Two-dimensional material solar hydrogen production efficiency prediction method and device based on machine learning

Abstract

The invention provides a two-dimensional material solar hydrogen production efficiency prediction method and device based on machine learning. The method comprises the following steps: reading a non-magnetic hexagonal crystal system single-layer material in a database to construct an initial data set; for samples with complete labels in the data set, training a single-layer material band edge position prediction model by adopting a feature engineering and gradient boosting decision tree algorithm; next, predicting and complementing samples with missing labels by using the trained model so as to generate a complete data set and further construct heterojunction double layers; carrying out data enhancement on a heterojunction sample of literature investigation through a CVAE-GAN hybrid architecture, and optimizing a feature vector; then training the new model by adopting a random forest algorithm to realize accurate prediction of the band edge position of the previously constructed heterojunction; and finally, the heterojunction type is judged in combination with a work function and a band edge position, the solar hydrogen production efficiency is calculated in batches by utilizing a program, and a high-efficiency heterojunction material is screened out.

Description

TECHNICAL FIELD

[0001] The application relates to the field of two-dimensional material performance prediction, and in particular to a two-dimensional material solar hydrogen production efficiency prediction method and device based on machine learning. BACKGROUND

[0002] With the rapid development of new energy materials and semiconductor technology, two-dimensional materials have shown broad application prospects in the fields of photoelectricity, catalysis and energy storage due to their unique physical and chemical properties. Especially in photocatalysis, the band edge position and work function of the material affect its performance. Therefore, constructing an efficient prediction model for accurate prediction of single-layer materials is of great significance for material screening and application. Traditional experimental methods usually face problems such as high cost and high time consumption when obtaining the performance of materials, and although the first-principle calculation method is accurate, it requires a large amount of computing resources and is difficult to meet the demand of large-scale material screening. SUMMARY

[0003] The purpose of the present application is to provide a two-dimensional material solar hydrogen production efficiency prediction method and device based on machine learning, to solve the problem that traditional experimental and calculation methods cannot efficiently screen high solar hydrogen production efficiency materials.

[0004] The present application provides a two-dimensional material solar hydrogen production efficiency prediction method based on machine learning, which comprises the following steps:

[0005] Step one, read the non-magnetic hexagonal single-layer material in the database to construct an initial data set;

[0006] Step two, for the samples with complete labels in the initial data set, use feature engineering and gradient boosting decision tree algorithm to train the prediction model of the band edge position of single-layer materials, then use the trained prediction model to predict and complete the samples with missing labels, to generate a complete data set and further construct a heterojunction bilayer;

[0007] Step three, use the CVAE-GAN hybrid architecture to perform data enhancement on the heterojunction samples researched in the literature as the training set and optimize the feature vector;

[0008] Step four, use the random forest algorithm to train a new model to accurately predict the band edge position of the previously constructed heterojunction;

[0009] Step five, combine the work function and band edge position to determine the heterojunction type, and then use the program to batch calculate the solar hydrogen production efficiency, and screen out high-efficiency heterojunction materials.

[0010] Further, the hexagonal single-layer material in step one includes crystal structure, constituent elements, Fermi energy level, vacuum energy level and band edge position parameters under HSE accuracy.

[0011] Further, the method of step two for training the prediction model of the edge position of the single-layer material band includes:

[0012] Based on the Magpie atomic characterization algorithm, 132-dimensional element descriptor is extracted to quantify the electronegativity intrinsic property of the constituent elements;

[0013] Then, 7-dimensional structure features are introduced to reveal the geometric structure and interlayer coupling effect of the two-dimensional material, and a 139-dimensional feature vector is constructed for each single-layer two-dimensional material;

[0014] Then, LASSO regularization is used to compress the feature dimension, and the optimal penalty coefficient λ=0.012 is determined through 10-fold cross-validation;

[0015] Then, the principal component features with a variance contribution rate between 90% and 95% are retained, so as to compress the feature dimension while retaining effective information, and ensure the robustness and prediction performance of the model training process;

[0016] The 7-dimensional structure features include formation energy, work function, and average electronegativity.

[0017] Further, the method of step two for training the prediction model of the edge position of the single-layer material band includes:

[0018] A series of decision trees are trained iteratively, and their prediction results are accumulated to approximate the target feature;

[0019] During training, the DFT band edge position provided by the C2DB database is taken as the true label y, and the extracted feature vector x is taken as the input. The gradient boosting process gradually optimizes the model's prediction of the band edge position by minimizing the loss. The mean absolute error, root mean square error, and determination coefficient are selected to evaluate the model, so as to screen out the single-layer material.

[0020] Further, the data enhancement method includes:

[0021] Through the CVAE-GAN hybrid architecture, 16 groups of clean heterojunctions selected from the investigated heterojunction samples are data enhanced, and an additional 180 virtual heterojunction sample data are generated.

[0022] Further, the parameters of the random forest algorithm of step four are 200 decision trees, a maximum depth of 10 layers, and a random seed of 42.

[0023] Further, the high-efficiency heterojunction material in step five refers to a heterojunction with a solar hydrogen production efficiency greater than 10% and a band edge position requirement of -4.44eV≤CBM≤-3.92eV and -6.59eV≤VBM≤-5.67eV. The heterojunction type is determined according to the energy band arrangement characteristics of type I and type II heterojunctions, and the method is determined by combining the energy band arrangement of the heterojunction and the direction of the built-in electric field caused by the work function difference between each single layer. That is, the CBM, VBM and Fermi level of one single layer material in the heterojunction are all higher than those of another single layer material, which meets the direct Z-type heterojunction. Finally, 196 type I, 170 type II and 50 type Z candidate heterojunctions are screened out.

[0024] Further, the average absolute error calculation formula is:

[0025]

[0026] wherein MAE is the average absolute error, n is the total number of samples, xi is the i-th real value, xi is the i-th predicted value, denotes the absolute error of the predicted value and the real value.

[0027] The root mean square error calculation formula is:

[0028]

[0029] wherein RMSE is the root mean square error, n is the total number of samples, xi is the i-th real value, xi is the i-th predicted value, denotes the square of the prediction error.

[0030] The determination coefficient calculation formula is:

[0031]

[0032] wherein, R is the determination coefficient, n is the total number of samples, xi is the i-th real value, xi is the i-th predicted value, is the mean of all real values, is the residual sum of squares, is the total sum of squares.

[0033] The application also provides a two-dimensional material solar hydrogen production efficiency prediction device based on machine learning, characterized by comprising:

[0034] A reading unit is configured to read initial data sets of non-magnetic hexagonal single-layer materials in a database;

[0035] A feature engineering and data completion unit is configured to screen the most representative features by using a LASSO regularization algorithm, screen a 41-dimensional feature vector for model training and screening of a single-layer material band edge position, screen a 266-dimensional feature vector for training of a heterojunction band edge position model, then complete the label missing samples to generate a complete data set, and further construct a heterostructure double layer.

[0036] A data enhancement unit is configured to enhance the heterojunction samples researched in the literature by using a CVAE-GAN hybrid architecture.

[0037] A model training unit is configured to train a single-layer material band edge position prediction model by using a gradient boosting decision tree, and train a heterojunction band edge position prediction model by using a random forest algorithm.

[0038] A screening unit is configured to screen single-layer materials with high solar hydrogen production efficiency and heterojunction materials with high solar hydrogen production efficiency by calculating solar hydrogen production efficiency.

[0039] Further, the 41-dimensional feature vector screened in the feature engineering unit is based on the Magpie atomic feature extraction algorithm to extract 132-dimensional element descriptor to quantify the intrinsic properties such as electronegativity of constituent elements; secondly, 7-dimensional structure features such as formation energy, work function and average electronegativity are introduced to reveal the geometric structure and interlayer coupling effect of two-dimensional materials based on the prior knowledge of material crystallography. After the above processing, a 139-dimensional feature vector is constructed for each single-layer two-dimensional material. The LASSO regularization is used to compress the feature dimension, the optimal penalty coefficient is determined by 10-fold cross-validation, and the principal component features with a variance contribution rate of 90% to 95% are reserved, so as to compress the feature dimension while retaining effective information, and ensure the robustness and prediction performance of the model training process.

[0040] Further, the 266-dimensional feature vector screened in the feature engineering unit is to extract 132-dimensional Magpie element features (including electronegativity, atomic radius, ionization energy and other physical parameters) of two single-layer materials, forming 264-dimensional basic features; secondly, the mean vector (132-dimensional) and the absolute difference vector (132-dimensional) of the Magpie features of the two materials are calculated to capture the interface electron coupling effect; finally, the band edge positions (VBM1, CBM1, VBM2, CBM2) of the single-layer materials are spliced to form a total of 532-dimensional feature vectors, and the LASSO is used to calculate the feature correlation, and the highly correlated redundant features are removed, and finally the 266-dimensional feature vector is selected for model training.

[0041] Further, the hybrid architecture of the data enhancement unit uses a fusion conditional variational autoencoder CVAE-GAN, which introduces variational inference constraints on the latent variable space, while retaining the generation ability of adversarial training, so that the statistical distribution of the generated data is highly consistent with the real data, thereby improving the feature correlation of the regression model. BRIEF DESCRIPTION OF DRAWINGS

[0042] The present application can be better understood by reference to the following description in conjunction with the accompanying drawings.

[0043] Figure 1 A flowchart of a two-dimensional material solar hydrogen production efficiency method based on machine learning provided by an embodiment of the present application;

[0044] Figure 2 129 unlabeled missing data selected by the present application;

[0045] Figure 3 The feature importance ranking of the conduction band minimum (CBM) and valence band maximum (VBM) predicted by the feature engineering of the present application, which corresponds to the correlation heat map of the selected features for predicting CBM and VBM, respectively;

[0046] Figure 4 The contour plot of the hyperparameter search of the CBM training model and the VBM training model of the present application;

[0047] Figure 5 The present application has high prediction accuracy for the band gap and band edge position of single-layer two-dimensional materials based on the gradient boosting decision tree model;

[0048] Figure 6 The virtual heterojunction data distribution generated by the data enhancement of the present application;

[0049] Figure 7 The feature importance analysis of the heterojunction prediction of the present application;

[0050] Figure 8 The four machine learning algorithms used by the present application, random forest regression (RF), gradient boosting decision tree (HGB), LassoCV and Ridge, obtain comparative results through repeated k-fold cross-validation;

[0051] Figure 9 The precision evaluation of the heterojunction prediction model based on the random forest model of the present application;

[0052] Figure 10 The present application obtains 1296 non-metallic single-layer materials for constructing heterojunctions from two-dimensional materials with a band gap greater than zero;

[0053] Figure 11The diagrams in the middle are the solar hydrogen production efficiency and parameter distribution diagrams of the present invention, (a) to (c) are the predicted solar hydrogen production efficiency and parameter distribution diagrams of type I, type II and Z-type heterojunctions, and (d) to (f) are the solar hydrogen production efficiency and parameter distribution diagrams of the finally screened type I, type II and Z-type heterojunctions;

[0054] Figure 12 It is a schematic structural diagram of the device of the present invention. DETAILED DESCRIPTION

[0055] The following will describe exemplary embodiments of the present invention in conjunction with the accompanying drawings. To ensure clarity and conciseness, the specification does not cover all features of the actual implementation. However, it should be understood that in the specific implementation process, developers need to make many targeted design decisions based on actual needs to achieve the intended goals. In addition, to avoid affecting the core understanding of the present invention due to too many details, the accompanying drawings only show the device structure and / or processing steps that are closely related to the solution of the present invention, and omit other content that is less relevant to the present invention.

[0056] The present invention first discusses how to read the non-magnetic hexagonal single-layer materials in the database to construct an initial data set, then uses feature engineering to screen out 139-dimensional feature vectors for the single-layer materials, then constructs a heterojunction based on the predicted single-layer materials, performs data enhancement based on samples from literature research, and constructs a heterojunction band edge position prediction model based on the random forest algorithm. The heterojunction is predicted using the trained model, and the solar hydrogen production efficiency of the heterojunction is calculated according to the program. Finally, heterojunctions with high solar hydrogen production efficiency are screened out. The present invention has certain reference value and practical significance for further research on the prediction of solar hydrogen production efficiency of two-dimensional materials.

[0057] like Figure 1 As shown, step 1, read the non-magnetic hexagonal single-layer materials in the database to construct an initial data set; step 2, for the samples with complete labels in the initial data set, use feature engineering and gradient boosting decision tree algorithm to train the single-layer material band edge position prediction model, and then use the trained prediction model to predict and complete the samples with missing labels, so as to generate a complete data set and further construct a heterojunction double layer; step 3, use the CVAE-GAN hybrid architecture to perform data enhancement on the heterojunction samples surveyed in the literature as a training set and optimize the feature vector; step 4, use the random forest algorithm to train a new model to achieve accurate prediction of the previously constructed heterojunction band edge position; step 5, combine the work function and band edge position to determine the heterojunction type, and then use the program to batch calculate its solar hydrogen production efficiency to screen out high-efficiency heterojunction materials.

[0058] In one embodiment of the present invention, a method for acquiring a data set includes:

[0059] C2DB database as a data source, the data is read through the ASE database interface and the labels and features used for model training are extracted therefrom, including the crystal structure, constituent elements, Fermi level, vacuum level, and band edge position under HSE accuracy of each material attribute, and then the data is cleaned according to the target task of the training model conduction band minimum (CBM) and valence band maximum (VBM). In order to make the heterojunction in the construction process of the heterojunction more stable, and to exclude the interference of lattice mismatch, interface defects and unstable compounds on the prediction of photocatalytic activity, the following screening strategies are used in the single-layer material processing process: first, based on symmetry analysis, hexagonal system materials are retained to optimize the interface lattice matching; second, semiconductor materials with a band gap greater than 0 are screened to meet the non-metallic material; in particular, magnetic materials are excluded to avoid the complex interface coupling effect caused by spin polarization effect, which is difficult to predict by machine learning model; at the same time, samples missing key physical property parameters are removed, and finally 362 kinds of non-magnetic single-layer two-dimensional material data set with clear structure and complete physical properties are obtained, which are used to construct a high-credibility photocatalytic activity prediction model, as shown in Figure 2 The transition metal sulfide, oxide material category is well represented.

[0060] In an embodiment of the present application, the feature engineering of single-layer materials is based on the Magpie atomic characterization algorithm to extract 132-dimensional element descriptors to quantify the electronegativity and other intrinsic properties of constituent elements; secondly, combined with the prior knowledge of material crystallography, 7-dimensional structure features such as formation energy, work function and average electronegativity are introduced to reveal the geometric structure and interlayer coupling effect of two-dimensional materials. After the above processing, a 139-dimensional feature vector is constructed for each single-layer two-dimensional material, covering structure, composition and basic electronic state information, etc. LASSO regularization is used to compress the feature dimension, and the optimal penalty coefficient λ = 0.012 is determined through 10-fold cross-validation, and then the principal component features with a variance contribution rate of 90% to 95% are retained, so as to retain effective information while compressing the feature dimension. The original feature data is processed by the principal component analysis (PCA) method.

[0061] Firstly, the original feature data is standardized to make the mean value of each feature dimension 0 and the variance 1, so as to eliminate the influence of different dimensions on the results; secondly, the covariance matrix of the standardized data is calculated, and the eigenvalue decomposition of the covariance matrix is performed to obtain the principal components and their corresponding variance contribution rates; then, the principal components are sorted according to the variance contribution rate, the cumulative variance contribution rate of each principal component is calculated, and the principal component set with a cumulative variance contribution rate of 90% to 95% is selected as the compressed feature space; finally, the original data is mapped to the selected principal component space to generate new low-dimensional feature data for subsequent model training, ensuring the robustness and prediction performance of the model training process.

[0062] The importance scores of all features were finally counted, and the top 10 features were selected to draw the feature importance bar chart. As shown in Figure 3 (a) and 3(c) respectively show the 10 most important features and their importance for predicting the conduction band minimum (CBM) and valence band maximum (VBM), respectively, Figure 3 (b) and 3(d) respectively show the corresponding feature correlation heat map, and it can be seen that there is no obvious collinearity or redundancy between different features, which proves the rationality of the selected features. The feature names and their numbers are shown in Table 1.

[0063] Table 1

[0064]

[0065] In an embodiment of the present application, the feature engineering of the two-dimensional heterojunction material is to extract 132-dimensional Magpie element features (including electronegativity, atomic radius, ionization energy, etc. Physical parameters) from two single-layer materials respectively, forming 264-dimensional basic features; secondly, the mean vector (132-dimensional) and the absolute difference vector (132-dimensional) of the Magpie features of the two materials are calculated to capture the interface electronic coupling effect; finally, the band edge positions (VBM1, CBM1, VBM2, CBM2) of the single-layer materials are spliced to form a total of 532-dimensional feature vectors. Therefore, the present study uses LASSO to calculate the feature correlation, and finally selects 266-dimensional feature vectors for model training. According to the feature importance ranking, the most important 10 features are selected according to different target tasks, as shown in Figure 4 (a) and 4(c). The actual feature names corresponding to the feature numbers are shown in Table 2. Figure 4 (b) and 4(d) show the correlation heat map of the 10 most important CBM and VBM features selected, and it can be seen that there is no obvious collinearity between the features after removing redundant features, which proves the rationality of the features used for model training.

[0066] Table 2

[0067] In an embodiment of the present application, the training and construction of the single-layer material model is realized based on a Gradient Boosting Decision Tree (GBDT) algorithm, which iteratively trains a series of decision trees and accumulates their prediction results to approximate the target function. Compared with the traditional GBDT, the HGBDT significantly reduces the calculation complexity of the split point by binning technology to discretize continuous features into histograms, which improves the training speed by 3-5 times while maintaining the prediction accuracy. Compared with a simple linear model, GBDT can fit more complex nonlinear relationships, and is therefore very suitable for processing highly nonlinear mapping relationships between material structure and performance, for example, training a large GBDT model with more samples and features, which has a significant advantage in processing mesoscale data sets with 41-dimensional features. With the band edge position and work function as the target function, the DFT band edge position provided by the C2DB database is used as the true label y, and the extracted feature vector x is used as the input during training. The loss function selects the mean square error (MSE), and the gradient boosting process gradually optimizes the model's prediction of the band edge position by minimizing the loss.

[0068] Next, the important hyperparameters of the GBDT model are optimized, such as the number of decision trees (n_estimators), the maximum depth (max_depth), the learning rate (learning_rate), and the subsample rate (subsample). These parameters are used to control the model capacity, limit the feature interaction complexity, adjust the gradient update step, and introduce randomness to enhance the generalization ability, respectively. By optimizing these key hyperparameters, overfitting can be prevented and the generalization ability can be improved. Figure 4 As shown in FIG. 5, after global optimization of the hyperparameters through grid search (Grid Search) and 5-fold cross-validation, the results show that when the n_estimators of the CBM prediction model is 160, the max_depth is 7, the learning_rate is 0.045, the n_estimators of the VBM prediction model is 150, the max_depth is 7, and the learning_rate is 0.055, the model achieves the optimal bias-variance balance on the validation set. To prevent overfitting, the early stopping method (Early Stopping) is used to dynamically monitor the validation set loss, and if the loss does not decrease for 20 consecutive iterations, the training is terminated.

[0069] In one embodiment of the present invention, the single-layer model evaluation is to comprehensively evaluate the performance of the single-layer material electronic structure prediction model. Three indicators, namely, mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²), are selected to test the model. MAE reflects the average level of the difference between the predicted value and the true value, RMSE pays more attention to larger errors to measure the robustness of the prediction, and R² indicates the degree to which the model explains the data variance (1 indicates a perfect prediction and 0 indicates a meaningless prediction). On the independent test set, the fitting plot of the true value and the predicted value of the CBM and VBM prediction models is shown in Figure 2. Figure 5 As shown in (a) and (b), the MAE of the single-layer material electronic structure prediction model for CBM and VBM are 0.113eV and 0.119eV, respectively, and the RMSE are 0.152eV and 0.167eV, respectively. 2 are 0.985 and 0.986 respectively. Figure 5 (c) shows that the model has a high prediction accuracy for the band gap and band edge positions of single-layer two-dimensional materials. The MAE is 42% lower than that of the traditional random forest model, which is due to the HGBDT algorithm's efficient modeling of nonlinear feature interactions. Through horizontal comparison, it is found that the model evaluation results are better than similar works reported in the literature (MAE is 0.3078, RMSE is 0.3953, R 2 is 0.9434), which is due to the physical constraints of the model added by the prior knowledge.

[0070] In one embodiment of the present invention, the method of data enhancement includes: investigating the band edge position data of heterojunctions and their two monolayer materials in 31 documents to generate an initial data set, but the models trained using this data set in subsequent studies generally have low determination coefficients, indicating that the data noise completely obscures the inherent correlation of physical laws. To this end, an analysis of the literature data from these surveys found that some of the data showed significant outliers in the characteristic distribution due to diverse material phases (such as β-AsP and other materials) or complex interface characteristics (such as porous graphite PG and other materials). To solve this problem, the initial data set was cleaned to exclude materials with complex phases and interface characteristics, and finally 16 groups of clean heterojunction data were selected to form the data set, as shown in Table 3. Table 3

[0071]

[0072] Due to the scarcity of heterojunction data, the dataset is expanded using data augmentation techniques. Generative Adversarial Network (GAN) is a classic data augmentation method that has been widely used in computer vision. GAN consists of a generator and a discriminator, which learn from each other to generate synthetic data that matches the distribution of real data. Using GAN, more "realistic" heterojunction samples are generated based on the known distribution of the existing small amount of heterojunctions, so that the training set size is expanded and the coverage is more comprehensive. Each heterojunction sample is represented as a high-dimensional vector, containing descriptors obtained by fusing the features of two single-layer materials. Finally, the trained generator generates an additional 180 heterojunction sample data, as shown in Figure 6 .

[0073] In one embodiment of the present application, the method for developing and verifying a two-dimensional heterojunction prediction model comprises: after preparing the expanded heterojunction data, a random forest regression algorithm is used to construct a heterojunction band edge position prediction model. The input features are generated through strict multi-level fusion: first, 132-dimensional Magpie element features (covering electronegativity, atomic radius, ionization energy, etc.) are extracted from two single-layer materials to form 264-dimensional basic features; second, the mean vector (132-dimensional) and absolute difference vector (132-dimensional) of the Magpie features of the two materials are calculated to capture the interface electron coupling effect; finally, the band edge positions (VBM1, CBM1, VBM2, CBM2) of the single-layer materials and the lattice mismatch ratio are spliced to form a total of 533-dimensional feature vectors. As shown in Figure 7 , feature importance analysis shows that the lattice mismatch ratio (mismatch) has the highest contribution to CBM prediction (weight 18.7%), while the single-layer VBM mean ((vbm1+vbm2) / 2) dominates the VBM evolution (weight 22.3%). As shown in Figure 7 , (b) and (d) show the correlation heat map of the top 10 most important CBM and VBM features, and there is no obvious collinearity between the non-aggregation function features.

[0074] To make the prediction effect of the heterojunction band edge position prediction model better, we use four algorithms, random forest regression (RF), gradient boosting decision tree (HGB), LassoCV and Ridge, to model the heterojunction band edge position, and compare them through cross-validation. As shown in Figure 8As shown, the RF model achieved the highest average coefficient of determination (R², approximately 0.827) for CBM predictions, while the RF model also performed best for VBM predictions (average R², approximately 0.804). Taking into account both CBM and VBM metrics, as well as model stability and generalization, the random forest regression model was ultimately selected as the final prediction model. Key parameters optimized using GridSearchCV were set as follows: 200 decision trees, a maximum depth of 10 layers, and a random seed of 42. Testing showed that this set of hyperparameters led to faster model convergence and higher accuracy. The augmented literature data was randomly split into a training set of 157 samples and a test set of 39 samples in an 8:2 ratio. Model stability was monitored during training using the out-of-bag error (OOB) test.

[0075] During the model development process, this paper conducted rigorous validation and error analysis on the random forest model. Using the test set data, the three evaluation parameters of the model were calculated, and the model was fitted based on the actual values ​​and the corresponding predicted values. The distribution of the predicted values ​​and the actual values ​​was compared through scatter plots. The evaluation results are as follows: Figure 9 The CBM prediction MAE is 0.091eV (RMSE = 0.115eV, R² = 0.931), and the VBM prediction MAE is 0.081eV (RMSE = 0.1eV, R² = 0.929). This prediction accuracy exceeds some existing heterojunction prediction work. The random forest model training results show that the model can accurately capture the potential correlation between the heterojunction band edge position arrangement and the constituent materials in the literature data, thus providing strong technical support for the prediction of the energy band of candidate heterojunctions.

[0076] In one embodiment of the present invention, the method for screening heterojunctions with high solar hydrogen production efficiency includes: the band edge positions of single-layer two-dimensional materials predicted by the single-layer material electronic structure prediction model of the present application provide a large amount of data for the construction of efficient photocatalytic heterojunctions. After the single-layer material labels are completed in the previous work, all samples have CBM and VBM labels. Next, the C2DB database is traversed again to extract the lattice constants and work functions of the candidate single-layer materials, and all sample data are integrated into a DataFrame. Screen the two-dimensional materials with band gaps greater than zero to obtain 1296 non-metallic single-layer materials for constructing heterojunctions, such as Figure 10 shown.

[0077] From the large amount of data generated by the single-layer material prediction model, 513 two-dimensional single-layer materials with photocatalytic potential were selected, then two by two vertically stacked to form heterojunctions, and during the construction process, the lattice mismatch rate was calculated according to the lattice constants of the two materials to screen the heterojunctions with a lattice mismatch ratio less than 1%, and finally 5086 heterojunctions were obtained. The combined features generated by the feature engineering are substituted into the heterojunction energy band prediction model for prediction.

[0078] In an embodiment of the present application, the program for calculating the theoretical solar hydrogen production efficiency limit of the heterojunction photocatalyst includes: selecting the global total spectrum irradiance as the original data for calculation, using the xlrd library to read the AM1.5G spectrum data from the Excel file, and converting the spectrum data into energy units for use in the integral calculation in the formula. Then import the numpy library to calculate the minimum theoretical limit of the heterojunction photocatalyst solar hydrogen production efficiency between data points, so as to calculate the interpolation integral. After interpolation, based on the overpotential of the heterojunction (for type I and type II heterojunctions) or the band gap values of the two components (for type Z heterojunctions), the solar hydrogen production efficiency of the heterojunction is calculated using the fitted data and the calculated interpolation. By inputting the CBM and VBM of type I and type II heterojunctions, or the band gap values of the two components of type Z heterojunction, the program can calculate the hydrogen evolution reaction and oxygen evolution reaction overpotential [χ(H2) and χ(O2)] under different pH conditions, light absorption efficiency (ηabs), carrier utilization rate (ηcu) and solar hydrogen production efficiency (ηsolar hydrogen production).

[0079] According to the characteristics of the energy band arrangement of type I and type II heterojunctions, the type of the heterojunction is determined, and the method of combining the energy band arrangement of the heterojunction and the direction of the built-in electric field caused by the difference in work function between the single layers is used, that is, the CBM, VBM and Fermi level of one single-layer material in the heterojunction are all higher than those of the other single-layer material, which meets the direct type Z heterojunction, and 1704 type I, 886 type II and 843 type Z heterojunctions are screened out, as shown in Figure 11 (a)-(c). The band edge position required for heterojunctions with solar hydrogen production efficiency greater than 10% is screened to be -4.44eV≤CBM≤-3.92eV and -6.59eV≤VBM≤-5.67eV, and the solar hydrogen production efficiency of these candidate heterojunctions is calculated using the program, and finally 1452 type I, 708 type II and 383 type Z heterojunctions are screened out, as shown in Figure 11Type I and Type II heterojunctions are easier to reach the limit required band edge position, so there are dozens of Type I and Type II heterojunctions predicted to have a limit of 17.12% solar hydrogen production efficiency, such as CdCl2 / ZrSe2, PtS2 / GeSe2 and PdS2 / PtI2; while Z-type heterojunctions are difficult to achieve the required conditions to reach the limit, so only a few Z-type heterojunctions are predicted to have a limit of 37.83% solar hydrogen production efficiency, such as NiTe2 / BSb.

[0080] In an embodiment of the present application, the method for verifying the accuracy of the prediction model comprises:

[0081] According to the prediction results of the heterojunction photocatalysts with high solar hydrogen production efficiency, the following structures MoS2 / WS2, MoO2 / WO2, MoS2 / Hf2CO2, Hf2CO2 / WS2, PtSe2 / WS2, PtS2 / GaSe and MoSe2 / Ti2CO2 are compared with the prediction results to verify the accuracy of the prediction model in actual situation. The results show that most of the predicted heterojunctions with high solar hydrogen production efficiency have not been studied and prepared, so only a small number of heterojunctions with literature research and their band edge positions are sorted out, as shown in Table 4. It can be seen that the error between the predicted value and the actual literature is within a reasonable range, proving the accuracy and application value of the high-efficiency heterojunction photocatalyst prediction model in this study, and also proving the realizability of the potential high-efficiency heterojunction photocatalysts predicted by the model.

[0082] Table 4

[0083]

[0084] This embodiment proposes a two-dimensional material solar hydrogen production efficiency prediction device based on machine learning, as shown in Figure 12 The device comprises:

[0085] A reading unit for reading the initial data set of non-magnetic hexagonal single-layer materials in the database;

[0086] A feature engineering and data completion unit for screening the most representative features using LASSO regularization algorithm, screening 41-dimensional feature vectors for single-layer material band edge position model training and screening 266-dimensional feature vectors for heterojunction band edge position model training, then completing the label missing samples to generate a complete data set, and further constructing a heterostructure double layer;

[0087] A data enhancement unit for enhancing the literature research heterojunction samples through CVAE-GAN hybrid architecture;

[0088] The model training unit is configured to train a single-layer material band edge position prediction model using gradient boosting decision trees and a heterojunction band edge position prediction model using a random forest algorithm.

[0089] The screening unit is configured to screen single-layer materials with high solar hydrogen production efficiency and heterojunction materials with high solar hydrogen production efficiency by calculating solar hydrogen production efficiency.

[0090] In an embodiment of the present application, the 266-dimensional feature vector screened in the feature engineering unit is obtained by extracting 132-dimensional Magpie element features (including electronegativity, atomic radius, ionization energy, etc.) of two single-layer materials, forming a 264-dimensional basic feature, calculating the mean vector (132-dimensional) and the absolute difference vector (132-dimensional) of the Magpie features of the two materials to capture the interface electron coupling effect, splicing the band edge positions (VBM1, CBM1, VBM2, CBM2) of the single-layer materials to form a total of 532-dimensional feature vector, and using LASSO to calculate the feature correlation and remove highly correlated redundant features, and finally selecting a 266-dimensional feature vector for model training.

[0091] In an embodiment of the present application, the CVAE-GAN hybrid architecture enhances the band edge position data obtained from literature research, constrains the latent variable space by introducing variational inference, while retaining the generation ability of adversarial training, so that the statistical distribution of the generated data is highly consistent with the real data, thereby improving the feature correlation of the regression model.

[0092] Although the present application has been described in terms of limited number of embodiments, those skilled in the art, with the benefit of the above description, will appreciate that other embodiments are possible within the scope of the application described herein. Furthermore, it should be noted that the language used in the specification has primarily been chosen for readability and instructional purposes, and can not have been selected to convey an exclusive or exhaustive description of the subject matter of the present application. Therefore, many modifications and variations of the present application are possible in light of the above teachings without departing from the scope and spirit of the appended claims. The disclosure of the present application is illustrative and not restrictive, and the scope of the present application is defined by the appended claims.

Claims

1. A method for predicting the efficiency of solar hydrogen production from two-dimensional materials based on machine learning, characterized in that: The following steps are involved: Step 1: Read the non-magnetic hexagonal single-layer material in the database to build an initial data set; Step 2: For the samples with complete labels in the initial data set, feature engineering and gradient boosting decision tree algorithm are used to train a prediction model for the edge position of the single-layer material. The trained prediction model is then used to predict and complete the samples with missing labels, thereby generating a complete data set and further constructing the heterojunction double layer. Step 3: Use the CVAE-GAN hybrid architecture to perform data augmentation on the heterojunction samples from the literature survey as a training set and optimize the feature vector; Step 4: Use the random forest algorithm to train a new model to accurately predict the position of the heterojunction band edge constructed previously; Step 5: Determine the heterojunction type based on the work function and band edge position, and then use the program to batch calculate its solar hydrogen production efficiency to screen out high-efficiency heterojunction materials.

2. The method for predicting the efficiency of solar hydrogen production from two-dimensional materials based on machine learning according to claim 1, characterized in that: The hexagonal single-layer material in step 1 includes crystal structure, constituent elements, Fermi level, vacuum energy level, and band edge position parameters under HSE precision.

3. The method for predicting the efficiency of solar hydrogen production from two-dimensional materials based on machine learning according to claim 1, characterized in that: The method of using feature engineering to train a prediction model for the edge position of a single-layer material in step 2 includes: Based on the Magpie atom characterization algorithm, a 132-dimensional element descriptor is extracted to quantify the electronegativity intrinsic properties of the constituent elements; The 7-dimensional structural features are then introduced to reveal the geometric structure and interlayer coupling effect of the 2D material, and a 139-dimensional feature vector is constructed for each single-layer 2D material. Then, LASSO regularization is used to compress the feature dimension, and the optimal penalty coefficient λ=0.012 is determined through 10-fold cross validation; Then, the principal component features with variance contribution rates between 90% and 95% are retained, thereby compressing the feature dimensions while retaining effective information, ensuring the robustness of the model training process and predictive performance; The 7-dimensional structural characteristics include formation energy, work function and average electronegativity.

4. The method for predicting the efficiency of solar hydrogen production from two-dimensional materials based on machine learning according to claim 1, characterized in that: The method of using the gradient boosting decision tree algorithm to train the prediction model of the edge position of the single-layer material in step 2 includes: Approximate the target feature by iteratively training a series of decision trees and accumulating their predictions; During training, the DFT band edge position provided by the C2DB database is used as the true label y, and the extracted feature vector x is used as input. The gradient boosting process gradually optimizes the model's prediction of the band edge position by minimizing the loss. The mean absolute error, root mean square error, and determination coefficient are selected to evaluate the model, thereby screening out single-layer materials.

5. The method for predicting the efficiency of solar hydrogen production from two-dimensional materials based on machine learning according to claim 1, characterized in that: The data enhancement method includes: The CVAE-GAN hybrid architecture is used to enhance the data of 16 groups of clean heterojunctions selected from the heterojunction samples under investigation, and an additional 180 virtual heterojunction sample data are generated.

6. The method for predicting the efficiency of solar hydrogen production from two-dimensional materials based on machine learning according to claim 1, characterized in that: The parameters of the random forest algorithm described in step 4 are 200 decision trees, 10 maximum depth, and 42 random seeds.

7. The method for predicting the efficiency of solar hydrogen production from two-dimensional materials based on machine learning according to claim 1, characterized in that: The high-efficiency heterojunction material described in step 5 refers to a heterojunction with a solar hydrogen production efficiency greater than 10% and a band edge position requirement of -4.44eV≤CBM≤-3.92eV and -6.59eV≤VBM≤-5.67eV.

8. The method for predicting the efficiency of solar hydrogen production from two-dimensional materials based on machine learning according to claim 4, characterized in that: The mean absolute error calculation formula is: Among them, MAE is the mean absolute error, is the total number of samples, is the i-th true value, is the i-th predicted value, Indicates the absolute error between the predicted value and the true value; The root mean square error calculation formula is: Among them, RMSE is the root mean square error, is the total number of samples, is the i-th true value, is the i-th predicted value, represents the square of the prediction error; The determination coefficient calculation formula is: in, is the coefficient of determination, is the total number of samples, is the i-th true value, is the i-th predicted value, is the mean of all true values, is the residual sum of squares, is the total sum of squares.

9. A two-dimensional material solar hydrogen production efficiency prediction device based on machine learning, characterized in that: The device comprises: A reading unit, used for reading non-magnetic hexagonal single-layer materials from the database to construct an initial data set; The feature engineering and data completion unit is used to select the most representative features using the LASSO regularization algorithm, screen out 41-dimensional feature vectors for model training of the band edge position of single-layer materials and 266-dimensional feature vectors for training the band edge position model of heterojunctions. It then completes the missing label samples to generate a complete data set and further constructs the heterostructure double layer. Data augmentation unit, used to enhance heterojunction samples from literature research through the CVAE-GAN hybrid architecture; A model training unit is used to train a single-layer material band edge position prediction model using a gradient boosting decision tree and a heterojunction band edge position prediction model using a random forest algorithm; The screening unit is used to screen out single-layer materials with high solar hydrogen production efficiency and heterojunction materials with high solar hydrogen production efficiency by calculating the solar hydrogen production efficiency.

Images