Gypsum-based material band gap prediction method based on machine learning

Abstract

The invention relates to the technical field of dielectric properties of gypsum-based materials, in particular to a gypsum-based material band gap prediction method based on machine learning. According to the method, a model is established through the material characteristics, the band gap, the dielectric constant and other key performance parameters, the band gap of the gypsum-based material can be rapidly predicted through a data driving method, and the optimal material combination is found out through an optimization algorithm. According to the method, reliable and efficient band gap prediction on the gypsum-based material is achieved, compared with a traditional method, it is not needed to prepare an actual sample for measurement or obtain the band gap value through theoretical calculation, the material development period is greatly shortened, the performance of the material can be improved, and therefore time and resources can be saved, and the cost of the preparation and calculation process is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of dielectric properties of gypsum-based materials, and in particular to a method for predicting band gaps of gypsum-based materials based on machine learning. Background Art

[0002] Radome is a key material for protecting radar, and its performance has a significant impact on the penetration and anti-interference ability of radar signals. However, existing radome materials have defects such as high dielectric constant and loss, and severe signal attenuation in high-frequency working environments, which limits the performance improvement of millimeter-wave radar.

[0003] The size of a material's bandgap is closely related to the high-frequency dielectric constant of the material. Generally speaking, materials with larger bandgaps tend to exhibit lower dielectric constants. This is because materials with large bandgaps are more difficult to polarize, which reduces dielectric loss and signal attenuation. Therefore, bandgap engineering plays an important role in the development of low-dielectric-constant radome materials. By adjusting the bandgap of a material, its dielectric constant can be effectively controlled, making it more suitable for working in a high-frequency environment, thereby meeting the stringent requirements of millimeter-wave radar for signal transmission.

[0004] As a cheap and abundant calcium sulfate mineral, gypsum shows potential as a radome material due to its unique physical and chemical properties. The natural characteristics of gypsum give it certain advantages in dielectric properties, especially after properly adjusting its band gap, it is expected to show a low dielectric constant and excellent wave transmission performance.

[0005] However, developing low-dielectric materials suitable for millimeter-wave radar requires a lot of time and resources. Traditional material research and development methods often require repeated experiments to adjust the material composition and structure, which is time-consuming and laborious. Therefore, using machine learning technology to accelerate the screening and optimization of materials has become an effective solution. Summary of the invention

[0006] The purpose of the present invention is to propose a band gap prediction method for gypsum-based materials based on machine learning in view of the above-mentioned deficiencies in the prior art.

[0007] A method for predicting band gap of gypsum-based materials based on machine learning of the present invention comprises the following specific steps:

[0008] S1. Obtain material structure data with band gap values ​​in the range of 3-10 eV through the Materials Project database;

[0009] S2. Use feature hashing conversion on the text features in the material structure data, and finally convert all features into numerical types;

[0010] S3, the features converted in step S2 are continued to be used with Magpie machine learning descriptors to generate 153 features, and a pair of features whose absolute value of the Pearson correlation coefficient between the features is greater than 0.9 is found and any feature is deleted;

[0011] S4, randomly divide the data set characterized in step S3 into a training set and a test set in a ratio of 4:1;

[0012] S5, using the XGBoost machine learning algorithm and the training set obtained in step S4 to build a model, and using Bayesian hyperparameter adjustment to optimize the model to obtain a trained XGBoost model;

[0013] S6. Use the SHAP model interpretation software package to integrate with the trained XGBoost model obtained in step S5 to perform feature weight analysis; retain the top 10 features with the weight value as input data;

[0014] S7, inputting the doping metal type, doping amount and predefined crystal system of the gypsum-based material into the software to generate structural information of different gypsum-based materials;

[0015] S8. The features converted from the text features in the structural data of these gypsum-based materials continue to use the Magpie machine learning descriptor to generate 153 features, and select the 10 features obtained from the analysis in step S6 as input data and input them into the XGBoost model obtained in step S5 to predict the band gap of the gypsum-based materials.

[0016] Furthermore, in step S1, 20141 material structure data with band gap values ​​in the range of 3-10 eV are obtained.

[0017] Furthermore, in step S7, the doping metal type is Ba, Mn, Sr, Ti, and Fe.

[0018] Furthermore, in step S7, the doping amount is 10-50 wt.%.

[0019] Furthermore, in step S7, the step length is 10.

[0020] Further, in step S7, the predefined crystal system is set to be square, orthorhombic, hexagonal, and monoclinic.

[0021] Furthermore, in step S7, the software is Matlab or Python.

[0022] Furthermore, in step S6, the top 10 features with the highest weight values ​​are retained, which respectively represent the range of filled d orbitals, the formation energy of a single atom, the maximum value of filled d orbitals, the number of electronegativity modes, the average covalent radius, the average deviation of the covalent radius, the average number of filled d orbitals, the volume of a unit cell, the number of structural symmetries, and the maximum electronegativity.

[0023] The present invention establishes a model by combining material properties with key performance parameters such as band gap and dielectric constant, and can quickly predict the band gap of gypsum-based materials through a data-driven method, and find the best material combination through an optimization algorithm.

[0024] The present invention realizes reliable and efficient bandgap prediction for gypsum-based materials. Compared with traditional methods, it does not require the preparation of actual samples for measurement or the acquisition of bandgap values ​​through theoretical calculations, which greatly shortens the material development cycle and improves the performance of the material, thereby saving time and resources and reducing the cost of the preparation and calculation process. This convenience provides an effective tool for researchers and developers to predict the bandgap of gypsum-based wave-transmitting materials for screening and rapid evaluation of the dielectric properties of materials. It provides an efficient and low-cost solution for practical applications such as radar covers, and provides a new idea for the design and optimization of low-dielectric materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 A flowchart of the bandgap prediction method for gypsum-based materials based on machine learning;

[0026] Figure 2 are the R2, MSE and RMSE values ​​of each model;

[0027] Figure 3 is the Pearson correlation coefficient between the 10 features;

[0028] Figure 4 The predicted and actual band gaps of the generated AH gypsum-based materials were compared. DETAILED DESCRIPTION

[0029] The following are specific embodiments of the present invention and the accompanying drawings to further describe the technical solution of the present invention, but the present invention is not limited to these embodiments.

[0030] The band gap prediction method of gypsum-based materials based on machine learning provided by the present invention is as follows Figure 1 shown.

[0031] Step S1, obtaining 20141 material structure data with band gap values ​​in the range of 3-10 eV through the Materials Project database as a data set sample for machine learning;

[0032] Step S2, use feature hashing conversion on the text features in the material structure data, and finally convert all features into numerical types.

[0033] The data set obtained in step S1 is preprocessed to achieve higher modeling accuracy. Specifically, in addition to digital features, the data set also contains some text features, such as chemical composition, crystal system, space group, whether the material is centrosymmetric, etc. For these classification features, feature hashing is used to convert the text features, and finally all features are converted to numerical types to better train the model;

[0034] Step S3, the numerical features of each material converted in step S2 are used to generate 153 features using Magpie machine learning descriptors. Here, any feature with a high correlation (i.e., Pearson correlation coefficient greater than 0.90) between the two features needs to be deleted because they may provide the same information for model training;

[0035] Step S4, randomly dividing the data set characterized in step S3 into a training set (80%) and a test set (20%);

[0036] Step S5, using the XGBoost machine learning algorithm and the training set obtained in step S4 to build a model, and using Bayesian hyperparameter adjustment to optimize the model to obtain a trained XGBoost model.

[0037] Here, the training set and test set in step S4 are used for training multiple machine learning algorithms, including random forest (RF), multilayer perceptron regression (MLPR), K-nearest neighbor (KNN), support vector regression (SVR), gradient boosting (GBoost), extreme gradient boosting (XGBoost), decision tree gradient boosting (LightGBM) and deep learning (DL).

[0038] Bayesian hyperparameter tuning was used to optimize the model. Cross-validation was used to prevent overfitting during hyperparameter tuning.

[0039] By evaluating the determination coefficient (R2), mean square error (MSE) and root mean square error (RMSE), the optimal band gap prediction model is the XGBoost model, such as Figure 2 shown.

[0040] Step S6, in order to evaluate the impact of input features on the band gap, the SHAP model interpretation software package is integrated with the optimal ML obtained in step 5 to perform feature weight analysis; the top 10 features with the highest weight value are retained as input data;

[0041] The top 10 features with the highest weight in step S7 are retained and the Pearson correlation coefficients between these 10 features are calculated, such as Figure 3As shown, the meanings of numbers 1-10 are shown in Table 1.

[0042] S7, input the doping metal type, doping amount and predefined crystal system of the gypsum-based material into Matlab or Python to generate structural information of different gypsum-based materials;

[0043] S8. The features converted from the text features in the structural data of these gypsum-based materials continue to use the Magpie machine learning descriptor to generate 153 features, and select the 10 features obtained from the analysis in step S6 as input data and input them into the XGBoost model obtained in step S5 to predict the band gap of the gypsum-based materials.

[0044] The doping metal type is set to Ba, Mn, Sr, Ti, Fe, the doping amount (wt.%) is set to 10-50, the step size is 10, and the predefined crystal system is set to square, orthorhombic, hexagonal, and monoclinic, and the structural information of about 10,000 AH materials is constructed. Using the above 10 main features as input data, load the trained XGBoost model to predict the band gap of the above-generated AH gypsum-based materials. Divide the original data into 80% training data group and 20% verification data group. After the model is trained, it is verified with the verification data group. The model will generate band gap data based on the structural characteristics of the verification data group. This is the predicted value. The true value is the theoretically calculated band gap value originally included in the verification data. These two sets of data are compared, such as Figure 4 shown.

[0045] This example uses an efficient machine learning method to model the band gap data of the material and successfully constructs an efficient prediction model for the band gap of gypsum-based materials. The determination coefficient of the model reaches 0.85, the MSE is 0.59, and the RMSE is 0.35, showing a high prediction accuracy.

[0046] This embodiment uses material samples in the database to establish an accurate and efficient prediction model, which has the advantages of low cost, green and high efficiency. This method based on machine learning achieves high-precision prediction of band gaps, provides strong technical support for the development of low-dielectric, microwave-transparent gypsum-based materials, and accelerates the process of their design and application.

[0047] Table 1 The first 10 feature descriptors and their meanings

[0048]

[0049]

[0050] For matters not mentioned above, the prior art applies.

[0051] Although some specific embodiments of the present invention have been described in detail through examples, those skilled in the art should understand that the above examples are for illustration only and are not intended to limit the scope of the present invention. Those skilled in the art to which the present invention belongs may make various modifications or supplements to the specific embodiments described or replace them in a similar manner, but they will not deviate from the direction of the present invention or exceed the scope defined by the attached claims. Those skilled in the art should understand that any modifications, equivalent substitutions, improvements, etc. made to the above embodiments based on the technical essence of the present invention should be included in the protection scope of the present invention.

Claims

1. A bandgap prediction method for gypsum-based materials based on machine learning, characterized in that: The specific steps include: S1. Obtain material structure data with band gap values ​​in the range of 3-10 eV through the Materials Project database; S2. Use feature hashing conversion on the text features in the material structure data, and finally convert all features into numerical types; S3, the features converted in step S2 are continued to be used with Magpie machine learning descriptors to generate 153 features, and a pair of features whose absolute value of the Pearson correlation coefficient between the features is greater than 0.9 is found and any feature is deleted; S4, randomly divide the data set characterized in step S3 into a training set and a test set in a ratio of 4:1; S5, using the XGBoost machine learning algorithm and the training set obtained in step S4 to build a model, and using Bayesian hyperparameter adjustment to optimize the model to obtain a trained XGBoost model; S6. Use the SHAP model interpretation software package to integrate with the trained XGBoost model obtained in step S5 to perform feature weight analysis; retain the top 10 features with the weight value as input data; S7, inputting the doping metal type, doping amount and predefined crystal system of the gypsum-based material into the software to generate structural information of different gypsum-based materials; S8. The features converted from the text features in the structural data of these gypsum-based materials continue to use the Magpie machine learning descriptor to generate 153 features, and select the 10 features obtained from the analysis in step S6 as input data and input them into the XGBoost model obtained in step S5 to predict the band gap of the gypsum-based materials.

2. A method for predicting band gap of gypsum-based materials based on machine learning as claimed in claim 1, characterized in that: In step S1, 20141 material structure data with band gap values ​​in the range of 3-10 eV are obtained.

3. The method for predicting band gap of gypsum-based materials based on machine learning according to claim 1, characterized in that: In step S7, the doping metal type is Ba, Mn, Sr, Ti, and Fe.

4. The method for predicting band gap of gypsum-based materials based on machine learning according to claim 1, characterized in that: In step S7, the doping amount is 10-50 wt.%.

5. The method for predicting band gap of gypsum-based materials based on machine learning according to claim 1, characterized in that: In step S7, the step length is 10.

6. A method for predicting band gap of gypsum-based materials based on machine learning as claimed in claim 1, characterized in that: In step S7, the predefined crystal system is set to be square, orthorhombic, hexagonal, and monoclinic.

7. The method for predicting band gap of gypsum-based materials based on machine learning according to claim 1, characterized in that: In step S7, the software is Matlab or Python.

8. The method for predicting band gap of gypsum-based materials based on machine learning according to claim 1, characterized in that: In step S6, the top 10 features with the highest weight values ​​are retained, which respectively represent the range of filled d orbitals, the formation energy of a single atom, the maximum value of filled d orbitals, the number of electronegativity modes, the average covalent radius, the average deviation of the covalent radius, the average number of filled d orbitals, the volume of the unit cell, the number of structural symmetries, and the maximum electronegativity.

Images