Automated screening method of molecular additives in perovskite light emitting diodes and photovoltaic cells

Abstract

The application provides an automatic screening method for a molecular additive in a perovskite light-emitting diode and a photovoltaic cell. On the basis of full utilization of an existing additive data set, a generating model and a prediction model in the field of deep learning are combined, a new molecule similar to an existing additive molecular structure is generated through a development idea of "generating first and screening later", and then the generated molecule is screened for multiple rounds with an ideal additive as a target, so that the PeLED additive molecule is automatically and quickly developed.

Description

Technical Field

[0001] This invention relates to the field of computer-aided molecular design technology, specifically to an automated screening method for molecular additives in perovskite light-emitting diodes and photovoltaic cells. Background Technology

[0002] Perovskite materials, due to their excellent optoelectronic properties such as low exciton binding energy, long carrier lifetime, and tunable bandgap, have shown great promise in fields such as solar cells, light-emitting diodes (LEDs), and photodetectors, and are widely used. Perovskite light-emitting diodes (PeLEDs) have become a focal point of competition in the global field of new light-emitting and display technologies, while perovskite solar cells (PSCs) have performed exceptionally well in the photovoltaic field. Nevertheless, due to unavoidable internal and interface defects, the energy conversion efficiency of perovskite optoelectronic devices still has room for improvement. Adding suitable organic molecules to the precursor solution to reduce defects in the perovskite thin film and optimize the perovskite layer and its interface is considered an effective strategy to improve device performance. Therefore, developing highly efficient additives that can significantly improve the performance of perovskite optoelectronic devices is of great significance for promoting the development of photovoltaic power generation and lighting display technologies.

[0003] Molecular design is generally understood as the process of designing and synthesizing molecules with ideal properties that meet specific needs. The main challenge of this process lies in the extremely vast chemical space that needs to be explored, encompassing approximately 10... 60 This space contains numerous possible molecular structures, and is characterized by its non-uniform distribution and unstructured nature. Given the vast differences in the impact of different molecules on device performance, rapidly and effectively identifying molecular structures with desirable properties from this enormous chemical space is a key challenge in molecular development.

[0004] Due to the complexity of the influence of additive molecules on perovskite crystallization, morphology, point defects, and precursor chemistry, the development of additives for perovskite optoelectronic devices currently relies mainly on experimental trial-and-error methods, which are time-consuming and inefficient. With improved computing power, researchers have begun to use machine learning-based quantitative structure-activity relationship (QSAR) models to screen additives based on small molecule structures. However, this method is often limited by existing chemical databases, making it difficult to effectively explore unknown chemical spaces, thus the screened molecules often lack novelty. With the further development of artificial intelligence, deep generative models, by modeling the probability distribution behind molecular sets, can explore the chemical space more comprehensively, thereby generating new molecular structures. However, how to effectively correlate these generated molecular structures with target properties still needs to be adjusted according to specific circumstances. In the additive development process of this patent, the labeled data obtained through experimental testing and characterization is relatively limited, and the potential information contained within inevitably has limitations. How to maximize the use of limited datasets for additive development is another problem that needs to be solved. Based on such a small sample dataset, exploring the molecular chemical space as much as possible to find additive molecular structures that can significantly improve device performance is the main challenge currently facing the development of perovskite optoelectronic device additives. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes an automated screening method for molecular additives in perovskite light-emitting diodes and photovoltaic cells, based on a deep generative model and a virtual screening approach based on a QSAR prediction model, thereby enabling the automated development of high-performance additives.

[0006] Since the perovskite films of PeLEDs and PSCs have the same ABX3 crystal structure and similar defect types, they face similar challenges in additive development. The automated additive screening scheme proposed in this invention differs in technical implementation details for PeLEDs and PSCs, but tends to be unified in logical ideas and processes. Therefore, this invention chooses to specifically demonstrate the development of high-efficiency additives in PeLEDs.

[0007] The proposed method for automated screening of molecular additives in perovskite light-emitting diodes and photovoltaic cells first utilizes a generative model fully trained on a large molecular dataset to generate a set of molecules. Then, multiple deep learning-based QSAR prediction models are used to perform multiple rounds of screening on the generated results based on the relevant properties of ideal additives, and finally, candidate molecules are obtained for device fabrication and experimental testing.

[0008] The technical solution of this invention is as follows:

[0009] The automated screening method for molecular additives in perovskite light-emitting diodes and photovoltaic cells includes the following steps:

[0010] Step 1: Establish a database of PeLED additive molecules, molecular donor number (DN), and hydrogen bond acceptor strength (β);

[0011] Step 2: Train the effectiveness prediction model, the donor quantity DN value prediction model, and the hydrogen bond acceptor strength β prediction model using the PeLED additive molecular database and the molecular donor number DN and hydrogen bond acceptor strength β databases, respectively.

[0012] Step 3: Use a pre-trained molecular generation model to generate new molecules similar to those in the PeLED additive molecular database to obtain a set of generated molecules;

[0013] Step 4: Based on the effectiveness prediction model, donor number DN value and hydrogen bond acceptor strength β prediction model trained in Step 2, perform multiple rounds of screening on the generated molecule set obtained in Step 3 with the ideal additive as the target.

[0014] Furthermore, in step 1, the PeLED additive molecular database is obtained through the following method:

[0015] We collected and established reported additive molecular structures and experimental data, including additive molecular structure names and device EQE efficiency values. Based on the molecular structures, we used the computational chemistry library Pybel and the RDKit information chemistry package to calculate the corresponding molecular descriptors and molecular fingerprint data, which together formed the original PeLED additive molecular database.

[0016] Furthermore, in step 2, the model structure of the effectiveness prediction model is a multi-input single-output multi-layer network structure; the model input consists of molecular descriptors from the original PeLED additive molecule database obtained in step 1, as well as molecular fingerprints and the implicit codes of molecules in the VAE model, which together constitute the representation features of the additive molecules; and 0 and 1 labels are assigned as label inputs according to the effectiveness results of the molecules; the model output is the effectiveness prediction probability of the molecules.

[0017] Furthermore, the model input for the effectiveness prediction model is obtained through the following process:

[0018] The CalcDescriptors() method in the RDKit library is used to calculate the values ​​of all descriptors in the Descriptors._descList list for all molecules in the PeLED additive database obtained in step 1. Descriptors with values ​​of 0 and those with standard deviations less than a set value are removed. Then, Pearson correlation coefficients are calculated for each pair of the remaining descriptors, and descriptors with correlations higher than the set value are removed. The remaining descriptors, molecular fingerprints, and the implicit codes of molecules in the VAE model are used together as the representation features of the additive molecules as input to the effectiveness prediction model.

[0019] Furthermore, in step 2, the MACCS molecular fingerprint data and molecular descriptors including the molecular most negative atomic charge, molecular HOMO energy level, and molecular HUMO-LUMO band gap from the molecular donor number DN dataset obtained in step 1 are used as model inputs, the molecular predicted DN value is used as output, and the DN value measured by standard experimental methods is used as the target variable to establish a donor number DN value prediction model.

[0020] Furthermore, in step 2, the donor number DN value prediction model is trained using five-fold cross-validation, and the average value of the model predictions in the five cross-validations is used as the final model performance.

[0021] Furthermore, in step 3, the pre-trained molecular generation model is obtained by pre-training on the ZINC dataset using a variational autoencoder.

[0022] Furthermore, the input of the variational autoencoder is the SMILES string corresponding to each molecular structure in the PeLED additive molecular database. After the SMILES string is converted into a one-hot vector, it is further input into a fully connected layer through multiple 1D convolutional layers. After standardization, a hidden code is obtained. After applying random perturbation to the hidden code, a series of new perturbation codes are obtained. After the new perturbation codes are decoded by the decoder, a new SMILES string similar to the molecular structure in the PeLED additive molecular database can be obtained.

[0023] Furthermore, the filtering in step 4 includes the following steps:

[0024] Step 4.1: Structural rationality screening:

[0025] The MolFromSmiles() method in the RDKit library is used to process the new SMILES string obtained by decoding. The SMILES string is converted into a Mol object in RDKit. During the conversion process, the chemical rationality of SMILES is checked and unreasonable parts are removed.

[0026] Step 4.2: Validity Screening:

[0027] Input the SMIELS string that can be successfully converted in step 4.1 into the effectiveness prediction model obtained in step 2 to obtain the effectiveness prediction probability value between 0 and 1. Select the molecular structures with effectiveness prediction probabilities greater than the set threshold to form the first candidate molecular set.

[0028] Step 4.3: Screening based on donor number (DN) and hydrogen bond acceptor strength (β):

[0029] The molecules in the first candidate molecule set obtained in step 4.2 are respectively input into the donor number DN value prediction model and the hydrogen bond acceptor strength β prediction model obtained in step 2, and the DN prediction value and β prediction value of each molecule are output. Molecular structures with DN values ​​and β values ​​within a certain range are selected to form the second candidate molecule set.

[0030] Step 4.4: Availability and Purchaseability Screening:

[0031] Use the PubChem API to obtain the supplier information of the molecules in the second candidate molecule set obtained in step 4.3, and select molecules with complete supplier information and prices within the set range to form the third candidate molecule set.

[0032] Furthermore, for molecules in the third candidate molecule set, device fabrication and experimental testing are carried out to obtain the test results of the corresponding devices; and the test results are added to the PeLED additive molecule database to update the prediction model training, similar molecule generation, and multiple rounds of screening aimed at ideal additive properties.

[0033] Beneficial effects

[0034] (1) As Figure 2 The present invention combines a deep generative model with a discriminative prediction model. By first generating similar structures based on existing datasets, and then performing multi-level judgment and screening on the generated results, it makes full use of the existing PeLEDs additive molecule database and achieves effective development of PeLED additive molecules under the condition of limited experimental data.

[0035] (2) The development framework based on deep learning algorithms, including the molecular property prediction model and the variational autoencoder generation model, not only showed good performance in prediction and generation tasks, but also significantly reduced time and economic costs compared with traditional experimental trial and error methods.

[0036] (3) In the screening section for high-performance additives, a method is proposed to screen high-performance additives by combining the number of Gutmann donors (DN) and the strength of Kamlet and Taft hydrogen bond acceptors (β), which can be used to quickly identify and screen high-performance additive molecules.

[0037] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0038] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0039] Figure 1 This is a flowchart of the automated screening process for molecular additives in perovskite light-emitting diodes and photovoltaic cells based on artificial intelligence, as proposed in this invention.

[0040] Figure 2 This is a detailed full-process framework diagram of the automated screening of molecular additives in perovskite light-emitting diodes and photovoltaic cells based on artificial intelligence, as proposed in this invention.

[0041] Figure 3 This is the input structure diagram of the DN value prediction model.

[0042] Figure 4 This is a schematic diagram of the process of generating similar molecules using the VAE model proposed in this invention.

[0043] Figure 5 This is a framework diagram of the multi-process screening for ideal additives.

[0044] Figure 6 This is a flowchart of the multi-process screening for ideal additives.

[0045] Figure 7 These are the performance results of different models in the process of modeling the number of molecular donors (DN) and the strength of hydrogen bond acceptors (β).

[0046] Figure 8 This is a scatter plot of molecular DN and β distributions with different efficiencies after predicting molecules in the original PeLED additive molecular database using a predictive model of donor number DN and hydrogen bond acceptor strength β. Detailed Implementation

[0047] The embodiments of the present invention are described in detail below. These embodiments are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0048] like Figure 1 As shown, the automated screening method for molecular additives in perovskite light-emitting diodes and photovoltaic cells based on artificial intelligence proposed in this invention includes the following steps:

[0049] Step 1: Build the database, including the following steps:

[0050] S100, establish a database of PeLED additive molecules, molecular donor number (DN), and hydrogen bond acceptor strength (β):

[0051] We collected and established reported additive molecular structures and experimental data by reviewing literature and other methods, including additive molecular structure names and device EQE efficiency values. Based on the molecular structures, we used the computational chemistry library Pybel and the RDKit information chemistry package to calculate the corresponding molecular descriptors, molecular fingerprints and other data, including 208 molecular descriptors and MACCS, ECFP and FP2 fingerprints, which together form the original PeLED additive molecular database.

[0052] Gutmann donor number (DN) data are provided in the literature. Due to historical reasons, the donor number (DN) value is an empirical parameter, initially defined as the negative enthalpy of the 1:1 adduct formed by a Lewis base and the standard Lewis acid SbCl5 (antimony pentachloride) in a dilute solution of the noncoordinate solvent 1,2-dichloroethane (DN is zero). Later, it was discovered that it could be obtained by linear fitting of the chemical shift using nuclear magnetic resonance (NMR). Therefore, the database contains DN values ​​obtained through these two methods, as well as descriptors and MACCS fingerprint data calculated using the RDKit chemistry package and PSI4 calculation program to form the solvent molecule donor number (DN) database. Specifically, 59 molecular descriptors, including the most negative atomic charge, HOMO level, and HOMO-LUMO band gap, and 167-bit MACCS molecular fingerprint data, were calculated and selected using Psi4 and the RDKit chemistry toolkit to construct the solvent molecule donor number (DN) dataset. The Kamlet and Taft hydrogen bond acceptor strength β aims to measure the strength of the hydrogen bond basicity or hydrogen bond acceptor strength of the solvent; the solvent hydrogen bond acceptor strength β dataset is obtained in a similar manner.

[0053] Step 2: Prediction model training:

[0054] S200 uses the original PeLED additive molecular database and the molecular donor number DN and hydrogen bond acceptor strength β databases to train the effectiveness prediction model, the donor number DN value, and the hydrogen bond acceptor strength β prediction model, respectively.

[0055] (1) Training effectiveness prediction model:

[0056] Experience shows that devices without any additives can achieve an efficiency of around 10%. Therefore, molecules that improve device efficiency by more than 10% after being added are considered effective additives. First, we use the fingerprints and descriptors of several different molecules as input features to establish an effectiveness prediction model to determine whether the additive is effective.

[0057] The effectiveness prediction model has a multi-input, single-output, multi-layer network structure.

[0058] The model input consists of molecular descriptors from the original PeLED additive molecule database obtained in step 1, along with molecular fingerprints and the implicit codes of the molecules in the VAE model, forming the representation features of the additive molecules. The molecules are labeled with 0 or 1 based on their effectiveness. The model output is the predicted effectiveness probability of the molecules.

[0059] During model training, the loss function used is binary crossentropy, and the optimization algorithm is the standard backpropagation algorithm, which aims to minimize the error between the predicted value and the label.

[0060] Specifically, in this embodiment, the CalcDescriptors() method in the RDKit library is used to calculate the values ​​of all descriptors in the Descriptors._descList list for all molecules in the original PeLED additive database obtained in step 1, and descriptors with values ​​of 0 in all additive molecules are removed. Further, descriptors with a standard deviation that is too small (less than 0.2) are removed. Then, the Pearson correlation coefficient is calculated between each pair of the remaining descriptors, and descriptors with a correlation higher than a set value are removed. The remaining descriptors, along with the molecule's MACCS, ECFP fingerprints, and the molecule's hidden code in the VAE model, are used together as the representation features of the additive molecules as input to the predictive neural network model.

[0061] The label input involves assigning "effective" and "ineffective" labels to the collected additives based on their device efficiency, using "1" and "0" as the labels. After training the model on an existing additive database, its accuracy on the validation set exceeded 90%. The trained effectiveness prediction model was then used to predict the generated molecules, and the more likely effective molecules were retained based on the prediction results.

[0062] (2) Training the predictive models for donor number DN and hydrogen bond acceptor strength β:

[0063] The prevalent halogen vacancies and other defects in perovskite films are the main reason for the limited EQE of PeLEDs, leading to nonradiative degradation. Small organic molecules containing lone pairs of electrons can control the crystallization kinetics and reduce the halogen vacancy density by forming strong coordination with lead. Studies have shown that the coordination strength between additive molecules and lead ions is related to their own electron-donating ability; molecules with strong electron-donating abilities have stronger coordination with lead ions, thus forming high-quality films. Furthermore, the strong hydrogen bonds formed between additives and organic cations affect the crystallization and passivation processes, leading to device efficiency degradation. Given the significant impact of the electron-donating ability and hydrogen-bonding ability of additives on device efficiency, we considered whether we could further screen the generated molecules based on the quantification of these two properties. We chose to use the Gutmann donor number (DN) parameter to represent the electron-donating ability of additive molecules and the Kamlet and Taft hydrogen bond acceptor strength (β) to measure the hydrogen-bonding ability of molecules. Predictive models for these two molecular properties were established, hoping to play a role in the ideal additive screening process. The model training details are as follows:

[0064] Using the MACCS molecular fingerprint data and molecular descriptors including the molecular most negative atom charge, molecular HOMO energy level, and molecular HUMO-LUMO band gap from the solvent molecule donor number DN dataset obtained in step 1 as model input, the molecular predicted DN (donor number) value as output, and the DN value measured by standard experimental methods as the target variable, a donor number DN value prediction model is established.

[0065] Specifically, in this embodiment, as Figure 3 As shown, for organic molecular solvents, the CalcDescriptors() method in the RDKit library is used to calculate descriptor values. Descriptors with values ​​of 0 in all molecules are removed, as are those with a standard deviation that is too small (not greater than 0.15). Then, correlation deduplication is performed, leaving 56 descriptors, including the maximum EState index (MaxEStateIndex), the minimum EState index (MinEStateIndex), and molecular weight. Then, the PSI4 chemical information calculation program is used to calculate three molecular property indices: the most negative atomic charge, the donor orbital energy (HOMO level), and the HOMO-LUMO band gap. These are then added to the molecular descriptor data. Normalization and interpolation are then performed to limit the numerical range of all descriptors to between -1 and 1, ensuring that the influence of each feature on the model is on the same order of magnitude. The normalized descriptor data is combined with the molecular MACCS fingerprint as the molecular representation feature.

[0066] For the established donor number DN prediction model, five-fold cross-validation was used for model training, and the average value of the model predictions in the five cross-validations was used as the final model performance.

[0067] The specific steps include: uniformly dividing the solvent molecule donor number DN dataset into five independent subsets; selecting one subset as the test set each time, and merging the other four subsets to use as the training set; training the model on the training set and evaluating its performance on the corresponding test set, repeating this process five times, each time selecting a different subset as the test set; finally, averaging the outputs of the five models to produce the final prediction output. By integrating model predictions from different test sets, the influence of random errors from a single model can be effectively reduced, thereby improving the overall stability and accuracy of the prediction results.

[0068] In this embodiment, as Figure 7 As shown in b, training was conducted on models such as Ridge Regression, Random Forest, Support Vector Machine, Gradient Boosting, and Neural Network, and the training results were observed. It was found that the prediction model based on neural networks performed best on the validation set. Specifically, the regression determination coefficient R0 was [value missing]. 2 The mean absolute error (MAE) was 0.84, the mean absolute error (MAE) was 4.21, the mean square error (MSE) was 33.59, and the root mean square error (RMSE) was 5.64. The trained neural network-based DN value prediction model was used to predict the DN values ​​of new solvents in the collected test set. The performance of the trained gradient boosting prediction model and the neural network prediction model in predicting the DN values ​​of molecules in the test set is shown in Table 1. The calculated MSE was 8.1, which is acceptable.

[0069] Table 1 shows the DN value prediction results for the molecular prediction model on the test set.

[0070]

[0071] The predictive modeling of hydrogen bond strength in solvent molecules is similar to the modeling process for the donor number (DN), using... Figure 3 A similar method was used to obtain molecular features and input them into the model, and cross-validation was also performed to evaluate and obtain the best predictive model, such as... Figure 7As shown in d, training comparisons were conducted on various architectures including Ridge Regression, Random Forest, Support Vector Machine, Gradient Boosting, eXtreme Gradient Boosting, and Neural Network. The eXtreme Gradient Boosting model performed best, with the lowest root mean square error (RMSE) of approximately 0.13.

[0072] Step 3: Generate similar new molecules:

[0073] S300 uses a pre-trained molecular generation model to generate new molecules similar to those in the original PeLED additive molecular database, thereby obtaining a set of generated molecules.

[0074] The pre-trained molecular generative model employs a variational autoencoder (VAE). A VAE is a generative model that generates data by encoding input data into a latent space and decoding it from that space to reconstruct the input data. The VAE model in this embodiment was pre-trained on the large ZINC dataset.

[0075] Based on the pre-trained variational autoencoder (VAE) model, the SMILES string corresponding to each molecular structure in the original PeLED additive molecule database is input into the encoder of the autoencoder model. After encoder encoding and compression processing, the hidden code of the original additive molecule in the 196-dimensional hidden space is obtained. Random perturbation is applied to the hidden code to obtain the perturbation code.

[0076] like Figure 4 As shown, the SMILES string is first converted into a one-hot vector, and then further input into a fully connected layer containing 196 neurons through multiple 1D convolutional layers. After standardization, a 196-dimensional hidden code is obtained. After applying random small perturbations to the hidden code, a series of new perturbation codes are obtained. After decoding the new perturbation codes, a new SMILES string with a molecular structure similar to the original PeLED additive molecular database can be obtained.

[0077] The implementation is as follows: The input implicit code is repeated multiple times to form a new data array Z. The number of repetitions is determined by the number of decoding attempts. Random noise is added by drawing values ​​from a standard normal distribution (mean 0, standard deviation 1) to ensure that each generated value is random and independent. Normalization is performed on the generated noise, including calculating the norm of the noise array and dividing each noise element by this norm. This normalization step ensures that the overall norm of the noise vector is standardized to 1, allowing adjustment of its amplitude without changing the data direction. A fixed-amplitude noise is applied to all data copies, with each copy receiving the same amount of noise. After these processing steps, the perturbed array Z is fed into the decoder, where the decoding process generates a new molecule string, SMILES.

[0078] Step 4: Perform multiple rounds of screening on the generated results with the goal of achieving ideal additive properties:

[0079] S400, the generated molecular set obtained in step 3 undergoes a multi-round screening process targeting ideal additives. For example... Figure 5 and Figure 6 The steps shown are as follows:

[0080] S111, Structural rationality screening:

[0081] The MolFromSmiles() method from the RDKit library is used to process the newly decoded SMILES string. This method converts the SMILES string into a Mol object in RDKit. During the conversion process, the chemical rationality of the SMILES is initially checked and unreasonable parts are removed.

[0082] Specifically, during the conversion process, the generated SMILES string is first checked to ensure it follows correct SMILES syntax and rules. Secondly, the connections between each atom are checked to ensure they conform to normal chemical valence states. For example, oxygen atoms typically form two bonds, carbon atoms can form four bonds, and so on. If the number of connections for an atom exceeds its common valence state, the SMILES parsing process will fail.

[0083] S112, Validity Screening:

[0084] Input the SMIELS string that can be successfully converted in S111 into the effectiveness prediction model obtained in step 2 to obtain the effectiveness prediction probability value between 0 and 1. Select molecular structures with an effectiveness prediction probability greater than 0.7, and consider these molecules to be more likely to be effective molecules that can improve device performance, thus forming the first candidate molecule set.

[0085] S113, donor number (DN) and hydrogen bond acceptor strength (β) screening:

[0086] The molecules in the first candidate molecule set obtained in S112 are input into the donor number DN value prediction model and the hydrogen bond acceptor strength β prediction model obtained in step 2, respectively, and the DN prediction value and β prediction value of each molecule are output. Molecular structures with DN values ​​and β values ​​within a certain range are selected to form the second candidate molecule set.

[0087] Since the DN value describes the electron-donating ability of a molecule, the larger the DN value of the additive, the stronger its binding ability with lead ions, which in turn affects the film morphology and crystallization process. In addition, the solvent DN value affects the stability of the precursor solution; if the DN value is too small, it is difficult to form a stable precursor. At the same time, the strong hydrogen bonds formed between the additive and the organic cations will affect the crystallization and passivation processes, leading to a deterioration in device efficiency. Therefore, the hydrogen bonding ability of the additive molecules also needs to be within a suitable range.

[0088] The properties of molecules in the original PeLED additive molecular database were predicted using a predictive model based on donor number (DN) and hydrogen bond acceptor strength (β), and the results were obtained, such as... Figure 8 The figure shows a scatter plot with the predicted donor number (DN) and hydrogen bond acceptor strength (β dB) for each molecule as the x and y axes. The color of the dots represents the corresponding device efficiency; the darker the color, the lower the device efficiency. As can be seen from the figure, high-performance additives with EQE efficiencies greater than 20% are concentrated in a certain area in the upper right corner of the figure, while the EQE efficiencies of additives in other areas are relatively low. To address this phenomenon, considering the error of the prediction model, we propose to define the region representing DN values ​​of 30 to 50 and β values ​​of 0.8 to 1.2 as the possible distribution range of high-performance additives, and eliminate generated molecules whose prediction results are not within this range. This method enables rapid screening of high-performance additives.

[0089] Therefore, in order to screen out possible high-performance additives, molecular structures with DN and β values ​​outside the specified range were selected and eliminated, and the remaining molecules formed a second candidate molecule set.

[0090] S114, Availability and Purchaseability Filter:

[0091] Using the PubChem API, obtain the supplier information of the molecules in the second candidate molecule set obtained in step S113. Select molecules with complete supplier information and prices within the set range to form the third candidate molecule set, with the final number being approximately 10.

[0092] The molecules from the obtained third candidate molecule set were used for device fabrication and experimental testing, yielding JV curves and other test results for the corresponding devices. These test results were then added to the original PeLED additive molecule database to update the prediction model training, similar molecule generation, and multiple rounds of screening aimed at achieving ideal additive properties.

[0093] In summary, this invention can be used for the development and design of PeLED additive molecules. The well-known implementation methods and characteristics described above are not elaborated upon here. It should be noted that those skilled in the art can make several modifications without departing from this invention, and these modifications should also be considered within the scope of protection of this invention. These modifications will not affect the implementation effect or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of the claims, and the specific embodiments described in the specification are used to interpret the content of the claims.

Claims

1. An automated screening method for molecular additives in perovskite light-emitting diodes and photovoltaic cells, characterized in that: Includes the following steps: Step 1: Establish a database of PeLED additive molecules, molecular donor number (DN), and hydrogen bond acceptor strength (β); Step 2: Train the effectiveness prediction model, the donor quantity DN value prediction model, and the hydrogen bond acceptor strength β prediction model using the PeLED additive molecular database and the molecular donor number DN and hydrogen bond acceptor strength β databases, respectively. Step 3: Using a pre-trained molecular generation model, similar new molecules are generated from molecules in the PeLED additive molecular database to obtain a set of generated molecules. The molecular generation model uses a variational autoencoder. The input of the variational autoencoder is the SMILES string corresponding to each molecular structure in the PeLED additive molecular database. After the SMILES string is converted into a one-hot vector, it is further input into a fully connected layer through multiple 1D convolutional layers. After standardization, a hidden code is obtained. After applying random perturbation to the hidden code, a series of new perturbation codes are obtained. After decoding the new perturbation codes, a new SMILES string similar to the molecular structure in the PeLED additive molecular database can be obtained. Step 4: Based on the effectiveness prediction model, donor number DN value and hydrogen bond acceptor strength β prediction model trained in Step 2, perform multiple rounds of screening on the generated molecule set obtained in Step 3 with the ideal additive as the target. The screening process includes the following steps: Step 4.1: Structural rationality screening: The MolFromSmiles() method in the RDKit library is used to process the new SMILES string obtained by decoding. The SMILES string is converted into a Mol object in RDKit. During the conversion process, the chemical rationality of SMILES is checked and unreasonable parts are removed. Step 4.2: Validity Screening: Input the SMIELS string that can be successfully converted in step 4.1 into the effectiveness prediction model obtained in step 2 to obtain the effectiveness prediction probability value between 0 and 1. Select the molecular structures with effectiveness prediction probabilities greater than the set threshold to form the first candidate molecular set. Step 4.3: Screening based on donor number (DN) and hydrogen bond acceptor strength (β): The molecules in the first candidate molecule set obtained in step 4.2 are respectively input into the donor number DN value prediction model and the hydrogen bond acceptor strength β prediction model obtained in step 2, and the DN prediction value and β prediction value of each molecule are output. Molecular structures with DN values ​​and β values ​​within a certain range are selected to form the second candidate molecule set. Step 4.4: Availability and Purchaseability Screening: Use the PubChem API to obtain the supplier information of the molecules in the second candidate molecule set obtained in step 4.3, and select molecules with complete supplier information and prices within the set range to form the third candidate molecule set.

2. The automated screening method for molecular additives in perovskite light-emitting diodes and photovoltaic cells according to claim 1, characterized in that: In step 1, the PeLED additive molecular database is obtained in the following way: We collected and established reported additive molecular structures and experimental data, including additive molecular structure names and device EQE efficiency values. Based on the molecular structures, we used the computational chemistry library Pybel and the RDKit information chemistry package to calculate the corresponding molecular descriptors and molecular fingerprint data, which together formed the original PeLED additive molecular database.

3. The automated screening method for molecular additives in perovskite light-emitting diodes and photovoltaic cells according to claim 1, characterized in that: In step 2, the effectiveness prediction model has a multi-input, single-output multi-layer network structure. The model input consists of molecular descriptors from the original PeLED additive molecule database obtained in step 1, as well as molecular fingerprints and the implicit codes of molecules in the VAE model, which together constitute the representation features of the additive molecules. Furthermore, the effectiveness results of the molecules are assigned 0 or 1 labels as label inputs. The model output is the effectiveness prediction probability of the molecules.

4. The automated screening method for molecular additives in perovskite light-emitting diodes and photovoltaic cells according to claim 3, characterized in that: The model input for the effectiveness prediction model is obtained through the following process: The CalcDescriptors() method in the RDKit library is used to calculate the values ​​of all descriptors in the Descriptors._descList list for all molecules in the PeLED additive database obtained in step 1. Descriptors with values ​​of 0 and those with standard deviations less than a set value are removed. Then, Pearson correlation coefficients are calculated for each pair of the remaining descriptors, and descriptors with correlations higher than the set value are removed. The remaining descriptors, molecular fingerprints, and the implicit codes of molecules in the VAE model are used together as the representation features of the additive molecules as input to the effectiveness prediction model.

5. The automated screening method for molecular additives in perovskite light-emitting diodes and photovoltaic cells according to claim 1, characterized in that: In step 2, the MACCS molecular fingerprint data and molecular descriptors including the molecular most negative atomic charge, molecular HOMO energy level, and molecular HUMO-LUMO band gap from the molecular donor number DN dataset obtained in step 1 are used as model inputs, the molecular predicted DN value is used as output, and the DN value measured by standard experimental methods is used as the target variable to establish a donor number DN value prediction model.

6. The automated screening method for molecular additives in perovskite light-emitting diodes and photovoltaic cells according to claim 1, characterized in that: In step 2, the donor number DN value prediction model is trained using five-fold cross-validation, and the average value of the model predictions in the five cross-validations is used as the final model performance.

7. The automated screening method for molecular additives in perovskite light-emitting diodes and photovoltaic cells according to claim 1, characterized in that: In step 3, the molecular generation model is pre-trained on the ZINC dataset.

8. The automated screening method for molecular additives in perovskite light-emitting diodes and photovoltaic cells according to claim 1, characterized in that: For molecules in the third candidate molecule set, device fabrication and experimental testing are carried out to obtain the test results of the corresponding devices; the test results are then added to the PeLED additive molecule database to update the prediction model training, similar molecule generation, and multiple rounds of screening aimed at ideal additive properties.

Images