An artificial intelligence hierarchical classification-regression combined prediction method for multi-class materials

Abstract

The application provides a kind of artificial intelligence hierarchical classification-regression combination prediction method for multi-class materials, belongs to material informatics and artificial intelligence technical field.The method obtains the historical sample data set of material chemical composition and crystal structure information, based on the principle of equal width, the band gap value is divided into multiple intervals, and the interval label is generated.By extracting key features and training SVM classification model to predict the band gap interval, then use multilayer neural network regression model to predict the band gap value.If the highest confidence is higher than the threshold, directly call the corresponding interval multilayer neural network regression model;if it is lower than the threshold, then multiple interval parallel regression and weighted fusion prediction are carried out.The application adopts the above-mentioned artificial intelligence hierarchical classification-regression combination prediction method for multi-class materials, which can effectively improve the precision and robustness of band gap prediction, especially when the high and low band gap sample distribution is uneven, and provides a precise and reliable solution for material band gap prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of material informatics and artificial intelligence, and particularly relates to an artificial intelligence hierarchical classification-regression combined prediction method for multi-class materials. BACKGROUND

[0002] The band gap is a core index for measuring the electronic structure and functional characteristics of materials, but due to the highly nonlinear influence of element combination, crystal structure and local environment, traditional single regression model is often difficult to balance the global fitting accuracy and generalization ability.

[0003] In a complex, multi-modal and significantly heteroscedastic data set, a globally trained regression network is prone to systematic deviation at the interval boundary and sparse area. Therefore, a hybrid learning framework that simultaneously handles the "global separability and local fitting" is needed to improve the accuracy and stability of band gap prediction. SUMMARY

[0004] The purpose of the present application is to provide an artificial intelligence hierarchical classification-regression combined prediction method for multi-class materials, which can improve the accuracy and robustness of material band gap prediction.

[0005] To achieve the above purpose, the present application provides an artificial intelligence hierarchical classification-regression combined prediction method for multi-class materials, comprising the following steps:

[0006] Step S1, obtaining a historical sample data set containing material chemical composition and crystal structure information, dividing the band gap value into a plurality of non-overlapping intervals based on the equal-width principle, and generating corresponding interval labels for each sample;

[0007] Step S2, extracting element attribute statistical features and crystal geometric features from the chemical composition and crystal structure information, and obtaining a key feature set after removing redundant features;

[0008] Step S3, training an SVM classification model based on the key feature set and the corresponding interval label, predicting the band gap interval of the to-be-tested sample, and outputting the class label and the confidence of each interval;

[0009] Step S4, dividing the historical sample data set into a plurality of subsets according to the interval label, and independently training a multi-layer neural network regression model for each interval subset;

[0010] Step S5, for the to-be-predicted sample, calling the classification model to obtain its band gap interval and confidence; if the highest confidence is higher than a preset threshold, calling the multi-layer neural network regression model of the band gap interval to output the prediction value; if the highest confidence is lower than the preset threshold, performing multi-interval parallel regression, and performing weighted fusion according to the confidence of each band gap interval;

[0011] Step S6, add the misclassified samples or samples with regression error exceeding the set threshold in the prediction process to the training set, and retrain the SVM classification model and the multi-layer neural network regression model.

[0012] Preferably, the interval division in step S1 is specifically: the samples with a band gap value of 0 eV are divided into a first interval, the samples with a band gap value of (0, 1] eV are divided into a second interval, the samples with a band gap value of (1, 2] eV are divided into a third interval, the samples with a band gap value of (2, 3] eV are divided into a fourth interval, and the samples with a band gap value greater than 3 eV are divided into a fifth interval.

[0013] Preferably, in step S1, each sample in the historical sample data set includes material chemical composition and crystal structure information, the material chemical composition includes chemical formula, element type and atomic ratio, and the crystal structure information includes lattice constant, volume, space group number, symmetry parameter and coordination number.

[0014] Preferably, the element attribute statistical features in step S2 include: average atomic number, valence electron number, electronegativity, atomic radius, first ionization energy and atomic volume; and the crystal geometric features include: lattice parameter, unit cell volume, space group number, symmetry parameter and coordination environment.

[0015] Preferably, the method for eliminating redundant features in step S2 is: calculating the Pearson correlation coefficient between the features, and eliminating one of the features with a Pearson correlation coefficient greater than 0.95.

[0016] Preferably, the kernel function of the SVM classification model in step S3 is a radial basis function, the hyperparameter C=10, and γ=0.01.

[0017] Preferably, the structure of the multi-layer neural network model in step S4 is: input layer, hidden layer 1, hidden layer 2, hidden layer 3 and output layer; wherein, the hidden layer 1 contains 128 neurons, the hidden layer 2 contains 64 neurons, and the hidden layer 3 contains 32 neurons, all the hidden layers use ReLU activation function, the output layer contains one neuron and uses linear activation function; the loss function for training the multi-layer neural network model is mean square error, the optimizer is Adam, the learning rate is set to 0.001, the batch size is set to 128, and the maximum number of iterations is set to 5000 times.

[0018] Preferably, the preset threshold in step S5 is 0.6; and the specific formula of weighted fusion is:

[0019]

[0020] wherein, represents the prediction value obtained by final fusion, K represents the number of intervals, p i represents the prediction probability of the i-th interval, represents the band gap prediction value output by the i-th interval neural network model.

[0021] Preferably, the incremental learning optimization in step S6 is specifically:

[0022] Error sample collection: during the prediction process, misclassified samples and samples with regression prediction errors exceeding a set threshold are collected and stored in the feedback database;

[0023] Optimization trigger condition: when the number of new samples in the feedback database exceeds a preset number, or a preset time interval has elapsed since the last optimization, the hyperparameter optimization process is automatically triggered;

[0024] Joint hyperparameter optimization: using the Bayesian optimization framework, the hyperparameters of the SVM classification model and the interval multi-layer neural network regression model are jointly adjusted to minimize the cross-validation mean absolute error or maximize the classification accuracy and F1-score;

[0025] The Bayesian optimization uses a Gaussian process as a proxy model, and uses expected improvement as a sampling strategy to select the parameter set to be evaluated;

[0026] The optimized hyperparameters include: the penalty coefficient C and the kernel function width γ of the SVM classification model, the learning rate, the number of hidden layer neurons, and the L2 regularization coefficient of the neural network regression model;

[0027] Model update: after the optimization process is completed, the SVM classification model and the interval multi-layer neural network regression model are retrained using the optimal hyperparameter combination to update the model performance.

[0028] Therefore, the present application adopts the above-mentioned artificial intelligence hierarchical classification-regression combined prediction method for multi-class materials, based on the strategy of "classification first, then regression", especially in the case of uneven distribution of high and low band gap samples, the fitting ability of the model for different interval samples can be significantly improved. At the same time, this method also has good generalization and interpretability, providing a more accurate and reliable solution for material band gap prediction. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 Flowchart of the present application, an artificial intelligence hierarchical classification-regression combined prediction method for multi-class materials;

[0030] Figure 2 Illustration of interval division of material band gap of the present application;

[0031] Figure 3 Comparison chart of neural network prediction performance without interval division;

[0032] Figure 4A comparison chart of the prediction performance of the regression model of each interval after the band gap interval division. DETAILED DESCRIPTION

[0033] The technical solutions of the present application are further described below by means of the drawings and examples.

[0034] Unless otherwise defined, the technical terms or scientific terms used in the present application shall have the usual meaning understood by a person with ordinary skills in the art to which the present application belongs.

[0035] Example 1

[0036] The material structure information used in this embodiment comes from the CIF file of the crystal material, and the data set contains 30000 historical sample data of known band gap values, and the data comes from the MaterialProjects database. Each historical sample data includes material chemical composition (chemical formula, element type, atomic ratio) and crystal structure information (lattice constant, volume, space group number, symmetry parameter, coordination number, etc.). The CIF file is parsed and feature engineered by the Matminer package of Python, the element attribute statistical features (including average atomic number, valence electron number, electronegativity, atomic radius, first ionization energy, atomic volume) and crystal geometry features (including lattice parameter, unit cell volume, space group number, symmetry parameter, coordination environment) are extracted, and the statistical methods (such as average value, standard deviation, range, maximum and minimum value, median, etc.) are used to summarize the element attribute statistical features contained in the data set, thereby generating a feature set representing the entire chemical formula, with hundreds of original features. Subsequently, data cleaning is performed to eliminate missing values and physically abnormal samples.

[0037] As shown in Figure 1 , a multi-class material-oriented artificial intelligence hierarchical classification-regression combined prediction method includes the following steps:

[0038] Step S1, interval division.

[0039] As shown in Figure 2 , according to the band gap value range, all samples are divided into 5 non-overlapping intervals:

[0040] The sample with a band gap value of 0eV is divided into the first interval, the sample with a band gap value of (0, 1]eV is divided into the second interval, the sample with a band gap value of (1, 2]eV is divided into the third interval, the sample with a band gap value of (2, 3]eV is divided into the fourth interval, and the sample with a band gap value greater than 3eV is divided into the fifth interval.

[0041] The interval division adopts the equal width principle, so that the number of samples in each interval is roughly balanced, and the performance degradation caused by class imbalance of the model is reduced.

[0042] Step S2, feature extraction.

[0043] All features were standardized, and redundant features with Pearson correlation coefficient greater than 0.95 were removed to reduce multicollinearity among features. Subsequently, principal component analysis (PCA) was used to reduce the dimensionality of the high-dimensional feature space, retaining 95% of the cumulative variance, and finally reducing the feature dimension to about 50 principal components. The principal components retained after dimensionality reduction are linear combinations of the original multidimensional physical features, no longer having a single physical meaning, but retaining the main information reflecting the material composition and structural variability in the original data, effectively reducing model complexity, suppressing overfitting, and improving learning efficiency.

[0044] Step S3, interval classification.

[0045] The support vector machine (SVM) classification model was trained using the feature vectors after PCA dimensionality reduction, with a radial basis function (RBF) kernel function, and the hyperparameters C = 10 and γ = 0.01 were selected by five-fold cross-validation. As shown in Table 1, the bandgap interval classification accuracy of the SVM classification model on the test set was about 65%, better than that of random forest (about 60%) and gradient boosting tree (about 58%).

[0046] Table 1 Test results

[0047] Model name Classification accuracy Cross-validation average F1 SVM 65.0％ 0.631 Random forest 60.1％ 0.598 Gradient boosting tree 56.3％ 0.581

[0048] Step S4, interval regression.

[0049] The training set was divided into 5 subsets according to the interval label, and independent multilayer neural network regression models were trained.

[0050] The network structure is: input layer (corresponding to the number of features after PCA dimensionality reduction) - hidden layer 1 (128 neurons, ReLU activation) - hidden layer 2 (64 neurons, ReLU activation) - hidden layer 3 (32 neurons, ReLU activation) - output layer (1 neuron, linear activation). The loss function is mean squared error (MSE), the optimizer is Adam (learning rate 0.001, batch size 128), the maximum number of iterations is 5000, and the model converges at about 600 iterations.

[0051] Step S5, combined reasoning.

[0052] In the prediction phase, first use the SVM classification model to predict the bandgap interval label; when the classification confidence is ≥0.6, directly call the corresponding interval neural network regression model to output the accurate bandgap value; when the confidence is <0.6, trigger multi-interval parallel regression, and according to the confidence of each interval, weighted fusion is performed to improve the prediction stability.

[0053] The specific formula for weighted fusion is:

[0054]

[0055] in, p represents the final fused predicted value, K represents the number of intervals, which is 5 in this invention; i This represents the predicted probability of the i-th interval, expressed as the confidence level output by the SVM. This represents the bandgap prediction value output by the neural network model for the i-th interval.

[0056] In the actual prediction phase, if the classifier's interval judgment of the sample does not have sufficient confidence (e.g., less than 0.6), the system will call all (or several high-confidence) neural network regression models in parallel and output the bandgap prediction values ​​respectively. And use the interval probability p corresponding to the sample i As a weighting factor, calculate the final fused output.

[0057] Step S6: Performance evaluation and feedback optimization.

[0058] like Figure 3 As shown, the R-value of the global multilayer neural network model without interval partitioning on the test set is... 2 It is only about 0.45, and the MAE is about 0.60 eV.

[0059] like Figure 4 As shown, after adopting the "classification first, regression later" method in this embodiment, the average R-squared value of the five interval regression models is... 2 The MAE increased to approximately 0.60, while the average MAE decreased to approximately 0.40 eV. The MAE distribution across all intervals ranged from 0.38 to 0.42 eV. R 2 It is distributed between 0.58 and 0.62.

[0060] After prediction is completed, the system will automatically record misclassified samples (incorrectly labeled predicted intervals) and high-error samples (bandgap prediction error MAE > 0.6 eV) to the feedback database. When the set conditions are met—such as the number of newly added error samples exceeding 500, or the time interval exceeding 7 days—the system will automatically trigger the Bayesian optimization process to jointly adjust the hyperparameters of the classification model and each interval regression model.

[0061] The optimization objective is to minimize the mean absolute error (MAE) under cross-validation, or to maximize classification accuracy and the F1 score. The optimization method is based on a Bayesian optimization framework, using a Gaussian process to construct a surrogate model and fitting the objective function with the current parameters and historical evaluation values. By using expected improvement as a sampling strategy, the system selects the next set of parameters to be evaluated in each round, gradually approaching the optimal solution.

[0062] Typical optimization parameters include C value and kernel function width gamma in support vector machine, and learning rate, hidden layer size and L2 regularization coefficient in neural network. The entire optimization process is usually set to no more than 30 rounds, and each round is evaluated on the basis of 5-fold cross-validation. After optimization, the system will save the optimal parameter combination and retrain the model to complete a performance self-update, ensuring that the prediction accuracy of the model on new material data continues to improve.

[0063] An artificial intelligence hierarchical classification-regression combined prediction system for multi-class materials, comprising:

[0064] A data preprocessing module: used to obtain a historical sample data set containing material chemical composition and crystal structure information, and divide the band gap value into a plurality of non-overlapping intervals according to the equal width principle, generate the corresponding interval label for each sample; meanwhile, extract element attribute statistical features and crystal geometric features from the chemical composition and crystal structure information, and obtain a key feature set after eliminating redundant features;

[0065] A classification model training module: training an SVM classification model based on the key feature set and the corresponding interval label, used to predict the band gap interval of the sample to be tested, and output the class label and the confidence of each band gap interval;

[0066] A regression model training module: dividing the historical sample data set into a plurality of subsets according to the interval label, and independently training a multi-layer neural network regression model for each interval subset;

[0067] A prediction module: for a sample to be predicted, calling the classification model to obtain the band gap interval and the confidence; if the highest confidence is higher than a preset threshold, calling the multi-layer neural network regression model of the band gap interval to output the predicted value; if the highest confidence is lower than the preset threshold, performing multi-interval parallel regression, and performing weighted fusion according to the confidence of each band gap interval;

[0068] An incremental learning optimization module: adding the misclassified samples or samples with regression error exceeding a set threshold generated in the prediction process to the training set, and retraining the SVM classification model and the multi-layer neural network regression model.

[0069] It is worth noting that the contents not elaborated in the present application are all prior art and are well known to those skilled in the art.

[0070] Therefore, the present application adopts the above-mentioned artificial intelligence hierarchical classification-regression combined prediction method and system for multi-class materials, which can effectively improve the accuracy and robustness of band gap prediction, especially when the distribution of high and low band gap samples is uneven, and provides a precise and reliable solution for material band gap prediction.

[0071] It should be pointed out finally that the above examples are only used to illustrate the technical solutions of the present application but not to limit it, and although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can still be modified or replaced equivalently, and these modifications or equivalent replacements should not make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. An artificial intelligence-based classification-regression combined prediction method for multiple categories of materials, characterized in that, Includes the following steps: Step S1: Obtain a historical sample dataset containing information on the chemical composition and crystal structure of the material. Divide the band gap value into multiple non-overlapping intervals based on the principle of equal width, and generate a corresponding interval label for each sample. The chemical composition of materials includes chemical formula, element types and atomic ratios, and crystal structure information includes lattice constant, volume, space group number, symmetry parameter and coordination number. Step S2: Extract elemental attribute statistical features and crystal geometric features from chemical composition and crystal structure information, and obtain the key feature set after removing redundant features; Step S3: Train an SVM classification model based on the key feature set and the corresponding interval labels to predict the band gap interval of the test sample and output the class label and the confidence score of each interval. Step S4: Divide the historical sample dataset into multiple subsets according to interval labels, and train a multi-layer neural network regression model independently for each interval subset; The structure of the multilayer neural network model is as follows: input layer, hidden layer 1, hidden layer 2, hidden layer 3, and output layer; where hidden layer 1 contains 128 neurons, hidden layer 2 contains 64 neurons, and hidden layer 3 contains 32 neurons, all of which use the ReLU activation function. The output layer contains 1 neuron and uses a linear activation function. The loss function for training the multilayer neural network model is mean squared error, the optimizer is Adam, the learning rate is set to 0.001, the batch size is set to 128, and the maximum number of iterations is set to 5000. Step S5: For the sample to be predicted, call the classification model to obtain its band gap interval and confidence level; if the highest confidence level is higher than the preset threshold, call the multi-layer neural network regression model of the band gap interval to output the predicted value; if the highest confidence level is lower than the preset threshold, perform parallel regression of multiple intervals and perform weighted fusion according to the confidence level of each band gap interval. Step S6: Add the misclassified samples or samples with regression errors exceeding the set threshold generated during the prediction process to the training set, and retrain the SVM classification model - multilayer neural network regression model.

2. The artificial intelligence-based classification-regression combined prediction method for multi-category materials according to claim 1, characterized in that, The specific interval division in step S1 is as follows: the samples with a band gap value of 0 eV are divided into the first interval, the samples with a band gap value of (0, 1] eV are divided into the second interval, the samples with a band gap value of (1, 2] eV are divided into the third interval, the samples with a band gap value of (2, 3] eV are divided into the fourth interval, and the samples with a band gap value greater than 3 eV are divided into the fifth interval.

3. The artificial intelligence-based classification-regression combined prediction method for multi-category materials according to claim 1, characterized in that, The elemental property statistical characteristics in step S2 include: average atomic number, number of valence electrons, electronegativity, atomic radius, first ionization energy, and atomic volume; the crystal geometric characteristics include: lattice parameters, unit cell volume, space group number, symmetry parameters, and coordination environment.

4. The artificial intelligence-based classification-regression combined prediction method for multi-category materials according to claim 1, characterized in that, The method for removing redundant features in step S2 is to calculate the Pearson correlation coefficient between features and remove one of the features with a Pearson correlation coefficient greater than 0.

95.

5. The artificial intelligence-based classification-regression combined prediction method for multi-category materials according to claim 1, characterized in that, The kernel function of the SVM classification model in step S3 is the radial basis function, and the hyperparameters are... =10, =0.

01.

6. The artificial intelligence-based classification-regression combined prediction method for multi-category materials according to claim 1, characterized in that, The preset threshold in step S5 is 0.6; the specific formula for weighted fusion is: = ； in, This represents the predicted value obtained from the final fusion. Represents the interval number. Indicates the first The predicted probability for each interval, Indicates the first The bandgap prediction value output by the interval neural network model.

7. The artificial intelligence-based classification-regression combined prediction method for multi-category materials according to claim 1, characterized in that, The incremental learning optimization in step S6 is as follows: During the prediction process, misclassified samples and samples with regression prediction errors exceeding a set threshold are collected and stored in the feedback database. When the number of new samples in the feedback database exceeds the preset number, or when a preset time interval has elapsed since the last optimization, the hyperparameter optimization process will be automatically triggered. A Bayesian optimization framework is adopted to jointly adjust the hyperparameters of the SVM classification model and the multilayer neural network regression model in each interval, with the goal of minimizing the cross-validation mean absolute error or maximizing the classification accuracy and F1-score. Bayesian optimization uses a Gaussian process as a surrogate model and selects the parameter set to be evaluated based on the desired improvement as the sampling strategy. The optimized hyperparameters include: the penalty coefficient of the SVM classification model. With kernel function width The learning rate, number of hidden layer neurons, and L2 regularization coefficient of the neural network regression model; After the optimization process is completed, the SVM classification model and the multilayer neural network regression model for each interval are retrained using the optimal hyperparameter combination to update the model performance.

Images