Method for predicting p-type conductivity of doped metal oxide based on first principle calculation and machine learning

Abstract

The invention discloses a method for predicting p-type conductivity of doped metal oxide based on first principle calculation and machine learning. The method comprises the following steps: firstly, collecting and calculating Fermi level data after a certain metal oxide is doped, and constructing a feature vector by using feature engineering; thirdly, selecting a support vector machine, a random forest, a Gaussian process, gradient boosting, limit gradient boosting and other algorithms to train a machine learning model, selecting an optimal model to explore all feature combinations, and identifying key features influencing the Fermi level, so as to predict the Fermi level after the unknown doping element is doped into the metal oxide; and a potential p-type dopant candidate is screened out. The candidate elements are then validated by first principle calculations. By applying the method, experimental resources and time can be remarkably reduced, and development and application of metal oxides are promoted.

Description

Technical Field

[0001] The present invention relates to a method for predicting p-type conductive properties of doped metal oxides based on first-principles calculation and machine learning, and belongs to the interdisciplinary field of materials science and computer technology. Background Art

[0002] Metal oxides are an important class of materials with a wide range of applications in electronics, optoelectronics, sensors, etc. They exhibit excellent stability and performance under high temperatures, corrosive environments, and other extreme conditions, and thus play a key role in applications such as semiconductor devices, catalysts, gas sensors, and transparent conductive films.

[0003] Doping is an important means of modifying metal oxides. By introducing different doping elements, the electronic structure and conductive properties of metal oxides can be significantly changed, making them exhibit p-type or n-type conductive behavior. In semiconductor technology, the balance and regulation of p-type conductivity (i.e., hole conduction) and n-type conductivity (i.e., electron conduction) are crucial to the realization of efficient electronic devices. For example, in pn junction diodes, field effect transistors, and solar cells, precise regulation of the conductive properties of the p-type and n-type regions is the basis for the realization of high-performance devices.

[0004] However, p-type doping of many metal oxide materials has always been a challenge because the effective mass of holes is large and easily compensated, resulting in poor p-type conductivity. Traditionally, researchers have tried different doping elements and their concentrations one by one through experimental methods to screen out materials that can give metal oxides excellent p-type conductivity properties. This method is not only time-consuming and costly, but also difficult to cover all possible doping combinations, limiting the rapid development of new materials. Therefore, there is an urgent need for an efficient prediction method that can quickly and accurately predict the impact of doping elements on the p-type conductivity properties of metal oxides based on big data.

[0005] In recent years, with the development of machine learning technology, data-driven materials science research methods have gradually emerged. Trained models can quickly predict material properties, and the combination of machine learning and first-principles calculations can significantly improve the efficiency and accuracy of predictions. Summary of the invention

[0006] The purpose of the present invention is to provide a method for predicting the p-type conductive properties of doped metal oxides based on first-principles calculations and machine learning, thereby improving the efficiency of doping element screening and optimization and reducing experimental resources and time.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] Step 1: Obtain the Fermi level data of a metal oxide after doping through literature or first-principles calculations as the target attribute of the machine learning model;

[0009] Step 2: Collect doping element attributes and doping structure attributes as features, and remove redundancy through Pearson correlation analysis and feature importance analysis;

[0010] Step 3: Combine the Fermi level data from step 1 and the features obtained from step 2 as the data set for machine learning, input support vector machine, random forest, Gaussian process, gradient boosting, and extreme gradient boosting algorithms for training, and use multi-step optimization strategies to optimize hyperparameters, and use the machine learning algorithm with the best evaluation index to train the model;

[0011] Step 4: Select the algorithm with the best evaluation index in step 3, explore the feature combination in step 2, and identify the key features that affect the Fermi level;

[0012] Step 5: Use the algorithm with the best evaluation index in step 3 and the best feature explored in step 4 to train the model, and then predict the Fermi level of the unknown doping element doped into the metal oxide to obtain the p-type dopant candidate.

[0013] Step 6: Verify the dopant obtained in step 5 through first-principles calculations to evaluate its effect on the p-type conductivity of the metal oxide.

[0014] Furthermore, the metal oxide in step 1 is a metal oxide that is difficult to dope with P type, including gallium oxide and zinc oxide.

[0015] Furthermore, the data in step 1 must be obtained using the same calculation method to ensure uniformity.

[0016] Furthermore, the element properties selected in step 2 are: atomic number, atomic mass, electronegativity, melting point, boiling point, electron affinity, ionization energy, empirical radius, density, electrical conductivity, thermal conductivity, heat capacity and number of valence electrons, and the doping structure properties are: doping concentration and doping site.

[0017] Furthermore, the method for removing redundancy in step 2 is Pearson correlation analysis and feature importance analysis.

[0018] Furthermore, the data set obtained in step 3 is divided into a training set and a test set in a ratio of 4:1.

[0019] Furthermore, the multi-step optimization strategy in step 3 is: firstly, a random search is performed in the parameter space to quickly locate the areas where the model performs relatively well, and then a more detailed parameter search is performed in these areas using the BO algorithm.

[0020] Furthermore, the evaluation indicators in step 3 are mean absolute error MAE, root mean square error RMSE and determination coefficient R 2 .

[0021] Furthermore, the features in step 4 are combined into all subsets of the feature space.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] 1. By using machine learning algorithms to train a large amount of experimental data, the effects of different doping elements on the p-type conductivity properties of metal oxides can be accurately predicted in a short period of time, avoiding the inefficient process of testing doping element combinations one by one in traditional experimental methods.

[0024] 2. Machine learning algorithms can automatically identify key features that affect the p-type conductivity of metal oxides, which provides new perspectives and guidance for materials science research, helping researchers better understand and optimize material performance.

[0025] 3. Verify the selected p-type dopant candidates through first-principles calculations to ensure the reliability and scientificity of the prediction results and provide a solid foundation for further experimental research. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a flow chart of the present invention.

[0027] Figure 2 This is a heat map of the Pearson correlation coefficient of the features in the embodiment of the present invention.

[0028] Figure 3 This is a feature importance analysis diagram obtained by using an embodiment of the present invention.

[0029] Figure 4 This is a graph of the lowest MAE evaluation of the random forest model obtained by using an embodiment of the present invention for different numbers of features.

[0030] Figure 5 This is a graph showing the prediction results of the Fermi level of gallium oxide doped with the first 94 elements in the periodic table using the random forest model obtained by an embodiment of the present invention.

[0031] Figure 6 This is a comparison of the migration energy levels of Cd, Mg, and Zn doped gallium oxide.

[0032] Figure 7 Comparison of electronic structure and state density of Cd, Mg and Zn doped and intrinsic gallium oxide. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0034] See also Figure 1 Process, we take β-Ga2O3 material as an example to implement the method based on machine learning to predict the influence of doping elements on the p-type conductive properties of metal oxides:

[0035] Step 1: Collect 180 Fermi level data of doped β-Ga2O3 from the literature as the target attribute for machine learning.

[0036] Step 2: Collect the atomic number, atomic mass, electronegativity, melting point, boiling point, electron affinity, ionization energy, empirical radius, density, electrical conductivity, thermal conductivity, heat capacity and number of valence electrons of the doping elements, and use these 13 properties and two doping structure properties of doping concentration and doping site as features. The Pearson correlation analysis of these features is as follows: Figure 2 As shown in the figure, the importance analysis of the features is as follows Figure 3 shown.

[0037] Step 3: Divide the data set into training set and test set in a ratio of 4:1, select support vector machine, random forest, Gaussian process, gradient boosting, and extreme gradient boosting algorithms, use mean absolute error MAE, root mean square error RMSE, and determination coefficient R² to evaluate the performance of the model, and then use multi-step optimization strategy to optimize the hyperparameters. The model evaluation results before and after the multi-step optimization strategy are shown in Table 1. The performance of each model has been improved, among which the random forest model performs best.

[0038] Table 1 Performance evaluation results of each machine learning model before and after the multi-step optimization strategy

[0039]

[0040] Step 4: Use the random forest model to explore the 511 feature subsets in the feature space and use MAE to evaluate the performance of the model. Figure 4 As shown in Figure 2, the model has the best performance when seven features are used. Among them, the number of valence electrons and electron affinity have a great influence on the Fermi level.

[0041] Step 5: Use the best performing random forest model to predict the first 94 elements of the periodic table. The results are shown in Figure 5 The elements marked in blue are Mg, Zn, and Cd from left to right. The prediction results show that Cd is a new type of p-type dopant candidate.

[0042] Step 6: Use first-principles calculations to compare the migration energy levels, electronic structures, state density, and Mulliken charge layout of intrinsic gallium oxide and Cd, Mg, and Zn-doped β-Ga2O3. The comparison results are shown in Figure 6 , Figure 7 and Table 2, verifying that Cd is a candidate for p-type dopant for β-Ga2O3 material.

[0043] Table 2 Comparison of Mulliken charge layout of Cd, Mg and Zn doped gallium oxide.

[0044]

[0045] The above are merely exemplary embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting the p-type conductivity properties of doped metal oxides based on first-principles calculations and machine learning, Its characteristics are as follows: Step 1: Obtain the Fermi level data of a metal oxide after doping through literature or first-principles calculations as the target attribute of the machine learning model; Step 2: Collect doping element attributes and doping structure attributes as features, and remove redundancy through Pearson correlation analysis and feature importance analysis; Step 3: Combine the Fermi level data from step 1 and the features obtained from step 2 as the data set for machine learning. Use support vector machine, random forest, Gaussian process, gradient boosting, and extreme gradient boosting algorithms for training. Use multi-step optimization strategy to optimize hyperparameters. Use the machine learning algorithm with the best evaluation index to train the model. Step 4: Select the algorithm with the best evaluation index in step 3, explore the feature combination in step 2, and identify the key features that affect the Fermi level; Step 5: Use the algorithm with the best evaluation index in step 3 and the best feature explored in step 4 to train the model, and then predict the Fermi level of the unknown doping element doped into the metal oxide to obtain the p-type dopant candidate; Step 6: Verify the dopant obtained in step 5 through first-principles calculations to evaluate its effect on the p-type conductivity of the metal oxide.

2. The method according to claim 1, characterized in that: The metal oxide in step 1 is a metal oxide that is difficult to dope with p-type, such as gallium oxide and zinc oxide.

3. The method according to claim 1, characterized in that: The data in step 1 must be obtained using the same calculation method to ensure uniformity.

4. The method according to claim 1, characterized in that: The element properties selected in step 2 are: atomic number, atomic mass, electronegativity, melting point, boiling point, electron affinity, ionization energy, empirical radius, density, electrical conductivity, thermal conductivity, heat capacity and number of valence electrons, and the doping structure properties are: doping concentration and doping site.

5. The method according to claim 1, characterized in that: The method for removing redundancy in step 2 is Pearson correlation analysis and feature importance analysis.

6. The method according to claim 1, characterized in that: The data set obtained in step 3 is divided into a training set and a test set in a ratio of 4:

1.

7. The method according to claim 1, characterized in that: The multi-step optimization strategy in step 3 is: first perform a random search in the parameter space to quickly locate areas where the model performs relatively well, and then use the BO algorithm to perform a more detailed parameter search in these areas.

8. The method according to claim 1, characterized in that: The evaluation indicators in step 3 are mean absolute error MAE, root mean square error RMSE and determination coefficient R 2 .

9. The method according to claim 1, characterized in that: In step 4, the features are combined into all subsets of the feature space.

Images