meets multi-performance requirements by machine learning technology.

Summary of the Invention

[0011] In view of the above circumstances, the main object of the present invention is to provide a design method for multi-performance collaborative optimization high-entropy alloy components based on machine learning.

[0012] To achieve the above object, the technical means of the present invention are as follows.

[0013] Embodiments of the present invention provide a design method for multi-performance collaborative optimization high-entropy alloy components based on machine learning. This method includes creating a high-entropy alloy component-performance initial dataset based on the historical data information of high-entropy alloys, and Performing data cleaning on each performance parameter of the high-entropy alloy, and standardizing the cleaned data to obtain a high-entropy alloy performance training set; Training a plurality of machine learning models respectively for each target performance of the high-entropy alloy with the high-entropy alloy performance training set, and screening and determining a base model for predicting each performance of the high-entropy alloy in turn; Determining the expected improvement (EI) value of each performance of the high-entropy alloy obtained by searching in the component space with the base model; Determining the search space of the target components of the high-entropy alloy in the optimized high-entropy alloy performance training set, performing genetic optimization search with the EI value of each performance of the high-entropy alloy as the target, and determining the Pareto front of the alloy performance EI value in the search component space; Performing cluster analysis on the Pareto front of the EI value, screening to obtain the high-entropy alloy corresponding to the EI value and its related component information; Manufacturing a high-entropy alloy corresponding to the EI value-related component information by vacuum arc melting furnace melting method; Performing a characteristic evaluation test on each target performance of the manufactured high-entropy alloy, and updating the high-entropy alloy component-performance initial data set based on the test data when the test data does not meet the requirements; Including.

[0014] In the above embodiment, in the step of creating a high-entropy alloy component-performance initial data set based on the historical data information of the high-entropy alloy, specifically, conducting a literature survey on refractory high-entropy alloys, and creating a high-entropy alloy component-performance initial data set including 10 common elements of Al, Ti, V, Zr, Cr, Nb, Mo, Hf, Ta, and W, including alloy components, high-temperature yield strength, and room-temperature compressive plastic properties.

[0015] In the above embodiment, data cleaning is performed on each performance parameter of the high-entropy alloy. Specifically, abnormal values and missing values of the reported alloy performance are deleted. For the performance data reported by different documents for the same component, the relative error of each reported value is determined. Duplicate values with the relative error smaller than the error threshold are averaged, and duplicate values with the relative error larger than the error threshold are deleted to obtain the final alloy performance value. The alloy performance includes high-temperature yield strength and room-temperature plasticity. The cleaned alloy performance values are normalized to determine a high-temperature yield strength training dataset including a plurality of samples and a room-temperature plasticity training dataset including a plurality of samples.

[0016] In the above embodiment, in the step of training a plurality of machine learning models for each target performance of the high-entropy alloy by using the high-entropy alloy performance training set and sequentially screening and determining the base model for predicting each performance of the high-entropy alloy, specifically, the high-temperature yield strength training dataset and the room-temperature plasticity training dataset are evenly divided into N groups. A machine learning model is trained and fitted using the first group of data, the remaining N - 1 groups of training data are predicted to obtain their performance prediction values, the root mean square error of the first power of the actual value and the prediction value is determined, and the same prediction operation is sequentially performed on the data from the second group to the Nth group to obtain the predicted root mean square error of the second power. The cross-validation error of each machine learning model is determined based on the root mean square error of the first power and the root mean square error of the second power. The machine learning model with the minimum cross-validation error is the base model for predicting the high-temperature strength and room-temperature plasticity performance of the alloy, respectively.

[0017] In the above embodiment, in the step of obtaining the expected improvement (EI) value of each target performance of the high-entropy alloy obtained by searching in the component space using the base model, specifically, resampling the alloy performance training set by the bootstrap method to obtain M groups of resampled data sets, the number of samples included in each group of resampled data sets is the same as the number of the training set, training the base model with each group of resampled data sets respectively to obtain M performance prediction base models, performing performance prediction on the high-entropy alloy obtained by searching in the component space using the M performance prediction base models to obtain M predicted values, determining the average value μ and variance σ of the performance predicted values, and using the formula:

Number

[0018] In the above embodiment, in the step of determining the search space of the high-entropy alloy target components in the optimized high-entropy alloy performance training set, specifically, according to the types of alloy constituent elements in the alloy performance training set, determining the search space of the high-entropy target components composed of elements of the types of Al, Ti, V, Zr, Cr, Nb, Mo, Hf, Ta, W, limiting that the search alloy contains 3 to 6 types of elements, and the molar content of each element in the alloy is 5% or more and 35% or less, and the variable step size of the content is 1%.

[0019] In the above embodiment, in the step of performing genetic optimization search with the EI value of each performance of the high-entropy alloy as the target and determining the Pareto front of the EI value in the search component space, specifically, in the search space of the target components of the high-entropy alloy, each time a genetic search is executed, a Pareto front of one alloy EI value is obtained. The genetic search is executed N times to obtain N Pareto fronts. Non-dominated sorting is performed on the N Pareto front alloys, and finally the Pareto front of the EI value of the alloy performance obtained by genetic search is determined.

[0020] In the above embodiment, in the step of performing cluster analysis according to the Pareto front of the alloy performance EI value, screening to obtain the high-entropy alloy corresponding to the performance EI value and the related component information, specifically, the Pareto front of the alloy performance EI value is cluster-analyzed by the K-means clustering method, the optimal number of cluster centers is determined by the elbow method, and clustering and screening are performed on the Pareto front of the alloy performance EI value using this number to obtain the alloy corresponding to the EI value and the related component information.

[0021] In the above embodiment, the method further includes a step of ending the design when the test data meets the requirements, and obtaining a high-entropy alloy with multi-performance cooperative optimization.

[0022] In the above embodiment, the method further includes a step of designing each high-entropy alloy in the high-entropy alloy component-performance initial data set until the multi-performance of the designed alloy meets the requirement of being simultaneously optimized and improved or reaches the upper limit of the budget.

[0023] Compared with the prior art, according to the component design method of the present invention, rapid search of the target alloy in the high-dimensional component space can be realized, and multi-performance of the high-entropy alloy can be cooperatively improved.

Brief Description of the Drawings

[0024] The drawings described herein are for further understanding of the present invention and are part of the present invention. The schematic embodiments of the present invention and their descriptions are for interpreting the present invention and do not limit the present invention.

[0025]

Figure 1

Figure 2a

Figure 2b

Figure 3

Figure 4

Figure 5

Embodiments for Carrying Out the Invention

[0026] To make the objectives, technical means, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. The following specific embodiments are only for explaining the present invention and do not limit the present invention.

[0027] In this specification, terms such as "comprising", "containing" or other variations mean non-exclusive inclusion, that is, a process, article or device containing a series of elements is intended to include not only these elements but also other elements not explicitly listed, or elements inherent in this process, article or device. Without further limitation, an element limited by the term "comprising one..." does not exclude other elements included in the process, article or device containing this element.

[0028] Embodiments of the present invention provide a design method for a multi-performance collaborative optimization high-entropy alloy composition based on machine learning. As shown in FIG. 1, this method includes the following steps.

[0029] S1: Create a high-entropy alloy composition-performance initial dataset based on the historical data information of the high-entropy alloy.

[0030] Specifically, conduct a literature survey on refractory high-entropy alloys, and create a high-entropy alloy composition-performance initial dataset containing 10 common elements of Al, Ti, V, Zr, Cr, Nb, Mo, Hf, Ta, and W, including alloy composition, high-temperature yield strength, and room-temperature compression plastic properties.

[0031] S2: Perform data cleaning on each performance parameter of the high-entropy alloy, standardize the cleaned data, and obtain a high-entropy alloy performance training set.

[0032] Specifically, delete the abnormal values and missing values of the reported alloy performance. For the performance data reported by different literatures for the same component, determine the relative error of each reported value. Average the duplicate values where the relative error is smaller than the error threshold, and delete the duplicate values where the relative error is larger than the error threshold to obtain the final alloy performance value. The alloy performance includes high-temperature yield strength and room-temperature plasticity. Normalize the cleaned alloy performance values to determine a high-temperature yield strength training dataset containing multiple samples and a room-temperature plasticity training dataset containing multiple samples.

[0033] In some embodiments, the error threshold is 10%.

[0034] In some embodiments, the high-temperature yield strength training dataset includes 140 samples, and the room-temperature plasticity training dataset includes 50 samples.

[0035] S3: Use the high-entropy alloy performance training set to train multiple machine learning models for each target performance of the high-entropy alloy, and screen and determine the base models for predicting each performance of the high-entropy alloy in turn.

[0036] Specifically, the model input is the content of each constituent element of the high-entropy alloy, and the model output is each target performance of the high-entropy alloy.

[0037] For the high-temperature yield strength training dataset and the room-temperature plasticity training dataset, a plurality of machine learning models are respectively trained, with the content of each constituent element of the high-entropy alloy as the input of the machine learning model and the high-temperature strength and room-temperature plasticity performance as the output of the machine learning model, and the mean squared error index parameter of each machine learning model is obtained.

[0038] The data for the high-temperature yield strength training dataset and the room-temperature plasticity training dataset is evenly divided into N groups. The machine learning model is trained and fitted using the first group of data, the remaining N - 1 groups of training data are predicted to obtain their performance prediction values, the root mean square error of the first power of the actual value and the prediction value is determined, and the same prediction operation is sequentially performed on the data from the second group to the Nth group to obtain the predicted root mean square error of the second power.

[0039] In some embodiments, N is 10, that is, the data for the high-temperature yield strength training dataset and the room-temperature plasticity training dataset is evenly divided into 10 groups. The machine learning model is trained and fitted using the first group of data, the remaining 9 groups of training data are predicted to obtain their performance prediction values, the root mean square error of the second power of the actual value and the prediction value is determined, and the same prediction operation is sequentially performed on the data from the second group to the 10th group to obtain the predicted root mean square error of the second power, and finally the cross-validation error of the machine learning model is obtained. The model with the minimum error is the prediction base model for the high-temperature strength and room-temperature plasticity performance of the alloy. Figures 2a and 2b respectively show the performance evaluation diagrams of the high-temperature strength and room-temperature plasticity performance prediction models of the alloy. The high-temperature strength prediction base model is an SVR model, and the room-temperature plasticity prediction base model is a GR model.

[0040] The machine learning model adopts at least one of the following five types: support vector machine (SVR), k-nearest neighbor method (KNN), Gaussian regression (GR), neural network (NN), and kernel ridge regression (KRR).

[0041] S4: Determine the expected improvement (EI) value of each performance of the high-entropy alloy obtained by searching in the component space based on the base model.

[0042] Specifically, resample the alloy performance training set by the bootstrap method to obtain M groups of resampled data sets. The number of samples included in each group of resampled data sets is the same as the number of the training set. Train the base model with each group of resampled data sets to obtain M performance prediction base models. Perform performance prediction on the high-entropy alloy obtained by searching in the component space by the M performance prediction base models, obtain M predicted values, determine the average value μ and variance σ of the performance predicted values, and use the formula:

Equation

[0043] In some embodiments, resample the training set by the bootstrap method to obtain 1000 groups of resampled data sets. The number of samples included in each group of resampled data sets is the same as the number of the training set. Train the base model with each group of resampled data sets to obtain 1000 performance prediction base models. Perform performance prediction on the alloy in the component search space by the 1000 base models to obtain 1000 predicted values. Assuming that this prediction follows a normal distribution, calculate to obtain the average value μ and variance σ of the predicted values of each performance of the alloy. The formula:

Number

[0044] S5: Determine the search space of the target components of the optimized high-entropy alloy, target the EI value of each performance of the high-entropy alloy, perform genetic optimization search, and determine the Pareto front of the EI value of the alloy performance in the search component space.

[0045] Specifically, according to the types of alloy constituent elements in the alloy performance training set, determine a high-dimensional component search space composed of elements of the types of Al, Ti, V, Zr, Cr, Nb, Mo, Hf, Ta, W, limit the search alloy to contain 3-6 types of elements, the molar content of each element in the alloy is 5% or more and 35% or less, and the variable step size of the content is 1%.

[0046] In the search space of the target components of the high-entropy alloy, each time a genetic search is executed, a Pareto front of the EI value of one alloy is obtained. The genetic search is executed N times to obtain N Pareto fronts. Perform non-dominated sorting on all alloys related to the N Pareto fronts, and finally determine the Pareto front of the EI value of the alloy performance to be genetically searched.

[0047] In some embodiments, perform target component search by the NSGA-II non-dominated sorting genetic algorithm, select an initial population number of 500, a genetic evolution generation number of 20, a population crossover rate of 0.8, and a population mutation rate of 0.02, and perform genetic optimization search 100 times by randomizing the initial population, obtain the Pareto front of the EI values of the high-temperature yield strength and room-temperature plastic properties of 100 alloys, perform non-dominated sorting on all alloys related to the 100 Pareto fronts, and finally determine the Pareto front of the EI value of the high-entropy alloy performance in the component space.

[0048] S6: Perform cluster analysis according to the Pareto front of the alloy performance EI value, screen to obtain the high-entropy alloy corresponding to the EI value and its related component information.

[0049] Specifically, perform cluster analysis on the Pareto front of the alloy EI value by the K-means clustering method, determine the optimal number of cluster centers by the elbow method, and use this number to perform clustering and screening on the Pareto front of the alloy performance EI value to obtain the alloy corresponding to the EI value and its related component information.

[0050] Before cluster analysis, logarithmically transform the target performance EI value on the Pareto front surface, then perform normalization processing, and then calculate the change in the within-cluster sum of squares (that is, the sum of the squares of the distances from each sample point to its cluster center) with the change in the number of cluster centers. When the decrease in the within-cluster sum of squares is very slow with the change in the number of cluster centers, the elbow of the number of clusters appears. The number K of cluster centers corresponding to the elbow is the actual number of cluster centers to be adopted. Perform clustering and screening on the Pareto front of the performance EI value with the K value to obtain the alloy corresponding to the EI value and its related component information.

[0051] S7: Produce the high-entropy alloy corresponding to the EI value-related component information by the vacuum arc melting furnace melting method.

[0052] Specifically, the steps of synthesizing and manufacturing the high-entropy alloy are as follows. Step 1: Use a metal with a purity higher than 99.9% as the alloy manufacturing raw material, polish the surface of the metal raw material to remove the surface oxide film, perform ultrasonic cleaning, and place it in a drying box for drying. Step 2: According to the alloy components to be screened, calculate the mass of the required metal raw materials based on the molar content of its constituent elements, and further measure the weight of the metal raw materials ultrasonically cleaned in Step 1 and store them until use. Step 3: Smelt the alloy in a non-consumable vacuum arc furnace. Put the metal raw materials into the melting tank in the furnace in the order from low melting point to high melting point. Evacuate the sample chamber and introduce high-purity argon gas for protection. Step 4: Turn on the welding power supply of the non-consumable vacuum arc furnace. After generating an arc, melt the titanium sponge placed in the combustion chamber in advance and perform deoxidation in the furnace. Step 5: Refine the alloy 5 - 8 times repeatedly. After each smelting, cool the alloy and turn it over to obtain a high-entropy alloy ingot.

[0053] S8: Conduct characteristic evaluation tests for each target performance of the manufactured high-entropy alloy.

[0054] S9: Evaluate whether the performance of the designed material meets the requirements. If it meets, end the design and obtain a high-entropy alloy with multi-performance coordinated optimization. If it does not meet, execute S10.

[0055] S10: Feed back the new alloy components and performance information to the high-entropy alloy component - performance initial dataset. Repeat S4 - S9 until the designed alloy meets the requirements of multi-performance synchronous optimization and improvement or reaches the upper limit of the budget, and execute the closed-loop path of the multi-performance coordinated optimization high-entropy alloy component design.

[0056] Example 1 The embodiment of the present invention provides a design method for a multi-performance coordinated optimization high-entropy alloy component based on machine learning. As shown in Figure 1, it includes the following steps.

[0057] S1: Conduct a literature survey on refractory high-entropy alloys, create a refractory high-entropy alloy dataset containing 10 common elements of Al, Ti, V, Zr, Cr, Nb, Mo, Hf, Ta, and W, including alloy components, 1000 °C high-temperature yield strength, and room-temperature compression plastic properties.

[0058] S2: Remove the abnormal and missing values of the reported alloy properties. Calculate the relative error of each reported value for the performance data of the same component reported in different documents. Average the duplicate values with an error less than 10% to obtain the final alloy performance value, and remove the duplicate values with an error greater than 10%. Normalize the cleaned alloy performance values to determine a dataset for high-temperature yield strength training containing 140 alloy samples and a training dataset for alloy room-temperature plasticity containing 50 samples.

[0059] S3: Train five machine learning models, namely Support Vector Machine (SVR), K-Nearest Neighbor (KNN), Gaussian Regression (GR), Neural Network (NN), and Kernel Ridge Regression (KRR), on the high-temperature strength performance training set and the room-temperature plasticity performance training set. Use the content of each constituent element of the high-entropy alloy as the model input, and high-temperature strength and room-temperature plasticity as the model output. Adopt the grid search method to optimize each model parameter according to the mean squared error index parameter, and train the determined optimal parameter model using the training set.

[0060] Furthermore, divide the data for each of the two training sets evenly into 10 groups and independently perform the following model evaluation operations. Use the first group of data to fit the machine learning model, predict the remaining 9 groups of training data to obtain their performance prediction values, calculate the root mean square error of the squared difference between the actual value and the prediction value, and then perform the same prediction operation on the data from the second group to the tenth group in sequence to obtain the predicted root mean square error of the squared difference, and finally obtain the cross-validation error of the machine learning model. The model with the minimum error was the prediction base model for the high-temperature strength and room-temperature plasticity of the alloy. Figures 2a and 2b show the performance evaluation diagrams of the high-temperature strength and room-temperature plasticity performance prediction models of the alloy, respectively. The high-temperature strength prediction base model was selected as the SVR model, and the room-temperature plasticity prediction base model was selected as the GR model.

[0061] S4: For the target properties of the high-temperature strength and room-temperature plasticity of the alloy, the expected improvement (EI) values of the target properties were calculated by the following operations respectively. The training set was resampled by the bootstrap method to obtain 1000 groups of resampled data sets. The number of samples included in each group of resampled data sets was the same as that of the training set. Base models were trained respectively by each group of resampled data sets to obtain 1000 performance prediction base models. Performance predictions were made for the alloys within the component search space by these 1000 base models to obtain 1000 predicted values. Assuming that this prediction follows a normal distribution, the mean value μ and variance σ of the predicted values of each target property were calculated. Formula:

Number

[0062] S5: According to the types of alloy composition elements in the training set, a high-dimensional component search space consisting of 10 elements of Al, Ti, V, Zr, Cr, Nb, Mo, Hf, Ta, and W was determined. The search alloy was limited to contain 3 - 6 types of elements. The molar content of each element in the alloy was 5% or more and 35% or less, and the variable step size of the content was 1%.

[0063] According to the calculation method of the alloy performance EI value determined in S4, the target component search was performed by the NSGA-II non-dominated sorting genetic algorithm. The initial population number 500, the genetic evolution algebra 20, the population crossover rate 0.8, and the population mutation rate 0.02 were selected. Genetic optimization search was performed 100 times by randomizing the initial population to obtain 100 Pareto fronts of the EI values of the high-temperature strength and room-temperature plasticity properties of the alloy. Non-dominated sorting was performed for all alloys related to the 100 Pareto fronts to determine the Pareto front of the EI value of the high-entropy alloy performance in the component space.

[0064] For the Pareto front of the EI value of the high-entropy alloy performance determined by S6:S5, it was analyzed by the K-means clustering method, and the number of cluster centers was determined by the "elbow method". Specifically, the target performance EI values on the Pareto front surface were logarithmically transformed, then normalized, and then the change in the within-cluster sum of squares (that is, the sum of the squares of the distances from each sample point to its cluster center) associated with the change in the number of cluster centers was calculated. When the decrease in the within-cluster sum of squares was very slow with the change in the number of cluster centers, the elbow of the number of clusters appeared. The number K of cluster centers corresponding to the elbow was the number of cluster centers actually adopted. Clustering and screening were performed on the Pareto front of the performance EI value with the K value, and the alloy corresponding to the EI value and its related component information were obtained.

[0065] S7: The target alloys screened by S6 were synthesized and manufactured. The steps are as follows. Step 1: Metals with a purity higher than 99.9% were used as alloy manufacturing raw materials. The surfaces of the metal raw materials were polished to remove the surface oxide films, ultrasonically cleaned, and placed in a drying box for drying. Step 2: According to the alloy components to be screened, the masses of the required metal raw materials were calculated based on the molar contents of their constituent elements. Furthermore, the weights of the metal raw materials ultrasonically cleaned in Step 1 were measured and stored until use. Step 3: The alloy was smelted by a non-consumable vacuum arc furnace. The metal raw materials were sequentially placed into the melting tank in the furnace in the order from low melting point to high melting point. The sample chamber was evacuated, and high-purity argon gas was introduced for protection. Step 4: The welding power supply of the non-consumable vacuum arc furnace was turned on. After the arc was generated, the titanium sponge placed in the combustion chamber in advance was melted to perform in-furnace deoxidation. Step 5: The alloy was refined 6 times in succession. After each smelting, the alloy was cooled and turned over to obtain a high-entropy alloy ingot.

[0066] The high-entropy alloy manufactured in S8:S7 was machined to obtain test samples. Performance tests of high-temperature strength at 1000 °C and room-temperature plasticity were respectively carried out to obtain the component and target performance data information of the newly manufactured alloy. Furthermore, feedback was given to the initial data set to perform iterative collaborative optimization of the two target performances.

[0067] The information of two types of refractory high-entropy alloys designed and manufactured through three iterative experiments according to the above implementation steps was as follows. Designed alloy A-Mo 0.14 Nb 0.28 Ta 0.2 Hf 0.15 Zr 0.23 Designed alloy B-Mo 0.2 Nb 0.26 Ta 0.19 Hf 0.16 Zr 0.19

[0068] Figure 3 shows the high-temperature strength and room-temperature plasticity of the refractory high-entropy alloy designed by the method of the present invention. Compared with the performance of conventional alloys, it shows a significantly higher multi-performance collaborative optimization effect. Figure 4 shows the room-temperature mechanical property curves of the two designed alloys A and B. Figure 5 shows the high-temperature mechanical property curves of the two designed alloys A and B. As can be seen from Figures 4 and 5, the room-temperature plasticity of alloy A is 35.7%, and the high-temperature yield strength at 1000 °C is 894 MPa. The room-temperature plasticity of alloy B is 30.5%, and the high-temperature yield strength at 1000 °C is 974 MPa, indicating the high efficiency of the multi-performance collaborative optimization design for the high-entropy alloy according to the present invention.

[0069] The above description is only a preferred embodiment of the present invention and does not limit the protection scope of the present invention.

Claims

1. A method for designing a multi-performance cooperative optimization high-entropy alloy composition based on machine learning, comprising: creating a high-entropy alloy composition-performance initial dataset based on the historical data information of the high-entropy alloy; performing data cleaning on each performance parameter of the high-entropy alloy, and performing standardization processing on the cleaned data to obtain a high-entropy alloy performance training set; training a plurality of machine learning models for each target performance of the high-entropy alloy by using the high-entropy alloy performance training set, and screening and determining a base model for predicting each performance of the high-entropy alloy in sequence; determining the expected improvement (EI) value of each performance of the high-entropy alloy obtained by searching in the component space by using the base model; determining the search space of the target component of the high-entropy alloy in the optimized high-entropy alloy performance training set, performing genetic optimization search with the EI value of each performance of the high-entropy alloy as the target, and determining the Pareto front of the alloy performance EI value in the search component space; performing cluster analysis on the Pareto front of the EI value, screening to obtain the high-entropy alloy corresponding to the EI value and its related component information; manufacturing a high-entropy alloy corresponding to the component information related to the EI value by using a vacuum arc melting furnace melting method; performing a characteristic evaluation test on each target performance of the manufactured high-entropy alloy, and if the test data does not meet the requirements, updating the high-entropy alloy composition-performance initial dataset based on the test data. A method characterized by including the above steps.

2. In the step of creating a high-entropy alloy composition-performance initial dataset based on the historical data information of the high-entropy alloy, specifically, conducting a literature survey on refractory high-entropy alloys, and creating a high-entropy alloy composition-performance initial dataset containing 10 common elements of Al, Ti, V, Zr, Cr, Nb, Mo, Hf, Ta, and W, and including alloy composition, high-temperature yield strength, and room-temperature compression plastic performance. The method according to Claim 1.

3. Perform data cleaning on each performance parameter of the high-entropy alloy. Specifically, delete the abnormal values and missing values of the reported alloy performance. For the performance data reported by different literatures for the same component, determine the relative error of each reported value. Average the duplicate values where the relative error is smaller than the error threshold, delete the duplicate values where the relative error is larger than the error threshold, obtain the final alloy performance value. The alloy performance includes high-temperature yield strength and room-temperature plasticity. Normalize the cleaned alloy performance value, and determine a high-temperature yield strength training data set including a plurality of samples and a room-temperature plasticity training data set including a plurality of samples. The method according to claim 1 or 2, characterized in that.

4. In the step of training a plurality of machine learning models for each target performance of the high-entropy alloy by using the high-entropy alloy performance training set, and screening and determining the base model for predicting each performance of the high-entropy alloy in turn. Specifically, evenly divide the high-temperature yield strength training data set and the room-temperature plasticity training data set into N groups. Use the first group of data to train and fit the machine learning model, predict the remaining N - 1 groups of training data to obtain their performance prediction values, determine the first root mean square error of the actual value and the prediction value. Further, perform the same prediction operation on the data from the second group to the Nth group in turn to obtain the predicted second root mean square error. Determine the cross-validation error of each machine learning model based on the first root mean square error and the second root mean square error. The machine learning model with the minimum cross-validation error is the base model for predicting the high-temperature strength and room-temperature plasticity performance of the alloy respectively. The method according to claim 3, characterized in that.

5. In the step of obtaining the expected improvement (EI) value of each target performance of the high-entropy alloy obtained by searching in the component space using the base model, specifically, resampling the alloy performance training set by the bootstrap method to obtain M groups of resampled data sets, the number of samples included in each group of resampled data sets is the same as the number of the training set, training the base model with each group of resampled data sets respectively to obtain M performance prediction base models, performing performance prediction on the high-entropy alloy obtained by searching in the component space using the M performance prediction base models to obtain M predicted values, determining the average value μ and variance σ of the performance predicted values, and the formula: 【Number 1】 determine an expected improvement (EI) value of the alloy performance, where μ ＊ represents the optimal value of the alloy performance in the training set, φ represents the probability density function, and Φ represents the probability distribution function, the method according to claim 4.

6. In the step of determining the search space of the high-entropy alloy target components in the optimized high-entropy alloy performance training set, specifically, according to the types of alloy constituent elements in the alloy performance training set, determining the search space of the high-entropy target components composed of elements of the types of Al, Ti, V, Zr, Cr, Nb, Mo, Hf, Ta, W, limiting the search alloy to contain 3 to 6 types of elements, and the molar content of each element in the alloy is 5% or more and 35% or less, and the variable step size of the content is 1%, the method according to claim 5.

7. In the step of performing genetic optimization search with the EI value of each performance of the high-entropy alloy as the target and determining the Pareto front of the EI value in the search component space, specifically, in the search space of the target components of the high-entropy alloy, each time the genetic search is executed once, one Pareto front of the alloy EI value is obtained, the genetic search is executed N times to obtain N Pareto fronts, performing non-dominated sorting on the N Pareto front alloys, and finally determining the Pareto front of the EI value of the alloy performance to be genetically searched, the method according to claim 6.

8. In the step of performing cluster analysis according to the Pareto front of the alloy performance EI value and screening to obtain the high-entropy alloy corresponding to the performance EI value and the related component information, specifically, the Pareto front of the alloy performance EI value is cluster-analyzed by the K-means clustering method, the optimal number of cluster centers is determined by the elbow method, clustering and screening are performed on the Pareto front of the alloy performance EI value using this number, and the alloy corresponding to the EI value and the related component information are obtained. The method according to claim 7, characterized in that.

9. The method according to claim 8, further comprising the step of terminating the design and obtaining a high-entropy alloy in which multiple performances are cooperatively optimized when the test data meets the requirements.

10. The method according to claim 8, further comprising the step of designing each high-entropy alloy in the high-entropy alloy component-performance initial data set until the multiple performances of the designed alloy simultaneously meet the requirements of optimization and improvement or reach the upper limit of the budget.

Images