Material structured data-oriented generation and screening method and device

Abstract

The invention discloses a material structured data-oriented generation and screening method and device. The method comprises the steps of S1, acquiring material structured data; s2, according to the material structured data, judging whether treatment is needed or not by evaluating the ratio of the characteristic quantity to the sample quantity and the insight of a data set, and if treatment is needed, performing the step S3; s3, effectively generating the structural data of the material by using the depth generative model oriented to the structural data; s4, the generated samples are evaluated through multiple dimensions and weighted integration is carried out; and S5, according to an evaluation result, performing multi-dimensional high-quality sample screening by fusing domain knowledge. By adopting the technical scheme of the invention, effective generation of high-quality material structured data is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and in particular relates to a method and device for generating and screening material structured data. Background Art

[0002] Machine Learning (ML) has been widely used to reveal the structure-activity relationship of materials, thereby helping to design new target materials. Therefore, the feature-based heuristic research model has achieved certain success in various applications, such as property prediction and ionic conductivity screening. To effectively implement this model, researchers usually select descriptors (i.e., features) based on existing physicochemical knowledge and construct a structured material data set. ML models are then used to explicitly or implicitly reveal the trends between the human-understandable characteristics of the material and the target properties. It should be noted that the learning mechanism of ML relies on fitting the distribution of material data to explore the key factors that determine the target properties, which causes the performance of ML to be heavily dependent on the amount of input data. However, the method of obtaining structured material data mainly relies on tedious and complex experiments and simulation calculations, which often leads to small sample problems, thereby hindering the ML model from accurately capturing the intrinsic relationship between external conditions, internal interactions and target properties.

[0003] Data augmentation has proven to be an effective method for increasing the amount of data. Initially, researchers generated data by applying noise, such as image transformation and noise addition, rule-based transformation of text, and value-based and structure-based transformation of time series. While this augmentation approach has achieved some success across different types of data, it limits the diversity of augmented samples, resulting in a large number of samples and lacking any practical value. With the advancement of deep learning (DL), neural network-based methods such as variational autoencoders (VAEs) and generative adversarial networks (GANs) have been successfully applied to augment various types of data. However, compared to image, text, or time series data, structured data is often heterogeneous, resulting in dense numerical features and sparse categorical features. Furthermore, the correlation between features is weaker than that introduced through other types of spatial or semantic relationships. This can cause the learning patterns of such generative models to map the space through a Gaussian distribution, resulting in a significant loss of valuable information, which is considered the "last unconquered castle" of deep models. To address these issues, Park et al. (2018) applied a deep convolutional GAN ​​to tabular data, termed Table-GAN, in which the discrete features in the records are converted to square matrices to ensure compatibility with two-dimensional convolutional neural networks. However, it is unclear to what extent the inductive biases used for images are applicable to structured data. To adapt neural network-based generative models to structured data, Xu et al. (2018) proposed a Tabular GAN (TGAN) based on the GAN learning model to generate datasets with continuous and categorical columns. They introduced a Gaussian Mixture Model (GMM) with a fixed number of modes on each individual continuous column to accommodate non-Gaussian distributed data. However, due to the limitations of its learning model, TGAN still does not address the problems of the original GAN, such as mode collapse and hard convergence. To further improve TGAN, Xu et al. (2019) proposed a Conditional Tabular GAN (CTGAN), which uses a conditional GAN ​​architecture to enforce the learning of the correct conditional distribution for each column. To obtain categorical values ​​and allow backpropagation in the presence of categorical values, the Gumbel-SoftMax technique is used. Meanwhile, Xu et al. (2019) also proposed a VAE-based model called Tabular VAE (TVAE), which outperformed their proposed GAN approach. It is worth noting that no matter how powerful the generative model is, the quality of the generated samples is difficult to guarantee due to its inherent random generation mechanism.In this case, researchers usually define a proxy classification task and train a model for it on the real training set and another model on the generated dataset. Another approach is to visually inspect the modeled distribution of each feature, such as the cumulative distribution function, or compare the expected values ​​in the scatter plot. However, these studies only focus on a single measure to evaluate the quality of generation. This allows researchers to invest more energy in comprehensively and effectively filtering high-quality target samples, and even more seriously induces data pollution, especially in areas where data processing is difficult. To this end, Chundawat et al. (2024) proposed a general indicator called TabSynDex for the comprehensive evaluation of synthetic structured data. This indicator evaluates the similarity between synthetic data and real data through different component scores and helps to gain a deeper understanding of the generated material structured data, but it is still limited to screening generated samples that meet the needs of machine learning modeling from a statistical perspective, and does not take into account the uncertainty factors brought by practical applications.

[0004] In materials science, research on improving the amount of structured data is still in its infancy. For example, Lee et al. (2021) attempted to use conditional GANs to generate 200 new samples from 989 original samples in four stages, improving the accuracy of the ML model from 84.75% to 93.17%. To alleviate the demand for samples of ultra-high-performance concrete compressive strength, Marani et al. (2020) used TGANs to generate 6513 synthetic data from 810 experimental data. The ML model trained on the synthetic data achieved comparable performance to that trained on the experimental data. Although these studies achieved the best performance of ML models by generating collaborative data, their structured data failed to include categorical features, resulting in limited generalization. This is because categorical features (i.e., descriptors) are typically selected to represent composition, structure, and process, which largely determine the properties of the material. Furthermore, as mentioned above, the generative model mechanism diversifies the generated results and only considers data features, not physical mechanisms or prior knowledge. Therefore, it is necessary to provide a selection process to remove samples that violate the materials domain knowledge to overcome the contamination of low-quality synthetic material data. There is still a lack of an accessible process for effective material structured data enhancement and comprehensive assessment and screening of material structured data quality for machine learning. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method and device for generating and screening material structured data.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A method for generating and screening material structured data, comprising:

[0008] Step S1, obtaining material structured data;

[0009] Step S2: Based on the material structured data, determine whether governance is needed by evaluating the ratio of feature quantity to sample size and the insight of the data set. If governance is needed, proceed to step S3;

[0010] Step S3: effectively generate material structured data using a deep generative model for structured data;

[0011] Step S4: Evaluate the generated data in multiple dimensions and perform weighted integration;

[0012] Step S5: Based on the evaluation results, multi-dimensional high-quality sample screening is performed by integrating domain knowledge.

[0013] Preferably, in step S3, the material structured data is preprocessed and generated data is obtained using a generative model; the generated data is screened under the guidance of domain knowledge, and the effectiveness of the screening is judged by evaluating the prediction accuracy of the machine learning model on the extended data set obtained based on different screening rules, until the combination of screening rules with the optimal prediction accuracy is found.

[0014] As an example, in step S4, the generated data is measured for mean, median and standard deviation, logarithmic transformation correlation evaluation, S pMSE Index metrics, regularization support coverage metrics, and machine learning efficiency metrics are weighted and integrated.

[0015] Preferably, in step S5, five types of domain knowledge, namely, descriptor numerical constraints, descriptor relationship constraints, meta-feature constraints, and data distribution constraints, are introduced and encoded as screening rules to filter problem samples in a fine-grained manner to obtain high-quality material structured data that conforms to the domain knowledge.

[0016] The present invention also provides a device for generating and screening material structured data, comprising:

[0017] Acquisition module, used to obtain material structured data;

[0018] A data evaluation module is used to determine whether governance is needed based on the material structured data by evaluating the ratio of feature quantity to sample size and the insight of the data set. If governance is needed, step S3 is performed;

[0019] A data generation module is used to effectively generate material structured data using a deep generative model for structured data;

[0020] Generated data evaluation module, used to evaluate the generated data in multiple dimensions;

[0021] Generate a data screening module to perform multi-dimensional high-quality sample screening by integrating domain knowledge based on the evaluation results.

[0022] Preferably, the data generation module preprocesses the material structured data and obtains generated data using a generative model; the generated data is screened under the guidance of domain knowledge, and the effectiveness of the screening is judged by evaluating the prediction accuracy of the machine learning model on the extended data set obtained based on different screening rules, until the combination of screening rules with the optimal prediction accuracy is found.

[0023] As an optimal method, the generated data evaluation module measures the mean, median and standard deviation of the generated data, evaluates the correlation of logarithmic transformation, and pMSE The index metric, regularization support coverage metric, and machine learning efficiency metric are weighted and integrated.

[0024] Preferably, the data screening module is generated based on five types of domain knowledge, namely, descriptor numerical constraints, descriptor relationship constraints, meta-feature constraints, and data distribution constraints, and encoded as screening rules. The data evaluation module is used to generate the highest-scoring data set to further screen and filter problem samples to obtain high-quality material structured data that conforms to domain knowledge.

[0025] This paper proposes a universal domain knowledge-guided process for generating and screening structured material data. This process includes a data quantity assessment module for evaluating the necessity of implementing quantitative governance, a data generation module for generating structured target materials, a data evaluation module for comprehensively assessing the quality of generated data, and a data screening module for multi-dimensional, high-quality sample screening by integrating domain knowledge. This technical solution enables the efficient generation of high-quality structured material data, providing a high-quality data foundation and methodological support for driving machine learning and facilitating new material research and development. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0027] Figure 1 This is a flow chart of a method for generating and screening material structured data according to an embodiment of the present invention;

[0028] Figure 2 Schematic diagram for data screening;

[0029] Figure 3It is the histogram of the feature distribution of D35. DETAILED DESCRIPTION

[0030] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0031] 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.

[0032] Example 1:

[0033] like Figure 1 、 2 As shown, an embodiment of the present invention provides a method for generating and screening material structured data, including:

[0034] Step S1, obtaining material structured data;

[0035] Step S2: Based on the material structured data, determine whether governance is needed by evaluating the ratio of feature quantity to sample size and the dataset insight (i.e., the prediction accuracy of the high-performance ML model). If governance is needed, proceed to step S3.

[0036] Step S3: effectively generate material structured data using a deep generative model for structured data;

[0037] Step S4: Evaluate the generated data through multiple dimensions and perform weighted integration;

[0038] Step S5: Based on the evaluation results, multi-dimensional high-quality sample screening is performed by integrating domain knowledge.

[0039] As an implementation method of an embodiment of the present invention, in step S2, when performing a machine learning task, the size of the sample size directly affects the generalization ability and prediction performance of the model. Generally speaking, as the sample size increases, the model is often able to better learn the laws of the data, thereby improving the accuracy and reliability of the prediction. However, when the number of features is relatively small or there is a clear correlation between the features and the target variable, a smaller sample size may be sufficient to support the training and prediction of the model. In addition, data insight reveals the information and patterns in the data set through the prediction results of the machine learning model, and its analysis results are also often used to evaluate the relationship between the sample size and the performance of the machine learning model. Therefore, the embodiment of the present invention starts from the diversity of features and samples and the modeling accuracy of the data set on the target task, and on the basis of previous modeling experience, proposes an empirical rule for sample size evaluation based on balance and insight, as shown in formula (1).

[0040]

[0041] Where ratio represents the ratio of features to samples, and its definition is as shown in formula (2). n and s represent the number of features and samples in the dataset, respectively. Metric represents the result of data insight analysis, and its value is R on a single model among the candidate machine learning models. 2 Maximum value. When the feature-to-sample ratio is less than r1 or the data insight analysis result is greater than r2, the sample size relative to the number of features is sufficient for machine learning modeling or the model has already achieved high prediction accuracy. Data augmentation may not significantly improve the model, so sample size management is not recommended. Conversely, if the sample-to-feature ratio is large or the data insight accuracy is not high, data augmentation is recommended to achieve higher prediction accuracy. In this invention, r1 and r2 are set to 0.04 and 0.95 respectively. Please note that r1 and r2 are empirical values.

[0042] As an implementation method of an embodiment of the present invention, in step S3, in the data generator, since the characteristics of the material data set are composed of variables representing continuous changes in performance, structure, etc. and variables representing discontinuous changes in category, number, etc., in order to unify the representation and facilitate subsequent data analysis and modeling, different data representation strategies are adopted for different data types. Discrete columns are encoded using One-Hot, continuous columns are normalized based on a variational Gaussian model, and the processed data are used to train a generative model. At the same time, considering that discrete columns may have an uneven distribution, in order to ensure that all values ​​in the discrete columns can be taken into account when generating data, a conditional variable cond is defined to represent the category information of each generated data; the input of the generator is the noise variable z and the conditional variable cond, and the output is a generated sample, which is input into the discriminator together with the real sample, and the generator and discriminator are updated according to the discrimination result of the discriminator.

[0043] In the data filter, the embodiment of the present invention designs a decision maker for evaluating whether different filtering rules are effective, as shown in formula (3).

[0044] IF:met new <met min THEN:met min ←met new and continue; ELSE: skip (3)

[0045] Among them, met new Represents the evaluation index of the machine learning model, here it is the root mean square error (RMSE), which specifies the initial time met min is the RMSE on the original training set. The original generated samples are further screened under the guidance of the screening rules based on domain knowledge. Each time, the decision maker determines whether the screening is effective. If its RMSE is less than the current minimum value, it indicates that the generated samples with higher prediction accuracy can be obtained through this screening. This screening is effective, and the minimum RMSE obtained in this screening is assigned to met min , and continue to add the next screening rule based on the current screening; otherwise, it means that the current screening rule is invalid, skip the rule and enter the next screening rule. The specific execution process of the 4 screening rules is as follows Figure 2 As shown in the figure, there are 15 combinations of the four screening rules. Nodes of different colors represent different screening rules. The corresponding screening rule form is marked next to each node. Regardless of whether it is a single screening rule or a combination of multiple screening rules, the decision maker determines whether the current screening is effective each time. If it is effective, it proceeds to the next rule screening until the combination with the optimal RMSE and the corresponding RMSE are output, thereby adaptively finding the rule combination that generates high-quality samples for each data set.

[0046] The specific execution process of the method and device for generating and screening material structured data is shown in Algorithm 1. The algorithm is divided into two processes: data generation and data screening. In data generation, the material structured data is preprocessed, and the generated data is obtained by using the generative model. In data screening, the generated data is screened under the guidance of domain knowledge, and the effectiveness of the screening is judged by evaluating the prediction accuracy of the machine learning model on the extended data set obtained based on different screening rules, until the combination of screening rules with the optimal prediction accuracy is found.

[0047]

[0048] As an implementation of an embodiment of the present invention, in step S4, in order to measure the effectiveness of the generated tabular data, the embodiment of the present invention decided to focus on the quality of the data and the statistical similarity. Therefore, in addition to basic statistics, correlation scores, and machine learning efficiency metrics, the embodiment of the present invention also calculated pMSE and support coverage. All these scores combined can effectively represent the quality of the generated data. The minimum value of each indicator is 0 and the maximum value is 1. If a negative score is encountered, it is truncated to 0. This is because the embodiment of the present invention does not want the negative score to affect the good performance of other indicators. Similarly, the upper limit of the score is limited to 1. The final value is the average of all five scores. Different weights can be assigned to each score, but in ordinary research, the embodiment of the present invention assigns equal weights to them.

[0049] ① Basic statistical measurements

[0050] First, the three basic statistics of the data, namely the mean, median and standard deviation, are calculated to calculate the error between each column of the generated data and the real data, as shown in formulas (4)-(6).

[0051]

[0052] Among them, n c is the total number of samples in the column, R i and F i are statistical measures of the real data and generated data columns, respectively. The maximum value of the relative error is constrained to 1 to prevent the poor performance of a single column from causing misjudgment of the entire data.

[0053]

[0054] Where n is the number of columns, is the relative mean error of column i. The median and standard deviation scores are calculated in a similar way.

[0055]

[0056] ② Logarithmic transformation correlation

[0057] This statistic takes into account the correlation between the columns of each data set. Embodiments of the present invention use the Pearson correlation coefficient to measure the association between two continuous columns, the correlation ratio to measure the association between a categorical column and a continuous column, and Theil's U to measure the association between two categorical columns. This method produces two matrices containing a total of n2 values, where n is the number of columns in the data set. These matrices are used to calculate the correlation score. Similarly, embodiments of the present invention calculate the relative error of each entry in the correlation matrix of the real data, limit it to a maximum value of 1, and subtract the mean from 1 to obtain a rescaled score.

[0058]

[0059] Among them, sln(r ij )=sign(x)ln(|x|),r ij and f ij are the corresponding entries in the association matrix of the real data and the generated data, respectively. sign(x) is equal to 1 when x is positive and equal to -1 when x is negative.

[0060] ③S pMSE index

[0061] S pMSE The index is constructed based on the Propensity Score Mean Squared Error (PMSE), which is often used to analyze the difference between the distribution of real data and synthetic data. real samples, the synthetic data contains N syn samples, then the combined data contains N real +N syn samples, where the label Y represents real data as 0 and synthetic data as 1. The present embodiment trains a logistic regression model on the combined data for classification. When the number of columns is large, this makes calculating the similarity score very expensive. Therefore, this will make the metric unsuitable for training purposes because it needs to be calculated quickly at each epoch. In addition, the present embodiment uses a regression model to predict the probability of each sample being synthetic or real. The calculation formula for PMSE is as follows:

[0062]

[0063] Among them, N is the total number of samples after the synthetic samples and real samples are combined, that is, N = N real +N syn , is the predicted probability that the i-th sample is a synthetic sample, and c is the proportion of samples that are synthetic samples, that is,

[0064] Considering that the generated data cannot be distinguished from the real data, the improvement of pMSE is as follows:

[0065] S pmse =α |1-r| (9)

[0066]

[0067] Where k is the number of parameters in the logistic regression model and E(·) is the expectation.0 The closer r is to 1, the more difficult it is to distinguish fake data from real data.

[0068] ④Regularization support coverage

[0069] This embodiment of the present invention proposes a support coverage-based metric for comparing histograms between real and synthetic data. Support coverage measures the amount of support for a real data variable covered in the synthetic data. If a variable's rare class is poorly represented, this embodiment of the present invention imposes a maximum penalty on the overall support coverage score. For a single variable or column c, the metric is calculated as follows:

[0070]

[0071] Among them, n cat is the number of categories of the variable, is the number of samples of this category in the generated data, is the number of samples in the real data, and the scaling factor is the ratio of the number of real samples to the number of generated data samples. cat is the number of categories. For continuous data, n cat The number of intervals used to divide the data into equally spaced intervals for analysis. This embodiment uses 20 intervals, or a 5% range per interval. This ensures that the intervals are neither too large to ignore the distribution of values ​​within the interval nor too small to have classes with no samples.

[0072] For this score for a single variable, if is small, it is affected the most. This means that if the corresponding If it is slightly larger, the total score may be much larger than 1, and if it is slightly smaller, the total score will be much smaller than 1. Embodiments of the present invention hope to see a smaller score if a rare class is not well represented, but not ignore the representation of other classes if a single rare class is over-represented. Therefore, embodiments of the present invention limit this score to a maximum of β, which can be determined based on the representation of the required class. The final score is the average of the scores of all columns. The score of each column is limited to a maximum of 1. This is done to ensure that a good representation of one column does not cause the overall score to ignore the poor representation of other columns.

[0073]

[0074] Among them S cr is the score of the i-th column, and n is the number of columns.

[0075] ⑤Machine Learning Efficiency Metrics

[0076] This embodiment of the present invention trains a general model on both real and generated data. The trained model is then evaluated on real and synthetic test sets. This embodiment calculates the relative error relative to the RMSE of the regression model trained on the real data. Similarly, the F1-Score of the classifier trained on the real data is calculated. It is important to note that all scores are scaled within the range [0, 1], i.e., the mean of the different errors minus 1, as shown below:

[0077]

[0078] in, is the error of the i-th model. Four models are used in this embodiment of the present invention: logistic regression, random forest, decision tree, and multilayer perceptron. These models are trained for classification and regression tasks on the target column, using random forest, lasso regression, ridge regression, and ElasticNet, respectively.

[0079] The above five measurement methods evaluate different aspects of the similarity between real data and generated data. The embodiment of the present invention further combines all these measurement methods, and the calculation formula is as follows:

[0080]

[0081] Among them, α i is the weight information.

[0082] As an implementation method of an embodiment of the present invention, in step S5, five types of domain knowledge are introduced: descriptor value constraints, descriptor relationship constraints, meta-feature constraints, and data distribution constraints, and encoded as screening rules. Based on the generated data with the highest evaluation index in step S4, the problem samples are further screened and filtered to obtain high-quality material structured data that conforms to the domain knowledge. The details are as follows:

[0083] ①Symbolic representation of domain knowledge

[0084] Materials research often involves multi-scale modeling, so data sets are often composed of multi-scale features such as microstructure, crystal structure, and chemical composition, and there are usually complex influence mechanisms between features. In order to ensure that the generated samples conform to the laws of physicochemistry, it is necessary to screen the original generated samples based on prior material domain knowledge. The embodiment of the present invention proposes a series of screening rules based on intrinsic domain knowledge such as the value range of features, the correlation between features, data distribution, and the characteristics of the data (i.e., meta-features) to effectively screen the generated samples.

[0085] (1) Feature value range

[0086] Generated data should fall within a known range of feature values ​​to ensure consistency between the generated samples and real data. If the value range of the generated data differs significantly from the original data, this can lead to training bias in the model and prevent it from correctly capturing the data features. For example, in the D35 zeolite framework bulk modulus prediction dataset, the value ranges and meanings of all features are shown in Tables 1 and 2. Table 1 shows that some local features, such as SiOSi_average and SiOSi_skew, have relatively concentrated value ranges and low dispersion; whereas pore features such as ASA and AV have a wider range of variation and higher dispersion. Figure 3 The distribution of some features, the horizontal axis is the value of each sample, it can be clearly seen Figure 3 The values ​​of the features in (a)-(b) are close, and Figure 3 The values ​​of the feature in (c) are very sparse. Therefore, in order to improve the quality and stability of the generated samples, the embodiment of the present invention uses the standard deviation of the dataset as a benchmark to formulate a screening rule 1 based on the value range.

[0087] Table 1

[0088] feature Minimum Maximum feature Minimum Maximum SiOSi_average 130.691 175.54 SiO_var 0.00 0.001 SiOSi_gmean 129.360 175.50 density 0.734 1.964 SiOSi_hmean 128.174 175.45 volume 16.934 45.29 SiOSi_max 134.363 180.00 spg 11 230.00 SiOSi_mean 130.691 175.54 ASA 0.00 2602.24 SiOSi_min 116.446 170.43 NASA 0.00 950.25 SiOSi_skew -1.110 3.93 AV 0.00 0.641 SiOSi_std 0.00 21.93 NAV 0.00 0.068 SiOSi_var 0.00 480.84 VolFrac 0.00 0.470 SiO_average 1.606 1.64 largest_free_sphere 1.074 11.58 SiO_gmean 1.606 1.64 largest_included_sphere 2.884 15.27 SiO_hmean 1.606 1.64 largest_included_sphere_free 2.811 15.27 SiO_max 1.610 1.67 max_dim 0.00 3.00 SiO_mean 1.61 1.64 min_dim 0.00 3.00 SiO_min 1.59 1.64 uni_anis -1.77 5.67 SiO_skew -3.94 2.27 energy 0.91 99.19 SiO_std 0.00 0.03 k_vrh 7.22 112.72

[0089] Table 2

[0090]

[0091] Filtering rule 1: Assume that the original data set D = {d1, d2, ..., d n}, n is the total number of features; standard deviation set STD={std1,std2,…,std n}, std j Represents the standard deviation of the jth feature; the value range set R={r1,r2,…,r n}and is the empirical value range of the jth feature, and are the minimum and maximum values ​​of the empirical values, respectively. In order to ensure that the generated samples retain a certain degree of diversity, the embodiment of the present invention uses the variable k1 to slightly expand the value range. The value of k1 is shown in formula (15). When the standard deviation is greater than 1, it means that the difference between the minimum and maximum values ​​of the feature is large. Let k1 = 50; otherwise, let k1 = 5; generate the data set G = {g1, g2, ..., g M}.in, G={g1,g2,…,g m} represents the mth generated sample, represents the value of the i-th feature of the m-th sample, where M is the total number of samples in the dataset G. For each sample in the generated sample G, if formula (16) is satisfied, it means that the value of each feature is within the empirical range of domain knowledge and it is a generated sample that meets screening rule 1; otherwise, it is eliminated.

[0092] IF std(d m )>10THEN k1=50ELSE k1=5 (15)

[0093]

[0094] (2) Correlation between features

[0095] Correlation refers to a logical relationship of mutual dependence or correlation between features. These relationships can be direct or indirect, simple or complex, and they reflect the structure and regularity within the data. In material data sets, features usually cover multiple aspects such as chemical composition, crystal structure, and physical properties. These features often contain complex influencing mechanisms, resulting in certain correlations between features. These relationships may be linear or nonlinear. For example, as shown in formulas (17)-(20), there is a certain relationship between some features representing local properties or pore properties in D35; as shown in formula (10), the sum of the alloying element components in the D42 nickel-based single crystal high-temperature alloy material system should be 100%; as shown in formulas (22)-(25), some features in the prediction of ionic conductivity of D47 superionic conductors have simple calculation relationships.

[0096] SiOSi_std=sqrt(SiOSi_var) (17)

[0097] SiOSi_mean=SiOSi_average (18)

[0098] SiO_std=sqrt(SiO_var) (19)

[0099] AV+NAV<0.5 (20)

[0100] Ni+Al+Co+Cr+Mo+Re+Ru+Ta+W+Hf+Nb+Ti+Ir=100 (21)

[0101] Slog=Log10(S) (22)

[0102]

[0103] D1eff=D1ionicr*Na (24)

[0104] D2eff=D2ionicr*d2occu (25)

[0105] Therefore, the screening rule 2 based on the correlation between features is formulated as follows.

[0106] Filtering rule 2: Assume that the original data set D = {d1, d2, ..., d n}, n is the total number of features; for each data set, its correlation set C = {c1, c2, ..., c n}, j is the number of related relationships, c j The definition of is shown in formula (26), d a ,d b ,d i For features with correlation; generate data set G = {g1, g2, ..., g M}, represents the mth generated sample, represents the value of the i-th feature of the m-th sample, and M is the total number of samples in the dataset G. represents any +-×÷ operation, Y(·) equals 1 or an operator such as sqrt() or log10(), k2 and k3 are arbitrary values ​​and k2 ≤ k3. For any sample g in the generated data G, if the features in g satisfy formula (27), then the correlation between the features in the sample is correct and it is a generated sample that meets screening rule 2; otherwise, it is eliminated.

[0107]

[0108] (3) Meta-features

[0109] The value range and correlation are used to screen the generated samples from a single dimension and between dimensions, respectively. Meta-features are used to describe the characteristics of the dataset itself in machine learning or data mining tasks and to evaluate the integrity of the dataset. Due to the complex distribution of features in material datasets, kurtosis, a statistical concept that measures the steepness of the data distribution, is often used in data analysis in multiple fields due to its simplicity and intuitiveness. Its calculation process is shown in formula (17). Therefore, kurtosis is selected as an evaluation indicator for the overall distribution of generated samples. The similarity between the generated samples and the real samples is evaluated by calculating the correlation coefficient between them. Based on this, a screening rule based on meta-features is proposed.

[0110]

[0111] Filtering rule 3: Assume that the original data set D = {d1, d2, ..., d n}, n is the total number of features; generate data set G = {g1, g2, ..., g MFor the original dataset and the generated dataset, if they satisfy formula (18), it means that the meta-features of the generated sample are close to the original data and it is a valid sample.

[0112]

[0113] Here, Kurtosis(D) represents the kurtosis set of the original dataset D, Kurtosis(G) represents the kurtosis set of the generated sample G, PCC(·) represents the calculation of the Pearson correlation between Kurtosis(D) and Kurtosis(G), and k4 is the correlation threshold, which defaults to 0.2.

[0114] (4) Data distribution

[0115] The main purpose of screening generated samples is to ensure that they accurately reflect the data distribution of real samples. If the generated dataset and the original dataset follow a similar distribution, the generated samples are more likely to represent the characteristics of the original data, thus ensuring that the generated samples are interpretable and reliable to a certain extent. This leads to the proposal of screening rule 4 based on data distribution.

[0116] Filtering rule 4: Assume that the original data set D = {d1, d2, ..., d n}, n is the total number of features, generating a data set G = {g1, g2, ..., g n}, for any g-th sample in G, if it satisfies formula (19), it means that the sample obeys the data distribution of the original data set.

[0117]

[0118] Where μ and σ represent the mean and standard deviation of the original data set, respectively, and k5 is 0.3.

[0119] In response to the shortcomings of the existing technology, the embodiments of the present invention have constructed a method and device for generating and screening structured material data. By integrating domain knowledge to analyze the connotation characteristics of the material data volume, a multi-dimensional evaluation index system is established, and corresponding data volume improvement methods are implemented. First, an evaluation criterion is constructed based on the sample-feature balance and insight analysis results to determine whether the data set needs to be governed. Subsequently, the conditional table generative adversarial network (CTGAN) and the dimensional weighted variational autoencoder (DWVAE) are introduced as material data generators, and multiple statistical indicators are used to quantitatively evaluate the generated structured material samples. At the same time, the material domain knowledge is encoded into logical rules and embedded in the generation process to improve data quality. This process provides theoretical guidance and technical implementation paths for the acquisition of high-quality material data, and can promote the in-depth application of machine learning in the research and development of new materials.

[0120] To verify the effectiveness of the proposed method and apparatus for generating and screening structured material data, the present invention compared the performance of raw samples, GANs, CTGANs with and without sample screening, and DWVAE. Table 3 shows the prediction performance of ML models on expanded datasets based on different generative models (using RMSE as an example). The RMSE of the ML models for 16 datasets showed a decreasing trend with increasing training set size, while D38, D47, D54, and D59 did not follow this trend. The number of features in D38, D47, and D59 exceeded half or even more of the sample size. Therefore, each additional sample likely covered more feature space and data distribution, making it difficult for the ML model to accurately fit the model. Due to the small number of samples in D54 and the poor quality of the generated samples, the ML model's prediction accuracy was poor. Across 120 datasets (20 datasets x 6 training set sizes), CTGAN generated samples of superior quality than GAN in 71% (85 of the 85 datasets) and was well adapted to datasets of varying sample sizes. This demonstrates that CTGAN possesses high stability due to its powerful learning and ability to handle discrete features. CTGAN-filtered samples generated by CTGAN outperformed CTGAN on 88% (105 datasets). Compared to CTGAN, the ML model achieved a 7% average RMSE improvement on the extended dataset obtained from the filtered CTGAN (D38 was not included in the statistics due to its large RMSE fluctuations). Furthermore, the filtered DWVAE outperformed the filtered CTGAN on the vast majority of datasets due to its dimension-weighting capabilities based on domain knowledge. This demonstrates that higher-quality generated samples can be obtained through domain knowledge constraints, and also validates the effectiveness of prior domain knowledge in further filtering generated samples.

[0121] Table 3

[0122]

[0123]

[0124]

[0125] Example 2:

[0126] The embodiment of the present invention further provides a device for generating and screening material structured data, comprising:

[0127] Acquisition module, used to obtain material structured data;

[0128] A data evaluation module is used to determine whether governance is needed based on the material structured data by evaluating the ratio of feature quantity to sample size and the insight of the data set. If governance is needed, step S3 is performed;

[0129] A data generation module is used to effectively generate material structured data using a deep generative model for structured data;

[0130] Generated data evaluation module, used to evaluate the generated data in multiple dimensions and perform weighted integration;

[0131] Generate a data screening module to perform multi-dimensional high-quality sample screening by integrating domain knowledge based on the evaluation results.

[0132] As an implementation method of an embodiment of the present invention, the data generation module preprocesses the material structured data and obtains generated data using a generative model; the generated data is screened under the guidance of domain knowledge, and the effectiveness of the screening is judged by evaluating the prediction accuracy of the machine learning model on the extended data set obtained based on different screening rules, until the combination of screening rules with the optimal prediction accuracy is found.

[0133] As an implementation of an embodiment of the present invention, the generated data evaluation module performs mean, median and standard deviation measurement, logarithmic transformation correlation evaluation, S PMSE exponential metrics, regularized support coverage metrics, and machine learning efficiency metrics.

[0134] As an implementation method of an embodiment of the present invention, a data screening module is generated to generate a data evaluation module based on descriptor numerical constraints, descriptor relationship constraints, meta-feature constraints and data distribution constraints to further screen and filter problem samples to obtain high-quality material structured data that conforms to domain knowledge.

[0135] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for generating and screening material structured data, characterized in that: include: Step S1, obtaining material structured data; Step S2: Based on the material structured data, determine whether governance is needed by evaluating the ratio of feature quantity to sample size and the insight of the data set. If governance is needed, proceed to step S3; Step S3: effectively generate material structured data using a deep generative model for structured data; Step S4: Evaluate the generated data in multiple dimensions and perform weighted integration; Step S5: Based on the evaluation results, multi-dimensional high-quality sample screening is performed by integrating domain knowledge.

2. The method for generating and screening material-oriented structured data according to claim 1, characterized in that: In step S3, the material structured data is preprocessed, and the generated data is obtained using the generative model. The generated data is screened under the guidance of domain knowledge, and the effectiveness of the screening is judged by evaluating the prediction accuracy of the machine learning model on the extended data set obtained based on different screening rules, until the combination of screening rules with the optimal prediction accuracy is found.

3. The method for generating and screening material structured data according to claim 2, wherein: In step S4, the generated data are measured for mean, median and standard deviation, logarithmic transformation correlation evaluation, S pMSE Index metrics, regularization support coverage metrics, and machine learning efficiency metrics are weighted and integrated.

4. The method for generating and screening material structured data according to claim 3, wherein: In step S5, five types of domain knowledge, including descriptor numerical constraints, descriptor relationship constraints, meta-feature constraints, and data distribution constraints, are introduced and encoded as screening rules to further screen and filter problem samples and obtain high-quality material structured data that conforms to domain knowledge.

5. A device for generating and screening material structured data, characterized in that: include: Acquisition module, used to obtain material structured data; A data evaluation module is used to determine whether governance is needed based on the material structured data by evaluating the ratio of feature quantity to sample size and the insight of the data set. If governance is needed, step S3 is performed; A data generation module is used to effectively generate material structured data using a deep generative model for structured data; Generated data evaluation module, used to evaluate the generated data in multiple dimensions; Generate a data screening module to perform multi-dimensional high-quality sample screening by integrating domain knowledge based on the evaluation results.

6. The device for generating and screening material structured data according to claim 5, characterized in that: The data generation module preprocesses the material structured data and uses the generative model to obtain generated data; the generated data is screened under the guidance of domain knowledge, and the effectiveness of the screening is judged by evaluating the prediction accuracy of the machine learning model on the extended data set obtained based on different screening rules, until the combination of screening rules with the optimal prediction accuracy is found.

7. The device for generating and screening material structured data according to claim 6, wherein: The generated data evaluation module measures the mean, median and standard deviation of the generated data, evaluates the correlation of logarithmic transformation, and pMSE exponential metrics, regularized support coverage metrics, and machine learning efficiency metrics.

8. The device for generating and screening material structured data according to claim 7, wherein: The data screening module is generated based on five types of domain knowledge: descriptor numerical constraints, descriptor relationship constraints, meta-feature constraints, and data distribution constraints, and encoded into screening rules to further screen and filter problem samples to obtain high-quality material structured data that conforms to domain knowledge.

Images