Method for predicting dielectric tensor and dielectric breakdown strength and database website thereof

Abstract

The invention discloses a method for predicting dielectric tensor and dielectric breakdown strength and a database website thereof, and belongs to the technical field of dielectric material performance prediction. According to the invention, dielectric databases which are from three different sources and are calculated based on a first principle are integrated, and materials with dielectric constants in a range of 30-50 are covered. And meanwhile, the dielectric tensor is predicted by adopting an open-source isovariant graph neural network framework, so that the method has key technical advantages. The method comprises the following steps: on the basis of a only model learned from experimental data, two key input parameters of the model are maximum phonon frequencies of a band gap and a gamma point, and a band gap value comes from a high-precision open source database; the maximum phonon frequency parameter of the gamma point is obtained by adopting the most advanced general machine learning force field model, so that the dielectric breakdown strength of the material can be efficiently predicted. Two key material parameters of dielectric tensor and dielectric breakdown strength are integrated to construct a database, and a corresponding website is developed. The method facilitates rapid data acquisition, promotes data sharing and communication, and has a promotion effect on development of related fields.

Description

Technical Field

[0001] The present invention relates to the technical field of dielectric material performance prediction, and in particular to a method for predicting dielectric tensor and dielectric breakdown strength and a database website thereof. Background Art

[0002] A dielectric is an electrical insulator that can be polarized by an applied electric field and is widely used in electronic devices, capacitors, high-voltage equipment, and other fields. The dielectric constant is an important physical quantity that characterizes the polarization ability of a dielectric in an electric field, while electrical breakdown or dielectric breakdown refers to the process in which an insulating material suddenly becomes a conductor and conducts electricity after being subjected to a sufficiently high voltage. The dielectric constant is an extremely important material parameter. A high dielectric constant is required in the dielectric layer of a transistor and energy storage devices to improve performance. At the same time, the dielectric breakdown strength is an important indicator for measuring device stability. In particular, in high-voltage equipment (such as transformers, capacitors, and cables), the breakdown strength of insulating materials is the key to ensuring the safe operation of the equipment. Insulating materials must have sufficient dielectric breakdown strength to prevent breakdown failure in high electric field environments. However, since there is usually an inverse relationship between the band gap and the dielectric constant [3], finding materials with both a large band gap and a high dielectric constant remains a challenge. Generally speaking, the larger the band gap of a material, the smaller the contribution of electronic polarization, resulting in a lower dielectric constant. Therefore, the design of materials with large band gap and high dielectric constant becomes a difficult problem in research.

[0003] The dielectric constant ε is a second-order tensor. At present, some literature has reported the use of density functional perturbation theory (DFPT) and first-principles high-throughput calculations to obtain dielectric tensor data of materials. However, in existing open source databases (such as Materials Project), materials with large band gaps and containing dielectric tensor data are still very limited. In order to reduce the cost of first-principles calculations, some researchers numerically average the dielectric tensor, treat it as a scalar, and predict it through machine learning. Although this method reduces the computational complexity, it may ignore the anisotropic information of the material. Recently, Ganose et al. developed a model based on equivariant graph neural network that can directly predict the dielectric tensor and retain the anisotropic characteristics of the material. However, due to the small number of samples of high dielectric constant materials in the training set, their model can only accurately predict materials with a dielectric constant less than 20. At present, there is still a lack of models that can directly, efficiently and accurately predict high dielectric constant (ε>30) materials.

[0004] The dynamic process of dielectric breakdown strength is extremely complex and is affected by many factors. Various phenomenological models have been developed experimentally to describe the process and mechanism of dielectric breakdown. However, accurate calculation and efficient prediction of dielectric breakdown strength based on first principles are still huge challenges. In addition, theoretical research has also used machine learning methods to obtain a phenomenological model for predicting dielectric breakdown strength by performing regression analysis on experimental data. However, this model requires parameters such as band gap and maximum phonon frequency at Γ point as input. When using this phenomenological model to predict dielectric breakdown strength, a major problem faced is that the maximum phonon frequency at Γ point is difficult to obtain. In fact, whether using the finite displacement method or density functional perturbation theory, the calculation of the phonon spectrum of the crystal is very time-consuming and consumes a lot of computing resources, which makes it difficult to obtain the breakdown strength of dielectric materials on a large scale.

[0005] In addition, there is a lack of material databases for dielectric tensor and dielectric breakdown strength. This limits the research and development of high-performance dielectric materials. Therefore, developing a method that can quickly and accurately predict dielectric constant and dielectric breakdown strength and establishing a related material database are of great significance for large-scale screening of high-performance dielectric materials. Summary of the invention

[0006] The purpose of the present invention is to provide a method for quickly and accurately predicting the dielectric constant and dielectric breakdown strength, and to establish a related material database to solve the problems in the prior art and meet the needs of research and development of high-performance dielectric materials.

[0007] To achieve the above object, the present invention provides a method for predicting dielectric tensor, comprising the following steps: The training set was constructed by using three open source databases based on first-principles calculations to determine the material dielectric constants that effectively contain data in the range of 30 to 50. The data set was divided into training set, validation set, and test set using a random partitioning method. Model training and evaluation: The open source Matten tensor prediction program is used for model training. The entire training process is carried out with the help of GPU acceleration technology. The test set and validation set are used to evaluate the performance of the training model. Dielectric tensor prediction and verification,The model was used to predict the dielectric tensor of materials in the SNUMAT database, which contains high-precision bandgap data.

[0008] Preferably, the specific steps of constructing the training set are: The data set was constructed by using three open source databases based on first-principles calculations to determine the material dielectric constants that effectively cover the range of 30 to 50; The original dataset is finally obtained by removing low-quality data from the basic dataset, and the dataset is randomly divided into a training set, a validation set, and a test set, each accounting for 80%, 10%, and 10% respectively.

[0009] Preferably, the model is evaluated using a test set and a validation set, mainly using the coefficient of determination R 2 And the root mean square error RMSE as the performance indicator: in, y i and f i are the i-th true value and predicted value respectively, It is the average of all true values. Generally, the closer the coefficient of determination is to 1, the better the model performance is, while the closer the RMSE is to 0, the smaller the model error is.

[0010] Preferably, the specific steps of dielectric tensor prediction and verification are: After training and evaluation, the model is applied to the SNUMAT database containing high-precision bandgap data to predict the dielectric tensor; 4341 crystal structures with band gaps greater than or equal to 3 eV were selected from the SNUMAT database, and the dielectric tensor was predicted; The model identifies the degree of crystal symmetry of the material, verifying the importance of anisotropy prediction of the dielectric tensor.

[0011] A method for predicting dielectric breakdown strength comprises the following steps: Phenomenological Model Dielectric breakdown strength F b Predictions are made where the band gap E g Derived from the high-precision SNUMAT database, Maximum phonon frequency at a point ω max Use the open source general machine learning force field model MACE to obtain; The MACE model was verified by comparing the phonon spectrum predicted by the MACE model with the phonon spectrum database MDR calculated by first principles. The validated MACE model is used to calculate the phonon spectra of crystal structures selected from the SNUMAT database, and the phenomenological model is applied in combination with the band gap value to predict the dielectric breakdown strength of the material.

[0012] A database website for predicting dielectric tensor and dielectric breakdown strength, including a database data construction module for constructing a training set training model using relevant links, the trained model is used to predict a second data set to obtain dielectric tensor properties, and MACE is used to predict the phonon spectrum of the second data set, and the phenomenological model is used to obtain the dielectric breakdown strength of the crystal structure in combination with the band gap information provided by the database; The database website development module uses the Vue framework to complete front-end development. Basic data is stored in Github, and data is crawled through the FetchAPI to load and build web pages.

[0013] Preferably, the database website development module provides three compound search methods: chemical formula search, component search and element search. Users can view the corresponding data details by clicking on the list items, including crystal structure visualization results, crystal lattice constants, bond angles and dielectric tensors.

[0014] Therefore, the present invention adopts a method for predicting dielectric tensor and dielectric breakdown strength and a database website thereof using the above structure, which has the following beneficial effects: (1) The present invention uses an open-source equivariant graph neural network framework to predict dielectric properties. Compared with traditional machine learning methods that can only predict scalars, this scheme can effectively predict the dielectric tensor. In addition, compared with the first-principles density functional perturbation theory calculation method, the calculation time is significantly shortened. The two key material parameters, dielectric tensor and dielectric breakdown strength, are integrated into a database, and a corresponding website is developed (https: / / tmm-group.github.io / diedatabase).

[0015] (2) The present invention integrates dielectric databases based on first-principles calculations from three different sources, covering materials with dielectric constants ranging from 30 to 50. This is different from other literature that usually only uses a single database for training. Through training with multi-source databases, the prediction accuracy of the model on high dielectric capacity materials is improved.

[0016] (3) This paper uses the most advanced general machine learning force field model to obtain the maximum phonon frequency at the Γ point ω max The parameters are efficient and accurate. No need to rely on supercomputers, a personal laptop equipped with a 4060 GPU card can complete the calculation. Compared with the first-principle phonon spectrum calculation, the calculation cost is greatly reduced and the calculation time is significantly shortened.

[0017] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A schematic diagram of a dielectric tensor prediction scheme of a method for predicting dielectric tensor and dielectric breakdown strength and a database website thereof of the present invention; Figure 2 A method for predicting dielectric tensor and dielectric breakdown strength and its training set, verification set and test set of the database website of the present invention ε poly Schematic diagram of distribution range and its quantity statistics; Figure 3 A schematic diagram of the performance of a method for predicting dielectric tensor and dielectric breakdown strength and a model of a database website of the present invention on a training set; Figure 4 A schematic diagram of the performance of a method for predicting dielectric tensor and dielectric breakdown strength and a model of a database website of the present invention on a validation set; Figure 5 A schematic diagram of the performance of a method for predicting dielectric tensor and dielectric breakdown strength and a model of a database website of the present invention on a test set; Figure 6 A schematic diagram of the verification of three large band gap and high dielectric constant materials of a method for predicting dielectric tensor and dielectric breakdown strength and a database website of the present invention; Figure 7 A schematic diagram of a method for predicting dielectric tensor and dielectric breakdown strength and a dielectric breakdown strength prediction scheme of a database website of the present invention; Figure 8 A schematic diagram of the prediction performance of the method for predicting dielectric tensor and dielectric breakdown strength and the MACE model of the database website of the present invention on the maximum phonon frequency of the phonon Γ point in the MDR database; Fig. 9 A scatter diagram of the relationship between the band gap and the predicted dielectric breakdown strength of a method for predicting dielectric tensor and dielectric breakdown strength and a database website of the present invention; Fig.10 A schematic diagram of a method for predicting dielectric tensor and dielectric breakdown strength and database data construction of a database website thereof according to the present invention; Fig.11 A schematic diagram of a database website search page of a method for predicting dielectric tensor and dielectric breakdown strength and a database website thereof of the present invention; Fig.12 A schematic diagram of various attribute display pages of a database website of a method for predicting dielectric tensor and dielectric breakdown strength and a database website of the present invention. DETAILED DESCRIPTION

[0019] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0020] Unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0021] Example like Figure 1 As shown, the present invention provides a method for predicting dielectric tensors, using an open source tensor prediction program Matten (Digital Discovery, 2024, 3.5: 869-882), which is based on the e3nn equivariant graph neural network framework (https: / / github.com / e3nn / e3nn), and can efficiently predict various physical tensors. In order to build an accurate dielectric tensor prediction model, three open source databases based on first-principles calculations were first used. The material dielectric constants in these databases effectively contain data in the range of 30 to 50. Subsequently, with the help of GPU acceleration technology, the model was fully trained and evaluated to improve the efficiency and accuracy of the prediction. After the training is completed, the model is used to predict the dielectric tensor of materials in the SNUMAT database (Scientific Data, 2020, 7.1: 387) containing high-precision band gap data. The details are as follows: Training set construction: The training set of the present invention comes from three open source databases based on first-principles calculations, the links are as follows: https: / / github.com / janosh / dielectrics https: / / github.com / takahashi-akira-36m / oxi_diel_db https: / / hackingmaterials.lbl.gov / matminer / dataset_summary.html like Figure 2By removing low-quality data (such as negative dielectric constant) from the basic data set, we finally obtained an original data set with a data volume of 2415. Using a random partitioning method, this data set is divided into a training set, a validation set, and a test set, accounting for 80%, 10%, and 10% respectively.

[0022] Model training and evaluation: Based on the aforementioned training set, the open source Matten tensor prediction program is used for model training. In order to speed up the training and improve the training efficiency of the model, the entire training process is carried out with the help of GPU acceleration technology. In order to comprehensively evaluate the performance of the training model, the model is evaluated using the test set and validation set, mainly using the determination coefficient R 2 And the root mean square error RMSE as the performance indicator: in, y i and f i are the i-th true value and predicted value respectively, It is the average of all true values. Generally, the closer the coefficient of determination is to 1, the better the model performance is, while the closer the RMSE is to 0, the smaller the model error is. R 2 In this training, as an indicator to determine whether the model training is complete, the model is trained until it performs well on the test set and the validation set. R 2 Until it is higher than 0.6. Finally, the trained equivariant graph neural network obtained 0.98, 0.64 and 0.7 on the training set, validation set and test set respectively. R 2 The results are as follows Figure 3-5 shown.

[0023] After completing the training and evaluation of the model, the model was further applied to the SNUMAT database containing high-precision band gap data to predict the dielectric tensor. In order to consider practical applications, 4341 crystal structures with band gaps greater than or equal to 3 eV were selected from the SNUMAT database and the dielectric tensor was predicted. In order to verify the accuracy of the prediction model, three representative materials were selected and first-principles calculations were performed. Figure 6 It can be seen that the model can accurately identify the crystal symmetry of the material, which is crucial for predicting the anisotropy of the dielectric tensor. For example, the dielectric constants of the three materials in different crystal directions can accurately match the first-principles calculation results, verifying the reliability of the model when dealing with materials with complex symmetry.

[0024] like Figure 7As shown, a method for predicting dielectric breakdown strength is as follows: The present invention adopts a phenomenological model to analyze the dielectric breakdown strength F b Make predictions: The key parameter of this model is the band gap E g , which comes from the high-precision SNUMAT database. In order to obtain another key parameter, the maximum phonon frequency at the Γ point ω max , the present invention uses the open source general machine learning force field model MACE. First, we use the first-principles phonon spectrum database MDR (https: / / doi.org / 10.48505 / nims.4197) to compare and analyze the phonon spectra predicted by the MACE model. Figure 8 The verified MACE model is used to calculate the phonon spectra of the crystal structures selected from the SNUMAT database. The extracted maximum phonon frequency at the Γ point is combined with the band gap value provided by the database, and the phenomenological model is used for numerical calculation to predict the dielectric breakdown strength of the material as shown in Figure 2. Fig. 9 shown.

[0025] like Fig.10 As shown, a dielectric tensor and dielectric breakdown strength database website is built, as shown below: Database data construction: First, use the link provided in the figure to build a training set for training the model. The trained model is then used to predict the second data set to obtain the dielectric tensor properties required for the database. At the same time, MACE is used to predict the phonon spectrum of the second data set, and the phenomenological model is used to obtain the dielectric breakdown strength of the crystal structure in combination with the band gap information provided by the database.

[0026] Database website development: The database website uses the Vue framework to complete the front-end development. The basic data is stored in Github, and the data is crawled through the FetchAPI, thereby realizing the loading and construction of the web page. The constructed website is https: / / tmm-group.github.io / diedatabase.

[0027] The initial interface of the web page is as follows Fig.11 As shown, the Table button is the periodic table display button. Clicking this button will pop up a periodic table bound to the search input bar (as indicated by the arrow ③) for element selection. After the selection is completed, press Enter or click the magnifying glass to search, and the results will appear as a list as indicated by the arrow ②. There are three ways to search for compounds: Chemical formula search: NaCl, this will directly search for compounds with the corresponding chemical formula.

[0028] Component search: Na-Cl, all compounds consisting only of Na and Cl will be listed.

[0029] Element search: Na-Cl-, all compounds containing Na and Cl elements will be listed.

[0030] Select and click any item in the list to pop up the corresponding data details (such as Fig.12 ), including crystal structure visualization results, crystal lattice constants, bond angles and dielectric tensors. The crystal structure visualization program is implemented using the tool library plotly. When the mouse hovers over an atom, the type of the corresponding element will be displayed in a pop-up box; and when hovering over the chemical bond connection between two elements, the corresponding bond length will pop up and display. The visualization window also provides visualization settings and download functions for the corresponding crystal structure. Click the folder button in the upper left corner, and the crystal structure acquisition method will be listed. The user can choose to copy the crystal structure string to the clipboard or directly download the corresponding POSCAR; the gear button under the folder button is the visualization setting button. Through the visualization setting button, the user can set the boundary of the visualization structure, the element chemical bond connection threshold and the boundary extension threshold.

[0031] Therefore, the present invention adopts the above-mentioned method for predicting dielectric tensor and dielectric breakdown strength and its database website, which can effectively predict dielectric tensor, and significantly shorten the calculation time compared with the first principle density functional perturbation theory calculation method. Through the training of multi-source database, the prediction accuracy of the model on high dielectric capacity materials is improved. There is no need to rely on supercomputers, and a personal notebook equipped with a 4060 GPU card can complete the calculation. Compared with the first principle phonon spectrum calculation, the calculation cost is greatly reduced and the calculation time is significantly shortened, which has the characteristics of high efficiency and accuracy.

[0032] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for predicting dielectric tensor, characterized in that: The following steps are involved: The training set was constructed by using three open source databases based on first-principles calculations to determine the material dielectric constants that effectively contain data in the range of 30 to 50. The data set was divided into training set, validation set, and test set using a random partitioning method. Model training and evaluation: The open source Matten tensor prediction program is used for model training. The entire training process is carried out with the help of GPU acceleration technology. The test set and validation set are used to evaluate the performance of the training model. Dielectric tensor prediction and verification,The model was used to predict the dielectric tensor of materials in the SNUMAT database, which contains high-precision bandgap data.

2. A method for predicting dielectric tensor according to claim 1, characterized in that: Specific steps for training set construction: The data set was constructed by using three open source databases based on first-principles calculations to determine the material dielectric constants that effectively cover the range of 30 to 50; The original dataset is finally obtained by removing low-quality data from the basic dataset, and the dataset is randomly divided into a training set, a validation set, and a test set, each accounting for 80%, 10%, and 10% respectively.

3. A method for predicting dielectric tensor according to claim 1, characterized in that: The model is evaluated using the test set and validation set, mainly using the determination coefficient R 2 And the root mean square error RMSE as the performance indicator: in, y i and f i are the i-th true value and predicted value respectively, It is the average of all true values. Generally, the closer the coefficient of determination is to 1, the better the model performance is, while the closer the RMSE is to 0, the smaller the model error is.

4. The method for predicting dielectric tensor and dielectric breakdown strength and its database website according to claim 1, characterized in that: Specific steps for dielectric tensor prediction and verification: After training and evaluation, the model is applied to the SNUMAT database containing high-precision bandgap data to predict the dielectric tensor; 4341 crystal structures with band gaps greater than or equal to 3 eV were selected from the SNUMAT database, and the dielectric tensor was predicted; The model identifies the degree of crystal symmetry of the material, verifying the importance of anisotropy prediction of the dielectric tensor.

5. A method for predicting dielectric breakdown strength, characterized in that: The following steps are involved: Phenomenological Model Dielectric breakdown strength F b Predictions are made where the band gap E g Derived from the high-precision SNUMAT database, Maximum phonon frequency at a point ω max Use the open source general machine learning force field model MACE to obtain; The MACE model was verified by comparing the phonon spectrum predicted by the MACE model with the phonon spectrum database MDR calculated by first principles. The validated MACE model is used to calculate the phonon spectra of crystal structures selected from the SNUMAT database, and the phenomenological model is applied in combination with the band gap value to predict the dielectric breakdown strength of the material.

6. A database website for predicting dielectric tensor and dielectric breakdown strength, characterized in that: It includes a database data construction module, which is used to build a training set training model using relevant links. The trained model is used to predict the second data set to obtain dielectric tensor properties. At the same time, MACE is used to predict the phonon spectrum of the second data set. Combined with the band gap information provided by the database, the phenomenological model is used to obtain the dielectric breakdown strength of the crystal structure; The database website development module uses the Vue framework to complete front-end development. Basic data is stored in Github, and data is crawled through the FetchAPI to load and build web pages.

7. A database website for predicting dielectric tensor and dielectric breakdown strength according to claim 6, characterized in that: The database website development module provides three compound search methods: chemical formula search, component search and element search. Users can view the corresponding data details by clicking on the list items, including crystal structure visualization results, crystal lattice constants, bond angles and dielectric tensors.

Images