Method for predicting dielectric property of modified perovskite based on deep learning integrated model

Abstract

The invention belongs to the technical field of intersection of artificial intelligence and material science, and particularly relates to a method for predicting the dielectric property of modified perovskite based on a deep learning integration model, which comprises a data acquisition module, a feature engineering module, a model construction module and a result output module, a deep learning prediction model GBR-DNN-CNN integrated model is constructed through a model construction module, deep feature learning is carried out on a data set, and regression prediction is carried out on the perovskite dielectric property by adopting a regression analysis technology; according to the method, the dielectric property of the modified perovskite is accurately predicted by utilizing the powerful data processing capability of deep learning, so that the prediction efficiency and accuracy are improved, and the method is beneficial to accelerating the screening and optimizing process of materials, improving the product performance, reducing the research and development cost and reducing the resource consumption in the research and development stage.

Description

Technical Field

[0001] The present invention belongs to the technical field of intersection of artificial intelligence and material science, and specifically relates to a method for predicting the dielectric properties of modified perovskite based on a deep learning integrated model. Background Art

[0002] Perovskite structural materials, with their unique crystal symmetry and excellent electronic properties, have shown the potential for wide application in the semiconductor field. In order to obtain excellent matrix materials that enhance the energy storage capacity of perovskites, doping modification is generally used to increase their dielectric constant to enhance energy storage density.

[0003] However, the dielectric properties of perovskite materials are often affected by many factors, especially when their properties are regulated by modification, predicting their dielectric properties becomes particularly complex and difficult. At the same time, due to the high cost of experimental research on perovskite materials, experimental data are scattered and sparse, and the sample size is small, making it difficult to collect and unify data sets, which seriously hinders in-depth research on their dielectric properties and energy storage density.

[0004] Traditional experimental methods rely on repeated trials and density functional theory first-principles calculations. In order to obtain the target material, a large number of calculations must be carried out, which is time-consuming and expensive. In addition, when dealing with small data samples, problems such as low prediction accuracy, slow convergence speed, and poor generalization ability may occur.

[0005] With the development of artificial intelligence technology, machine learning (ML) technology has made significant breakthroughs in the discovery and design of materials, and is expected to solve many deficiencies in existing technologies. However, the training of performance prediction models usually requires a large number of samples, which are often difficult to obtain, resulting in problems such as low prediction accuracy, slow convergence, and poor generalization ability. Deep learning (DL) can quickly and effectively discover new materials. It is mainly aimed at establishing small sample data sets and relying on reliable experimental data. Neural networks can describe the functional composition and process parameters of material properties in high-dimensional space, significantly improving the accuracy and convergence rate of prediction models, thereby obtaining more accurate prediction results. Therefore, there is an urgent need for a perovskite property prediction method based on deep learning. Summary of the invention

[0006] The purpose of the present invention is to provide a method for predicting the dielectric properties of modified perovskite based on a deep learning integrated model, so as to solve the problems raised in the background technology.

[0007] In order to achieve the above technical objectives, the technical solution adopted by the present invention is as follows:

[0008] A method for predicting the dielectric properties of modified perovskite based on a deep learning integrated model includes a data acquisition module, a feature engineering module, a model building and prediction module, and a result output module. The specific steps are as follows:

[0009] S1: Collect a dataset containing target attributes;

[0010] S2: Perform feature engineering to improve data training efficiency;

[0011] S3: Construct a deep learning prediction model GBR-DNN-CNN integrated model, conduct in-depth feature learning of the data set, and use regression analysis technology to perform regression prediction on the dielectric properties of perovskite;

[0012] S4: Display the prediction results and generate customized reports.

[0013] Preferably, in step S1, the required data is collected from an open source database, and strategic doping and data partitioning are adopted to provide a solid data foundation for the deep learning model to improve the accuracy and generalization ability of the model.

[0014] Preferably, in step S2, normalization processing and Pearson correlation coefficient are used to effectively process the data, and the data is comprehensively optimized to achieve optimization of feature selection and data dimensionality reduction, thereby improving the generalization ability and prediction efficiency of the model, while reducing dependence on computing resources.

[0015] Preferably, in step S3, gradient boosting regression (GBR) is used to effectively process smaller data sets, and the performance of the model is improved by gradually adding weak learners, thereby reducing the overfitting of the deep learning model on the small data set and providing an initial prediction for the subsequent model.

[0016] Preferably, in step S3, a deep neural network (DNN) is used to perform feature learning and efficient processing on the preprocessed data, and the composition of input neurons, hidden neurons, and output neurons enhances the generalization of the pretrained model.

[0017] Preferably, in step S3, a convolutional neural network (CNN) is used to extract spatial features of the data to improve the parameter explosion problem in the DNN model.

[0018] Preferably, in step S3, the model is applied to the data set to test the effectiveness of the model on new data outside the data set. The mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R 2 ) evaluation indicators are used to conduct a comprehensive performance evaluation of the prediction results of the deep learning model to enhance the accuracy and credibility of the prediction.

[0019] Preferably, in step S4, visualization technology is used to display the prediction results in real time to fully understand the dielectric properties of the material. Reports are prepared through database integration to efficiently manage the prediction data and provide in-depth performance trend analysis. Automated anomaly detection and multi-level alarm systems are implemented to promptly notify potential problems, combined with manual re-inspection to ensure the accuracy of the prediction and improve the robustness of the system.

[0020] The present invention uses the powerful data processing capabilities of deep learning to accurately predict the dielectric properties of modified perovskites, thereby improving prediction efficiency and accuracy. This method not only helps to accelerate the material screening and optimization process and improve product performance, but also reduces R&D costs and reduces resource consumption in the R&D stage. It shows great application value and market potential in the fields of semiconductor electronic device design, new energy material development, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The present invention is further illustrated by means of the non-limiting examples given in the accompanying drawings.

[0022] Figure 1 It is a structural schematic diagram of a method for predicting dielectric properties of modified perovskite based on a deep learning integrated model of the present invention;

[0023] Figure 2 It is a block diagram of a data acquisition module in the modified perovskite dielectric property prediction method based on a deep learning integrated model of the present invention;

[0024] Figure 3 It is a block diagram of a feature engineering module in a modified perovskite dielectric property prediction method based on a deep learning integrated model of the present invention;

[0025] Figure 4 It is a block diagram of a model building module and a prediction module in the modified perovskite dielectric property prediction method based on a deep learning integrated model of the present invention;

[0026] Figure 5 It is a block diagram of the result output module in the modified perovskite dielectric property prediction method based on the deep learning integrated model of the present invention. DETAILED DESCRIPTION

[0027] In order to enable those skilled in the art to better understand the present invention, the technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0028] like Figure 1-5 As shown, a method for predicting the dielectric properties of modified perovskite based on the GBR-DNN-CNN deep learning integrated model of the present invention;

[0029] The data acquisition module is used to provide comprehensive and balanced dielectric performance data support for the prediction model. The required data is identified and collected from the open source database, the standardized conversion of chemical ratios is achieved, and key material properties are screened based on the doping strategy. The hierarchical clustering method is used to optimize the data set division and the generalization ability of the model training. Specifically, the following steps are included;

[0030] Step 1: Identify data sources, using the open source toolkit Matminer module, as well as publicly available experimental data and literature research, identification, and data collection;

[0031] Step ②: Data collection: collect dielectric constant data of doped and modified perovskite materials to ensure the representativeness and diversity of the data. The data only uses material composition, adjusts the stoichiometric ratio between dopants, and converts it into a unified molar ratio expressed in mole percentage (mol%);

[0032] Step 3: Dopant selection: ABO3 perovskite is used as the material, and Sr is doped at the A position. 2+ , Ba 2+ , Ca 2+ , Cs + 、CH3NH3 + Equivalent cations, doping Ti at the B site 4+ 、Zr 4+ 、Al 3+ 、Zn 2+ Sn 2+ Cations such as cations, and rare earth elements Nd 3+ 、Sm 3+ , Pr 3 + 、Dy 3+ etc., changing the key aspects of perovskite performance prediction such as energy band characteristics, optimal composition, dielectric properties, band gap energy and dielectric breakdown strength;

[0033] Step ④: Dataset partitioning strategy: According to the hierarchical clustering framework, the dataset is divided into a training set (80%) and a test set (20%) to avoid situations where memorization or overfitting hinders new information, reduce the reuse of the same categories, and ensure the diversity of the dataset, thereby promoting the model's ability to learn new information.

[0034] The feature engineering module consists of two parts: data preprocessing and feature selection. Its specific steps include:

[0035] Data preprocessing unit: Use exploratory data analysis techniques, normalization processing and visual analysis tools to ensure the consistency and analyzability of the data set. The specific steps are as follows:

[0036] Exploratory data analysis: Use statistical charts and visualization tools (distribution, skewness) to understand and analyze the basic characteristics of data;

[0037] Normalization: The mass ratio of elements to doping materials is converted into molar ratio, and the Min-MaxScaler algorithm is used to normalize the data set to eliminate the dimension and scale differences between different features to achieve data standardization. The normalization formula is:

[0038]

[0039] Among them, x is the characteristic value of each sample, min(x) is the minimum value of each characteristic, min(x) is the maximum value, X normalize is the normalized eigenvalue;

[0040] Data visualization: Visualization tools such as box plots and density plots are used to deeply analyze the characteristic distribution of data sets, so as to intuitively display the statistical characteristics of the data such as the median, quartiles, extreme values, and outliers, and reveal the distribution characteristics of the data such as skewness;

[0041] Feature selection unit: Pearson correlation coefficient is used to quantify the linear correlation between features and target variables, reduce feature dimensions, optimize model performance and generalization ability, and achieve efficient feature representation;

[0042] Specifically, two variables X and Y are set, where and represent the means of their respective variables;

[0043] The calculated Pearson correlation coefficient ranges from -1 to 1; values ​​close to 1 indicate a positive correlation between the two variables, values ​​close to -1 indicate a negative correlation, and values ​​close to 0 indicate a lack of significant correlation between the two variables;

[0044] During the initial screening process, features with high correlation with the target variable are identified and eliminated to reduce multicollinearity between features; features with low correlation are retained as inputs to the machine learning model to eliminate the reliance on large amounts of computing resources and time-consuming feature selection processes, thus lowering the threshold for achieving high-performance prediction models;

[0045] The characteristics of elements with lower Pearson correlation coefficients with the molar ratio of doping elements (such as Sr, Ti, O) are further retained as inputs to the ML model; the calculation formula for the Pearson correlation coefficient is:

[0046]

[0047] Among them, X and Y represent the feature and target variable respectively, n is the number of samples, and is their mean, σ X and σ Y is the standard deviation and cov(X,Y) is their covariance.

[0048] Collect some data that fail to increase or even decrease the dielectric constant, add them into the database and use them for model training;

[0049] Gradient Boosting Regression (GBR) model: By minimizing the algorithm generated during the training process, the prediction accuracy of the final regression result is improved, providing an initial prediction for subsequent models to improve the accuracy and robustness of the overall prediction; the GBR initial model is obtained from the Python language and scikit-learn library, and the following improvements are made to the model:

[0050] Model architecture optimization: The GBR model integrates multiple weak predictors (such as decision trees) through algorithm optimization to form a strong predictor. Each weak predictor is trained based on the residual of the previous round of iterations to enhance the model's ability to capture complex data relationships and gradually reduce prediction errors.

[0051] Loss function definition: Use mean square error (MSE) as the loss function, which is used to quantify the error of model prediction by the average value of the square difference between the predicted value and the observed value. The formula of MSE is:

[0052]

[0053] Where N is the number of samples, E i is the actual value, and is the predicted value of the model;

[0054] Hyperparameter tuning: Optimize the hyperparameters of the GBR model to adapt to the small dataset. The specific steps include:

[0055] The upper limit of iteration number is set to 2000;

[0056] Set the learning rate parameter to 0.1 to adjust the magnitude of model adjustment in each iteration and control the iteration step size;

[0057] The maximum depth of the decision tree is limited to 2 layers to avoid overfitting and maintain the generalization ability of the model;

[0058] Weight update strategy: In each round of iteration, the weight of the new weak predictor is determined by the loss function minimization process. The weight update formula is:

[0059]

[0060] Where w is the weight of the new weak predictor, Fn-1 (X i ) is the prediction of the current model, and f(X i ,w) is the prediction of the new weak predictor;

[0061] Model training and verification: The preprocessed data was input to train and predict the improved GBR model, verifying the effectiveness and accuracy of the model on a small data set to improve the prediction accuracy of the final regression results;

[0062] Deep Neural Network (DNN) Model: Utilize the high flexibility of deep neural networks (DNN) to capture patterns in data, automatically extract useful features from raw data, predict the energy storage density of materials with high accuracy, and provide fast and effective decision support for material design and optimization;

[0063] In addition, DNN models can predict unknown situations by learning from existing data to obtain experimental data that is difficult or costly to obtain;

[0064] Model architecture: Build a deep neural network (DNN) consisting of an input layer, several hidden layers, and an output layer.

[0065] Input layer: consists of 18 neurons representing independent input features.

[0066] Output layer: It consists of a single neuron and uses the linear activation function purelin to generate the predicted output.

[0067] Hidden layers: The number of hidden layers and neuron configuration are determined according to specific empirical formulas to ensure that the model remains simple while achieving the required prediction accuracy.

[0068] The specific training steps are as follows:

[0069] ① Using existing models and literature experience, we initialized the model's parameters such as the number of hidden layers, the activation function of each layer, the number of nodes per layer, etc., established the optimal DNN model with 1-5 hidden layers, and compared the prediction accuracy of different numbers of hidden layers.

[0070] ② The DNN model is trained using an empirical method based on determining the number of hidden layers and corresponding nodes. The empirical function formula is:

[0071]

[0072] Among them, m is the number of input layer nodes, n is the number of output layer nodes, and a is generally an integer between 1 and 10;

[0073] The number of hidden layers ranges from 1 to 5, and the activation functions are Y = tansig(x), Y = purelin(x), Y = logsig(x), and ReLu;

[0074] ③ The trainlm optimization algorithm is used to pre-train the DNN, with the preset number of training cycles being 1000 and the learning rate being set to 0.01, in order to minimize the error to 0.000001. The mean square error (MSE) of the validation set is used as a measure of the prediction accuracy in the pre-training phase, and the optimal hidden layer neuron configuration is determined accordingly;

[0075] ④ Use the data trained by the GRB model to pre-train the DNN model, initialize the parameters set by the existing knowledge, and carefully adjust the model parameters according to the training results, including the number of hidden layer nodes, to obtain the optimal model configuration parameters and reconstruct the DNN model;

[0076] Convolutional Neural Network (CNN) Model: Use convolutional neural network (CNN) to extract spatial features of data, focusing on improving the parameter explosion problem in the DNN model. Build a suitable network architecture. The CNN model covers convolutional layer (ConvolutionalLayer), pooling layer (Pooling Layer), fully connected layer (Fully Connected Layer) and output layer (OutputLayer) to ensure the efficiency and effectiveness of feature extraction. The preferred ones are:

[0077] The input data matrix is ​​sent to the one-dimensional convolution layer to perform a one-dimensional convolution operation, and the convolution kernel is used to perform a sliding window operation on the data to effectively extract local features. The convolution kernel is multiplied element by element with the local area of ​​the data matrix and then summed to capture local dependencies, and then a nonlinear transformation is performed with the help of an activation function. The convolution operation formula is:

[0078] (f*g)[i]=∑ k f[k]·g[ki]

[0079] Among them, f is the input data, g is the convolution kernel, and * represents the convolution operation;

[0080] The ReLu activation function formula is:

[0081] ReLu=max(0,x)

[0082] The maximum pooling operation is used to reduce the feature dimension and compress the number of parameters, effectively preventing overfitting. It also filters out the most representative information, ignores unimportant noise, and enhances the generalization and robustness of the model.

[0083] The function prototype of the maximum pooling output is:

[0084]

[0085] The fully connected layer is used to merge the information of the previous layers. The output of the pooling layer is sent to one or more fully connected layers. Each neuron in these layers is connected to the activation value of the previous layer. A linear transformation is performed in the fully connected layer. The formula is:

[0086] y=σ(ω*x+b)

[0087] Among them, y is the output, ω is the weight matrix, b is the bias term, and σ is the activation function;

[0088] The last layer is the regression layer, which first calculates the raw prediction value (logits) of each category through the fully connected layer, and then passes these unprocessed outputs (logits) to the softmax activation function to calculate the predicted probability distribution of each possible category. On this basis, the category with the highest probability is selected as its final prediction output, and the formula is:

[0089]

[0090] Among them, x is the input of the fully connected layer, W T is the weight, b is the bias term, is the probability of softmax output, and softmax is calculated as follows:

[0091]

[0092] Among them, z i is the i-th element of the fully connected layer output vector, corresponding to the unnormalized probability of the i-th category;

[0093] Use optimization algorithms such as gradient descent (or its variants) to adjust the weights and biases of the network to minimize the loss function and improve the parameter explosion problem;

[0094] The adaptive moment estimation (ADAM) optimizer is used for model training, and the initial learning rate is set to 0.005. The learning rate and batch size are dynamically adjusted during the iteration process to accelerate the training process and improve efficiency.

[0095] Regression prediction unit: Deploy the model on the data set to verify its generalization ability and prediction accuracy outside the data set, and make predictions on the test set in the original data set. The preferred ones are:

[0096] Model prediction: The established GBR-DNN-CNN model is used to perform regression prediction on the test part of the dataset.

[0097] Performance evaluation: Three standard error indicators are used to evaluate the prediction accuracy of each machine learning model, including: Mean absolute error (MAE): The average value of the absolute error between the predicted value and the observed value is evaluated. The formula is:

[0098]

[0099] Root mean square error (RMSE): It is used to evaluate the sample standard deviation of the difference (residual) between the predicted value and the observed value. It has a penalizing effect on large errors and is also useful when the error is particularly large and undesirable. The formula is:

[0100]

[0101] Coefficient of determination (R 2 ): It is used to measure the degree of fit between the measured data and the regression line and is measured in percentage. The formula is:

[0102]

[0103] Among them, RSE is the relative square error;

[0104] in, is the prediction result of variable E; is the mean of the sum of n samples, i.e., i = 1, 2, …, n;

[0105] Error sensitivity analysis: MAE is more sensitive to smaller prediction errors, while RMSE is more sensitive to larger prediction errors. Both are expressed in variable units and are negative indicators. The lower the value, the better the model performance.

[0106] Model interpretability metrics: R 2 The indicator quantifies the degree to which the model explains the variability of the data, and its value range is between 0 and 1. 2 The closer the value is to 1, the better the model fits the data and the stronger its explanatory power;

[0107] Comprehensively evaluate the performance indicators of each model and its integrated combination, select the model with the best performance for practical application, and improve the accuracy and credibility of the prediction.

[0108] Result output module: In a comprehensive display and analysis platform, the dielectric property prediction of perovskite materials is intuitively presented, allowing users to monitor, evaluate and deeply analyze data in real time, generate customized reports, and perform data management and anomaly detection functions on the prediction results to ensure the accuracy of the prediction information and the responsiveness of the model. The preferred:

[0109] Results display: An interactive interface is designed to visualize the performance prediction results. The prediction results, including dielectric performance indicators and predicted values, can be displayed in real time on the display interface of the central control system. By integrating interactive analysis tools, users can dig deep into the data and dynamically adjust model settings, thus providing solid data-driven support for the decision-making process. The following steps and techniques are used:

[0110] 1) User interface design: Develop an interactive user-friendly interface that integrates real-time forecast result charts, performance indicator statistics, and historical forecast data. The interface supports interactive functions such as data point query, trend line analysis, etc., and supports functions such as zooming, panning, and detailed information query to facilitate users to conduct in-depth analysis.

[0111] 2) Real-time visualization of prediction results: Using advanced visualization technology, dynamic charts are used to display the changing trends of dielectric constant and dielectric loss over time or temperature in real time. By using color coding and graphic symbols to distinguish the dielectric performance indicators and predicted values ​​of different element materials, the prediction results of the regression model are intuitively displayed, allowing users to intuitively and quickly obtain material performance.

[0112] 3) Visualize the model decision path: The model’s inference process is converted into an intuitive graphical interface through graphical means, such as decision trees or flow charts, to visualize the model’s decision process and enable users to clearly track the generation process of the prediction results.

[0113] 4) Dynamic adjustment: Parameters such as the constituent elements, crystal structure and environmental conditions of the perovskite material can be adjusted in real time through the interactive interface to observe the dynamic changes of the prediction results and improve the application flexibility of the model.

[0114] Detailed performance indicator display: present key dielectric parameters and influencing factors through an intuitive interface, use sorting and charting tools to highlight influencing factors, and use historical comparison functions to help users understand the evolution of performance over time. The specific steps and techniques are as follows:

[0115] 1) Performance Index Panel: Displays the currently predicted detailed dielectric performance indicators on the side of the interface or in a pop-up window, including dielectric constant, dielectric loss, frequency dependence, and possible key performance influencing factors.

[0116] 2) Dielectric performance characteristics importance ranking: Use machine learning algorithms to evaluate and quantify the impact of each feature on the prediction of perovskite dielectric properties. Use dynamic charts (such as bar charts or Pareto charts) to show the contribution of each factor to the prediction results, and intuitively present the factors that have the greatest impact on dielectric properties.

[0117] 3) Historical data comparison: Through the timeline or comparison view, users can review historical prediction results and analyze the changing trends of material dielectric properties over time.

[0118] Data management and analysis: Store forecast data in an integrated data repository, compress and efficiently retrieve data, generate customizable reports, and identify long-term trends and potential problems to assist in in-depth performance analysis and process optimization. The specific steps and techniques are as follows:

[0119] 1) Database integration: Build a database to store prediction results, including material parameters, dielectric performance indicators, timestamps, and experimental conditions. The database system places special emphasis on data access efficiency and optimization of data management.

[0120] 2) Data management and optimization: Use data compression technology to reduce storage requirements and implement efficient data access strategies to ensure data security as well as fast retrieval and long-term storage capabilities.

[0121] 3) Generate customized reports: Automatically generate reports that provide detailed information on prediction results, including dielectric performance statistics, change trends, and anomaly detection, and support classification analysis by time, material batch, or performance type. Support operators to customize report content according to needs, and allow data export, providing strong data support for scientific research and engineering decision-making.

[0122] 4) Trend analysis: Utilize automated data analysis tools to identify long-term trends and periodic changes in dielectric properties as well as potential problems from prediction results, providing data support for material optimization and process improvement.

[0123] 5) Performance prediction and optimization suggestions: The model prediction provides future predictions of the dielectric properties of materials based on historical data and current predictions, and provides optimization suggestions for process flows such as perovskite material synthesis and processing based on the prediction results.

[0124] Early warning and response mechanism: Intelligent detection quickly identifies abnormal data, triggers a multi-level alarm system, and implements real-time notification. Manual rechecks verify the accuracy of predictions, and records alarm events for continuous improvement, enhancing the robustness of the model and the user's responsiveness. The specific steps and techniques are as follows:

[0125] 1) Automatic anomaly detection: Use automated statistical detection methods, such as the quartile-based IQR method or Z-score, to identify potential outliers in the data. Automatically identify and mark abnormal changes in dielectric properties, indicating potential data or model problems.

[0126] 2) Automatic alarm: When the prediction results show that the dielectric properties of the material may be below the standard or there is a potential risk, the system will automatically trigger a warning signal and ensure that the information can be conveyed to relevant personnel in a timely manner through a variety of communication channels (such as visual interface, audio message, email, etc.).

[0127] 3) Multi-level alarm system: Based on the severity of the performance indicators, the system will implement a multi-level early warning strategy to ensure that appropriate measures are taken in a timely manner.

[0128] 4) Manual review: For prediction results that require further confirmation, the system will prompt researchers to conduct manual review to ensure the accuracy and reliability of the prediction.

[0129] 5) Alarm records and continuous improvement: Record and store detailed information of all warning events, including time, performance indicators and treatment measures, for subsequent analysis to discover the optimization potential of the warning mechanism and enhance the prediction accuracy and response efficiency of the system. Based on alarm records and user feedback, continuously optimize the prediction model and output module.

[0130] In summary, the present invention uses the powerful data processing capabilities of deep learning to accurately predict the dielectric properties of modified perovskites, thereby improving the prediction efficiency and accuracy. This method not only helps to accelerate the material screening and optimization process and improve product performance, but also reduces R&D costs and reduces resource consumption in the R&D stage. It shows great application value and market potential in the fields of semiconductor electronic device design, new energy material development, etc.

[0131] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the technology may modify or change the above embodiments without violating the spirit and scope of the present invention. Therefore, all equivalent modifications or changes made by those with ordinary knowledge in the technical field without departing from the spirit and technical ideas disclosed by the present invention should still be covered by the claims of the present invention.

Claims

1. A method for predicting the dielectric properties of modified perovskite based on a deep learning integrated model, characterized in that: It includes data acquisition module, feature engineering module, model building module and result output module. The specific steps are as follows: S1: Collecting a data set containing target attributes through a data collection module; S2: Improve data training efficiency through feature engineering through feature engineering module; S3: Build a deep learning prediction model GBR-DNN-CNN integrated model through the model building module, conduct in-depth feature learning of the data set, and use regression analysis technology to perform regression prediction on the dielectric properties of perovskite; The GBR-DNN-CNN integrated model consists of a GBR model, a DNN model, and a CNN model, where GBR is a gradient boosting regression model, DNN is a deep neural network model, and CNN is a convolutional neural network model; The gradient boosting regression model processes smaller data sets, improves the performance of the model by gradually adding weak learners, reduces the overfitting of the deep learning model on small data sets, and provides an initial prediction for subsequent models; The deep neural network model performs feature learning and efficient processing on preprocessed data. The composition of input neurons, hidden neurons, and output neurons enhances the generalization of the pretrained model. The convolutional neural network model extracts the spatial features of the data and improves the parameter explosion problem in the DNN model; Regression prediction applies the model to the data set, tests the effectiveness of the model on new data outside the data set, and uses mean absolute error, root mean square error, and determination coefficient evaluation indicators to conduct a comprehensive performance evaluation of the prediction results of the deep learning model to enhance the accuracy and credibility of the prediction; S4: Display the prediction results and generate customized reports through the result output module.

2. The method for predicting the dielectric properties of modified perovskite based on a deep learning integrated model according to claim 1, characterized in that: In step S1, the data acquisition module collects the required data from the open source database, adopts strategic doping and data partitioning, and provides a solid data foundation for the deep learning model to improve the accuracy and generalization ability of the model, which specifically includes the following steps: Step 1: Identify data sources, using the open source toolkit Matminer module, as well as publicly available experimental data and literature research, identification, and data collection; Step 2: Data collection: collect dielectric constant data of doped and modified perovskite materials to ensure the representativeness and diversity of the data; Step 3: Dopant selection: ABO3 perovskite is used as the material, and Sr is doped at the A position. 2+ , Ba 2+ , Ca 2+ , Cs + 、CH3NH3 + Cation, doping Ti at B site 4+ 、Zr 4+ 、Al 3+ 、Zn 2+ Sn 2+ Cations, and rare earth elements Nd 3+ 、Sm 3+ , Pr 3+ 、Dy 3+ , changing the energy band characteristics, optimal composition, dielectric properties, band gap energy and dielectric breakdown strength Key aspects of perovskite performance prediction; Step ④: Dataset partitioning strategy: divide the dataset into training and test sets according to the hierarchical clustering framework to avoid situations where memorization or overfitting hinders new information, reduce the reuse of the same categories, ensure the diversity of the dataset, and thus promote the model's ability to learn new information.

3. The method for predicting the dielectric properties of modified perovskite based on a deep learning integrated model according to claim 1, characterized in that: In step S2, the feature engineering module consists of two parts: a data preprocessing unit and a feature selection unit; The data preprocessing unit uses exploratory data analysis techniques, normalization processing and visual analysis tools to ensure the consistency and analyzability of the data set; Among them, exploratory data analysis techniques use statistical charts and visualization tools to understand and analyze the basic characteristics of data; Among them, the normalization process converts the mass ratio of elements to doping materials into a molar ratio, and uses the Min-MaxScaler algorithm to normalize the data set to eliminate the dimension and scale differences between different features to achieve data standardization. The normalization formula is: x is the feature value of each sample, min(x) is the minimum value of each feature, min(x) is the maximum value, X normalize is the normalized eigenvalue; Among them, data visualization uses box plots and density map visualization tools to deeply analyze the characteristic distribution of data sets, intuitively display the median, quartiles, extreme values, and outlier statistical characteristics of the data, and reveal the skewed distribution characteristics of the data; The feature selection unit uses the Pearson correlation coefficient to quantify the linear correlation between features and target variables, reduce feature dimensions, and optimize model performance and its generalization ability.

4. The method for predicting the dielectric properties of modified perovskite based on a deep learning integrated model according to claim 1, characterized in that: In step S3, the gradient boosting regression model improves the prediction accuracy of the final regression result by minimizing the algorithm generated during the training process, provides an initial prediction for the subsequent model to improve the accuracy and robustness of the overall prediction, obtains the GBR initial model from the Python language and scikit-learn library, and improves the model as follows; Model architecture optimization: The GBR model integrates multiple weak predictors to form a strong predictor through algorithm optimization. Each weak predictor is trained based on the residual of the previous round of iterations to enhance the model's ability to capture complex data relationships and gradually reduce prediction errors. Loss function definition: Use mean square error as the loss function, which is used to quantify the error of model prediction by taking the average of the squared differences between the predicted value and the observed value; Hyperparameter tuning: Optimize the hyperparameters of the GBR model to adapt to the small dataset; Weight update strategy: In each iteration, the weight of the new weak predictor is determined by the loss function minimization process; Model training and verification: The preprocessed data was input to train and predict the improved GBR model, verifying the effectiveness and accuracy of the model on a small data set to improve the prediction accuracy of the final regression results.

5. The method for predicting the dielectric properties of modified perovskite based on a deep learning integrated model according to claim 4, characterized in that: In step S3, the high flexibility of the deep neural network model is used to capture patterns in the data, automatically extract useful features from the raw data, and predict the energy storage density of the material with high accuracy, providing fast and effective decision support for material design and optimization; The deep neural network model architecture consists of an input layer, several hidden layers, and an output layer; Among them, the input layer: consists of 18 neurons representing independent input characteristics; Among them, the output layer: consists of a single neuron, using the linear activation function purelin to generate the predicted output; Among them, hidden layer: the number of hidden layers and neuron configuration are determined according to specific empirical formulas to ensure that the model achieves the required prediction accuracy while maintaining simplicity; The specific training steps of the deep neural network model are as follows: ① Using existing models and literature experience, initialize the number of hidden layers, activation functions of each layer, and number of nodes of each layer. Establish the best DNN model with 1-5 hidden layers, and compare the prediction accuracy of different numbers of hidden layers; ② The DNN model is trained using an empirical method based on determining the number of hidden layers and corresponding nodes. The empirical function formula is: Among them, m is the number of input layer nodes, n is the number of output layer nodes, and a is generally an integer between 1 and 10; The number of hidden layers ranges from 1 to 5, and the activation functions are Y = tansig(x), Y = purelin(x), Y = logsig(x), and ReLu; ③ The trainlm optimization algorithm is used to pre-train the DNN. The number of training cycles is preset to 1000 and the learning rate is set to 0.01, in order to minimize the error to 0.000001. The mean square error of the validation set is used as a measure of the prediction accuracy in the pre-training stage, and the optimal hidden layer neuron configuration is determined accordingly. ④ Use the data trained by the GRB model to pre-train the DNN model, initialize the training of the parameters set by the existing knowledge, and carefully adjust the model parameters according to the training results, including the number of hidden layer nodes, to obtain the optimal model configuration parameters and reconstruct the DNN model.

6. The method for predicting the dielectric properties of modified perovskite based on a deep learning integrated model according to claim 5, characterized in that: In the step S3, a convolutional neural network (CNN) is used to extract spatial features of the data, focusing on improving the parameter explosion problem in the DNN model and building a suitable network architecture. The CNN model includes a convolutional layer, a pooling layer, a fully connected layer, and an output layer to ensure the efficiency and effectiveness of feature extraction; The input data matrix is ​​sent to the one-dimensional convolution layer to perform a one-dimensional convolution operation. The convolution kernel is used to perform a sliding window operation on the data to effectively extract local features. The convolution kernel is multiplied element by element with the local area of ​​the data matrix and then summed to capture local dependencies. Then, the activation function is used for nonlinear transformation. The convolution operation formula is: (f*g)[i]=∑ k f[k]·g[k-i] Among them, f is the input data, g is the convolution kernel, and * represents the convolution operation; The ReLu activation function formula is: ReLu=max(0,x) Use the maximum pooling operation to reduce feature dimensions and compress the number of parameters, effectively prevent overfitting, filter out the most representative information, ignore unimportant noise, and enhance the generalization and robustness of the model; The function prototype of the maximum pooling output is: The fully connected layer is used to merge the information of the previous layers. The output of the pooling layer is sent to one or more fully connected layers. Each neuron in these layers is connected to the activation value of the previous layer. A linear transformation is performed in the fully connected layer. The formula is: y=σ(ω*x+b) Among them, y is the output, ω is the weight matrix, b is the bias term, and σ is the activation function; The last layer is the regression layer, which first calculates the original prediction value of each category through the fully connected layer. Then, these unprocessed output raw prediction values ​​are passed to the softmax activation function to calculate the predicted probability distribution of each possible category. On this basis, the category with the highest probability is selected as its final prediction output. The formula is: Among them, x is the input of the fully connected layer, W T is the weight, b is the bias term, is the probability of softmax output, and softmax is calculated as follows: Among them, z i is the i-th element of the fully connected layer output vector, corresponding to the unnormalized probability of the i-th category; Use the gradient descent optimization algorithm to adjust the weights and biases of the network to minimize the loss function and improve the parameter explosion problem; An adaptive moment estimation optimizer is used for model training. The initial learning rate is set to 0.

005. The learning rate and batch size are dynamically adjusted during the iteration process in order to accelerate the training process and improve efficiency.

7. The method for predicting the dielectric properties of modified perovskite based on a deep learning integrated model according to claim 1, characterized in that: In step S4, the dielectric property prediction of the perovskite material is intuitively presented in a comprehensive display and analysis platform, allowing users to monitor, evaluate and deeply analyze data in real time, generate customized reports, and perform data management and anomaly detection functions on the prediction results to ensure the accuracy of the prediction information and the responsiveness of the model.

Images