Dielectric constant prediction method and system, and network training method and system

Abstract

The invention discloses a dielectric constant prediction method and system and a network training method and system. The method comprises the following steps: acquiring material characteristics of a to-be-tested material; constructing a plurality of descriptors according to the material characteristics of the to-be-tested material; screening the plurality of descriptors through a genetic algorithm, determining M main descriptors, and determining the descriptors except the main descriptors as other descriptors; and inputting the main descriptor and other descriptors into a preset dielectric constant prediction model, and outputting dielectric constant prediction data. The dielectric constant of the microwave high-frequency material is efficiently and accurately predicted through the machine learning technology. Compared with a traditional neural network model and descriptor selection method, the prediction precision, the model efficiency and the generalization ability are remarkably improved, and the problem that high-dimensional descriptors are not matched with small data sets is solved.

Description

Technical Field

[0001] This application relates to the fields of materials science and artificial intelligence technology, and more specifically, to a method and system for predicting dielectric constant, and a method and system for training a network. Background Art

[0002] The dielectric constant is an important physical quantity that describes the relative ability of a dielectric to store electrostatic energy in an electric field. Its research is of crucial significance in the fields of energy, electronics, communication, and new material design. For example, in microwave high-frequency devices, low-dielectric-loss materials have higher signal transmission efficiency, while in energy storage materials, high-dielectric-constant materials can significantly increase the energy density. However, despite the important position of the dielectric constant in materials science, the process of experimentally measuring the dielectric constant is often time-consuming, costly, and greatly affected by equipment, environment, etc. In addition, existing theoretical calculation methods (such as density functional theory, DFT) also face many challenges in accurately predicting the dielectric constant.

[0003] Currently, the research on the dielectric constant mainly relies on two methods: experimental measurement and theoretical calculation. The experimental method can provide high-precision dielectric constant data, but it is limited by many factors such as sample preparation, test environment, and equipment accuracy, resulting in low repeatability and predictability of experimental results. In addition, for microwave high-frequency materials, it is often difficult to obtain accurate dielectric constants through experimental measurement. Theoretical calculation methods (such as simulation studies based on Debye dielectric theory) predict the dielectric constant through analytical formulas, but their results rely on multiple assumptions and approximations. Especially when the intermolecular interactions in the material are strong or the material has a complex lattice structure, the error of theoretical calculation increases significantly.

[0004] In recent years, with the rapid development of machine learning and artificial intelligence technologies, data-driven models have been widely used in the field of materials science. By learning the laws from a large amount of experimental data or computational data, machine learning technologies can efficiently construct the mapping relationship between the dielectric constant and material descriptors. These methods can not only significantly reduce the computational cost but also achieve high-precision prediction on a limited data set. In particular, deep learning models (such as neural networks) have gradually become an important tool for studying material properties due to their strong fitting ability and non-linear feature capture ability. However, existing machine learning models still face the following challenges when dealing with the dielectric constant problem: Data scarcity: The experimental data in the field of materials science is limited, especially the dielectric constant data of microwave high-frequency materials is scarce, resulting in the difficulty for existing models to fully learn the data characteristics during training.

[0005] The contradiction between high-dimensional descriptors and small data sets: Constructing a material property prediction model requires a large number of descriptors, which may include element properties, structural parameters, etc. However, high-dimensional descriptors often lead to overfitting of the model and reduce the generalization ability.

[0006] Single model structure: The existing neural network structures are difficult to effectively process complex high-dimensional data, resulting in insufficient prediction accuracy.

[0007] Therefore, developing an efficient and accurate dielectric constant prediction tool that can overcome data scarcity and high-dimensional problems while improving the generalization ability of the model is of great significance for accelerating the discovery and optimization of new materials. Summary of the Invention

[0008] In view of the above problems, the object of the present invention is to provide a dielectric constant prediction method and system, and a network training method and system, which can efficiently and accurately predict the dielectric constant of microwave high-frequency materials through machine learning technology. Compared with traditional neural network models and descriptor selection methods, the prediction accuracy, model efficiency and generalization ability are significantly improved, and the problem of mismatch between high-dimensional descriptors and small data sets is solved.

[0009] Adopt the hierarchical neural network (HNN) architecture of the statistical ensemble. Through multi-level sub-model decomposition and integration, the information in the small data set is fully mined, and the generalization ability of the model is significantly improved.

[0010] Combined with the genetic algorithm (GA), 909 basic descriptors are screened and optimized, and finally 6 main descriptors that are most important for dielectric constant prediction are refined: whether it is a d-block metal, atomic weight, number of p-block electrons, polarizability, Wigner-Seitz electron cloud density and electronegativity. This dimensionality reduction method based on the importance ranking of descriptors significantly reduces the model complexity while retaining key information.

[0011] By introducing a hierarchical neural network architecture, the complex descriptor space is decomposed into multiple sub-models, and through layer-by-layer optimization and integration, the prediction accuracy and generalization ability are significantly improved.

[0012] The first aspect of the present invention provides a dielectric constant prediction method, including: Obtain the material characteristics of the material to be measured; Construct multiple descriptors according to the material characteristics of the material to be measured; Screen the multiple descriptors through the genetic algorithm to determine M main descriptors, and determine the descriptors other than the main descriptors as other descriptors; the main descriptors include whether it is a d-block metal, atomic weight, number of p-block electrons, polarizability, electron cloud density and electronegativity; Input the main descriptors and the other descriptors into a preset dielectric constant prediction model, and output dielectric constant prediction data.

[0013] In this solution, the constructing multiple descriptors according to the material characteristics of the material to be measured includes: Calculating the configurational entropy of materials: ; where ΔS con is the configurational entropy of the material, k B is the Boltzmann constant, T is the Kelvin temperature, and c i is the proportion of element i among all elements; Calculate the ratios of the s, p, d, and f valence electrons to the total sum of the s, p, d, and f valence electrons in the elements presented by the material respectively to determine the electronic structure attributes; Determine the ionic compound attributes according to the electronegativity of the material; Statistically analyze the properties of multiple elements of the material through multiple preset statistical methods to determine the statistical properties of the element properties of the material; Determine the configurational entropy, electronic structure attributes, ionic compound attributes, and statistical properties of element properties of the material as descriptors.

[0014] In this solution, screening the multiple descriptors through a genetic algorithm to determine M main descriptors, including: Step 1, create individuals randomly or based on some prior knowledge of the problem to construct an initial population; Step 2, evaluate the fitness of each individual in each population through a preset evaluation criterion; the preset evaluation criterion includes the coefficient of determination R 2 , mean square error MSE, and root mean square error RMSE; Step 3, select individuals from each population according to the fitness of the individuals to create a new population; Step 4, select individuals from the new population for genetic operations to generate new individuals; the genetic operations include selection, mutation, and crossover; Step 5, replace the old population with the new population; Continuously repeat steps 2 - 5 until the evaluation criterion converges, and determine the descriptors corresponding to the converged descriptor dimension and the population with a high coefficient of determination R 2 as the main descriptors.

[0015] In this solution, it also includes: The preset dielectric constant prediction model is a hierarchical neural network structure composed of multiple sub - networks with the same structure and arranged hierarchically. Each sub - network includes m input nodes and 1 output node, and the output data of the nth - layer sub - network is the input data of the (n + 1)th - layer sub - network; input the M main descriptors and m - M other descriptors into each sub - network in the first layer respectively, and the other descriptors input into each sub - network are determined by random selection.

[0016] A second aspect of the present invention provides a method for training a network, where the network is used to predict the dielectric constant of a material, and the method includes: Obtain the dielectric constant data of the sample material; Clean the dielectric constant data of the sample material to obtain a sample data data set; Construct multiple descriptors according to the material characteristics of the sample materials in the sample data data set; Screen the multiple descriptors through a genetic algorithm to determine M main descriptors, and determine the descriptors other than the main descriptors as other descriptors; the main descriptors include whether it is a d-block metal, atomic weight, number of p-block electrons, polarizability, electron cloud density, and electronegativity; Train the main descriptors and the other descriptors through a hierarchical neural network to establish a preset dielectric constant prediction model.

[0017] In this solution, the constructing multiple descriptors according to the material characteristics of the sample materials in the sample data data set includes: Calculate the configurational entropy of the material: ; where, ΔS con is the configurational entropy of the material, k B is the Boltzmann constant, T is the Kelvin temperature, c i is the proportion of element i among all elements; Calculate the ratios of the s, p, d, and f valence electrons to the total sum of the s, p, d, and f valence electrons of the elements presented by the material respectively to determine the electronic structure attributes; Determine the ionic compound attributes according to the electronegativity of the material; Statistically analyze the properties of various elements of the material through multiple preset statistical methods to determine the statistical properties of the element properties of the material; Determine the configurational entropy, electronic structure attributes, ionic compound attributes, and statistical properties of the element properties of the material as descriptors.

[0018] In this solution, the screening the multiple descriptors through a genetic algorithm to determine M main descriptors includes: Step 1, create individuals randomly or based on some prior knowledge of the problem to construct an initial population; Step 2, evaluate the fitness of each individual in each population through a preset evaluation criterion; the preset evaluation criterion includes the coefficient of determination R 2 , mean square error MSE, and root mean square error RMSE; Step 3, select individuals from each population according to the fitness of the individuals to create a new population; Step 4: Select individuals from the new population for genetic operations to generate new individuals; the genetic operations include selection, mutation, and crossover; Step 5: Replace the old population with the new population; Continuously repeat Steps 2 - 5 until the evaluation criteria converge, and determine the descriptor dimensions and high determination coefficients R 2 of the population corresponding to the converged ones as the main descriptors.

[0019] In this solution, it further includes: The preset dielectric constant prediction model is a hierarchical neural network structure composed of multiple sub - networks with the same structure arranged hierarchically. Each sub - network includes m input nodes and 1 output node. The output data of the nth - layer sub - network is the input data of the (n + 1)th - layer sub - network; input M main descriptors and m - M other descriptors into each sub - network in the first layer respectively, and the other descriptors input into each sub - network are determined by random selection.

[0020] The third aspect of the present invention provides a dielectric constant prediction system configured to predict the dielectric constant of a material, including: At least one storage medium storing at least one instruction set; and At least one processor communicatively connected to the at least one storage medium, wherein when the prediction system runs, the at least one processor reads the at least one instruction set and executes the method described in any one of the dielectric constant prediction methods according to the instructions of the at least one instruction set.

[0021] The fourth aspect of the present invention provides a training system configured to train a prediction network for predicting the dielectric constant of a material, including: At least one storage medium storing at least one instruction set; and At least one processor communicatively connected to the at least one storage medium, wherein when the training system runs, the at least one processor reads the at least one instruction set and executes the method described in any one of the network training methods according to the instructions of the at least one instruction set.

[0022] The present invention discloses a method and system for predicting dielectric constant, and a method and system for training a network. The method includes: obtaining material characteristics of a material to be measured; constructing a plurality of descriptors according to the material characteristics of the material to be measured; screening the plurality of descriptors through a genetic algorithm to determine M main descriptors, and determining the descriptors other than the main descriptors as other descriptors; inputting the main descriptors and other descriptors into a preset dielectric constant prediction model to output dielectric constant prediction data. The present invention efficiently and accurately predicts the dielectric constant of microwave high-frequency materials through machine learning technology. Compared with traditional neural network models and descriptor selection methods, the prediction accuracy, model efficiency and generalization ability are significantly improved, and the problem of mismatch between high-dimensional descriptors and small data sets is solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 FIG. shows a flowchart of a method for predicting dielectric constant provided by the present invention; Figure 2 FIG. shows a flowchart of a method for training a network provided by the present invention; Figure 3 FIG. shows a schematic diagram of material characteristics provided by the present invention; Figure 4 FIG. shows a flowchart of the hierarchical neural network model iteratively working on descriptors; Figure 5 FIG. shows a schematic diagram of the prediction effect of a dielectric constant prediction model built by a traditional neural network provided by the present invention; Figure 6 FIG. shows a schematic diagram of the prediction effect of a dielectric constant prediction model built by a hierarchical neural network provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] In order to more clearly understand the above objects, features and advantages of the present invention, the present invention will be further described in detail below with reference to the drawings and specific embodiments. It should be noted that, without conflict, the embodiments of the present application and the features in the embodiments may be combined with each other.

[0025] In the following description, many specific details are set forth in order to fully understand the present invention. However, the present invention may be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited by the specific embodiments disclosed below.

[0026] Figure 1 FIG. shows a flowchart of a method for predicting dielectric constant provided by the present invention.

[0027] As Figure 1 shown, the present invention discloses a method for predicting dielectric constant, including: S102, obtaining material characteristics of a material to be measured; S104. Construct multiple descriptors according to the material characteristics of the material to be measured; S106. Screen the multiple descriptors through a genetic algorithm to determine M main descriptors, and determine the descriptors other than the main descriptors as other descriptors; the main descriptors include whether it is a d-block metal, atomic weight, number of p-block electrons, polarizability, electron cloud density, and electronegativity; S108. Input the main descriptors and other descriptors into a preset dielectric constant prediction model, and output dielectric constant prediction data.

[0028] According to an embodiment of the present invention, as Figure 3 shown, there are 53 material characteristics of the material. The material characteristics of the material to be measured can be obtained through one or more detection methods preset by the system (such as X-ray detection, look-up table method, etc.). The system analyzes the material characteristics of the material to be measured by selecting a preset descriptor calculation method, and determines 909 descriptors including 1 configurational entropy, 4 electronic structure attributes, 3 ionic compound attributes, and 901 elemental property statistical properties. Screen the 909 descriptors through a genetic algorithm to determine the main descriptors that are most important for predicting the dielectric constant. There are 6 main descriptors in total, which are whether it is a d-block metal, atomic weight, number of p-block electrons, polarizability, electron cloud density, and electronegativity. The 6 main descriptors have clear physical meanings and are closely related to the physical nature of the dielectric constant: Whether it is a d-block metal: represents the charge separation ability of the material. Metal oxides in the d-block usually have a relatively high dielectric constant.

[0029] Atomic weight and number of p-block electrons: are closely related to the chemical composition and electronic structure of the material, and affect the magnitudes of the polarizability and dielectric constant.

[0030] Polarizability and electron cloud density: directly reflect the response ability of the material under the action of an electric field and are the key determinants of the dielectric constant.

[0031] Electronegativity: characterizes the ability of an atom to attract electrons and affects the electrostatic energy storage ability of the material.

[0032] Input the main descriptors and other descriptors into a preset dielectric constant prediction model. Input all the main descriptors into each sub-network of the first layer of the preset dielectric constant prediction model respectively, and randomly input other descriptors through the remaining idle input nodes. Take the output data of each layer as the input data of the next layer. Through hierarchical analysis, combined with the dielectric constant data and material characteristics of the sample materials in the database, finally output the prediction data of the corresponding dielectric constant.

[0033] According to an embodiment of the present invention, constructing multiple descriptors according to the material characteristics of the material to be measured includes: Calculating the configurational entropy of the material: ; where ΔS con is the configurational entropy of the material, k B is the Boltzmann constant, T is the Kelvin temperature, and c i is the proportion of element i among all elements; Calculate the ratios of the s, p, d, and f valence electrons to the total sum of the s, p, d, and f valence electrons of the elements presented in the material, respectively, to determine the electronic structure property; Determine the ionic compound property according to the electronegativity of the material; Perform statistics on various element properties of the material through multiple preset statistical methods to determine the statistical properties of the element properties of the material; Determine the configurational entropy, electronic structure property, ionic compound property, and statistical property of element properties of the material as descriptors.

[0034] It should be noted that there are 909 descriptors in total, including 1 configurational entropy, 4 electronic structure properties, 3 ionic compound properties, and 901 statistical properties of element properties. First, determine the element i contained in the material to be tested and the proportion c i , of element i among all elements, and input the proportion of element i among all elements into the preset configurational entropy calculation formula of the system to determine the configurational entropy of the material. The configurational entropy can represent the spatial arrangement form of atoms in the material. Considering the configurational entropy in the descriptor can characterize the key properties of the material to be tested and improve the prediction accuracy of the dielectric constant of the material to be tested.

[0035] The calculation formula for the electronic structure property is: ; where F x represents the electronic structure property corresponding to the xth type of valence electron, c i is the proportion of the ith element among all elements, and E x represents the number of the xth type of valence electron (s, p, d, or f valence electron) of the ith element in the material to be tested, and E n represents the total number of s, p, d, and f valence electrons of the ith element.

[0036] The calculation formula for the ionic compound property is: ; where I is the ionic compound property, e is the natural constant, c i is the proportion of the ith element among all elements, f i is the electronegativity corresponding to the ith element, is the maximum or average value of the electronegativities corresponding to each element. When When it is the maximum value among the electronegativities corresponding to each element, the first ionic compound property is determined; when it is the average value among the electronegativities corresponding to each element, the second ionic compound property is determined. Meanwhile, the third ionic compound property is determined based on the second ionic compound property. When I > 1.7, the third ionic compound property is 1; when I ≤ 1.7, the third ionic compound property is 0.

[0037] The preset statistical methods include calculating the minimum value, maximum value, and range size of element properties; the (weighted) minimum value, maximum value, and range size after weighted sorting; the (weighted) minimum value, maximum value, range size, and mean of the absolute percentage of element properties, etc., a total of 17 statistical methods. Respectively through the above 17 preset statistical methods, Figure 3 the 53 element properties shown in are statistically analyzed to determine 901 element property statistical properties.

[0038] According to the embodiments of the present invention, multiple descriptors are screened through a genetic algorithm to determine M main descriptors, including: Step 1, create individuals randomly or based on some prior knowledge of the problem to construct an initial population; Step 2, evaluate the fitness of each individual in each population through a preset evaluation criterion; the preset evaluation criterion includes the coefficient of determination R 2 , mean squared error MSE, and root mean squared error RMSE; Step 3, select individuals from each population according to the fitness of the individuals to create a new population; Step 4, select individuals from the new population for genetic operations to generate new individuals; the genetic operations include selection, mutation, and crossover; Step 5, replace the old population with the new population; Continuously repeat steps 2 - 5 until the evaluation criterion converges, and determine the descriptors corresponding to the converged descriptor dimension and high coefficient of determination R 2 as the main descriptors.

[0039] It should be noted that the genetic algorithm (Genetic Algorithm, GA) is classified in the family of evolutionary algorithms. Its theory draws on the idea of "survival of the fittest" in the biological world, and generates high-quality solutions to optimization and search problems through biologically inspired operators (selection, mutation, and crossover). The schematic diagram of the genetic algorithm is as shown in Figure 4 a part of. The basic steps of the genetic algorithm are as follows: Initialization: Based on the system settings, randomly create individuals with a preset number of individuals or based on some prior knowledge of the problem (each individual includes one descriptor or multiple descriptors) to construct an initial population. Each individual represents a certain combination of descriptors, which is represented by binary codes. Among them, 1 represents that the descriptor at its corresponding position is selected, while 0 is the opposite.

[0040] Evaluation: Evaluate the fitness of each individual in the population. Fitness represents the ability of an individual to solve the problem, usually determined by a fitness function that quantifies the goal or cost of the solution. For a neural network, the coefficient of determination R 2 , the mean squared error MSE, and the root mean squared error RMSE, etc. can be used as the evaluation criteria for the combination of descriptors represented by an individual.

[0041] Selection: Select individuals from the population according to the fitness of the individuals to create a new population. Individuals with higher fitness are more likely to be selected, mimicking the concept of "survival of the fittest".

[0042] Reproduction: Through genetic operations such as crossover (recombination) and mutation on the selected individuals, new individuals (offspring) are generated. Crossover involves combining the genetic material of two individuals to create new solutions, while mutation introduces random changes to maintain the diversity of the population.

[0043] Replacement: Replace the old population with the new population. The new population includes both the selected individuals and the newly generated offspring.

[0044] Termination: Continuously repeat the above steps until the evaluation criteria converge. In specific practice, use the descriptor dimension and the coefficient of determination R 2 as the evolution direction of the population, and finally obtain a population with a converged descriptor dimension and a high coefficient of determination R 2 , that is, determine the main descriptors.

[0045] According to an embodiment of the present invention, it further includes: The preset dielectric constant prediction model is a hierarchical neural network structure composed of multiple sub-networks with the same structure and arranged hierarchically. Each sub-network includes m input nodes and 1 output node. The output data of the nth layer sub-network is the input data of the n + 1th layer sub-network; input M main descriptors and m - M other descriptors into each sub-network in the first layer, and the other descriptors input into each sub-network are determined by random selection.

[0046] It should be noted that the other descriptors input into each sub-network can be It is represented that, among them, 909-M represents other descriptors, that is, descriptors other than the main descriptor; m-M represents the number of other descriptors that can be input into the sub-network in addition to the main descriptor. m-M other descriptors are randomly (without order) selected from 909-M other descriptors and input into the sub-network together with M main descriptors. The other descriptors input into each sub-network are randomly selected, so the other descriptors input into each sub-network are not exactly the same.

[0047] In the hierarchical neural network, the number of sub-networks included in each hierarchical network gradually decreases, that is, the number of sub-networks included in the current level is less than the number of sub-networks included in the previous level, and the highest level only includes one sub-network. The output data of each sub-network in the current level is the input data of each sub-network in the next level. The network structure of each sub-network is the same, and those skilled in the art can also modify the network structure of any sub-network in any level according to actual needs.

[0048] Figure 2 The flowchart of a network training method provided by the present invention is shown.

[0049] As Figure 2 shown, the second aspect of the present invention provides a network training method, and the network is used to predict the dielectric constant of materials. The method includes: S202, obtaining the dielectric constant data of the sample materials; S204, cleaning the dielectric constant data of the sample materials to obtain a sample data dataset; S206, constructing a plurality of descriptors according to the material characteristics of the sample materials in the sample data dataset; S208, screening the plurality of descriptors by a genetic algorithm to determine M main descriptors, and determining the descriptors other than the main descriptors as other descriptors; the main descriptors include whether it is a d-block metal, atomic weight, number of p-block electrons, polarizability, electron cloud density, and electronegativity; S210, training the main descriptors and other descriptors through a hierarchical neural network to establish a preset dielectric constant prediction model.

[0050] It should be noted that 1124 sample materials mainly composed of microwave high-frequency materials are collected through literature and other means, and the dielectric constant data and material characteristics of each sample material are determined. Among them, there are 53 kinds of material characteristics, specifically as Figure 3 shown. The training process of the preset dielectric constant prediction model is as Figure 4As shown in the figure, first, data cleaning is performed on the dielectric constant data of the obtained sample materials. The data cleaning steps include duplicate value deletion, missing value supplementation, and outlier processing, so as to obtain a high-quality sample data dataset (a total of 1124 sample data). Each sample data in the sample data dataset is analyzed in turn to construct corresponding descriptors. The main descriptors are selected through a genetic algorithm, and a hierarchical neural network (HNN) architecture based on a convolutional neural network (CNN) is adopted. 880 data are selected from the sample data dataset as the training set, and 224 data are used as the test set. Through hierarchical training with different combinations of input descriptors (including all main descriptors and some other descriptors), the model prediction ability is gradually improved, and a preset dielectric constant prediction model is established. At the same time, 880 data are selected as the training set, and 224 data are used as the test set to train a traditional convolutional neural network model for a comparative experiment. The traditional convolutional neural network model uses 145 descriptors to build a dielectric constant prediction model, and its training results are as Figure 5 shown, the training set R 2 = 0.990, the test set R 2 = 0.789. The hierarchical neural network adopted in the present invention uses 909 descriptors to build a dielectric constant prediction model, and its training results are as Figure 6 shown, the training set R 2 = 0.994, the test set R 2 = 0.903.

[0051] The results show that the prediction accuracy of the HNN model on the test set has increased by 14.5%, reaching a high level of prediction accuracy. At the same time, by optimizing the model structure layer by layer through the hierarchical neural network, the complex non-linear relationship between the descriptors and the dielectric constant can be fully explored, the prediction results are more reliable, and the error is smaller.

[0052] In addition, the present invention screens descriptors through a genetic algorithm (GA), optimizes the original 909 descriptors into 6 main descriptors, greatly reduces the model complexity, and at the same time retains the information crucial for dielectric constant prediction. The optimized model shows the following advantages: Improved computational efficiency: After descriptor screening, the model parameters are significantly reduced, and the training time is reduced by an average of 30% - 40%; the genetic algorithm reduces redundant descriptors and avoids the burden of high-dimensional input on model training.

[0053] The physical meanings of the main descriptors are clear: The six main descriptors selected by the genetic algorithm, including whether it is a d-block metal, atomic weight, number of p-block electrons, polarizability, electron cloud density, and electronegativity, are all closely related to the physical nature of the dielectric constant. These descriptors show a high contribution degree in model training, ensuring the accuracy and reliability of model prediction.

[0054] In addition, the hierarchical neural network model of the present invention significantly enhances the generalization ability of the model on small data sets through hierarchical input and statistical ensemble optimization, effectively solving the problem of data scarcity: Solution to the problem of data scarcity: Under the condition of only using 1124 data samples (880 training sets and 224 test sets), the R² value of the HNN model on the test set reaches 0.903, fully verifying its excellent performance on small data sets.

[0055] Avoiding overfitting: Through hierarchical input of descriptors and layer-by-layer optimization, the model effectively avoids the overfitting problem caused by the mismatch between high-dimensional descriptors and small data sets; the ensemble statistical method integrates the training results multiple times in the output of each layer of the model, further improving the robustness and generalization ability of the model.

[0056] The third aspect of the present invention provides a dielectric constant prediction system configured to predict the dielectric constant of a material, including: At least one storage medium storing at least one instruction set; and At least one processor communicatively connected to the at least one storage medium, wherein, when the prediction system runs, the at least one processor reads the at least one instruction set and executes the method of any one of a dielectric constant prediction method according to the instructions of the at least one instruction set.

[0057] The fourth aspect of the present invention provides a training system configured to train a prediction network for predicting the dielectric constant of a material, including: At least one storage medium storing at least one instruction set; and At least one processor communicatively connected to the at least one storage medium, wherein, when the training system runs, the at least one processor reads the at least one instruction set and executes the method of any one of a network training method according to the instructions of the at least one instruction set.

[0058] The information involved in this application (including but not limited to user device information, user personal information, etc.), data (including but not limited to data for analysis, stored data, displayed data, etc.), and signals (including but not limited to signals transmitted between user terminals and other devices, etc.) are all authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant laws, regulations, and standards of relevant countries and regions. For example, the "material characteristics of the material to be measured" and "dielectric constant data of the sample material" involved in this disclosure are all obtained under full authorization.

[0059] The present invention discloses a dielectric constant prediction method and system, and a network training method and system. The method includes: obtaining the material characteristics of the material to be measured; constructing a plurality of descriptors according to the material characteristics of the material to be measured; screening the plurality of descriptors through a genetic algorithm to determine M main descriptors, and determining the descriptors other than the main descriptors as other descriptors; inputting the main descriptors and other descriptors into a preset dielectric constant prediction model to output dielectric constant prediction data. The present invention efficiently and accurately predicts the dielectric constant of microwave high-frequency materials through machine learning technology. Compared with traditional neural network models and descriptor selection methods, the prediction accuracy, model efficiency, and generalization ability are significantly improved, and the problem of mismatch between high-dimensional descriptors and small data sets is solved.

[0060] In several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only illustrative. For example, the division of the units is only a logical function division, and there can be other division methods in actual implementation. For example, multiple units or components can be combined, or can be integrated into another system, or some features can be ignored, or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces. The indirect coupling or communication connection of devices or units can be electrical, mechanical, or other forms.

[0061] The units described above as separate components may or may not be physically separated. The components shown as units may or may not be physical units; they can be located in one place or distributed to multiple network units; some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0062] In addition, each functional unit in the embodiments of the present invention can be all integrated in a processing unit, or each unit can be separately used as a unit, or two or more units can be integrated in a unit; the above integrated unit can be implemented in the form of hardware, or in the form of a combination of hardware and software functional units.

[0063] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions. The foregoing program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps including those of the above method embodiments; and the foregoing storage medium includes: various media that can store program codes, such as removable storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs.

[0064] Alternatively, if the above integrated units of the present invention are implemented in the form of software function modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the embodiments of the present invention essentially or the part that contributes to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the various embodiments of the present invention. And the foregoing storage medium includes: various media that can store program codes, such as removable storage devices, ROM, RAM, magnetic disks, or optical discs.

Claims

1. A method for predicting dielectric constant, characterized in that, Including: Obtaining the material characteristics of the material to be tested; Constructing a plurality of descriptors according to the material characteristics of the material to be tested; Screening the plurality of descriptors by a genetic algorithm to determine M main descriptors, and determining the descriptors other than the main descriptors as other descriptors; the main descriptors include whether it is a d-block metal, atomic weight, number of p-block electrons, polarizability, electron cloud density, and electronegativity; Inputting the main descriptors and the other descriptors into a preset dielectric constant prediction model to output dielectric constant prediction data.

2. The method for predicting the dielectric constant according to claim 1, wherein The constructing a plurality of descriptors according to the material characteristics of the material to be tested includes: Calculating the configurational entropy of the material: ； Among them, ΔS con is the configurational entropy of the material, k B is the Boltzmann constant, T is the Kelvin temperature, and c i is the proportion of element i among all elements; Respectively calculating the ratios of s, p, d, and f valence electrons to the total sum of s, p, d, and f valence electrons in the elements presented by the material to determine the electronic structure attributes; Determining the ionic compound attributes according to the electronegativity of the material; Statistically analyzing the properties of various elements of the material by a variety of preset statistical methods to determine the statistical properties of the elemental properties of the material; Determining the configurational entropy, electronic structure attributes, ionic compound attributes, and statistical properties of elemental properties of the material as descriptors.

3. The dielectric constant prediction method according to claim 1, characterized in that The screening the plurality of descriptors by a genetic algorithm to determine M main descriptors includes: Step 1, creating individuals randomly or based on some prior knowledge of the problem to construct an initial population; Step 2, evaluate the fitness of each individual in each group through a preset evaluation criterion; the preset evaluation criterion includes the coefficient of determination R 2 , the mean squared error MSE and the root mean squared error RMSE; Step 3, selecting individuals from each population according to the fitness of the individuals to create a new population; Step 4, selecting individuals from the new population for genetic operations to generate new individuals; the genetic operations include selection, mutation, and crossover; Step 5, replacing the old population with the new population; Repeat steps 2 - 5 continuously until the evaluation criteria converge, and determine the descriptors corresponding to the converged descriptor dimensions and high determination coefficients R 2 of the population as the main descriptors.

4. The dielectric constant prediction method according to claim 1, characterized in that Also including: The preset dielectric constant prediction model is a hierarchical neural network structure composed of a plurality of sub-networks with the same structure and arranged hierarchically. Each sub-network includes m input nodes and 1 output node, and the output data of the nth layer sub-network is the input data of the n+1th layer sub-network; Inputting M main descriptors and m-M other descriptors into each sub-network in the first layer respectively, and the other descriptors input into each sub-network are determined by random selection.

5. A training method of a network, the network is used to predict the dielectric constant of a material, and the method includes: Obtaining the dielectric constant data of the sample material; Performing data cleaning on the dielectric constant data of the sample material to obtain a sample data data set; Constructing a plurality of descriptors according to the material characteristics of the sample material in the sample data data set; Screening the plurality of descriptors by a genetic algorithm to determine M main descriptors, and determining the descriptors other than the main descriptors as other descriptors; the main descriptors include whether it is a d-block metal, atomic weight, number of p-block electrons, polarizability, electron cloud density, and electronegativity; Training the main descriptors and the other descriptors by a hierarchical neural network to establish a preset dielectric constant prediction model.

6. The training method according to claim 5, wherein The constructing a plurality of descriptors according to the material characteristics of the sample material in the sample data data set includes: Calculating the configurational entropy of the material: ； Among them, ΔS con is the configurational entropy of the material, k B is the Boltzmann constant, T is the Kelvin temperature, and c i is the proportion of element i among all elements; Respectively calculating the ratios of s, p, d, and f valence electrons to the total sum of s, p, d, and f valence electrons in the elements presented by the material to determine the electronic structure attributes; Determine the properties of ionic compounds according to the electronegativity of materials; Statistically analyze the properties of multiple elements of the material respectively through a variety of preset statistical methods to determine the statistical properties of the elemental properties of the material; Determine the configurational entropy, electronic structure properties, ionic compound properties and elemental property statistical properties of the said material as descriptors.

7. The training method according to claim 5, wherein Screen the multiple descriptors through a genetic algorithm to determine M main descriptors, including: Step 1, create individuals randomly or based on some prior knowledge of the problem to construct an initial population; Step 2, evaluate the fitness of each individual in each group through a preset evaluation criterion; the preset evaluation criterion includes the coefficient of determination R 2 , mean squared error MSE, and root mean squared error RMSE; Step 3, select individuals from each population according to the fitness of the individuals to create a new population; Step 4, select individuals from the new population for genetic operations to generate new individuals; the genetic operations include selection, mutation and crossover; Step 5, replace the old population with the new population; Repeat steps 2 - 5 continuously until the evaluation criteria converge, and determine the descriptor dimensions and population with a high coefficient of determination R 2 corresponding to the population as the main descriptors.

8. The training method according to claim 5, wherein Also includes: The preset dielectric constant prediction model is a hierarchical neural network structure composed of multiple sub-networks with the same structure and arranged hierarchically. Each sub-network includes m input nodes and 1 output node, and the output data of the nth layer sub-network is the input data of the n+1th layer sub-network; Input M main descriptors and m-M other descriptors into each sub-network in the first layer respectively. The other descriptors input into each sub-network are determined by random selection.

9. A dielectric constant prediction system, characterized in that, Configured to predict the dielectric constant of materials, including: At least one storage medium storing at least one instruction set; and At least one processor communicatively connected to the at least one storage medium, wherein, when the prediction system runs, the at least one processor reads the at least one instruction set and executes the method according to any one of claims 1-4 according to the instructions of the at least one instruction set.

10. A training system, characterized in that, Configured to train a prediction network for predicting the dielectric constant of materials, including: At least one storage medium storing at least one instruction set; and At least one processor communicatively connected to the at least one storage medium, wherein, when the training system runs, the at least one processor reads the at least one instruction set and executes the method according to any one of claims 5-8 according to the instructions of the at least one instruction set.

Images