Dielectric material screening method based on machine learning

Abstract

The invention discloses a dielectric material screening method based on machine learning, and the method comprises the steps: inputting a to-be-screened ferroelectric hysteresis loop image into a trained classification model, and obtaining a corresponding classification result; the training steps of the classification model are as follows: performing data enhancement and preprocessing on acquired historical ferroelectric hysteresis loop image data to obtain a training gray level image data set and a test gray level image data set; performing feature extraction and tagging processing on the two data sets to obtain a corresponding training set and a test set; performing hyper-parameter optimization on a set machine learning classifier by adopting a grid search and cross validation strategy to obtain an optimal hyper-parameter of the machine learning classifier; and based on the optimal hyper-parameter, performing model training according to the training set to obtain an optimal machine learning model as a classification model, and performing testing according to the test set. The method not only realizes rapid and accurate screening of dielectric material types, but also can strain ferroelectric hysteresis loop images with different image qualities and drawing styles.

Description

Technical Field

[0001] The present invention relates to the technical field of dielectric materials, and more specifically to a dielectric material screening method based on machine learning. Background Art

[0002] Polarization refers to the physical phenomenon that the centers of positive and negative charges do not overlap inside a material. This phenomenon often occurs inside a dielectric material when it is exposed to an applied electric field. Because of the polarization behavior exhibited under an applied electric field, it may have dielectric, piezoelectric, pyroelectric, and ferroelectric properties, and these properties can have broad application prospects in many fields such as dielectric capacitors, transformers, infrared detectors, and memory, making it one of the most important functional materials.

[0003] Initially, researchers explored dielectric materials mainly through scientific experience and "trial and error", but this process often requires a lot of experiments and time costs. In 2011, people were inspired by the Materials Genome Project and combined experiments, theories, and calculations through high-throughput methods. A large number of samples can be prepared and characterized in a short period of time, generating a huge database of material properties. Among them, the hysteresis loop shows the changing relationship between the polarization intensity of the material and the applied electric field. According to this changing relationship, dielectric materials can be divided into: linear dielectrics, ferroelectric materials, relaxor ferroelectric materials, and antiferroelectric materials. Each dielectric material has its corresponding hysteresis loop, similar to a fingerprint.

[0004] However, as the field of materials enters the era of big data, the amount of data that needs to be analyzed is also increasing significantly. Continuing to use traditional manual methods to extract material information requires a lot of time and labor costs. In addition, for real signals and artifacts in the database, professionals in this field usually rely on scientific experience or other measurement methods to assist in confirming whether they are real signals. This may lead to misjudgment for professionals in other fields.

[0005] In summary, how to quickly and accurately identify the type of hysteresis loop, reduce the time and labor costs consumed by people in analyzing massive data, and distinguish the real signal and artifacts in the hysteresis loop are problems that technical personnel in this field urgently need to solve. Summary of the invention

[0006] In view of the above problems, the main purpose of the present invention is to provide a dielectric material screening method based on machine learning to at least solve some of the technical problems mentioned in the above background technology.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A dielectric material screening method based on machine learning, comprising:

[0009] Inputting the hysteresis loop image to be screened into the trained classification model to obtain the classification result corresponding to the hysteresis loop image to be screened;

[0010] The training steps of the classification model include:

[0011] Acquire historical hysteresis loop image data, perform data enhancement processing and preprocessing on the historical hysteresis loop image data, and obtain a training grayscale image data set and a test grayscale image data set;

[0012] Performing feature extraction and labeling processing on the training grayscale image data set and the test grayscale image data set to obtain corresponding training sets and test sets;

[0013] Set up a machine learning classifier and use grid search and cross-validation strategies to optimize the hyperparameters of the machine learning classifier to obtain the optimal hyperparameters of the machine learning classifier;

[0014] Based on the optimal hyperparameters, model training is performed according to the training set to obtain an optimal machine learning model as a classification model; and the trained classification model is tested according to the test set.

[0015] Furthermore, the obtaining of historical hysteresis loop image data specifically includes:

[0016] Obtain historical hysteresis loop experimental data and hysteresis loop literature data;

[0017] The historical hysteresis loop experimental data and the hysteresis loop literature data are analyzed and screened according to material properties to obtain historical hysteresis loop image data. The categories of the historical hysteresis loop image data include non-ferroelectrics, linear dielectrics, antiferroelectric materials, relaxor ferroelectric materials and ferroelectric materials.

[0018] Furthermore, the data enhancement processing specifically includes:

[0019] A data enhancement process is formulated based on five conversion methods, namely, geometric transformation, mean filtering, noise injection, color processing, and random occlusion, and data enhancement processing is performed on the historical hysteresis loop image data to obtain hysteresis loop image enhanced data.

[0020] Furthermore, the preprocessing specifically includes:

[0021] Combination processing: combining the historical hysteresis loop image data and the hysteresis loop image enhancement data to obtain a hysteresis loop image data set;

[0022] Data division processing: dividing the hysteresis loop image data set according to a preset ratio to obtain a training image data set and a test image data set;

[0023] Image preprocessing: The training image dataset and the test image dataset are both subjected to image grayscale conversion and image size unification operations to achieve the conversion of the RGB three-channel hysteresis loop image to a single-channel grayscale image and the unification of the image size, thereby obtaining a training grayscale image dataset and a test grayscale image dataset.

[0024] Furthermore, in the image grayscale operation, the weighted average calculation formula of the three RGB components is as follows:

[0025] L=0.299R+0.587G+0.114B

[0026] Among them, L is the pixel value of the single-channel grayscale image; R is the pixel value of the hysteresis loop image of the RGB three-channel under the R channel; G is the pixel value of the hysteresis loop image of the RGB three-channel under the G channel; B is the pixel value of the hysteresis loop image of the RGB three-channel under the B channel.

[0027] Furthermore, the feature extraction process specifically includes:

[0028] A pixel reading operation is performed on each image in the training grayscale image dataset and the testing grayscale image dataset to obtain a training feature value dataset and a testing feature value dataset corresponding to the training grayscale image dataset and the testing grayscale image dataset.

[0029] Furthermore, the labeling process specifically includes:

[0030] The eigenvalues ​​in the training eigenvalue data set and the test eigenvalue data set are labeled with labels of the hysteresis loop image categories to which they belong, so as to obtain corresponding training sets and test sets.

[0031] Furthermore, the machine learning classifier is a convolutional neural network classifier.

[0032] Furthermore, the grid search and cross-validation strategies are used to optimize the hyperparameters of the machine learning classifier to obtain the optimal hyperparameters of the machine learning classifier, specifically including:

[0033] Setting multiple hyperparameter combinations, and repeatedly performing model training multiple times using the training set under each set of hyperparameters to obtain multiple initial models under each set of hyperparameters;

[0034] During each round of training, s% of the data in the training set is used to train the model, and (100-s)% of the data is used to estimate the ability of the model in identifying the types of hysteresis loops;

[0035] Finally, the average performance of multiple initial models under each set of hyperparameters is used as the evaluation index to obtain the optimal hyperparameters of the machine learning classifier.

[0036] Furthermore, the range of the variable s is specifically 70≤s≤80.

[0037] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a method for screening dielectric materials based on machine learning, which specifically includes the following beneficial effects:

[0038] (1) The present invention can quickly identify the type of hysteresis loop, reduce the time and labor costs consumed by people in analyzing massive data, and greatly improve the efficiency of people in analyzing data.

[0039] (2) The present invention can accurately distinguish between real signals and artifacts in the hysteresis loop, so that people in other fields with less professional knowledge can also extract effective information from the hysteresis loop, and has strong practicality.

[0040] (3) The present invention can adapt the hysteresis loop image effects under different image qualities and drawing styles, as well as the changes of similar hysteresis loop images under different backgrounds, and has strong versatility.

[0041] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0043] Figure 1 A schematic diagram of the classification model training steps provided in an embodiment of the present invention.

[0044] Figure 2 A schematic diagram of the prediction performance evaluation of the K-nearest neighbor (KNN) model obtained in Example 1 provided in an embodiment of the present invention.

[0045] Figure 3 A schematic diagram of the prediction performance evaluation of the decision tree (DT) model obtained in Example 2 provided in an embodiment of the present invention.

[0046] Figure 4 A schematic diagram of the prediction performance evaluation of the support vector machine (SVM) model obtained in Example 3 provided in an embodiment of the present invention.

[0047] Figure 5 A schematic diagram of the prediction performance evaluation of the convolutional neural network (CNN) model obtained in Example 4 provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] Embodiment 1:

[0050] Embodiment 1 of the present invention discloses a method for screening dielectric materials based on machine learning, comprising: inputting an image of a hysteresis loop to be screened into a trained classification model to obtain a classification result corresponding to the image of the hysteresis loop to be screened;

[0051] Among them, the training steps of the classification model are as follows Figure 1 As shown, the specific steps include:

[0052] S1. Acquire historical hysteresis loop image data, perform data enhancement processing and preprocessing on the historical hysteresis loop image data, and obtain a training grayscale image data set and a test grayscale image data set;

[0053] S2, performing feature extraction and labeling processing on the training grayscale image data set and the test grayscale image data set to obtain corresponding training sets and test sets;

[0054] S3. Set a machine learning classifier and use grid search and cross-validation strategies to optimize the hyperparameters of the machine learning classifier to obtain the optimal hyperparameters of the machine learning classifier;

[0055] S4. Based on the optimal hyperparameters, the model is trained according to the training set to obtain the optimal machine learning model as the classification model; and the trained classification model is tested according to the test set.

[0056] Next, each training step of the above classification model is explained respectively.

[0057] In the above step S1, the following contents are specifically included:

[0058] (1) Obtaining historical hysteresis loop image data, including:

[0059] Obtain historical hysteresis loop experimental data and hysteresis loop literature data;

[0060] According to the material properties, the historical hysteresis loop experimental data and the hysteresis loop literature data are analyzed and screened to obtain the historical hysteresis loop image data; the categories of the historical hysteresis loop image data include non-ferroelectric, linear dielectric, antiferroelectric material, relaxor ferroelectric material and ferroelectric material;

[0061] Among them, the hysteresis loop belonging to the non-ferroelectric category is an artifact that exists due to non-ferroelectric factors. By obtaining a training set based on historical hysteresis loop image data containing non-ferroelectrics, the trained classification model can accurately distinguish the categories of real signals and artifacts; in addition, since the historical hysteresis loop image data contains hysteresis loop images with different image qualities and drawing styles, it is helpful for the trained classification model to classify hysteresis loop images with different image qualities and drawing styles.

[0062] (2) Perform data enhancement processing, including:

[0063] A data enhancement process is formulated based on five conversion methods, namely, geometric transformation, mean filtering, noise injection, color processing, and random occlusion, and data enhancement processing is performed on the historical hysteresis loop image data to obtain hysteresis loop image enhanced data.

[0064] (3) Preprocessing, including:

[0065] Combined processing: The historical hysteresis loop image data and the hysteresis loop image enhancement data are combined to obtain the hysteresis loop image dataset; the classification model trained based on this can further transform the hysteresis loop image effects under different image qualities and drawing styles, as well as the changes of similar hysteresis loop images under different backgrounds;

[0066] Data partitioning processing: the hysteresis loop image dataset is divided according to a preset ratio to obtain a training image dataset and a test image dataset; that is, the hysteresis loop image dataset is partitioned by stratified sampling. For example, 80% of the data in the hysteresis loop image dataset is randomly divided into a training image dataset, and 20% of the data is randomly divided into a test image dataset, thereby ensuring the randomness and class balance of the data in the training image dataset and the test image dataset;

[0067] Image preprocessing: For both the training image dataset and the test image dataset, image grayscale conversion and image size unification operations are performed to achieve the conversion of RGB three-channel hysteresis loop images to single-channel grayscale images and unify the image sizes, thus obtaining the training grayscale image dataset and the test grayscale image dataset;

[0068] Among them, in the image grayscale operation, the weighted average calculation formula of the three components of RGB is as follows:

[0069] L=0.299R+0.587G+0.114B

[0070] Where, L is the pixel value of the single-channel grayscale image; R is the pixel value of the hysteresis loop image of the RGB three-channel under the R channel; G is the pixel value of the hysteresis loop image of the RGB three-channel under the G channel; B is the pixel value of the hysteresis loop image of the RGB three-channel under the B channel;

[0071] In the image size unification operation, the pixel size of the image is unified to 224×224.

[0072] In the above step S2, the following contents are specifically included:

[0073] (1) performing feature extraction processing on the training grayscale image dataset and the test grayscale image dataset, specifically including: performing a pixel reading operation on each image in the training grayscale image dataset and the test grayscale image dataset to obtain a training feature value dataset and a test feature value dataset corresponding to the training grayscale image dataset and the test grayscale image dataset.

[0074] (2) Labeling the training eigenvalue dataset and the test eigenvalue dataset, specifically including: labeling the eigenvalues ​​in the training eigenvalue dataset and the test eigenvalue dataset with the hysteresis loop image category to which they belong, to obtain the corresponding training set and test set.

[0075] In the above step S3, the following contents are specifically included:

[0076] (1) Setting a machine learning classifier, such as a K-nearest neighbor (KNN) classifier;

[0077] The specific settings of the KNN classifier are: n_neighbors (the number of nearest neighbors to be selected when predicting the sample category) = 7, weights (weights) = 'uniform', metric (distance metric) = 'euclidean'.

[0078] (2) The grid search and cross-validation strategies are used to optimize the hyperparameters of the KNN classifier to obtain the optimal hyperparameters of the KNN classifier, including:

[0079] Set multiple hyperparameter combinations, and repeat model training multiple times with the training set under each set of hyperparameters to obtain multiple initial models under each set of hyperparameters; for example, under each set of hyperparameters, obtain 5 initial models by repeatedly using the training set 5 times;

[0080] In each round of training, s% of the data in the training set is used to train the model, and (100-s)% of the data is used to estimate the prediction performance of the model; where the range of the variable s is specifically 70≤s≤80. For example, each time 70% of the data in the training set is randomly used to train the model, and 30% of the data is used to estimate the model's ability to identify hysteresis loop categories.

[0081] Finally, the average performance of multiple initial models under each set of hyperparameters is used as the evaluation indicator (for example, the average performance of the 5 initial models under each set of hyperparameters is used as the evaluation indicator), and the optimal hyperparameters of the KNN classifier are obtained: n_neighbors (the number of nearest neighbors required to predict the sample category) = 11, weights (weight) = 'distance', metric (distance metric) = 'manhattan'.

[0082] In the above step S4, based on the optimal hyperparameters obtained above, model training is performed according to the training set to obtain the optimal KNN model as the classification model; and the trained classification model is tested according to the test set;

[0083] The model training specifically includes: setting the hyperparameters of the KNN classifier to the optimal hyperparameters in step S3, that is, n_neighbors (the number of nearest neighbors to be selected when predicting the sample category) = 11, weights (weights) = 'distance', metric (distance metric) = 'manhattan', and then training the model based on the training set to obtain the optimal KNN model.

[0084] The model prediction specifically includes: inputting the test feature value data set in the test set into the result of model training, that is, the optimal KNN model, to obtain the classification result of the test set.

[0085] like Figure 2 The figure shows the prediction performance evaluation diagram of the KNN model in Example 1. Figure 2 Figure (a) shows the confusion matrix obtained by comparing the predicted labels and true labels output by the KNN model. Obviously, through calculation, it is known that the accuracy of the KNN model is 66.6667%; Figure 2 Figure (b) shows the receiver operating characteristic (ROC) curve and the corresponding area under the curve (AUC) of the KNN model on the test set. The macro-average AUC value presented by the KNN model is 0.8804, which is far from the ideal AUC value (AUC = 1); the lower accuracy and smaller macro-average AUC value indicate that the KNN model has poor predictive performance in identifying hysteresis loop categories.

[0086] Embodiment 2:

[0087] Embodiment 2 of the present invention discloses a method for screening dielectric materials based on machine learning, comprising: inputting an image of a hysteresis loop to be screened into a trained classification model to obtain a classification result corresponding to the image of the hysteresis loop to be screened;

[0088] Among them, the training steps of the classification model specifically include:

[0089] S1. Acquire historical hysteresis loop image data, perform data enhancement processing and preprocessing on the historical hysteresis loop image data, and obtain a training grayscale image data set and a test grayscale image data set;

[0090] S2, performing feature extraction and labeling processing on the training grayscale image data set and the test grayscale image data set to obtain corresponding training sets and test sets;

[0091] S3. Set a machine learning classifier and use grid search and cross-validation strategies to optimize the hyperparameters of the machine learning classifier to obtain the optimal hyperparameters of the machine learning classifier;

[0092] S4. Based on the optimal hyperparameters, the model is trained according to the training set to obtain the optimal machine learning model as the classification model; and the trained classification model is tested according to the test set.

[0093] Next, each training step of the above classification model is explained respectively.

[0094] In the above step S1, the following contents are specifically included:

[0095] (1) Obtaining historical hysteresis loop image data, including:

[0096] Obtain historical hysteresis loop experimental data and hysteresis loop literature data;

[0097] According to the material properties, the historical hysteresis loop experimental data and the hysteresis loop literature data are analyzed and screened to obtain the historical hysteresis loop image data; the categories of the historical hysteresis loop image data include non-ferroelectrics, linear dielectrics, antiferroelectric materials, relaxor ferroelectric materials and ferroelectric materials.

[0098] (2) Perform data enhancement processing, including:

[0099] A data enhancement process is formulated based on five conversion methods, namely, geometric transformation, mean filtering, noise injection, color processing, and random occlusion, and data enhancement processing is performed on the historical hysteresis loop image data to obtain hysteresis loop image enhanced data.

[0100] (3) Preprocessing, including:

[0101] Combination processing: combining the historical hysteresis loop image data and the hysteresis loop image enhancement data to obtain the hysteresis loop image data set;

[0102] Data partitioning processing: the hysteresis loop image dataset is divided according to a preset ratio to obtain a training image dataset and a test image dataset; that is, the hysteresis loop image dataset is partitioned by stratified sampling. For example, 80% of the data in the hysteresis loop image dataset is randomly divided into a training image dataset, and 20% of the data is randomly divided into a test image dataset, thereby ensuring the randomness and class balance of the data in the training image dataset and the test image dataset;

[0103] Image preprocessing: For both the training image dataset and the test image dataset, image grayscale conversion and image size unification operations are performed to achieve the conversion of RGB three-channel hysteresis loop images to single-channel grayscale images and unify the image sizes, thus obtaining the training grayscale image dataset and the test grayscale image dataset;

[0104] Among them, in the image grayscale operation, the weighted average calculation formula of the three components of RGB is as follows:

[0105] L=0.299R+0.587G+0.114B

[0106] Where, L is the pixel value of the single-channel grayscale image; R is the pixel value of the hysteresis loop image of the RGB three-channel under the R channel; G is the pixel value of the hysteresis loop image of the RGB three-channel under the G channel; B is the pixel value of the hysteresis loop image of the RGB three-channel under the B channel;

[0107] In the image size unification operation, the pixel size of the image is unified to 224×224.

[0108] In the above step S2, the following contents are specifically included:

[0109] (1) performing feature extraction processing on the training grayscale image dataset and the test grayscale image dataset, specifically including: performing a pixel reading operation on each image in the training grayscale image dataset and the test grayscale image dataset to obtain a training feature value dataset and a test feature value dataset corresponding to the training grayscale image dataset and the test grayscale image dataset.

[0110] (2) Labeling the training eigenvalue dataset and the test eigenvalue dataset, specifically including: labeling the eigenvalues ​​in the training eigenvalue dataset and the test eigenvalue dataset with the hysteresis loop image category to which they belong, to obtain the corresponding training set and test set.

[0111] In the above step S3, the following contents are specifically included:

[0112] (1) Setting a machine learning classifier, such as a decision tree (DT) classifier;

[0113] The specific settings of the DT classifier are: max_features (maximum number of features) = 0.5, max_leaf_nodes (maximum number of leaf nodes) = 4.

[0114] (2) The grid search and cross-validation strategies are used to optimize the hyperparameters of the DT classifier to obtain the optimal hyperparameters of the DT classifier, including:

[0115] Set multiple hyperparameter combinations, and repeat model training multiple times with the training set under each set of hyperparameters to obtain multiple initial models under each set of hyperparameters; for example, under each set of hyperparameters, obtain 5 initial models by repeatedly using the training set 5 times;

[0116] In each round of training, s% of the data in the training set is used to train the model, and (100-s)% of the data is used to estimate the prediction performance of the model; where the range of the variable s is specifically 70≤s≤80. For example, each time 70% of the data in the training set is randomly used to train the model, and 30% of the data is used to estimate the model's ability to identify hysteresis loop categories.

[0117] Finally, the average performance of multiple initial models under each set of hyperparameters is used as the evaluation indicator (for example, the average performance of the 5 initial models under each set of hyperparameters is used as the evaluation indicator), and the optimal hyperparameters of the DT classifier are obtained: max_features (maximum number of features) = 0.9, max_leaf_nodes (maximum number of leaf nodes) = 25.

[0118] In the above step S4, based on the optimal hyperparameters obtained above, model training is performed according to the training set to obtain the optimal DT model as the classification model; and the trained classification model is tested according to the test set;

[0119] Among them, model training specifically includes: setting the hyperparameters of the DT classifier to the optimal hyperparameters in step S3, that is, max_features (maximum number of features) = 0.9, max_leaf_nodes (maximum number of leaf nodes) = 25, and then training the model based on the training set to obtain the optimal DT model.

[0120] Among them, model prediction specifically includes: inputting the test feature value data set in the test set into the result of model training, that is, the optimal DT model, to obtain the classification result of the test set.

[0121] like Figure 3 The figure shows the prediction performance evaluation diagram of the DT model in Example 2. Figure 3Figure (a) shows the confusion matrix obtained by comparing the predicted labels and true labels output by the DT model. Obviously, through calculation, it is known that the accuracy of the DT model is 68.6667%; Figure 3 Figure (b) shows the receiver operating characteristic (ROC) curve and the corresponding area under the curve (AUC) of the DT model on the test set. The macro-average AUC value presented by the DT model is 0.8474, which is far from the ideal AUC value (AUC = 1); the lower accuracy and smaller macro-average AUC value indicate that the DT model has poor predictive performance in identifying hysteresis loop categories.

[0122] Example 3

[0123] Embodiment 3 of the present invention discloses a method for screening dielectric materials based on machine learning, comprising: inputting the hysteresis loop image to be screened into a trained classification model to obtain a classification result corresponding to the hysteresis loop image to be screened;

[0124] Among them, the training steps of the classification model specifically include:

[0125] S1. Acquire historical hysteresis loop image data, perform data enhancement processing and preprocessing on the historical hysteresis loop image data, and obtain a training grayscale image data set and a test grayscale image data set;

[0126] S2, performing feature extraction and labeling processing on the training grayscale image data set and the test grayscale image data set to obtain corresponding training sets and test sets;

[0127] S3. Set a machine learning classifier and use grid search and cross-validation strategies to optimize the hyperparameters of the machine learning classifier to obtain the optimal hyperparameters of the machine learning classifier;

[0128] S4. Based on the optimal hyperparameters, the model is trained according to the training set to obtain the optimal machine learning model as the classification model; and the trained classification model is tested according to the test set.

[0129] Next, each training step of the above classification model is explained respectively.

[0130] In the above step S1, the following contents are specifically included:

[0131] (1) Obtaining historical hysteresis loop image data, including:

[0132] Obtain historical hysteresis loop experimental data and hysteresis loop literature data;

[0133] According to the material properties, the historical hysteresis loop experimental data and the hysteresis loop literature data are analyzed and screened to obtain the historical hysteresis loop image data; the categories of the historical hysteresis loop image data include non-ferroelectrics, linear dielectrics, antiferroelectric materials, relaxor ferroelectric materials and ferroelectric materials.

[0134] (2) Perform data enhancement processing, including:

[0135] A data enhancement process is formulated based on five conversion methods, namely, geometric transformation, mean filtering, noise injection, color processing, and random occlusion, and data enhancement processing is performed on the historical hysteresis loop image data to obtain hysteresis loop image enhanced data.

[0136] (3) Preprocessing, including:

[0137] Combination processing: combining the historical hysteresis loop image data and the hysteresis loop image enhancement data to obtain the hysteresis loop image data set;

[0138] Data partitioning processing: the hysteresis loop image dataset is divided according to a preset ratio to obtain a training image dataset and a test image dataset; that is, the hysteresis loop image dataset is partitioned by stratified sampling. For example, 80% of the data in the hysteresis loop image dataset is randomly divided into a training image dataset, and 20% of the data is randomly divided into a test image dataset, thereby ensuring the randomness and class balance of the data in the training image dataset and the test image dataset;

[0139] Image preprocessing: For both the training image dataset and the test image dataset, image grayscale conversion and image size unification operations are performed to achieve the conversion of RGB three-channel hysteresis loop images to single-channel grayscale images and unify the image sizes, thus obtaining the training grayscale image dataset and the test grayscale image dataset;

[0140] Among them, in the image grayscale operation, the weighted average calculation formula of the three components of RGB is as follows:

[0141] L=0.299R+0.587G+0.114B

[0142] Where, L is the pixel value of the single-channel grayscale image; R is the pixel value of the hysteresis loop image of the RGB three-channel under the R channel; G is the pixel value of the hysteresis loop image of the RGB three-channel under the G channel; B is the pixel value of the hysteresis loop image of the RGB three-channel under the B channel;

[0143] In the image size unification operation, the pixel size of the image is unified to 224×224.

[0144] In the above step S2, the following contents are specifically included:

[0145] (1) performing feature extraction processing on the training grayscale image dataset and the test grayscale image dataset, specifically including: performing a pixel reading operation on each image in the training grayscale image dataset and the test grayscale image dataset to obtain a training feature value dataset and a test feature value dataset corresponding to the training grayscale image dataset and the test grayscale image dataset.

[0146] (2) Labeling the training eigenvalue dataset and the test eigenvalue dataset, specifically including: labeling the eigenvalues ​​in the training eigenvalue dataset and the test eigenvalue dataset with the hysteresis loop image category to which they belong, to obtain the corresponding training set and test set.

[0147] In the above step S3, the following contents are specifically included:

[0148] (1) Setting a machine learning classifier, such as a support vector machine (SVM) classifier;

[0149] The SVM classifier is specifically set as follows: kernel (kernel function type)='linear', C (penalty coefficient)=1.

[0150] (2) The grid search and cross-validation strategies are used to optimize the hyperparameters of the SVM classifier to obtain the optimal hyperparameters of the SVM classifier, including:

[0151] Set multiple hyperparameter combinations, and repeat model training multiple times with the training set under each set of hyperparameters to obtain multiple initial models under each set of hyperparameters; for example, under each set of hyperparameters, obtain 5 initial models by repeatedly using the training set 5 times;

[0152] In each round of training, s% of the data in the training set is used to train the model, and (100-s)% of the data is used to estimate the prediction performance of the model; where the range of the variable s is specifically 70≤s≤80. For example, each time 70% of the data in the training set is randomly used to train the model, and 30% of the data is used to estimate the model's ability to identify hysteresis loop categories.

[0153] Finally, the average performance of multiple initial models under each set of hyperparameters is used as the evaluation index (for example, the average performance of the five initial models under each set of hyperparameters is used as the evaluation index), and the optimal hyperparameters of the SVM classifier are obtained: kernel (kernel function type) = 'linear', C (penalty coefficient) = 4.

[0154] In the above step S4, based on the optimal hyperparameters obtained above, model training is performed according to the training set to obtain the optimal SVM model as the classification model; and the trained classification model is tested according to the test set;

[0155] The model training specifically includes: setting the hyperparameters of the SVM classifier to the optimal hyperparameters in step S3, namely kernel (kernel function type) = 'linear', C (penalty coefficient) = 4, and then training the model based on the training set to obtain the optimal SVM model.

[0156] The model prediction specifically includes: inputting the test feature value data set in the test set into the result of model training, that is, the optimal SVM model, to obtain the classification result of the test set.

[0157] like Figure 4 Shown is a diagram of the prediction performance evaluation of the SVM model in Example 3, Figure 4 Figure (a) shows the confusion matrix obtained by comparing the predicted labels and true labels output by the SVM model. Obviously, through calculation, it is known that the accuracy of the SVM model reaches 88.6667%; Figure 4 Figure (b) shows the receiver operating characteristic (ROC) curve and the corresponding area under the curve (AUC) of the SVM model on the test set. Obviously, the SVM model presents a larger macro-average AUC value (AUC = 0.9714); the higher accuracy and larger macro-average AUC value indicate that the SVM model has good predictive performance in identifying hysteresis loop categories.

[0158] Embodiment 4:

[0159] Embodiment 4 of the present invention discloses a method for screening dielectric materials based on machine learning, comprising: inputting the hysteresis loop image to be screened into a trained classification model to obtain a classification result corresponding to the hysteresis loop image to be screened;

[0160] Among them, the training steps of the classification model specifically include:

[0161] S1. Acquire historical hysteresis loop image data, perform data enhancement processing and preprocessing on the historical hysteresis loop image data, and obtain a training grayscale image data set and a test grayscale image data set;

[0162] S2, performing feature extraction and labeling processing on the training grayscale image data set and the test grayscale image data set to obtain corresponding training sets and test sets;

[0163] S3. Set a machine learning classifier and use grid search and cross-validation strategies to optimize the hyperparameters of the machine learning classifier to obtain the optimal hyperparameters of the machine learning classifier;

[0164] S4. Based on the optimal hyperparameters, the model is trained according to the training set to obtain the optimal machine learning model as the classification model; and the trained classification model is tested according to the test set.

[0165] Next, each training step of the above classification model is explained respectively.

[0166] In the above step S1, the following contents are specifically included:

[0167] (1) Obtaining historical hysteresis loop image data, including:

[0168] Obtain historical hysteresis loop experimental data and hysteresis loop literature data;

[0169] According to the material properties, the historical hysteresis loop experimental data and the hysteresis loop literature data are analyzed and screened to obtain the historical hysteresis loop image data; the categories of the historical hysteresis loop image data include non-ferroelectrics, linear dielectrics, antiferroelectric materials, relaxor ferroelectric materials and ferroelectric materials.

[0170] (2) Perform data enhancement processing, including:

[0171] A data enhancement process is formulated based on five conversion methods, namely, geometric transformation, mean filtering, noise injection, color processing, and random occlusion, and data enhancement processing is performed on the historical hysteresis loop image data to obtain hysteresis loop image enhanced data.

[0172] (3) Preprocessing, including:

[0173] Combination processing: combining the historical hysteresis loop image data and the hysteresis loop image enhancement data to obtain the hysteresis loop image data set;

[0174] Data partitioning processing: the hysteresis loop image dataset is divided according to a preset ratio to obtain a training image dataset and a test image dataset; that is, the hysteresis loop image dataset is partitioned by stratified sampling. For example, 80% of the data in the hysteresis loop image dataset is randomly divided into a training image dataset, and 20% of the data is randomly divided into a test image dataset, thereby ensuring the randomness and class balance of the data in the training image dataset and the test image dataset;

[0175] Image preprocessing: For both the training image dataset and the test image dataset, image grayscale conversion and image size unification operations are performed to achieve the conversion of RGB three-channel hysteresis loop images to single-channel grayscale images and unify the image sizes, thus obtaining the training grayscale image dataset and the test grayscale image dataset;

[0176] Among them, in the image grayscale operation, the weighted average calculation formula of the three components of RGB is as follows:

[0177] L=0.299R+0.587G+0.114B

[0178] Where, L is the pixel value of the single-channel grayscale image; R is the pixel value of the hysteresis loop image of the RGB three-channel under the R channel; G is the pixel value of the hysteresis loop image of the RGB three-channel under the G channel; B is the pixel value of the hysteresis loop image of the RGB three-channel under the B channel;

[0179] In the image size unification operation, the pixel size of the image is unified to 224×224.

[0180] In the above step S2, the following contents are specifically included:

[0181] (1) performing feature extraction processing on the training grayscale image dataset and the test grayscale image dataset, specifically including: performing a pixel reading operation on each image in the training grayscale image dataset and the test grayscale image dataset to obtain a training feature value dataset and a test feature value dataset corresponding to the training grayscale image dataset and the test grayscale image dataset.

[0182] (2) Labeling the training eigenvalue dataset and the test eigenvalue dataset, specifically including: labeling the eigenvalues ​​in the training eigenvalue dataset and the test eigenvalue dataset with the hysteresis loop image category to which they belong, to obtain the corresponding training set and test set.

[0183] In the above step S3, the following contents are specifically included:

[0184] (1) Setting a machine learning classifier, such as a convolutional neural network (CNN) classifier;

[0185] The specific settings of the CNN classifier are as follows: the initial architecture is the LeNet-5 network architecture, Epoch (number of iterations) = 50, batch_size (batch size) = 5, lr (learning rate) = 0.001, and the activation function is the S-type activation function.

[0186] (2) The grid search and cross-validation strategies are used to optimize the hyperparameters of the CNN classifier to obtain the optimal hyperparameters of the CNN classifier, including:

[0187] Set multiple hyperparameter combinations, and repeat model training multiple times with the training set under each set of hyperparameters to obtain multiple initial models under each set of hyperparameters; for example, under each set of hyperparameters, obtain 5 initial models by repeatedly using the training set 5 times;

[0188] In each round of training, s% of the data in the training set is used to train the model, and (100-s)% of the data is used to estimate the prediction performance of the model; where the range of the variable s is specifically 70≤s≤80. For example, each time 70% of the data in the training set is randomly used to train the model, and 30% of the data is used to estimate the model's ability to identify hysteresis loop categories.

[0189] Finally, the average performance of multiple initial models under each set of hyperparameters is used as the evaluation indicator (for example, the average performance of the 5 initial models under each set of hyperparameters is used as the evaluation indicator), and the optimal hyperparameters of the CNN classifier are obtained: based on the LeNet-5 network architecture, the activation function in the C1, C3, C5, and F6 layers is changed from the S-type activation function to the ReLu function, the convolution kernel step size in the C5 layer is changed from 1 to 4, and the Dropout layer is added after the C5 layer and the F6 layer respectively, Epoch (iteration number) = 36, batch_size (batch size) = 30, lr (learning rate) = 0.0001.

[0190] In the above step S4, based on the optimal hyperparameters obtained above, model training is performed according to the training set to obtain the optimal CNN model as the classification model; and the trained classification model is tested according to the test set;

[0191] Among them, the model training specifically includes: setting the hyperparameters of the CNN classifier to the optimal hyperparameters in step S3, that is, based on the LeNet-5 network architecture, changing the activation function in the C1, C3, C5, and F6 layers from the S-type activation function to the ReLu function, changing the convolution kernel step size in the C5 layer from 1 to 4, adding Dropout layers after the C5 layer and the F6 layer, Epoch (number of iterations) = 36, batch_size (batch size) = 30, lr (learning rate) = 0.0001, and then training the model based on the training set to obtain the optimal CNN model.

[0192] Among them, model prediction specifically includes: inputting the test feature value data set in the test set into the result of model training, that is, the optimal CNN model, to obtain the classification result of the test set.

[0193] like Figure 5 Shown is a diagram of the prediction performance evaluation of the CNN model in Example 4, Figure 5 Figure (a) shows the confusion matrix obtained by comparing the predicted labels and true labels output by the CNN model. Obviously, through calculation, it is known that the accuracy of the CNN model is as high as 95.3333%; Figure 5 Figure (b) shows the receiver operating characteristic (ROC) curve and the corresponding area under the curve (AUC) of the CNN model on the test set. Obviously, the macro-average AUC value (AUC=0.9902) presented by the CNN model is close to 1; the high accuracy and large macro-average AUC value indicate that the CNN model has excellent predictive performance in identifying the categories of hysteresis loops; further, it is calculated that the average time required for the CNN model to identify 150 hysteresis loops is 0.385s, which proves that the CNN model is efficient in identifying the categories of hysteresis loops.

[0194] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A dielectric material screening method based on machine learning, characterized in that: include: Inputting the hysteresis loop image to be screened into the trained classification model to obtain the classification result corresponding to the hysteresis loop image to be screened; The training steps of the classification model include: Acquire historical hysteresis loop image data, perform data enhancement processing and preprocessing on the historical hysteresis loop image data, and obtain a training grayscale image data set and a test grayscale image data set; Performing feature extraction and labeling processing on the training grayscale image data set and the test grayscale image data set to obtain corresponding training sets and test sets; Set up a machine learning classifier and use grid search and cross-validation strategies to optimize the hyperparameters of the machine learning classifier to obtain the optimal hyperparameters of the machine learning classifier; Based on the optimal hyperparameters, model training is performed according to the training set to obtain an optimal machine learning model as a classification model; and the trained classification model is tested according to the test set.

2. The method for screening dielectric materials based on machine learning according to claim 1, characterized in that: The obtaining of historical hysteresis loop image data specifically includes: Obtain historical hysteresis loop experimental data and hysteresis loop literature data; The historical hysteresis loop experimental data and the hysteresis loop literature data are analyzed and screened according to material properties to obtain historical hysteresis loop image data; the categories of the historical hysteresis loop image data include non-ferroelectrics, linear dielectrics, antiferroelectric materials, relaxor ferroelectric materials and ferroelectric materials.

3. The method for screening dielectric materials based on machine learning according to claim 1, characterized in that: The data enhancement processing specifically includes: A data enhancement process is formulated based on five conversion methods, namely, geometric transformation, mean filtering, noise injection, color processing, and random occlusion, and data enhancement processing is performed on the historical hysteresis loop image data to obtain hysteresis loop image enhanced data.

4. The method for screening dielectric materials based on machine learning according to claim 3, characterized in that: The preprocessing specifically includes: Combination processing: combining the historical hysteresis loop image data and the hysteresis loop image enhancement data to obtain a hysteresis loop image data set; Data division processing: dividing the hysteresis loop image data set according to a preset ratio to obtain a training image data set and a test image data set; Image preprocessing: The training image dataset and the test image dataset are both subjected to image grayscale conversion and image size unification operations to achieve the conversion of the RGB three-channel hysteresis loop image to a single-channel grayscale image and the unification of the image size, thereby obtaining a training grayscale image dataset and a test grayscale image dataset.

5. The method for screening dielectric materials based on machine learning according to claim 1, characterized in that: The feature extraction process specifically includes: A pixel reading operation is performed on each image in the training grayscale image dataset and the testing grayscale image dataset to obtain a training feature value dataset and a testing feature value dataset corresponding to the training grayscale image dataset and the testing grayscale image dataset.

6. The method for screening dielectric materials based on machine learning according to claim 5, characterized in that: The labeling process specifically includes: The eigenvalues ​​in the training eigenvalue data set and the test eigenvalue data set are labeled with labels of the hysteresis loop image categories to which they belong, so as to obtain corresponding training sets and test sets.

7. The method for screening dielectric materials based on machine learning according to claim 1, characterized in that: The machine learning classifier is a convolutional neural network classifier.

8. The method for screening dielectric materials based on machine learning according to claim 1, characterized in that: The grid search and cross-validation strategies are used to optimize the hyperparameters of the machine learning classifier to obtain the optimal hyperparameters of the machine learning classifier, specifically including: Setting multiple hyperparameter combinations, and repeatedly performing model training multiple times using the training set under each set of hyperparameters to obtain multiple initial models under each set of hyperparameters; During each round of training, s% of the data in the training set is used to train the model, and (100-s)% of the data is used to estimate the ability of the model in identifying the types of hysteresis loops; Finally, the average performance of multiple initial models under each set of hyperparameters is used as the evaluation index to obtain the optimal hyperparameters of the machine learning classifier.

9. The method for screening dielectric materials based on machine learning according to claim 8, characterized in that: The specific range of the variable s is 70≤s≤80.

Images