A method and system for predicting a high-performance new alloy

Abstract

The application discloses a kind of high-performance new alloy prediction method and system, method includes: obtaining historical alloy data set;The historical alloy data set is preprocessed, and the alloy element characteristic factor set of each historical alloy is obtained;Key feature factor combination is obtained by correlation analysis method, gradient descent method and SHAP method to alloy element characteristic factor set;According to the data situation, a large number of different model hyperparameter combinations are constructed using the Otuna algorithm, neural network model hyperparameter self-optimization and model self-selection are carried out in combination with key feature factor combination, whether the best model is obtained according to model generalization ability and test accuracy;The element characteristic factor data of all elements in periodic table of elements is obtained;According to element characteristic factor data and the best model, high-performance new alloy is predicted.The problem that the machine learning model generalization ability is weak in the process of manually adjusting hyperparameters, the accuracy is low due to model overfitting or underfitting.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of material design, and particularly relates to a prediction method and system of a high-performance new alloy. BACKGROUND

[0002] Metallic materials have been of great significance in human history since the Bronze Age, and have been rapidly updated along with the development needs of human civilization. Generally, the traditional alloy composition design strategy includes adding a relatively small amount of various elements in a main element, such as steel, aluminum alloy, titanium alloy, magnesium alloy and copper alloy. Therefore, the inherent properties of the alloy depend largely on the main element. However, since 2004, Cantor team and Ye Junwei team respectively proposed an alloy design idea of eliminating the distinction between main elements and trace elements, namely, alloy “high entropy”. Based on this idea, an alloy containing multiple main elements is prepared and named “high-entropy alloy”. Compared with ordinary alloys, such alloys exhibit more excellent strength, hardness, corrosion resistance and wear resistance, thereby leading a new wave of material research and attracting the attention of many material researchers.

[0003] At present, empirical trial and error and phase diagram theory are still the most common methods for developing new materials. Although methods such as molecular dynamics, thermodynamic calculation and high-throughput calculation can accelerate the process of alloy development to some extent, due to the large number of elements contained in high-entropy alloys, the composition design space constructed is huge, and the calculation cost is very high. Machine learning, as a multi-disciplinary subject involving probability theory and statistics, is essentially a method of using algorithms to enable computers to efficiently and autonomously construct a complex nonlinear relationship between input data and output targets from existing data, and use this relationship to build a model to predict unknown data.

[0004] Applying machine learning methods to the design of alloy materials involves first constructing a machine learning classification or regression model, adjusting the model training steps and learning factors manually to train and reduce the error value of the model, and then predicting the organization and performance of unknown alloys.

[0005] However, applying machine learning methods to the design of alloy materials does not consider the possibility of optimizing the machine learning classification or regression model itself. Each time the hyperparameters of the model are manually adjusted, a model with strong directionality is obtained, and then the model is locally optimized. Due to the limitations of manual adjustment, the model may not be the most suitable for the data at this time. Manually adjusting the hyperparameters is essentially an empirical trial and error method, and the model has a high contingency. SUMMARY

[0006] The application provides a prediction method and system of a high-performance new alloy, realizes a change from a traditional manual local optimization to an automatic global optimization mode based on coupling and linkage of a model and data in a model construction and optimization process, and solves the problem of low accuracy caused by weak generalization ability of a machine learning model, model overfitting or underfitting in a manual super parameter adjustment process.

[0007] To solve the above technical problems, the technical scheme adopted by the application is:

[0008] In a first aspect, a prediction method of a high-performance new alloy is provided, comprising:

[0009] Obtaining a historical alloy data set, the historical alloy data set containing alloy data of a plurality of historical alloys, each alloy having a plurality of alloy elements;

[0010] Data preprocessing is performed on the historical alloy data set to obtain an alloy element characteristic factor set of each historical alloy;

[0011] The alloy element characteristic factor set is screened by a correlation analysis method, a gradient descent method and a SHAP method to obtain a key characteristic factor combination;

[0012] A neural network model of a large number of different model super parameter combinations is constructed according to data conditions by using an Otuna algorithm, the key characteristic factor combination is put into the neural network model for super parameter self-optimization and model self-screening, and the best model is obtained according to model generalization ability and test accuracy;

[0013] Obtaining element characteristic factor data of all elements in the periodic table of elements;

[0014] Predicting a high-performance new alloy according to the element characteristic factor data and the best model.

[0015] Optionally, obtaining the historical alloy data set comprises:

[0016] Obtaining alloy data of a plurality of historical alloys, the alloy data including composition data, process data, organization data and performance data of each historical alloy, the composition data including element content and mole fraction of each alloy element, and the sum of mole fractions of all alloy elements of each historical alloy being equal to 100%;

[0017] Obtaining the historical alloy data set according to the alloy data of each historical alloy.

[0018] Optionally, data preprocessing is performed on the historical alloy data set to obtain an alloy element characteristic factor set of each historical alloy, comprising:

[0019] Extracting element content and mole fraction of each alloy element in each historical alloy from the historical alloy data set;

[0020] According to the element content and the mole fraction, the average characteristic factor, the mean square deviation characteristic factor and the deviation characteristic factor of each historical alloy are calculated;

[0021] According to the average characteristic factor, the mean square deviation characteristic factor and the deviation characteristic factor, the alloy element characteristic factor set of each historical alloy is obtained.

[0022] Optionally, the alloy element characteristic factor set is screened by a correlation analysis method, a gradient descent method and a SHAP method to obtain a key characteristic factor combination, including:

[0023] The pearson correlation analysis method is used to calculate the characteristic factor correlation coefficient value between any two characteristic factors in the alloy element characteristic factor set, and the target correlation coefficient value between each characteristic factor and target data, including organization data and performance data;

[0024] When the absolute value of the characteristic factor correlation coefficient value is not greater than a preset threshold value, the corresponding two characteristic factors are removed;

[0025] When the absolute value of the characteristic factor correlation coefficient value is greater than the preset threshold value, the target correlation coefficient values of the corresponding two characteristic factors are compared, the characteristic factor with the smaller target correlation coefficient value is removed, and the remaining characteristic factor is taken as a first characteristic factor;

[0026] The support vector model of the first characteristic factor and the target data is established by the gradient descent method;

[0027] The model accuracy of the support vector model at each gradient descent is calculated according to K-fold cross-validation;

[0028] The key characteristic factor combination is selected from the first characteristic factor according to the model accuracy;

[0029] The SHAP method is used to evaluate the feature importance of the key characteristic factor combination to obtain the best key characteristic factor combination.

[0030] Optionally, the neural network model of a large number of different model hyperparameter combinations is constructed by using the otuna algorithm according to the data situation, the key characteristic factor combination is put into the neural network model for hyperparameter self-optimization and model self-screening, and the best model is obtained according to the model generalization ability and test accuracy, including:

[0031] The neural network model framework of the key characteristic factor and the target data is initialized by the multilayer perception MLP algorithm;

[0032] The initialization-based neural network model framework constructs a plurality of different model hyperparameter combinations of neural network models according to preset hyperparameters through an Optuna algorithm;

[0033] The key feature factor combination is put into the neural network model for hyperparameter self-optimization and model self-screening.

[0034] The best model is obtained according to the model generalization ability and test accuracy.

[0035] Optionally, the preset threshold of the absolute value of the correlation coefficient of the feature factors is 0.9.

[0036] Optionally, the K-fold cross-validation is 5-fold or 10-fold.

[0037] Optionally, the different preset hyperparameters include the number of neural network layers, the number of neuron nodes, the optimizer, the learning rate and the loss rate, the number of neural network layers is 2 to 10, the number of neuron nodes is 8 to 1024, the optimizer is Adam, RMSprop or SGD, the learning rate is 10 -6 to 10 -1 , and the loss rate is 0.01 to 0.6.

[0038] In a second aspect, a prediction system for a high-performance new alloy is provided, comprising:

[0039] A data acquisition module is configured to acquire a historical alloy data set, wherein the historical alloy data set comprises alloy data of a plurality of historical alloys, and each alloy has a plurality of alloy elements.

[0040] A data processing module is configured to perform data preprocessing on the historical alloy data set to obtain an alloy element feature factor set of each historical alloy.

[0041] A feature factor screening module is configured to screen the alloy element feature factor set by using a correlation analysis method, a gradient descent method and a SHAP method to obtain a key feature factor combination.

[0042] A model optimization module is configured to construct a plurality of different model hyperparameter combinations of neural network models according to data conditions by using an Optuna algorithm, put the key feature factor combination into the neural network model for hyperparameter self-optimization and model self-screening, and obtain a best model according to a model generalization ability and a test accuracy.

[0043] An alloy prediction module is configured to acquire element feature factor data of all elements in a periodic table, and predict a high-performance new alloy according to the element feature factor data and the best model.

[0044] The present application has the following advantages:

[0045] Obtain a historical alloy data set, the historical alloy data set containing alloy data of a plurality of historical alloys, each alloy having a plurality of alloy elements; perform data preprocessing on the historical alloy data set to obtain an alloy element characteristic factor set of each historical alloy; filter the alloy element characteristic factor set through a correlation analysis method, a gradient descent method and a SHAP method to obtain a key characteristic factor combination; use an otuna algorithm to construct a neural network model of a large number of different model hyperparameter combinations according to data conditions, put the key characteristic factor combination into the neural network model to perform hyperparameter self-optimization and model self-filtering, and obtain an optimal model according to model generalization ability and test accuracy; obtain element characteristic factor data of all elements in the periodic table of elements; and predict a high-performance new alloy according to the element characteristic factor data and the optimal model. The key characteristic factor combination of the historical alloy obtained after filtering is used to filter and automatically adjust the neural network model of a large number of different preset hyperparameters to obtain the optimal model, thereby solving the problem of low accuracy caused by weak generalization ability of the machine learning model, model overfitting or underfitting in the process of manually adjusting the hyperparameters. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 A flowchart of the prediction method of the high-performance new alloy of the present application;

[0047] Figure 2 A process diagram of data preprocessing of the present application;

[0048] Figure 3 A structure diagram of the prediction system of the high-performance new alloy of the present application. DETAILED DESCRIPTION

[0049] The present application will be further described below in conjunction with the drawings. The following examples are only used to more clearly illustrate the technical solutions of the present application, and cannot be used to limit the protection scope of the present application.

[0050] As shown in Figure 1 A prediction method of a high-performance new alloy, comprising the following steps:

[0051] 101, obtaining a historical alloy data set;

[0052] The alloy data of a plurality of historical alloys is collected from published literature or experimental records, the alloy data including composition data, process data, structure data and performance data of each historical alloy, the composition data including element content and mole fraction of each alloy element, the sum of mole fractions of all alloy elements of each historical alloy being equal to 100%; an alloy data set is obtained according to the alloy data of each historical alloy. The performance data includes hardness, strength, elongation and electrical conductivity; and the properties of alloy elements include atomic size information, electronic information, physical information of atoms, chemical information of atoms, thermal information and magnetic information.

[0053] It should be noted that in the process of model training by using the historical alloy data set, generally, three stages of training, verification and testing need to be experienced, and then extraction needs to be performed according to the proportion, and the proportion of the training set, the verification set and the test set is 7:1:2, 6:1:3 or 5:1:4.

[0054] 102, data preprocessing is performed on the historical alloy data set to obtain an alloy element feature factor set of each historical alloy;

[0055] The data preprocessing process is shown in Figure 2 , and the specific process is as follows:

[0056] 201, the element content and mole fraction of each alloy element in each historical alloy are extracted from the historical alloy data set;

[0057] wherein c is the mole fraction of a certain alloy element, and is the element content of a certain alloy element.

[0058] 202, the average feature factor, the mean square deviation feature factor and the deviation feature factor of each historical alloy are calculated according to the element content and the mole fraction;

[0059] The average feature factor f is calculated by the expression of the average feature factor m , wherein i represents each feature quantity of a certain alloy element, j represents a certain alloy element, and n is the component number of the alloy, the mean square deviation feature factor f is calculated by the expression of the mean square deviation feature factor v , and the deviation feature factor f is calculated by the expression of the deviation feature factor z .

[0060] 203, the alloy element feature factor set of each historical alloy is obtained according to the average feature factor, the mean square deviation feature factor and the deviation feature factor.

[0061] Wherein, each historical alloy average characteristic factor, mean square deviation characteristic factor and deviation characteristic factor are collected to obtain an alloy element characteristic factor set of each historical alloy.

[0062] 103, the alloy element characteristic factor set is screened by correlation analysis method, gradient descent method and SHAP method to obtain a key characteristic factor combination.

[0063] Wherein, the pearson correlation analysis method is used to calculate the characteristic factor correlation coefficient value between any two characteristic factors in the alloy element characteristic factor set, and the target correlation coefficient value between each characteristic factor and the target data, the target data including the organization data and the performance data; the specific formula is r represents the correlation coefficient, and is the value of the two alloy characteristic factors corresponding to the alloy in the ith group of data, and is the average value of the two alloy characteristic factors in all data, when the absolute value of the characteristic factor correlation coefficient value is not greater than the preset threshold (set to 0.9), the corresponding two characteristic factors are removed; when the absolute value of the characteristic factor correlation coefficient value is greater than the preset threshold (set to 0.9), the target correlation coefficient value of the corresponding two characteristic factors is compared, the characteristic factor with smaller target correlation coefficient value is removed, and the remaining characteristic factor is taken as the first characteristic factor; It should be noted that the first characteristic factor here is not only a characteristic factor, but also a collection of the remaining characteristic factors;

[0064] The support vector model of the first characteristic factor and the target data is established by the gradient descent method;

[0065] In order to prevent overfitting or underfitting of the model, the model accuracy of the support vector model at each gradient descent is calculated according to K-fold cross validation, K-fold cross validation is 5-fold or 10-fold; the calculation method of the model accuracy is as follows:

[0066]

[0067] Wherein, represents the prediction result, represents the test result.

[0068] According to the SHAP method, the key characteristic factor combination is evaluated to obtain the best key characteristic factor combination.

[0069] 104, a large number of different model hyperparameter combinations of neural network models are constructed according to data conditions by using the Otona algorithm, the key feature factor combination is put into the neural network model for hyperparameter self-optimization and model self-screening, and the best model is obtained according to the model generalization ability and test accuracy;

[0070] The neural network model framework of the key feature factor and the target data is initialized by a Multilayer Perceptron (MLP) algorithm; the initialized neural network model framework is used to construct a neural network model according to preset hyperparameters by an Otona algorithm; different preset hyperparameters include the number of neural network layers, the number of neuron nodes, an optimizer, a learning rate and a loss rate, the number of neural network layers is 2 to 10, the number of neuron nodes is 8 to 1024, the optimizer is Adam, RMSprop or SGD, the learning rate is 10 -6 to 10 -1 , and the loss rate is 0.01 to 0.6. It should be noted that the range and definition of the preset hyperparameters are set in advance according to actual experience.

[0071] The key feature factor combination is put into the neural network model for hyperparameter self-optimization and model self-screening, and the best model is obtained according to the model generalization ability and test accuracy.

[0072] 105, obtaining element feature factor data of all elements in the periodic table;

[0073] The periodic table is known, and each element in the periodic table is extracted and converted into element feature factor data.

[0074] 106, predicting a high-performance new alloy according to the element feature factor data and the best model.

[0075] The best model is used for prediction, and a new high-performance alloy is predicted according to the best model, and the organization and performance of the high-performance alloy are detected, and the alloy composition is selected and prepared.

[0076] The implementation principle of the embodiment of the application is:

[0077] The historical alloy data set contains alloy data of a plurality of historical alloys, each alloy having a plurality of alloy elements; the historical alloy data set is subjected to data preprocessing to obtain an alloy element characteristic factor set of each historical alloy; the alloy element characteristic factor set is screened through a correlation analysis method, a gradient descent method and a SHAP method to obtain a key characteristic factor combination; an Otuna algorithm is used to construct a neural network model of a large number of different model hyperparameter combinations according to data conditions, the key characteristic factor combination is put into the neural network model for hyperparameter self-optimization and model self-screening, and the best model is obtained according to model generalization ability and test accuracy; element characteristic factor data of all elements in the periodic table of elements is obtained; and a high-performance new alloy is predicted according to the element characteristic factor data and the best model. The key characteristic factor combination of the historical alloy obtained after screening is used to screen and automatically adjust the large number of neural network models with different preset hyperparameters to obtain the best model, thereby solving the problem of low accuracy caused by weak generalization ability of the machine learning model and overfitting or underfitting of the model in the process of manually adjusting the hyperparameters.

[0078] In the above description, the prediction method of the high-performance new alloy is described in detail, and the prediction system of the high-performance new alloy is described through an embodiment as shown in Figure 3

[0079] The data acquisition module 301 is configured to acquire a historical alloy data set, and the historical alloy data set contains alloy data of a plurality of historical alloys, each alloy having a plurality of alloy elements.

[0080] The data processing module 302 is configured to perform data preprocessing on the historical alloy data set to obtain an alloy element characteristic factor set of each historical alloy.

[0081] The characteristic factor screening module 303 is configured to screen the alloy element characteristic factor set through a correlation analysis method, a gradient descent method and a SHAP method to obtain a key characteristic factor combination.

[0082] The model optimization module 304 is configured to construct a neural network model of a large number of different model hyperparameter combinations according to data conditions by using an Otuna algorithm, put the key characteristic factor combination into the neural network model for hyperparameter self-optimization and model self-screening, and obtain the best model according to model generalization ability and test accuracy.

[0083] The alloy prediction module 305 is configured to obtain element characteristic factor data of all elements in the periodic table of elements, and predict a high-performance new alloy according to the element characteristic factor data and the best model.

[0084] The implementation principle of the embodiment of the application is as follows:

[0085] ​The data acquisition module 301 acquires a historical alloy data set, the historical alloy data set containing alloy data of a plurality of historical alloys, each alloy having a plurality of alloy elements; the data processing module 302 performs data preprocessing on the historical alloy data set to obtain an alloy element characteristic factor set of each historical alloy; the characteristic factor screening module 303 screens the alloy element characteristic factor set by a correlation analysis method and a gradient descent method to obtain a key characteristic factor combination; the model optimization module 304 optimizes and self-selects a neural network model constructed with different preset hyperparameters according to the key characteristic factor combination to obtain an optimal model; the alloy prediction module 305 acquires element characteristic factor data of all elements in the periodic table of elements; and a high-performance new alloy is predicted according to the element characteristic factor data and the optimal model. The key characteristic factor combination of the historical high-performance new alloy obtained after screening is used to screen and automatically adjust a large number of neural network models constructed with different preset hyperparameters to obtain the optimal model, thereby solving the problem of low accuracy caused by weak generalization ability of a machine learning model in a manual hyperparameter adjustment process, model overfitting or underfitting.

[0086] The data processing procedures and methods of the modules in the system are consistent, and are not described repeatedly.

[0087] Based on the same technical solution, the application further discloses a computer readable storage medium storing one or more programs, the one or more programs including instructions that, when executed by a computing device, cause the computing device to perform the prediction method of the high-performance new alloy described above.

[0088] Based on the same technical solution, the application further discloses a computing device including one or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and configured to be executed by the one or more processors, and the one or more programs include instructions for performing the prediction method of the high-performance new alloy described above.

[0089] Those skilled in the art should understand that embodiments of the application can be provided as a method, a system, or a computer program product. Therefore, the application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0090] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks. Figure 1 one or more flow or blocks.

[0091] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 one or more flow or blocks. Figure 1 one or more flow or blocks.

[0092] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks. Figure 1 one or more flow or blocks.

[0093] The above merely provides an embodiment of the present application, but is not intended to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall fall within the scope of the claims of the present application.

Claims

1. A method for predicting high-performance novel alloys, characterized in that, include: Obtain a historical alloy dataset, which contains alloy data for a variety of historical alloys, each alloy having multiple alloying elements; The historical alloy dataset is preprocessed to obtain a set of alloy element characteristic factors for each historical alloy. Using Pearson correlation analysis, the correlation coefficient between any two characteristic factors in the alloy element characteristic factor set is calculated, and the target correlation coefficient between each characteristic factor and the target data is calculated. The target data includes microstructure data and performance data. When the absolute value of the correlation coefficient of the feature factor is not greater than a preset threshold, the two corresponding feature factors are removed. When the absolute value of the correlation coefficient of the feature factor is greater than a preset threshold, the target correlation coefficient values ​​of the two corresponding feature factors are compared, the feature factor with the smaller target correlation coefficient value is removed, and the remaining feature factor is taken as the first feature factor. A support vector model of the first feature factor and the target data is established using the gradient descent method. The model accuracy of the support vector model is calculated for each gradient descent step based on K-fold cross-validation. Based on the model accuracy, a combination of key feature factors is selected from the first feature factors; The key feature factor combination is evaluated using the SHAP method to obtain the optimal key feature factor combination. The neural network model framework for key feature factors and target data is initialized using the multilayer perceptron (MLP) algorithm. The initialization-based neural network model framework uses the Optograph algorithm to construct a large number of neural network models with different combinations of model hyperparameters based on preset hyperparameters; The key feature factors are combined and incorporated into the neural network model for hyperparameter self-optimization and model self-selection. The optimal model is obtained based on the model's generalization ability and test accuracy. Obtain the element characteristic factor data for all elements in the periodic table; High-performance novel alloys are obtained based on the elemental characteristic factor data and the optimal model prediction.

2. The method for predicting high-performance novel alloys according to claim 1, characterized in that, The acquisition of historical alloy datasets includes: Acquire alloy data for multiple historical alloys, including composition data, process data, microstructure data, and property data for each historical alloy. The composition data includes the element content and mole fraction of each alloying element, and the sum of the mole fractions of all alloying elements in each historical alloy equals 100%. A historical alloy dataset is obtained based on the alloy data of each historical alloy.

3. The method for predicting high-performance novel alloys according to claim 2, characterized in that, The data preprocessing of the historical alloy dataset yields a set of alloy element characteristic factors for each historical alloy, including: The element content and mole fraction of each alloy element in each historical alloy were extracted from the historical alloy dataset. Based on the element content and the mole fraction, the average characteristic factor, root mean square characteristic factor, and deviation characteristic factor of each historical alloy are calculated. Based on the average characteristic factor, the root mean square error characteristic factor, and the deviation characteristic factor, the set of alloying element characteristic factors for each historical alloy is obtained.

4. The method for predicting high-performance novel alloys according to claim 1, characterized in that, The preset threshold for the absolute value of the correlation coefficient of the feature factor is 0.

9.

5. The method for predicting high-performance novel alloys according to claim 1, characterized in that, The K-fold cross-validation is either 5-fold or 10-fold.

6. The method for predicting high-performance novel alloys according to any one of claims 1-5, characterized in that, Different preset hyperparameters include the number of neural network layers, the number of neurons, the optimizer, the learning rate, and the dropout rate. The number of neural network layers is 2 to 10, the number of neurons is 8 to 1024, the optimizer is Adam, RMSprop, or SGD, and the learning rate is 10. -6 Up to 10 -1 The loss rate is 0.01 to 0.

6.

7. A prediction system for high-performance novel alloys, characterized in that, include: The data acquisition module is used to acquire historical alloy datasets, which contain alloy data of various historical alloys, each alloy having multiple alloying elements; The data processing module is used to preprocess the historical alloy dataset to obtain the alloy element characteristic factor set for each historical alloy. The feature factor screening module is used to calculate the feature factor correlation coefficient between any two feature factors in the alloy element feature factor set using Pearson correlation analysis, and to calculate the target correlation coefficient between each feature factor and target data, where the target data includes microstructure data and performance data. When the absolute value of the feature factor correlation coefficient is not greater than a preset threshold, the corresponding two feature factors are removed. When the absolute value of the feature factor correlation coefficient is greater than the preset threshold, the target correlation coefficient values ​​of the corresponding two feature factors are compared, the feature factor with the smaller target correlation coefficient value is removed, and the remaining feature factor is taken as the first feature factor. A support vector model of the first feature factor and the target data is established using gradient descent; the model accuracy of the support vector model is calculated for each gradient descent step using K-fold cross-validation; and a combination of key feature factors is selected from the first feature factor based on the model accuracy. The key feature factor combination is evaluated using the SHAP method to obtain the optimal key feature factor combination. The model optimization module is used to initialize the neural network model framework of key feature factors and target data using the multilayer perceptron (MLP) algorithm; based on the initialized neural network model framework, the Optograph algorithm is used to construct a large number of neural network models with different combinations of model hyperparameters according to preset hyperparameters; the key feature factor combinations are put into the neural network models for hyperparameter self-optimization and model self-selection; the optimal model is obtained based on the model's generalization ability and test accuracy. The alloy prediction module is used to predict high-performance novel alloys based on the elemental characteristic factor data and the optimal model.

Images