Descriptor screening based on material genetic engineering and method for predicting physical properties of crystal materials

Abstract

The application discloses a descriptor screening and crystal material physical property prediction method based on material genetic engineering, and the method comprises the following steps: collecting crystal descriptor by calculating and collecting part of crystal material target property data, then carrying out cleaning and comprehensive screening on the descriptor by virtue of variance information, four correlation coefficients and a random sequential feature selection method, then automatically optimizing and adjusting the hyperparameter setting in the model building process by using intelligent optimization thinking, finally applying the trained optimal model to the physical property prediction of the crystal material to obtain the prediction result, and overcoming the non-standard in the model building process of the material physical property prediction model, and improving the learning efficiency and prediction effect of the model.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of material science research and development, and particularly relates to a descriptor screening and crystal material property prediction method based on material genetic engineering. BACKGROUND

[0002] Material genetic engineering is a new material research and development mode combining high-throughput calculation, high-throughput experiment, big data technology and the like. It innovates material discovery means, shortens research and development cycle and cost, and promotes the development of new material industry. Among them, material property prediction based on big data is one of the important contents of material genetic engineering. Material research and development under big data technology mainly realizes the prediction of the properties of unknown materials by training a model of known material properties through an artificial intelligence model.

[0003] Due to the development of computer technology and the improvement of computing power, more density functional theory calculation results and molecular dynamics simulation results with prediction performance are generated and stored in databases. By combining these data as objective functions with machine learning methods, the prediction and classification of target physical properties can be realized.

[0004] For crystal material property prediction under big data technology, the key of the model lies in the data (descriptors). The category, quantity and quality of the descriptors play a crucial role in material data machine learning. At present, there are many characterization technologies and data processing methods to measure and create descriptors for the composition, crystal structure and other characteristics of crystal materials. However, the crystal material properties (such as thermal conductivity, electrical conductivity, etc.) as the screening target are usually difficult to establish a quantitative relationship with the existing descriptors, and too much redundant data can easily lead to overfitting of the artificial intelligence model, reducing its generalization ability, which makes the screening of descriptors particularly important. In addition, in the training and building process of a typical machine learning model, the model prediction accuracy often needs to be improved by manually adjusting the hyperparameters, which makes the training process often consume a lot of computing power for model parameter adjustment work. SUMMARY

[0005] The purpose of the present application is to provide a descriptor screening and crystal material property prediction method based on material genetic engineering, to overcome the non-standardization in the process of building a material property prediction model, and to provide an efficient descriptor screening and crystal material property prediction method, thereby improving the learning efficiency and prediction effect of the model.

[0006] To solve the above problems, the technical scheme of the present application is as follows:

[0007] The descriptor screening and crystal material property prediction method based on material genetic engineering comprises the following steps:

[0008] P1: Collect material target physical property data and crystal descriptor data, generate objective function data set Y n Wherein n Collect the number of crystal materials, and obtain a large number of crystal descriptors related to the corresponding materials through existing databases to generate a crystal descriptor data set D mn Wherein D mn is m dimensional, containing n descriptors, m and n crystal materials corresponding to the Y n data set;

[0009] P2: Data cleaning and descriptor screening are performed on the created crystal descriptor data set D mn to generate a data set;

[0010] P3: Model hyperparameter tuning is carried out;

[0011] P4: Model prediction is carried out using a new model in combination with materials in the database for which the target physical properties are unknown, to obtain the final prediction result.

[0012] Preferably, the step P2 specifically comprises:

[0013] P21: Clean low-variance descriptors in D mn , and the formula is:

[0014]

[0015] Wherein S 2 Djn is the variance of the D j descriptor, and the cleaning threshold is set to 0.01 by default, and descriptors below the threshold are excluded, and a descriptor data set D P21 is generated, wherein D P21 is m P21 dimensional, containing n descriptors m P21 m P21 m , n crystal materials;

[0016] P22: Comprehensive screening of four correlation coefficients is carried out, and four calculation methods of Pearson correlation coefficient, Spearman correlation coefficient, distance correlation coefficient, and maximum mutual information are used to carry out comprehensive screening on the D P21 data set;

[0017] ​​P23: Conduct random sequential feature selection statistical screening. After completing the comprehensive screening of the four correlation coefficients in step P22, analyze the generated descriptor dataset D. P22 Conduct randomized sequential feature selection statistical screening.

[0018] Preferably, step P22 specifically includes:

[0019] P221: Regarding Pearson correlation coefficient R pearson Descriptor D P21 With objective function Y n The formula is as follows:

[0020]

[0021] P222: Regarding Spearman's correlation coefficient R spearman Descriptor D P21 With objective function Y n The formula is as follows:

[0022]

[0023] P223: Regarding the distance correlation coefficient R dist Descriptor D P21 With objective function Y n The formula is as follows:

[0024]

[0025] in Let the distance covariance be... For descriptor D P21 With objective function Y n The product of the standard deviation of the distance;

[0026] P224: Solving for the descriptor D for the maximum mutual information metric. P21 With objective function Y n of R MIC ;

[0027] P225: Set the four correlation coefficient screening thresholds for descriptor D. P21 With objective function Y n After calculating the four correlation coefficients, take their absolute values. Set the corresponding correlation coefficient screening threshold based on the calculation results. R set , Exceed R setIt is strong, recorded as 1, not more than the class of weak correlation, recorded as 0, the formula is as follows:

[0028]

[0029] Wherein represent the storage correlation decision value 0 or 1, Four correlation coefficient absolute value;

[0030] P226: the four correlation coefficient of Summed up , The value is 0 to 4 integer.

[0031] P227: according to The case of setting screening threshold , the range of values is 1 to 4 integer, finally reserved Greater than or equal to The result of the descriptor, generate descriptor data set D P22 , wherein D P22 For m P22 × n Dimension, containing m P22 Descriptor (D m P22 m P21 ), n Crystal material.

[0032] Preferably, the step P23 specifically comprises:

[0033] P231: set the number of random times T random , that is, set the sequential feature selection under the specific machine learning model of T random Times, before each sequential feature selection, D P22 With descriptor (D m P22 ) as a category is randomly sorted to break the local optimum caused by the inherent descriptor sequence combination, at the same time, determine the number of descriptors reserved T descriptor , that is, in each sequential feature selection, the descriptors of T descriptor Category are reserved;

[0034] P232: store the descriptor combination of each random sequential feature selection, and count the number of all screening results, for D P22 Descriptor data set, the number of times different descriptors are screened and reserved is​Tm P22 It is expressed in the form of Tm P22-1 , Tm P22-2 , …, Tm P22-i , i It refers to D P22 The first descriptor in the dataset i Class descriptor;

[0035] P233: Regarding Tm P22 Based on the statistics, set a cleaning threshold. Tm P22low Descriptors below this threshold are excluded, and the final descriptor dataset D is generated. model .

[0036] Preferably, step P3 specifically includes:

[0037] P31: Determine the condition variables and objective function for hyperparameter tuning and optimization, and perform R-skill optimization on the hyperparameters of the machine learning model being trained. 2 Or RMSE is the objective function f ( x )Right now y In intelligent optimization adjustments, the hyperparameters in the machine learning model serve as condition variables in the intelligent optimization process. x If the machine learning model contains character type hyperparameters Hyper_par This can be achieved through the condition variable corresponding to the hyperparameter. x str Set a threshold and substitute it into the mathematical symbol function sgn( x str The formula is as follows:

[0038]

[0039] in Hyper_par i This indicates the hyperparameter of the character type. Hyper_par The i-th character parameter, x i_low Represented as Hyper_par i Corresponding condition variable x str The minimum threshold value, similarly, x i_high Represented as Hyper_par i Corresponding condition variable x str The highest threshold value;

[0040] P32: carry out intelligent optimization parameter adjustment, save and output the model, hyperparameter content and model evaluation index obtained by training in each optimization process.

[0041] P33: screen and judge the optimization parameter adjustment result of P32, i.e. the model obtained by training and the model evaluation index, and finally confirm the most suitable crystal material property prediction model Model opt .

[0042] Preferably, the intelligent optimization adjustment in step P31 uses Bayesian optimization, genetic algorithm optimization, particle swarm optimization or simulated annealing optimization method.

[0043] Preferably, step P4 specifically comprises:

[0044] P41: according to the final descriptor data set D generated in step P233 model According to the requirements, the corresponding descriptors of other non-predicted crystal materials in the database are created to obtain a predicted material descriptor data set D predict ;

[0045] P42: using the crystal material property prediction model Model obtained in step P33 opt to carry out prediction on the predicted material descriptor data set D predict in step P41, and obtain a prediction result.

[0046] Compared with the prior art, the present application has the following advantages and positive effects:

[0047] The present application discloses a descriptor screening and crystal material property prediction method based on material genetic engineering. The target property data of part of the crystal materials are calculated and the crystal descriptors are collected, then the descriptors are cleaned and comprehensively screened by means of variance information, four correlation coefficients and random sequential feature selection method, then the hyperparameters in the model building process are automatically optimized and adjusted by intelligent optimization thinking, finally the optimal trained model is applied to the property prediction of the crystal materials, and the prediction result is obtained, so that the non-standard in the material property prediction model building process is overcome, and the learning efficiency and prediction effect of the model are improved. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 The present application discloses a descriptor screening and crystal material property prediction method based on material genetic engineering. The target property data of part of the crystal materials are calculated and the crystal descriptors are collected, then the descriptors are cleaned and comprehensively screened by means of variance information, four correlation coefficients and random sequential feature selection method, then the hyperparameters in the model building process are automatically optimized and adjusted by intelligent optimization thinking, finally the optimal trained model is applied to the property prediction of the crystal materials, and the prediction result is obtained, so that the non-standard in the material property prediction model building process is overcome, and the learning efficiency and prediction effect of the model are improved.

[0049] Figure 2 The present application discloses a descriptor screening and crystal material property prediction method based on material genetic engineering. The target property data of part of the crystal materials are calculated and the crystal descriptors are collected, then the descriptors are cleaned and comprehensively screened by means of variance information, four correlation coefficients and random sequential feature selection method, then the hyperparameters in the model building process are automatically optimized and adjusted by intelligent optimization thinking, finally the optimal trained model is applied to the property prediction of the crystal materials, and the prediction result is obtained, so that the non-standard in the material property prediction model building process is overcome, and the learning efficiency and prediction effect of the model are improved.

[0050] Figure 3A flowchart of a random sequential feature selection statistical screening process in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0051] The descriptor screening and crystal material property prediction method based on material genetic engineering provided by the present application will be described in further detail below in combination with the accompanying drawings and specific embodiments. The advantages and features of the present application will be more apparent according to the following description and claims.

[0052] Reference Figures 1 to 3 As shown, the embodiment provides a descriptor screening and crystal material property prediction method based on material genetic engineering, comprising the following steps:

[0053] P1: Collecting material target physical property data and crystal descriptor data; here, the phonon scattering phase space data in the DFT calculation process is taken as an example, and the target physical property data of 319 materials are collected. In addition, a large number of crystal descriptors related to the corresponding materials are obtained through an existing database to generate a crystal descriptor data set D mn , wherein D mn is a 310*319-dimensional matrix containing 310 descriptors and 319 Y n corresponding to the crystal materials in the data set;

[0054] Table 1 below shows the phonon scattering phase space data of 319 materials in the embodiment.

[0055] Table 1: Top 10 phonon scattering phase space data

[0056] Crystal chemical formula Phonon scattering phase space / x 10-4 cm BaTe 13.41 VN 2.35 InBi 6.98 ScC 4.77 GeC 1.08 SiC 4.63 TlSb 11.12 BAs 0.65 TeAs 8.37 BiB 3.20 …… ……

[0057] Table 2 below shows the crystal descriptor type data collected from the public database in the embodiment.

[0058] Table 2: Top 10 crystal descriptor type data

[0059] Serial number Crystal descriptor type 1 spacegroup 2 a 3 b 4 c 5 alpha 6 beta 7 gamma 8 volume 9 volume / aom 10 band_gap …… ……

[0060] P2: Data cleaning and descriptor screening are performed on the created descriptor data set to improve the fitting effect and prediction accuracy of model training, and a data set is generated.

[0061] P3: The adjustable hyperparameters are taken as independent variables, and the training accuracy is taken as the objective function to realize the function of hyperparameter optimization and adjustment, and the best training model is obtained.

[0062] P4: The new model is used to carry out model prediction on the materials with unknown target physical properties in the database to obtain the final prediction result.

[0063] Further, the process P2 specifically comprises:

[0064] P21: Cleaning D mn Low-Medium Variance Descriptor. Formula is:

[0065] where S 2 Djn is the variance of the descriptor, here the cleaning threshold j is set to 0.01 by default, descriptors below this threshold are excluded and the descriptor dataset D P21 P21 is reduced from 310 to 301 descriptors;

[0066] The following table 3 shows the variance values of the descriptors that did not reach the cleaning threshold in this example.

[0067] Table 3: Variance values of descriptors that did not reach the cleaning threshold

[0068] Crystal descriptor type Computed variance values alpha 0 beta 0 gamma 0 min_num_f_unfilled 0 var_en_ghosh 0.00000747 ave_en_ghosh 0.001445 min_en_ghosh 0.00159 max_en_ghosh 0.002421 min_gs_mag_moment 0.00861

[0069] P22: Comprehensive screening of the four correlation coefficients is carried out, as shown in Figure 2 The Pearson correlation coefficient, the Spearman correlation coefficient, the distance correlation coefficient, and the maximum mutual information are used to carry out comprehensive screening on the D P21 dataset, which specifically includes:

[0070] P221: For the Pearson correlation coefficient R pearson , the descriptor D P21 and the target function Y n The formula is as follows:

[0071]

[0072] P222: For the Spearman correlation coefficient R spearman , the descriptor D P21 and the target function Y n The formula is as follows:

[0073]

[0074] P223: For the distance correlation coefficient R dist , the descriptor D P21 and the target function Y n The formula is as follows:

[0075]

[0076] where​​​ Let the distance covariance be... For descriptor D P21 With objective function Y n The product of the distance and the standard deviation.

[0077] P224: Solving for the descriptor D for the maximum mutual information metric. P21 With objective function Y n of R MIC .

[0078] Table 4 below shows the calculated values ​​of the four correlation coefficients for the crystal descriptor type in this embodiment.

[0079] Table 4: Calculated values ​​of the four major correlation coefficients for the top 10 crystal descriptor types

[0080] Crystal descriptor type Rpearson Rspearman Rdist RMIC spacegroup 0.064954 0.177799 0.213682 0.265269 a 0.57972 0.633897 0.600688 0.413378 b 0.57972 0.633897 0.600688 0.413378 c 0.57972 0.633897 0.600688 0.413378 volume 0.563322 0.633897 0.601369 0.413378 volume / aom 0.498557 0.634657 0.57552 0.436177 band_gap 0.123743 0.124129 0.150079 0.205905 density -0.05098 0.013843 0.124107 0.21571 e_above_hull -0.08309 -0.17865 0.177218 0.194426 e_above_hull / atom -0.09187 -0.18197 0.181882 0.194426 …… …… …… …… ……

[0081] P225: Set the four correlation coefficient screening thresholds. For descriptor D... m With objective function Y n After calculating the four correlation coefficients, take their absolute values. Set the corresponding correlation coefficient screening threshold based on the calculation results. R set , Exceed R set A strong correlation is represented by a value of 1, while a weak correlation is represented by a value of 0 if the value is less than 1. The formula is as follows:

[0082]

[0083] in The storage relevance determination value is 0 or 1. The absolute values ​​of the four correlation coefficients.

[0084] P226: On the four types of correlation coefficients Summation yields . The possible values ​​are integers from 0 to 4.

[0085] Table 5 below shows the four correlation determination values ​​and summation values ​​for the crystal descriptor type in this embodiment.

[0086] Table 5: Four Major Correlation Determination Values ​​and Summation Values ​​for the Top 10 Crystal Descriptor Types

[0087] Crystal descriptor type Rpearson_0,1 Rspearman_0,1 Rdist_0,1 RMIC_0,1 R0,1,2,3,4 spacegroup 0 1 0 1 2 a 1 1 1 1 4 b 1 1 1 1 4 c 1 1 1 1 4 volume 1 1 1 1 4 volume / aom 1 1 1 1 4 band_gap 1 1 0 0 2 density 0 0 0 0 0 e_above_hull 0 1 0 0 1 e_above_hull / atom 0 1 0 0 1 …… …… …… …… …… ……

[0088] P227: According to Set a filtering threshold based on the situation Its range is from 1 to 4, and it is an integer. Finally, retain... Greater than or equal to The resulting descriptors generate a descriptor dataset D. P42m’n In this example, the settings are as follows: D P22 The dimensions are 55×319, and the number of descriptors is reduced from 304 to 55.

[0089] P23: Conduct statistical screening using randomized sequential feature selection, such as... Figure 3 As shown. After completing the comprehensive screening of the four major correlation coefficients on page 22, the generated descriptor dataset D was analyzed. P22 Conduct randomized sequential feature selection statistical screening.

[0090] P231: Setting the number of random attempts T random That is, to set up and carry out T random For each sequential feature selection under a specific machine learning model, D must be evaluated before each sequential feature selection. P22 With descriptor ( m P22 To break the local optima caused by the inherent combination of descriptor sequences, the categories are randomly sorted. Simultaneously, the number of descriptors to retain is determined. T descriptor That is, retaining in each sequential feature selection T descriptor Each category has its own descriptor. In this example, a random forest regression model is used, and the following parameters are set: T random =100, T descriptor =20.

[0091] P232: Store the descriptor combinations selected in each random sequential feature selection, and perform a count of all selection results for D. P22 The descriptor dataset, where the number of times different descriptors are filtered and retained is used. Tm P22 It is expressed in the form of Tm P22-1 , Tm P22-2 , …, Tm P22-i , i It refers to D P22 The first descriptor in the dataset i Class descriptor.

[0092] P233: Regarding Tm P22 Based on the statistics, a cleaning threshold is set here. TmP22low descriptors below the threshold are excluded and the final descriptor dataset D is generated model In this example, the cleaning threshold Tm P22low According to the number of reserved times statistics value Tm P22 The top 50% of descriptors in descending order are set, and finally D model is 27x319-dimensional, and the descriptors are reduced from 55 to 27.

[0093] Further, the flow P3 specifically comprises:

[0094] P31: Determine the condition variable and objective function of the hyperparameter adjustment optimization. For the hyperparameters of the machine learning model being trained, the R 2 is the objective function (the larger the better, i.e., close to 1) or the RMSE (the smaller the better, i.e., close to 0) f ( x ) is the intelligent optimization adjustment of y , wherein the intelligent optimization process can use Bayesian optimization, genetic algorithm optimization, particle swarm optimization, simulated annealing optimization, etc., and the hyperparameters in the machine learning model are the condition variables in the intelligent optimization process x . If there is a character type hyperparameter Hyper_par in the machine learning model, the threshold of the corresponding condition variable x str is set, and the mathematical symbol function sgn( x str ) is substituted. The formula is as follows:

[0095]

[0096] wherein Hyper_par i represents the i-th character parameter of the character type hyperparameter Hyper_par x i_low represents the threshold minimum value of the corresponding condition variable Hyper_par i x str , and similarly, x i_high represents the threshold maximum value of the corresponding condition variable Hyper_par i x str

[0097] ​​​​In this example, based on the random forest model, the Bayesian optimization parameter adjustment process is carried out, wherein n_estimators and max_depth are set as condition variables, and the adjustment ranges are set at 50 to 400 and 10 to 200, respectively. The test set R 2 is taken as the target function.

[0098] P32: Intelligent optimization parameter adjustment is carried out, and the model, hyperparameter content, and model evaluation index (including the training set and the test set) obtained by training in each optimization process are saved and output.

[0099] P33: The optimization parameter adjustment results of P32, i.e., the model and the model evaluation index obtained by training, are screened and judged, and the most suitable crystal material property prediction model Model opt is finally confirmed. In this example, the test set R 2 = 0.85523 and the training set R 2 = 0.91291 are taken as the final crystal material property prediction model Model opt .

[0100] Further, the process P4 specifically includes:

[0101] P41: According to the final descriptor data set D model generated in the process P233, the corresponding descriptor is created for other crystal materials in the database that have not been predicted, and the prediction material descriptor data set D predict is obtained.

[0102] P42: The most suitable crystal material phonon scattering phase space prediction model Model opt obtained by P33 is used to carry out prediction on the prediction material descriptor data set D predict in the process P41, and the phonon scattering phase space prediction result is obtained. Through the prediction result, the top 100 materials list of the phonon scattering phase space also contains some typical high-thermal-conductivity materials, such as diamond, SiC, GaN, BAs, Be2C, BN, etc., which achieves the expected effect of model building.

[0103] The embodiments of the application are described in detail above in combination with the drawings, but the application is not limited to the above-described embodiments. Even if various changes are made to the application, if the changes fall within the scope of the claims of the application and equivalent technologies, they still fall within the protection scope of the application.

Claims

1. A method for descriptor screening and crystalline material property prediction based on materials genome engineering, characterized in that, Includes the following steps: P1: Collect target physical property data and crystal descriptor data of the material to generate the objective function dataset Y. n ,in n To determine the number of crystalline materials collected, a large number of crystal descriptors related to the corresponding materials were obtained from existing databases, generating a crystal descriptor dataset D. mn D mn for m × n Dimension, including m One descriptor, n One and Y n The crystal material corresponding to the dataset; P2: For the created crystal descriptor dataset D mn Perform data cleaning and descriptor filtering to generate a dataset; P3: Conduct hyperparameter tuning of the model; P4: Using the new model in conjunction with materials in the database whose target physical properties are unknown, we can perform model predictions and obtain the final prediction results; Specifically, step P2 includes: P21: Cleaning D mn The formula for the low to medium variance descriptor is: Where S 2 Djn D j Descriptor variance, setting the cleaning threshold The default value is 0.

01. Descriptors below this threshold are excluded, and a descriptor dataset D is generated. P21 D P21 for m P21 × n Dimension, including m P21 One descriptor, m P21 < m , n Individual crystalline materials; P22: Conduct a comprehensive screening using four correlation coefficients, employing Pearson correlation coefficient, Spearman correlation coefficient, distance correlation coefficient, and maximum mutual information to evaluate D. P21 The dataset underwent comprehensive screening; P23: Conduct random sequential feature selection statistical screening. After completing the comprehensive screening of the four correlation coefficients in step P22, analyze the generated descriptor dataset D. P22 Conduct randomized sequential feature selection and statistical screening; Step P3 specifically includes: P31: Determine the condition variables and objective function for hyperparameter tuning and optimization, and perform R-skill optimization on the hyperparameters of the machine learning model being trained. 2 Or RMSE is the objective function f ( x )Right now y In intelligent optimization adjustments, the hyperparameters in the machine learning model serve as condition variables in the intelligent optimization process. x If the machine learning model contains character type hyperparameters Hyper_par This can be achieved through the condition variable corresponding to the hyperparameter. x str Set a threshold and substitute it into the mathematical symbol function sgn( x str The formula is as follows: in Hyper_par i This indicates the hyperparameter of the character type. Hyper_par The i-th character parameter, x i_low Represented as Hyper_par i Corresponding condition variable x str The minimum threshold value, similarly, x i_high Represented as Hyper_par i Corresponding condition variable x str The highest threshold value; P32: Conduct intelligent optimization and hyperparameter tuning. During each optimization process, the trained model, hyperparameter content, and model evaluation metrics are saved and output. P33: Based on the optimization and parameter tuning results from P32, i.e., the trained model and model evaluation metrics, we screen and judge to finally confirm the most suitable crystal material property prediction model. opt .

2. The method for descriptor screening and crystal material property prediction based on materials genome engineering according to claim 1, characterized in that, Step P22 specifically includes: P221: Regarding Pearson correlation coefficient R pearson Descriptor D P21 With objective function Y n The formula is as follows: P222: Regarding Spearman's correlation coefficient R spearman Descriptor D P21 With objective function Y n The formula is as follows: P223: Regarding the distance correlation coefficient R dist Descriptor D P21 With objective function Y n The formula is as follows: in Let the distance covariance be... For descriptor D P21 With objective function Y n The product of the standard deviation of the distance; P224: Solving for the descriptor D for the maximum mutual information metric. P21 With objective function Y n of R MIC ; P225: Set the four correlation coefficient screening thresholds for descriptor D. P21 With objective function Y n After calculating the four correlation coefficients, take their absolute values. Set the corresponding correlation coefficient screening threshold based on the calculation results. R set , Exceed R set A strong correlation is indicated by a value of 1, while a value less than 1 indicates a weak correlation, indicated by a value of 0. The formula is as follows: in The storage relevance determination value is 0 or 1. The absolute values ​​of the four correlation coefficients; P226: On the four types of correlation coefficients Summation yields , The possible values ​​are integers from 0 to 4; P227: According to Set a filtering threshold based on the situation Its range is integers from 1 to 4, and the final value is retained. Greater than or equal to The resulting descriptors generate a descriptor dataset D. P22 , where D P22 for m P22 × n Dimension, including m P22 One descriptor, m P22 < m P21 , n A crystalline material.

3. The method for descriptor screening and crystal material property prediction based on materials genome engineering according to claim 1, characterized in that, Step P23 specifically includes: P231: Setting the number of random attempts T random That is, to set up and carry out T random For each sequential feature selection under a specific machine learning model, D must be evaluated before each sequential feature selection. P22 With descriptor m P22 Randomly sort the categories to break the local optima caused by the inherent combination of descriptor sequences, and at the same time, determine the number of descriptors to retain. T descriptor That is, retaining in each sequential feature selection T descriptor Descriptors for each category; P232: Store the descriptor combinations selected in each random sequential feature selection, and perform a count of all selection results for D. P22 The descriptor dataset, where the number of times different descriptors are filtered and retained is used. Tm P22 It is expressed in the form of Tm P22-1 , Tm P22-2 , …, Tm P22-i , i It refers to D P22 The first descriptor in the dataset i Class descriptor; P233: Regarding Tm P22 Based on the statistics, set a cleaning threshold. Tm P22low Descriptors below this threshold are excluded, and the final descriptor dataset D is generated. model .

4. The method for descriptor screening and crystal material property prediction based on materials genome engineering according to claim 1, characterized in that, In step P31, intelligent optimization and adjustment uses Bayesian optimization, genetic algorithm optimization, particle swarm optimization, and simulated annealing optimization methods.

5. The method for descriptor screening and crystal material property prediction based on materials genome engineering according to claim 1, characterized in that, Step P4 specifically includes: P41: The final descriptor dataset D generated according to step P233 model The requirement is to create corresponding descriptors for other unpredicted crystalline materials in the database, resulting in a predicted material descriptor dataset D. predict ; P42: Using the crystal material property prediction model obtained in step P33 opt For the predicted material descriptor dataset D in step P41 predict Conduct forecasts and obtain forecast results.

Images