Perovskite solar cell photoelectric conversion efficiency model construction method and efficiency prediction method

Abstract

The invention belongs to the field of machine learning model construction, and relates to a construction method and an efficiency prediction method of a photoelectric conversion efficiency model of a perovskite solar cell, and the construction method comprises the steps: 1), building an original data set of additive molecules of the perovskite solar cell; 2) dividing the original data set constructed in the step 1) based on a molecular skeleton method to obtain a plurality of sub-data sets; 3) training the basic model by using the plurality of sub-data sets obtained in the step 2) to obtain a solar cell photoelectric conversion efficiency model; the basic model is a multi-input single-output regression model. The invention provides the construction method and the efficiency prediction method of the photoelectric conversion efficiency model of the perovskite solar cell, which can improve the pertinence and the efficiency of the model in a feature learning process and can improve the prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of machine learning model construction, and relates to a method for constructing a perovskite solar cell photoelectric conversion efficiency model and an efficiency prediction method, and in particular to a method for constructing a perovskite solar cell photoelectric conversion efficiency model modified by an additive and a method for predicting the photoelectric conversion efficiency of a solar cell based on the model. Background Art

[0002] Perovskite materials have attracted extensive attention from many researchers due to their excellent photoelectric properties such as wide absorption range, long carrier lifetime, and low exciton binding energy. The photoelectric conversion efficiency of perovskite solar cells has rapidly increased from the initial 3.8% to 26.7%, and is widely regarded as the most promising new generation of solar cells. However, defects inside the perovskite crystal can cause non-radiative recombination of carriers, thereby reducing the photoelectric conversion efficiency of the device. By using organic molecules as additives to regulate the surface and interface structure of perovskites, the quality of perovskite films can be significantly improved and their interface contact with the transport layer material can be optimized to improve the photoelectric conversion efficiency of solar cells. Therefore, the development of efficient additives is of great significance to improving the performance of perovskite solar cells.

[0003] At present, the development of additives for perovskite solar cells mainly relies on experimental trial and error, which is not only long-term but also inefficient. In recent years, machine learning has attracted much attention as a key technology to accelerate the discovery of new materials and reveal the relationship between material properties and performance. However, due to the scarcity of experimental data for additives, building a model that can accurately capture molecular features is particularly important for improving prediction accuracy and accelerating the screening of new materials. Previous studies have mostly used molecular descriptors or fingerprints to characterize molecular features and build models, but the prediction accuracy of these models is obviously limited. First, the data set is small and unevenly distributed, resulting in insufficient accuracy of the model when predicting new molecules, which in turn affects its generalization ability. In addition, commonly used molecular descriptors usually rely on human selection, which may ignore key molecular structure details and introduce bias. Therefore, the development of new methods that can more comprehensively capture the relationship between molecular structure and performance is not only crucial to improving the accuracy of model predictions, but also to accelerate the screening of efficient additives suitable for perovskite solar cells. Summary of the invention

[0004] In order to solve the above-mentioned technical problems existing in the background technology, the present invention provides a method for constructing a photoelectric conversion efficiency model of a perovskite solar cell and an efficiency prediction method, which can improve the pertinence and efficiency of the model in the feature learning process and improve the prediction accuracy.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] A method for constructing a photoelectric conversion efficiency model of a perovskite solar cell, characterized in that the method for constructing a photoelectric conversion efficiency model of a perovskite solar cell comprises the following steps:

[0007] 1) Establish an original data set of perovskite solar cell additive molecules;

[0008] 2) dividing the original data set constructed in step 1) based on the molecular skeleton method to obtain multiple sub-data sets;

[0009] 3) Using the multiple sub-data sets obtained in step 2) to train the basic model to obtain a solar cell photoelectric conversion efficiency model; the basic model is a regression model with multiple inputs and a single output.

[0010] Preferably, the specific implementation method of step 1) adopted in the present invention is:

[0011] 1.1) Collect and organize the molecular structures of additives reported in existing literature and the experimental data corresponding to the molecular structures of additives;

[0012] 1.2) Based on the molecular structure of the additive collected in step 1.1), the molecular fingerprint and molecular descriptor data of the molecular structure of the additive are calculated using the RDKit information chemistry package;

[0013] 1.3) Using the variational autoencoder model pre-trained on the ZINC dataset, the molecular structure of the additive collected in step 1.1) is encoded to obtain latent space encoding information;

[0014] 1.4) Based on the molecular descriptors and molecular fingerprints obtained in step 1.2) and the latent space encoding information obtained in step 1.3), an original dataset of perovskite solar cell additive molecules is constructed.

[0015] Preferably, the specific implementation method of step 1.2) adopted by the present invention is:

[0016] 1.2.1) Use the MACCSkeys.GenMACCSKeys method in the RDKit library to calculate the MACCS fingerprint of the additive molecular structure, and use the CalcDescriptors() method in the RDKit library to calculate the descriptor values ​​of all the additive molecular structures in the Descriptors._descList list;

[0017] 1.2.2) Screen the descriptors, remove the descriptors whose values ​​are 0 in all additive molecules, and then remove the descriptors whose standard deviation is lower than the set value;

[0018] 1.2.3) Calculate the Pearson correlation coefficient for each pair of the screened descriptors, and remove the descriptors with correlations higher than the set value to obtain the final descriptor set used for the model.

[0019] Preferably, the specific implementation method of step 1.3) adopted by the present invention is:

[0020] 1.3.1) In JTVAE, the additive molecular structure is first represented as a SMILES sequence and decomposed into a tree-like substructure;

[0021] 1.3.2) Using a recursive neural network to encode the tree substructure obtained in step 1.3.1) to generate an encoding of the substructure features;

[0022] 1.3.3) Use graph neural networks to encode the connectivity information of atoms and bonds in the tree substructure to obtain the encoding of global structural features;

[0023] 1.3.4) Integrate the encoding of the substructure features obtained in step 1.3.2) and the encoding of the global structure features obtained in step 1.3.3) to generate latent space encoding information of the additive molecular structure through a variational autoencoder.

[0024] Preferably, the specific implementation of step 2) adopted in the present invention is:

[0025] 2.1) using the MurckoScaffold.GetScaffoldForMol method in the RDKit library to calculate the skeleton of the additive molecule one by one for all the additive molecules in the original data set constructed in step 1), and obtain the skeleton of each additive molecule;

[0026] 2.2) analyzing the skeleton of each additive molecule obtained in step 2.1), clustering the additive molecules with similar skeletons, and classifying the additive molecules with similar structures into the same category;

[0027] 2.3) According to the categories obtained in step 2.2), the original data set obtained in step 1) is divided into multiple sub-data sets, each sub-data set containing additive molecules with the same or similar skeleton category.

[0028] Preferably, the specific implementation method of step 3) adopted by the present invention is:

[0029] 3.1) Randomly divide the sub-dataset obtained in step 2) into a training set and a test set;

[0030] 3.2) Using the data in the training set selected in step 3.1) as input parameters and the predicted photoelectric conversion efficiency of perovskite solar cells as output results, a multi-input single-output regression model is trained to ultimately obtain a solar cell photoelectric conversion efficiency model.

[0031] Preferably, the input parameters in step 3.2) adopted in the present invention include the molecular fingerprint and molecular descriptor data of the additive molecular structure calculated using the RDKit information chemistry package and the latent space encoding information generated by the variational autoencoder.

[0032] Preferably, in step 3.1) of the present invention, the ratio of the training set to the test set is 9:1.

[0033] A perovskite solar cell photoelectric conversion efficiency model constructed based on the perovskite solar cell photoelectric conversion efficiency model construction method as described above.

[0034] A method for predicting the photoelectric conversion efficiency of a perovskite solar cell based on the above-mentioned perovskite solar cell photoelectric conversion efficiency model, characterized in that the prediction method comprises the following steps:

[0035] 1) Obtain a realistic dataset of perovskite solar cell additive molecules;

[0036] 2) The actual data set obtained in step 1) is input as an input parameter into a solar cell photoelectric conversion efficiency model, and the photoelectric conversion efficiency of the perovskite solar cell is output through the solar cell photoelectric conversion efficiency model.

[0037] The beneficial effects of the present invention are:

[0038] The present invention provides a method for constructing a photoelectric conversion efficiency model of a perovskite solar cell, comprising: 1) establishing an original data set of perovskite solar cell additive molecules; 2) dividing the original data set constructed in step 1) based on a molecular skeleton method to obtain multiple sub-data sets; 3) using the multiple sub-data sets obtained in step 2) to train a basic model to obtain a solar cell photoelectric conversion efficiency model; the basic model is a regression model with multiple inputs and a single output. The present invention divides the data set based on the molecular skeleton, and uses the latent space encoding of the generation model Junction Tree Variational Autoencoder (JTVAE) fully trained in a large molecular data set as an input to improve the prediction accuracy of the perovskite additive efficiency model. The present invention calculates the Murcko skeleton through RDKit and classifies the molecular data, so that the molecular structure information can be more effectively extracted and utilized. The method focuses on the core skeleton of the molecule, removes the side chain and the modifying group, thereby enhancing the representativeness of the structural features. This strategy focusing on the core skeleton helps to extract key structural information and improve the pertinence and efficiency of the model in the feature learning process. By classifying molecules with similar skeletons, the model can more accurately grasp the relationship between molecular features and performance, thereby improving its generalization ability on new data. The present invention significantly enhances the representation capability of molecular structure information by introducing JTVAE latent space coding as model input. Latent space coding can effectively capture the complex structural features within the molecule and the quantitative relationship between the characteristics and performance, so that the model can more deeply analyze the influence of molecular structure and functional groups on the photoelectric conversion efficiency (PCE) during the training process. Obviously, the present invention can improve the accuracy of performance prediction, thereby providing a reliable basis for optimizing the design and screening of additives. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flow chart of dividing a data set based on a molecular skeleton proposed by the present invention.

[0040] Figure 2 It is a flow chart of calculation and screening of molecular descriptors of additives.

[0041] Figure 3 It is a flow chart of the model training test set and training set division method.

[0042] Figure 4 It is a graph of the prediction error results of different data partitioning methods during the modeling process.

[0043] Figure 5 It is the Spearman correlation coefficient result graph of the prediction results of different data partitioning methods in the modeling process.

[0044] Figure 6 It is a JTVAE encoding-decoding diagram.

[0045] Figure 7 This is a graph of the prediction error results of different models using JTVAE latent space encoding as model input.

[0046] Description of reference numerals:

[0047] M: MACCS molecular fingerprint; D: RDKit molecular descriptor; Z: JTVAE latent space encoding. DETAILED DESCRIPTION

[0048] The present invention is described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0049] The present invention provides a method for constructing a photoelectric conversion efficiency model of a perovskite solar cell, the method comprising the following steps:

[0050] Step 1: Construct a database, which is the original database of perovskite solar cell additive molecules:

[0051] By consulting literature and other channels, we collected and sorted out the existing molecular structures of additives and their experimental data, including the molecular structure names of additives and the photoelectric conversion efficiency (PCE) data of the corresponding devices. Based on these molecular structures, we used the RDKit chemical informatics toolkit to calculate the corresponding molecular fingerprints and molecular descriptors. At the same time, we used the variational autoencoder (JTVAE) pre-trained on the ZINC dataset to encode the molecular structure. Figure 1 As shown in the figure, in JTVAE, molecules are first represented as SMILES sequences and decomposed into tree-like substructures (such as rings and functional groups) to capture their core structural units and connection relationships. Subsequently, the recursive neural network (RNN) encodes these tree-like substructures to generate encodings of substructure features. At the same time, the graph neural network (GNN) is used to encode the connection information of atoms and bonds in the molecule to obtain global structural features. After these substructures and global feature representations are integrated, the latent space encoding of the molecule is generated by the variational autoencoder as its feature representation. Finally, a raw dataset of perovskite solar cell additive molecules containing 208 molecular descriptors, molecular fingerprints, and latent space encodings was constructed.

[0052] Step 2: Divide the data set based on the molecular skeleton to obtain multiple sub-data sets, which specifically includes the following steps:

[0053] like Figure 2As shown in the figure, it is a flow chart for dividing the data set based on the molecular skeleton. First, the MurckoScaffold.GetScaffoldForMol method in the RDKit library is used to calculate the molecular skeletons of all molecules in the perovskite solar cell additive data set obtained in step 1 one by one. This process first removes the side chains and functional groups of each molecule, retaining only the key ring structure and its basic connection mode in the molecule, so as to generate the skeleton characteristics of each molecule. After extracting the skeleton, these skeletons are clustered by analyzing the structural characteristics of the generated molecular skeletons. In the clustering process, the similarity of the skeleton shape, connection relationship and core structure is used as the clustering criteria, and molecules with similar structures are classified into the same category, so that the same category contains molecules with the same or similar skeletons. This processing step provides a clear structural basis for the division of the data set, making it possible to classify molecules with similar skeletons. Finally, through this clustering process, the entire data set is divided into 9 sub-datasets, each of which contains molecules with the same or similar skeleton structure.

[0054] Step 3: Using the original perovskite solar cell additive molecule dataset to train the efficiency prediction model, including the following steps:

[0055] The efficiency prediction model adopts a regression model with multiple inputs and a single output. The model input includes three main components: first, the MACCS molecular fingerprint data from the original dataset of perovskite solar cell additive molecules obtained in step 1; second, various molecular descriptors, including heavy atom molecular weight, LogP value, and maximum absolute electronic state index; finally, the latent space encoding generated by the variational autoencoder (JTVAE). In addition, the model also considers the combination of the above three types of input features to enhance the richness of feature representation. The model output is the predicted efficiency of perovskite solar cells after molecular modification, thereby establishing an efficient perovskite solar cell photoelectric conversion efficiency prediction model.

[0056] During the training process, the loss function used is the mean absolute error (MAE) to measure the gap between the model prediction and the true label, and the parameters are adjusted through the optimization algorithm to gradually reduce the error, thereby improving the accuracy and predictive ability of the model.

[0057] Specifically, if Figure 3As shown, in the present embodiment, by using the CalcDescriptors() method in the RDKit library, 208 descriptors are generated for the additive molecule. First, the descriptors that are constant to 0 in all molecules are removed, and 184 valid descriptors remain after screening. Further, these descriptors are subjected to maximum value normalization, and the data is scaled to within the range of [0, 1] to standardize the data value. After normalization, the standard deviation threshold is set to 0.2 to remove the descriptors whose standard deviation is too small (less than 0.2), thereby screening out 49 descriptors. Further, by calculating the Pearson correlation coefficient (PCC) between the remaining descriptors, the correlation between the descriptors is evaluated. For PCC with an absolute value greater than 0.9, the relevant descriptors are divided into a group, and a representative descriptor is retained in each group. Through this step, 5 highly correlated descriptors are removed, and 44 descriptors with significant discrimination and information content are finally retained for subsequent analysis and model building. The normalized descriptor data are combined with the molecular MACCS fingerprint as the representation feature of the molecule. The 44 descriptors, the MACCS fingerprint of the molecules, and the latent space encoding of the molecules generated by the JTVAE model, together with various combinations of the three types of input features, were input into the prediction model as the characterization features of the additive molecules. All input variables were min-max normalized during modeling to ensure that the feature values ​​were within the same scale range.

[0058] For the established efficiency prediction model, 200 repeated experiments were adopted, and the average value of the 200 model predictions was used as the final model performance.

[0059] Specific steps are as follows Figure 4As shown in the figure: According to the 9 sub-datasets divided by the additive molecules in step 2), the sample extraction method is stratified sampling to ensure the representativeness of each type of molecule in the data set. For each subset, due to the difference in the number of molecules, the extraction methods of the training set and the test set are different. Specifically, for the subsets with a small number of molecules (the number of molecules in each subset is less than or equal to 3), one molecule is randomly selected as the test set, and the remaining molecules constitute the training set. This strategy ensures that model training and evaluation can still be carried out effectively even under small sample conditions. Relatively speaking, for the subsets with a large number of molecules (the number of molecules in each subset is greater than 3), two molecules are randomly selected as the test set, and the remaining molecules are used as the training set to make full use of the large sample information, thereby enhancing the generalization ability of the model. Based on the above stratified sampling, the entire data set is finally divided into a training set and a test set according to the set ratio, and the specific ratio is 9:1. To ensure the stability of the results and the generalization ability of the model, this study conducted 200 repeated experiments during the training process. In each experiment, the allocation of the training set and the test set is kept consistent to more accurately evaluate the model performance. Finally, the final evaluation index of the model performance was obtained by taking the average of the prediction results of 200 tests, thereby ensuring the accuracy and robustness of the evaluation results.

[0060] Ultimately, an efficient prediction model for the photoelectric conversion efficiency of perovskite solar cells is established.

[0061] In addition, based on the above content, the present invention also provides a method for predicting the photoelectric conversion efficiency of solar cells based on the prediction model, and the prediction method includes the following steps:

[0062] 1) Obtain a realistic dataset of perovskite solar cell additive molecules;

[0063] 2) The actual data set obtained in step 1) is input as an input parameter into the constructed solar cell photoelectric conversion efficiency model, and the photoelectric conversion efficiency of the perovskite solar cell is output through the solar cell photoelectric conversion efficiency model.

[0064] Exemplarily, the prediction method may have the following specific results:

[0065] In this embodiment, based on the molecular skeleton partitioning data set, the molecular fingerprint and descriptor are first used as model input to compare with the traditional random partitioning method. Figure 5As shown in the figure, the molecular skeleton partitioning method is significantly better than random partitioning in terms of MAE and deviation, showing higher accuracy. Furthermore, the Leave-One-Group-Out (LOGO) method is used, that is, the model is trained on eight groups of data and tested on the remaining group. The results show that compared with the LOGO method, the molecular skeleton partitioning method obtains a smaller error, verifying its accuracy in model evaluation. Figure 6 As shown in Table 1, the model based on molecular skeleton partitioning obtained a higher Spearman correlation coefficient, which further highlights the significant advantage of this method in improving the accuracy of machine learning model evaluation. The model training results based on different data partitioning methods are detailed in Table 1.

[0066] Table 1 Model training results of different data partitioning methods

[0067]

[0068] Table 2 Model training results with different inputs

[0069]

[0070] Furthermore, based on the molecular skeleton partitioning method, JTVAE latent space encoding is introduced as model input, such as Figure 7 As shown in Figure 2, a variety of models such as Random Forest, Support Vector Machine and Gradient Boosting were used for training, and their performance was analyzed. The results show that the introduction of latent space encoding generated by JTVAE as model input significantly improves the prediction ability of the model, which is specifically reflected in the reduction of mean square error (MAE). This phenomenon shows that latent space encoding can effectively capture more comprehensive molecular features, thereby enhancing the accuracy of prediction. In addition, when molecular descriptors are combined with latent space encoding as model input, the prediction model based on gradient boosting achieves the best performance on the validation set, with a MAE value of 0.8297±0.2477, reflecting a lower prediction error. Using the molecular skeleton partitioning method, the model training results based on different inputs are detailed in Table 2.

[0071] In summary, the present invention can improve the prediction accuracy of the photoelectric conversion efficiency of additive-modified perovskite solar cells.

Claims

1. A method for constructing a photoelectric conversion efficiency model of a perovskite solar cell, characterized in that: The method for constructing the photoelectric conversion efficiency model of the perovskite solar cell comprises the following steps: 1) Establish an original data set of perovskite solar cell additive molecules; 2) dividing the original data set constructed in step 1) based on the molecular skeleton method to obtain multiple sub-data sets; 3) Using the multiple sub-data sets obtained in step 2) to train the basic model to obtain a solar cell photoelectric conversion efficiency model; the basic model is a regression model with multiple inputs and a single output.

2. The method for constructing a photoelectric conversion efficiency model of a perovskite solar cell according to claim 1, characterized in that: The specific implementation method of step 1) is: 1.1) Collect and organize the molecular structures of additives reported in existing literature and the experimental data corresponding to the molecular structures of additives; 1.2) Based on the molecular structure of the additive collected in step 1.1), the molecular fingerprint and molecular descriptor data of the molecular structure of the additive are calculated using the RDKit information chemistry package; 1.3) Using the variational autoencoder model pre-trained on the ZINC dataset, the molecular structure of the additive collected in step 1.1) is encoded to obtain latent space encoding information; 1.4) Based on the molecular descriptors and molecular fingerprints obtained in step 1.2) and the latent space encoding information obtained in step 1.3), an original dataset of perovskite solar cell additive molecules is constructed.

3. The method for constructing a photoelectric conversion efficiency model of a perovskite solar cell according to claim 2, characterized in that: The specific implementation method of step 1.2) is: 1.2.1) Use the MACCSkeys.GenMACCSKeys method in the RDKit library to calculate the MACCS fingerprint of the additive molecular structure, and use the CalcDescriptors() method in the RDKit library to calculate the descriptor values ​​of all the additive molecular structures in the Descriptors._descList list; 1.2.2) Screen the descriptors, remove the descriptors whose values ​​are 0 in all additive molecules, and then remove the descriptors whose standard deviation is lower than the set value; 1.2.3) Calculate the Pearson correlation coefficient for each pair of the screened descriptors, and remove the descriptors with correlations higher than the set value to obtain the final descriptor set used for the model.

4. The method for constructing a photoelectric conversion efficiency model of a perovskite solar cell according to claim 3, characterized in that: The specific implementation method of step 1.3) is: 1.3.1) In JTVAE, the additive molecular structure is first represented as a SMILES sequence and decomposed into a tree-like substructure; 1.3.2) Using a recursive neural network to encode the tree substructure obtained in step 1.3.1) to generate an encoding of the substructure features; 1.3.3) Use graph neural networks to encode the connectivity information of atoms and bonds in the tree substructure to obtain the encoding of global structural features; 1.3.4) Integrate the encoding of the substructure features obtained in step 1.3.2) and the encoding of the global structure features obtained in step 1.3.3) to generate latent space encoding information of the additive molecular structure through a variational autoencoder.

5. The method for constructing a photoelectric conversion efficiency model of a perovskite solar cell according to claim 4, characterized in that: The specific implementation method of step 2) is: 2.1) using the MurckoScaffold.GetScaffoldForMol method in the RDKit library to calculate the skeleton of the additive molecule one by one for all the additive molecules in the original data set constructed in step 1), and obtain the skeleton of each additive molecule; 2.2) analyzing the skeleton of each additive molecule obtained in step 2.1), clustering the additive molecules with similar skeletons, and classifying the additive molecules with similar structures into the same category; 2.3) According to the categories obtained in step 2.2), the original data set obtained in step 1) is divided into multiple sub-data sets, each sub-data set containing additive molecules with the same or similar skeleton category.

6. The method for constructing a photoelectric conversion efficiency model of a perovskite solar cell according to claim 5, characterized in that: The specific implementation method of step 3) is: 3.1) Randomly divide the sub-dataset obtained in step 2) into a training set and a test set; 3.2) Using the data in the training set selected in step 3.1) as input parameters and the predicted photoelectric conversion efficiency of perovskite solar cells as output results, a multi-input single-output regression model is trained to ultimately obtain a solar cell photoelectric conversion efficiency model.

7. The method for constructing a photoelectric conversion efficiency model of a perovskite solar cell according to claim 6, characterized in that: The input parameters in step 3.2) include the molecular fingerprint and molecular descriptor data of the additive molecular structure calculated using the RDKit information chemistry package and the latent space encoding information generated by the variational autoencoder.

8. The method for constructing a photoelectric conversion efficiency model of a perovskite solar cell according to claim 7, characterized in that: The ratio of the training set to the test set in step 3.1) is 9:

1.

9. A perovskite solar cell photoelectric conversion efficiency model constructed based on the method for constructing a perovskite solar cell photoelectric conversion efficiency model as described in any one of claims 1 to 8.

10. A method for predicting the photoelectric conversion efficiency of a perovskite solar cell based on the perovskite solar cell photoelectric conversion efficiency model as claimed in claim 9, characterized in that: The prediction method comprises the following steps: 1) Obtain a realistic dataset of perovskite solar cell additive molecules; 2) The actual data set obtained in step 1) is input as an input parameter into the perovskite solar cell photoelectric conversion efficiency model constructed according to claim 9, and the perovskite solar cell photoelectric conversion efficiency model constructed according to claim 9 outputs the photoelectric conversion efficiency of the perovskite solar cell.

Images