A material data set screening method and system based on active learning

Abstract

The present invention relates to a material data set screening method and system based on active learning, wherein the expansion method comprises: extracting different material data from an undetected data pool to expand an initial training data pool, obtaining multiple expanded training data pools and multiple trained baseline models as preliminary screening results; from the preliminary screening results, selecting the expanded training data pool corresponding to the trained baseline model with the best performance parameters and the trained baseline model with the best performance as the data set screening result. Based on the idea of ​​active learning, the present invention uses different types of expanded data for model training, and determines the quality of the expanded data based on the model training results, thereby achieving accurate screening of reliable data in literature data to expand experimental data, thereby improving the performance of the trained performance prediction model.

Description

Technical Field

[0001] The present invention relates to the technical field of new material research and development, and in particular to a material data set screening method and system based on active learning. Background Art

[0002] In materials science research, it is often necessary to predict the performance of newly designed materials (such as negative thermal expansion coefficient, etc.) to evaluate the practical application value of the material. In the process of such performance prediction in the field of materials science, the experiment of each new material is very expensive and it is difficult to obtain real experimental data. At present, machine learning methods can be used to predict the performance of unknown materials by fitting existing data, providing guidance for materials science. In order to achieve the amount of data required by machine learning models, the common method is to obtain it from literature data, which will inevitably lead to unreliable data. The traditional machine learning processing method is to clean the literature data, delete the points whose distribution is far different from the real experimental data, and then use the machine learning model to adjust the parameters for fitting, but such traditional methods generally show poor results. How to screen reliable data from literature data to expand experimental data and then improve the performance of the trained performance prediction model has become a technical problem that needs to be solved urgently. Summary of the invention

[0003] In view of this, the present invention provides a material data set screening method and system based on active learning, so as to accurately screen reliable data in literature data, expand experimental data, and further improve the performance of the trained performance prediction model.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] A material data set screening method based on active learning, the screening method comprising the following steps:

[0006] Extracting different material data from an untested data pool to expand the initial training data pool, and obtaining multiple expanded training data pools and multiple trained baseline models as preliminary screening results; the trained baseline model is obtained by training the baseline model using the expanded training data pool;

[0007] From the preliminary screening results, the trained baseline model with the best performance parameters and the expanded training data pool corresponding to the trained baseline model with the best performance are selected as the data set screening results.

[0008] Optionally, the extracting of different material data from the undetected data pool expands the initial training data pool to obtain multiple expanded training data pools and multiple trained baseline models as preliminary screening results; the trained baseline model is obtained by training the baseline model using the expanded training data pool, specifically including:

[0009] The sliding window j selects material data in the undetected data pool to obtain multiple data candidate blocks, and adds the multiple data candidate blocks to the candidate set;

[0010] Each data candidate block in the candidate set is added to the training data pool after the m-th expansion, to obtain multiple training data pools after the m+1-th expansion;

[0011] The baseline model is trained using the training data pool obtained by the m+1th expansion to obtain the baseline model after the m+1th training, and the performance score of the baseline model after the m+1th training is calculated to obtain the performance score of the training data pool after the m+1th expansion after n data candidate blocks are added;

[0012] Selecting the expanded training data pool with the highest performance score as the training data pool obtained by the m+1th expansion, selecting the trained baseline model with the highest performance score as the baseline model obtained by the m+1th training, and deleting the data candidate blocks added in the expanded training data pool with the highest performance score from the candidate set;

[0013] Determine whether the number of times the performance score of the baseline model obtained through training continuously decreases exceeds a first preset threshold, and obtain a first determination result;

[0014] If the first judgment result indicates no, then the value of m is increased by 1, the candidate set is cleared, and the training data pool expansion submodule is called;

[0015] If the first judgment result indicates yes, then selecting the trained baseline model with the highest performance score, the expanded training data pool and the performance score corresponding to the trained baseline model with the highest performance score as the kth preliminary screening result corresponding to window j;

[0016] Determine whether the value of k is less than a second preset threshold K, and obtain a second determination result;

[0017] If the second judgment result indicates yes, the order of the material data in the undetected data pool is disrupted, the value of k is increased by 1, the value of m is set to 0, the candidate set is cleared, and the process returns to the step of "sliding window j selects material data in the undetected data pool to obtain multiple data candidate blocks, and adding the multiple data candidate blocks to the candidate set";

[0018] If the second judgment result indicates no, then judging whether the value of j is less than a third preset threshold value J, and obtaining a third judgment result;

[0019] If the third judgment result indicates yes, then update window j to obtain window j+1, increase the value of j by 1, set the values ​​of m and k to 0, clear the candidate set, and return to the step of "sliding window j to select material data in the undetected data pool to obtain multiple data candidate blocks, and add the multiple data candidate blocks to the candidate set";

[0020] If the third judgment result indicates no, K×J preliminary screening results are output.

[0021] Optionally, the sliding window j selects material data from an undetected data pool to obtain a plurality of data candidate blocks, and adds the plurality of data candidate blocks to a candidate set, and the method also includes:

[0022] The baseline model is preliminarily trained using the material data in the initial training data pool to obtain a preliminarily trained baseline model.

[0023] Optionally, the formula for calculating the performance score of the trained baseline model is:

[0024] Score=P1*score1+P2*score2+P3*score3;

[0025] score1=1-Relavent error;

[0026]

[0027]

[0028] Among them, Score represents the performance score of the trained baseline model, and score1, score2, and score3 represent the relative error performance index, divergence index, and R of the trained baseline model, respectively. 2 Performance indicators; P1, P2 and P3 represent the weights of score1, score2 and score3 respectively;

[0029] Relavent error means relative error.

[0030] n represents the number of material data in the screened training data pool, represents the predicted value of the performance parameter output by the trained baseline model corresponding to the i-th material data, y (i) represents the actual value of the performance parameter in the i-th material data;

[0031] JS(·) represents the probability distribution similarity calculation formula, KL(·) represents the divergence calculation formula, R 2 Indicates the correlation between the predicted result and the actual result. Represents the average value of the performance parameters in all material data in the screened training data pool.

[0032] Optionally, the method further includes selecting, from the preliminary screening results, a trained baseline model with the best performance parameters and an expanded training data pool corresponding to the trained baseline model with the best performance as a data set screening result, and then further including:

[0033] Construct a machine learning model of the trained baseline model with the best performance parameters as a material property prediction model;

[0034] Initializing the parameters of the material property prediction model based on the parameters of the baseline model obtained through training with the best performance parameters;

[0035] The initialized material property prediction model is trained using the expanded training data pool corresponding to the baseline model obtained through training with the best performance, so as to obtain a trained material property prediction model;

[0036] The trained material performance prediction model is used to predict the material performance parameters corresponding to the material data to be tested.

[0037] Optionally, the method further comprises extracting different material data from the undetected data pool to expand the initial training data pool, obtaining a plurality of expanded training data pools and a plurality of trained baseline models as preliminary screening results; the trained baseline model is obtained by training the baseline model using the expanded training data pool, and the method further comprises:

[0038] Using the initial training data pool to train different types of machine learning models respectively, to obtain multiple trained machine learning models;

[0039] The trained machine learning model with the smallest relative error is selected as the baseline model.

[0040] A material data set screening system based on active learning, the screening system comprising:

[0041] A training data pool expansion module is used to extract different material data from an undetected data pool to expand the initial training data pool, and obtain multiple expanded training data pools and multiple trained baseline models as preliminary screening results; the trained baseline model is obtained by training the baseline model using the expanded training data pool;

[0042] The expanded training data pool screening module is used to select the trained baseline model with the best performance parameters and the expanded training data pool corresponding to the trained baseline model with the best performance from the preliminary screening results as the data set screening result.

[0043] Optionally, the training data pool screening module specifically includes:

[0044] In the first material data selection submodule, the sliding window j selects material data in the undetected data pool to obtain a plurality of data candidate blocks, and adds the plurality of data candidate blocks to the candidate set;

[0045] The training data pool expansion submodule adds each data candidate block in the candidate set to the training data pool after the mth expansion, and obtains multiple training data pools after the m+1th expansion;

[0046] The baseline model training submodule uses the training data pool obtained by the m+1th expansion to train the baseline model, obtains the baseline model after the m+1th training, and calculates the performance score of the baseline model after the m+1th training, and obtains the performance score of the training data pool after the m+1th expansion after n data candidate blocks are added;

[0047] The data pool updating submodule selects the expanded training data pool with the highest performance score as the training data pool obtained by the m+1th expansion, selects the trained baseline model with the highest performance score as the baseline model obtained by the m+1th training, and deletes the data candidate blocks added in the expanded training data pool with the highest performance score from the candidate set;

[0048] A first judgment submodule is used to judge whether the number of times the performance score of the trained baseline model continuously decreases exceeds a first preset threshold, and obtain a first judgment result;

[0049] A second material data selection submodule, for increasing the value of m by 1, clearing the candidate set, and calling the training data pool expansion submodule if the first judgment result indicates no;

[0050] an expanded training data pool selection submodule, for selecting, if the first judgment result indicates yes, the trained baseline model with the highest performance score, the expanded training data pool and the performance score corresponding to the trained baseline model with the highest performance score as the kth preliminary screening result corresponding to window j;

[0051] A second judgment submodule is used to judge whether the value of k is less than a second preset threshold K, and obtain a second judgment result;

[0052] a first calling submodule, for, if the second judgment result indicates yes, disrupting the order of the material data in the undetected data pool, increasing the value of k by 1, setting the value of m to 0, clearing the candidate set, and calling the first material data selection submodule;

[0053] A third judgment submodule, configured to judge whether the value of j is less than a third preset threshold value J if the second judgment result indicates no, and obtain a third judgment result;

[0054] The second calling submodule is used for updating window j to obtain window j+1 if the third judgment result indicates yes, increasing the value of j by 1, setting the values ​​of m and k to 0, clearing the candidate set, and calling the first material data selection submodule;

[0055] The screening result output submodule is used to output K×J preliminary screening results if the third judgment result indicates no.

[0056] Optionally, the screening system further comprises:

[0057] The preliminary training submodule is used to perform preliminary training on the baseline model using the material data in the initial training data pool to obtain a baseline model after preliminary training.

[0058] Optionally, the formula for calculating the performance score of the trained baseline model is:

[0059] Score=P1*score1+P2*score2+P3*score3;

[0060] score1=1-Relavent error;

[0061]

[0062]

[0063] Among them, Score represents the performance score of the trained baseline model, and score1, score2, and score3 represent the relative error performance index, divergence index, and R of the trained baseline model, respectively. 2 Performance indicators; P1, P2 and P3 represent the weights of score1, score2 and score3 respectively;

[0064] Relavent error means relative error.

[0065] n represents the number of material data in the screened training data pool, represents the predicted value of the performance parameter output by the trained baseline model corresponding to the i-th material data, y(i) represents the actual value of the performance parameter in the i-th material data;

[0066] JS(·) represents the probability distribution similarity calculation formula, KL(·) represents the divergence calculation formula, R 2 Indicates the correlation between the predicted result and the actual result. Represents the average value of the performance parameters of all material data in the expanded training data pool.

[0067] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0068] The present invention discloses a material data set screening method and system based on active learning, the screening method comprising the following steps: extracting different material data from an undetected data pool to expand an initial training data pool, obtaining multiple expanded training data pools and multiple trained baseline models as preliminary screening results; the trained baseline model is obtained by training a baseline model using the expanded training data pool; from the preliminary screening results, selecting the trained baseline model with the best performance parameters and the expanded training data pool corresponding to the trained baseline model with the best performance as the data set screening result. Based on the idea of ​​active learning, the present invention uses different types of expanded data for model training, and determines the quality of the expanded data based on the model training results, thereby achieving accurate screening of reliable data in literature data to expand experimental data, thereby improving the performance of the trained performance prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] 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 will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0070] Figure 1 A flow chart of a material data set screening method based on active learning provided in Example 1 of the present invention;

[0071] Figure 2 A flow chart of modeling using an active learning framework to predict material properties provided in Example 2 of the present invention;

[0072] Figure 3 A flowchart of the active learning framework operation process provided in Example 3 of the present invention;

[0073] Figure 4 A schematic diagram of a material data set screening method based on active learning provided in Example 3 of the present invention;

[0074] Figure 5 This is a distribution diagram of the performance alpha of material A provided in Example 2 of the present invention; wherein, Figure 5 (a) is the histogram of the original data performance alpha. The blue curve represents the fitting of the numerical distribution, and the black line represents the fitting of the standard normal distribution on the distribution interval. The relationship between the real data and the standard normal distribution can be intuitively evaluated. Figure 5 (b) is a QQ plot, which shows the comparison between the quantiles of alpha and the quantiles of the normal distribution. If the data conforms to the normal distribution, all the points will fall on the straight line. DETAILED DESCRIPTION

[0075] 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.

[0076] The purpose of the present invention is to provide a material data set screening method and system based on active learning. In order to solve the problem that unlabeled or labeled but imprecise data may be abundant or easy to obtain in the material research and development process, but labeled data is difficult, time-consuming or expensive to obtain, a material property prediction method based on an active learning framework is provided.

[0077] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0078] At this time, the concept of active learning was introduced. The core idea of ​​active learning is that if the machine learning algorithm is allowed to choose the data it will learn, it can achieve higher accuracy with less labeled data. Active learning is a subfield of machine learning and a complete framework that can be expressed as A = (G, L, S, Q, U), where the input refers to the unlabeled sample pool (U), the screening strategy (Q), the annotation expert (S), the training data set (L), and the target model (G). First, the selected model is used to score the training data set, and then the data is taken from the unlabeled sample pool according to the query strategy and annotated by the annotation expert, and then the annotated samples are put into the training data set. For material data, in most cases, the label is known, so the annotation expert can be omitted, and the query strategy can be used to directly take out the data item that is best for the model, which is equivalent to a screening strategy. The entire framework can finally obtain the data set that works best on the target model, and after the model is trained on this data set, it can be used to predict the performance of other new materials.

[0079] The present invention uses the data obtained from real experiments as the initial sample set, i.e., the training data pool, and the uncertain and unreliable data collected from literature and the like as the sample pool, i.e., the undetected data pool, and uses SVR as the classifier in the active learning framework. In the active learning process, the sample pool order is disrupted multiple times and several data items are taken out according to the sliding window. The current optimal data item is given according to the query strategy, and the multiple data items that finally make the performance score reach the highest are found, thereby obtaining the optimal data set and the corresponding model.

[0080] In the framework of active learning, the present invention proposes a query strategy based on hybrid error reduction, and through detailed parameter adjustment, determines the best sliding window size on the data pool in each screening process, as well as the optimal parameter range of SVR in grid search. At the same time, this method is compared with other traditional methods of outlier screening combined with machine learning algorithms, proving that this method has extremely high stability and robustness, and has high practical guiding significance for material performance prediction.

[0081] The query strategy is a screening strategy in the active learning process, and is further divided into a stream-based query strategy and a pool-based query strategy according to the different amounts of query data obtained each time. Based on the common pool-based query strategy, the present invention modifies the performance score function to obtain a query strategy based on relative error reduction in order to meet the requirement that the absolute error between the predicted value and the true value in the material data should be as small as possible.

[0082] Example 1

[0083] like Figure 1 As shown, the present invention provides a material data set screening method based on active learning, and the expansion method comprises the following steps:

[0084] Step 101, extract different material data from the undetected data pool to expand the initial training data pool, and obtain multiple expanded training data pools and multiple trained baseline models as preliminary screening results; the trained baseline model is obtained by training the baseline model using the expanded training data pool.

[0085] Step 101, before that, also includes:

[0086] Rough material data processing: clear suspected erroneous data samples, and clear or process data items containing null values;

[0087] Material data transformation: according to the needs of subsequent modeling, the data is transformed into smoothing, normalization, standardization, discretization, etc.;

[0088] Baseline selection: Use multiple machine learning models such as random forest and support vector regression to fit the material data and use R 2, mean absolute error and other indicators to select the best performing machine learning model as the baseline model, in order to compare the effects of subsequent active learning models.

[0089] The specific implementation steps of step 101 are:

[0090] Construct active learning elements: Material property prediction is a regression task. At present, there is no library that can be summarized and called, so it is necessary to manually construct active learning elements.

[0091] Undetected data pool U: unreliable literature data, unverified data, etc. collected.

[0092] Training data pool L: consists of real and reliable experimental data.

[0093] Machine Learning Model G: Considering that a large number of parameter adjustments and experiments are required to screen data during each active learning process, the present invention selects multiple baseline models to ensure that the highest score can be obtained for different data. The baseline models used in this framework are shown in Table 1.

[0094] Table 1 Candidate machine learning models for active learning

[0095]

[0096] Query strategy Q: The query strategy of the present invention combines the perspective of material science and data analysis to ensure the generalization and accuracy of the model, calculate multiple evaluation indicators and integrate them. That is, one or more samples<x,y> Take out the training data set from the data pool U The condition is that the regressor G in the dataset The performance score obtained is the largest. The present invention includes the following indicators.

[0097] Relative error:

[0098] The goal of combining material performance prediction is to make the predicted value as close to the true value as possible. First, the relative error index is added.

[0099]

[0100] Where n represents the number of data samples, and i represents the result of the model on the i-th item.

[0101] Next, to facilitate the implementation of the algorithm, the original problem is further transformed into a maximization problem.

[0102] score1=1-Relavent error

[0103] JS Divergence:

[0104] From the perspective of data statistics, JS is used to measure the similarity of the probability distribution of real data and predicted data, with a value of 0 to 1. The larger the value, the closer the probability of the real data and the predicted data is. The calculation method is as follows.

[0105]

[0106] R2:

[0107] From the perspective of evaluating model G, when evaluating the model under different dimensions, R2 eliminates the influence of the dimension. Its calculation method is as follows. The closer the value is to 1, the better the model prediction effect.

[0108]

[0109] Based on the above three results, a performance score function is designed and calculated as follows, where P1 to P3 pass in different parameters according to the actual task.

[0110] Score=P1*score1+P2*score2+P3*score3

[0111] Running the framework: The process of running a round of the framework is to disrupt the data pool U, slide a sliding window of size a in the data pool, and calculate Score on the regressor, take out the data with the highest score and put it into the training data and delete it from the data pool U until the performance score continues to decrease while continuing to add data to the training data.

[0112] Parameter tuning: Set GridSearch to adjust the hyperparameters of SVR, repeat the framework multiple times with different a values, and finally get the parameters that stably obtain the highest score.

[0113] Step 102: From the preliminary screening results, select the trained baseline model with the best performance parameters and the expanded training data pool corresponding to the trained baseline model with the best performance as the data set screening results.

[0114] Example 2

[0115] Now let’s make a performance prediction for material A. The features constructed using this material include 19 items. The numerical information of these features is calculated. The information of some features is shown below, and the number, mean, standard deviation, minimum value, quartile, maximum value, etc. of the data are statistically analyzed, as shown in Table 2.

[0116] Table 2 Some characteristic statistics of material A

[0117]

[0118] The target distribution that needs to be predicted is as follows Figure 5 As shown:

[0119] from Figure 5 (a) can intuitively evaluate the relationship between the real data and the standard normal distribution. Figure 5 (b) is a QQ graph, which shows the comparison between the quantile of alpha and the quantile of the normal distribution. If the data conforms to the normal distribution, all the points will fall on the straight line. From the distribution of alpha, it can be seen that the distribution of this feature deviates far from the normal distribution, and it is difficult for traditional machine learning algorithms to fit.

[0120] See also Figure 2 to Figure 4 , the performance prediction method based on the active learning framework in the material data is used to process and predict the performance of material A.

[0121] Reference Figure 2 , a material data set screening method based on active learning, wherein the process of processing and modeling the original material data includes the following steps:

[0122] Step 1: Clean up the material data: remove data samples suspected of errors and abnormalities, and delete data items containing missing values; perform correlation analysis on the data, such as chi-square test, correlation coefficient calculation, covariance analysis, etc., to facilitate the baseline model to fit;

[0123] Step 2: Material data transformation: According to the needs of subsequent modeling, the data is transformed into smoothing, normalization, standardization, discretization, etc.;

[0124] Step 3: Baseline selection: Use multiple machine learning models such as random forest and support vector regression to fit the material data and calculate the relative errors of these models in order to compare the effects of subsequent deep learning models. The best baseline model at this time is used as a candidate model for the active learning framework.

[0125] Step 4: Build an active learning framework: Build the target model based on the active learning framework;

[0126] Step 5: Training model: Select the model from step 4, perform multiple rounds of training on it, and adjust the hyperparameters to achieve the best results.

[0127] Step 6. Save the model and predict the sample: Save the optimal data items and corresponding models obtained after multiple rounds of screening, and use them to predict new samples designed artificially, which will provide guidance for the construction of high-performance materials.

[0128] Example 3

[0129] Reference Figure 3 , a material data set screening method based on active learning of the present invention comprises the following steps:

[0130] Step 1: Input the cleaned material data: the data does not contain obviously erroneous data or data with blank items.

[0131] Step 2: Establish a basic regressor: Take out 6 real data obtained from the experiment and set them as the initial training set L. Use the regressor SVR to establish a basic regressor on the initial training set L to obtain the initial performance score.

[0132] Step 3: Reference Figure 4 , building an active learning framework requires:

[0133] Construct 125 document data as the undetected data pool U, and design a query strategy Q based on relative error reduction. Select the machine learning model SVR as the regressor G, build a sliding window of size a, and use the preprocessed data as the training data pool.

[0134] Step 4: Run the framework: Use a sliding window of size a to extract the data pair 〈x,y〉 from the sample pool U, set P1 to 1 in the query strategy Q, and set the others to 0, calculate the performance score of the regressor G in the data set L∪〈x,y〉, continuously update the parameters of the regressor G, and finally select the a item with the highest performance score in U and add it to the training set L. Stop when the performance score after screening starts to drop three times in a row, and save the highest performance score in history and the corresponding data set;

[0135] Step 5: Parameter tuning: Modify the range of the Grid Search parameter tuning in the active learning framework to determine the range where the optimal parameters appear densely. Adjust the sliding window size a in the range of (1, 12).

[0136] Step 6: Save the results. Shuffle the undetected data pool U, return to step 4, and execute 20 times in total, saving the data set with the highest performance score and the corresponding model each time;

[0137] Step 7: After the framework is completed, the current optimal data set and corresponding model are obtained, and used to predict the performance of unknown artificially designed material samples.

[0138] Example 4

[0139] The present invention also provides a material data set screening system based on active learning, the screening system comprising:

[0140] The training data pool expansion module is used to extract different material data from the undetected data pool to expand the initial training data pool, and obtain multiple expanded training data pools and multiple trained baseline models as preliminary screening results; the trained baseline model is obtained by training the baseline model using the expanded training data pool.

[0141] The expanded training data pool selection module is used to select the trained baseline model with the best performance parameters and the expanded training data pool corresponding to the trained baseline model with the best performance from the preliminary screening results as the data set screening result.

[0142] The training data pool expansion module specifically includes: a first material data selection submodule, in which a sliding window j selects material data from an undetected data pool to obtain a plurality of data candidate blocks, and adds the plurality of data candidate blocks to a candidate set; a training data pool expansion submodule, in which each data candidate block in the candidate set is added to the training data pool after the mth expansion, to obtain a plurality of training data pools after the m+1th expansion; a baseline model training submodule, in which a baseline model is trained using the training data pool obtained after the m+1th expansion to obtain a baseline model after the m+1th training, and a performance score of the baseline model after the m+1th training is calculated to obtain a performance score of the m+1th expanded training data pool after adding n data candidate blocks; a data pool update submodule, in which an expanded training data pool with the highest performance score is selected as the training data pool obtained after the m+1th expansion, a trained baseline model with the highest performance score is selected as the baseline model obtained after the m+1th training, and the data candidate block added to the expanded training data pool with the highest performance score is deleted from the candidate set;

[0143] The second material data selection submodule is used to increase the value of m by 1, clear the candidate set, and call the training data pool expansion submodule if the first judgment result indicates no; the expanded training data pool selection submodule is used to select the trained baseline model with the highest performance score, the expanded training data pool corresponding to the trained baseline model with the highest performance score, and the performance score as the kth preliminary screening result corresponding to window j if the first judgment result indicates yes; the second judgment submodule is used to determine whether the value of k is less than a second preset threshold K to obtain a second judgment result; the first calling submodule is used to disrupt the unchecked data pool if the second judgment result indicates yes The order of material data in the measured data pool is determined, and the value of k is increased by 1, the value of m is set to 0, the candidate set is cleared, and the first material data selection submodule is called; the third judgment submodule is used to judge whether the value of j is less than the third preset threshold J if the second judgment result indicates no, and obtain the third judgment result; the second calling submodule is used to update the window j if the third judgment result indicates yes, obtain the window j+1, increase the value of j by 1, set the values ​​of m and k to 0, clear the candidate set, and call the first material data selection submodule; the screening result output submodule is used to output K×J preliminary screening results if the third judgment result indicates no.

[0144] The screening system also includes: a preliminary training submodule, which is used to perform preliminary training on the baseline model using the material data in the initial training data pool to obtain a baseline model after preliminary training.

[0145] The formula for calculating the performance score of the trained baseline model is:

[0146] Score=P1*score1+P2*score2+P3*score3;

[0147] score1=1-Relavent error;

[0148]

[0149]

[0150] Among them, Score represents the performance score of the trained baseline model, and score1, score2, and score3 represent the relative error performance index, divergence index, and R of the trained baseline model, respectively. 2 Performance indicators; P1, P2 and P3 represent the weights of score1, score2 and score3 respectively;

[0151] Relavent error means relative error.

[0152] n represents the number of material data in the screened training data pool, represents the predicted value of the performance parameter output by the trained baseline model corresponding to the i-th material data, y (i) represents the actual value of the performance parameter in the i-th material data;

[0153] JS(·) represents the probability distribution similarity calculation formula, KL(·) represents the divergence calculation formula, R 2 Indicates the correlation between the predicted result and the actual result. Represents the average value of the performance parameters of all material data in the expanded training data pool.

[0154] Beneficial effects of the present invention:

[0155] The present invention provides a material data set screening method and system based on active learning. The method is based on an active learning framework, and its input includes an unlabeled sample pool (U), a screening strategy (Q), a training data set (L), and a target model (G), which is used to deal with the situation that there are few real experimental data samples that often appear in material data, but many unreliable data samples are obtained from literature and other channels. The Score function designed to meet the material science is used as a screening strategy, and a variety of baseline models are used as candidates for the target model. First, this method uses uncertain data as a sample pool and real experimental samples as an initial training set. Through active learning, the uncertain samples that improve the performance score of the regression model G most significantly are selected, and the best data set and regression model are obtained after the entire process is completed. Then, this method is used to predict the performance of artificially designed materials to guide real material experiments.

[0156] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0157] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A material dataset screening method based on active learning, It is characterized in that The expansion method comprises the following steps: Extracting different material data from the untested data pool to expand the initial training data pool, obtaining multiple expanded training data pools and multiple trained baseline models as preliminary screening results; the trained baseline model is obtained by training the baseline model using the expanded training data pool; the untested data pool includes collected unverified literature data; The initial training data pool includes experimental data; From the preliminary screening results, selecting the trained baseline model with the best performance parameters and the expanded training data pool corresponding to the trained baseline model with the best performance as the data set screening result; The extracting of different material data from the undetected data pool to expand the initial training data pool, and obtaining multiple expanded training data pools and multiple trained baseline models as preliminary screening results, specifically includes: The sliding window j selects material data in the undetected data pool to obtain a plurality of data candidate blocks, and adds the plurality of data candidate blocks to the candidate set; Each data candidate block in the candidate set is added to the training data pool after the mth expansion, to obtain multiple training data pools after the m+1th expansion; wherein the training data pool after the 0th expansion is the initial training data pool; The baseline model is trained using the training data pool obtained by the m+1th expansion to obtain the baseline model after the m+1th training, and the performance score of the baseline model after the m+1th training is calculated to obtain the performance score of the training data pool after the m+1th expansion after n data candidate blocks are added; Selecting the expanded training data pool with the highest performance score as the training data pool obtained by the m+1th expansion, selecting the trained baseline model with the highest performance score as the baseline model obtained by the m+1th training, and deleting the data candidate blocks added in the expanded training data pool with the highest performance score from the candidate set; Determine whether the number of times the performance score of the baseline model obtained through training continuously decreases exceeds a first preset threshold, and obtain a first determination result; If the first judgment result indicates no, then the value of m is increased by 1, the candidate set is cleared, and the training data pool expansion submodule is called; If the first judgment result indicates yes, then selecting the trained baseline model with the highest performance score, the expanded training data pool and the performance score corresponding to the trained baseline model with the highest performance score as the kth preliminary screening result corresponding to window j; Determine whether the value of k is less than a second preset threshold K, and obtain a second determination result; If the second judgment result indicates yes, the order of the material data in the undetected data pool is disrupted, the value of k is increased by 1, the value of m is set to 0, the candidate set is cleared, and the process returns to step "sliding window j selects material data in the undetected data pool to obtain multiple data candidate blocks, and adds the multiple data candidate blocks to the candidate set"; If the second judgment result indicates no, then judging whether the value of j is less than a third preset threshold value J, and obtaining a third judgment result; If the third judgment result indicates yes, then update window j to obtain window j+1, increase the value of j by 1, set the values ​​of m and k to 0, clear the candidate set, and return to step "sliding window j to select material data in the undetected data pool to obtain multiple data candidate blocks, and add the multiple data candidate blocks to the candidate set"; If the third judgment result indicates no, then output K×J The preliminary screening results.

2. The material dataset screening method based on active learning according to claim 1, It is characterized in that The sliding window j selects material data in the undetected data pool to obtain multiple data candidate blocks, and adds the multiple data candidate blocks to the candidate set, which also includes: The baseline model is preliminarily trained using the material data in the initial training data pool to obtain a preliminarily trained baseline model.

3. The material dataset screening method based on active learning according to claim 1, It is characterized in that The formula for calculating the performance score of the trained baseline model is: ScoreP 1 Come to SCore1+P 2 *score2+P3*SC01e3; score11-Relaventerror; Among them, Score represents the performance score of the trained baseline model, and score1, score2, and score3 represent the relative error performance index, divergence index, and R of the trained baseline model, respectively. 2 Performance index; P 1 , P 2 and P 3 Represent the weights of score1, score2 and score3 respectively; Relavent error means relative error. n represents the number of material data in the screened training data pool, represents the predicted value of the performance parameter output by the trained baseline model corresponding to the i-th material data, and y(i) represents the actual value of the performance parameter in the i-th material data; JS(·) represents the probability distribution similarity calculation formula, KL(·) represents the divergence calculation formula, R 2 Indicates the correlation between the predicted result and the actual result. Represents the average value of the performance parameters of all material data in the expanded training data pool.

4. The material dataset screening method based on active learning according to claim 1, It is characterized in that The method further comprises: selecting, from the preliminary screening results, a baseline model obtained through training with the best performance parameters and an expanded training data pool corresponding to the baseline model obtained through training with the best performance as a data set screening result, and then: Construct a machine learning model of the trained baseline model with the best performance parameters as a material property prediction model; Initializing the parameters of the material property prediction model based on the parameters of the baseline model obtained through training with the best performance parameters; The initialized material property prediction model is trained using the expanded training data pool corresponding to the baseline model obtained through training with the best performance, so as to obtain a trained material property prediction model; The trained material performance prediction model is used to predict the material performance parameters corresponding to the material data to be tested.

5. The material dataset screening method based on active learning according to claim 1, It is characterized in that The extracting different material data from the undetected data pool expands the initial training data pool to obtain multiple expanded training data pools and multiple trained baseline models as preliminary screening results; The trained baseline model is obtained by training the baseline model using the expanded training data pool, and the method also includes: Use the training data pool to train different types of machine learning models respectively to obtain multiple trained machine learning models; The trained machine learning model with the smallest relative error is selected as the baseline model.

6. A material dataset screening system based on active learning, It is characterized in that The screening system comprises: A training data pool expansion module is used to extract different material data from an untested data pool to expand the initial training data pool, and obtain multiple expanded training data pools and multiple trained baseline models as preliminary screening results; the trained baseline model is obtained by training the baseline model using the expanded training data pool; the untested data pool includes collected unverified literature data; the initial training data pool includes experimental data; An expanded training data pool selection module is used to select, from the preliminary screening results, the trained baseline model with the best performance parameters and the expanded training data pool corresponding to the trained baseline model with the best performance as the data set screening result; The training data pool expansion module specifically includes: A first material data selection submodule is used for selecting material data in an undetected data pool using a sliding window j to obtain a plurality of data candidate blocks, and adding the plurality of data candidate blocks to a candidate set; A training data pool expansion submodule, used to add each data candidate block in the candidate set to the training data pool after the mth expansion, to obtain multiple training data pools after the m+1th expansion; wherein the training data pool after the 0th expansion is the initial training data pool; The baseline model training submodule is used to train the baseline model using the training data pool obtained by the m+1th expansion, obtain the baseline model after the m+1th training, and calculate the performance score of the baseline model after the m+1th training, and obtain the performance score of the training data pool after the m+1th expansion after n data candidate blocks are added; An updating submodule is used to select the expanded training data pool with the highest performance score as the training data pool obtained by the m+1th expansion, select the trained baseline model with the highest performance score as the baseline model obtained by the m+1th training, and delete the data candidate block added in the expanded training data pool with the highest performance score from the candidate set; A first judgment submodule is used to judge whether the number of times the performance score of the trained baseline model continuously decreases exceeds a first preset threshold, and obtain a first judgment result; A second material data selection submodule, for increasing the value of m by 1, clearing the candidate set, and calling the training data pool expansion submodule if the first judgment result indicates no; an expanded training data pool selection submodule, for selecting, if the first judgment result indicates yes, the trained baseline model with the highest performance score, the expanded training data pool and the performance score corresponding to the trained baseline model with the highest performance score as the kth preliminary screening result corresponding to window j; A second judgment submodule is used to judge whether the value of k is less than a second preset threshold K, and obtain a second judgment result; a first calling submodule, for, if the second judgment result indicates yes, disrupting the order of the material data in the undetected data pool, increasing the value of k by 1, setting the value of m to 0, clearing the candidate set, and calling the first material data selection submodule; A third judgment submodule, configured to judge whether the value of j is less than a third preset threshold value J if the second judgment result indicates no, and obtain a third judgment result; The second calling submodule is used for updating window j to obtain window j+1 if the third judgment result indicates yes, increasing the value of j by 1, setting the values ​​of m and k to 0, clearing the candidate set, and calling the first material data selection submodule; The screening result output submodule is used to output K×J preliminary screening results if the third judgment result indicates no.

7. The material dataset screening system based on active learning according to claim 6, It is characterized in that The screening system further comprises: The preliminary training submodule is used to perform preliminary training on the baseline model using the material data in the initial training data pool to obtain a baseline model after preliminary training.

8. The material dataset screening system based on active learning according to claim 6, It is characterized in that The formula for calculating the performance score of the trained baseline model is: Score=P 1 *score1+P 2 *score2+P 3 *score3; score1=1-Relavent error; Among them, Score represents the performance score of the trained baseline model, and score1, score2, and score3 represent the relative error performance index, divergence index, and R of the trained baseline model, respectively. 2 Performance index; P 1 , P 2 and P 3 Represent the weights of score1, score2 and score3 respectively; Relavent error means relative error. n represents the number of material data in the screened training data pool, represents the predicted value of the performance parameter output by the trained baseline model corresponding to the i-th material data, and y(i) represents the actual value of the performance parameter in the i-th material data; JS(·) represents the probability distribution similarity calculation formula, KL(·) represents the divergence calculation formula, R 2 Indicates the correlation between the predicted result and the actual result. Represents the average value of the performance parameters of all material data in the expanded training data pool.

Images