A method and system for predicting dihedral torsional potential of an organic solar cell material

Abstract

The present application belongs to the technical field of organic solar cell material data mining, in particular to a method and system for predicting dihedral angle torsional potential of organic solar cell material. The method comprises the following steps: S1: initial data set establishment; S2: data preprocessing; S3: machine learning algorithm model establishment and evaluation; S4: verification; S5: explainability analysis; S6: molecular design and dihedral angle torsional potential prediction. The present application uses a machine learning model to replace traditional DFT flexibility scanning, and the prediction time of dihedral angle torsional potential of a single molecule is shortened from several hours or days to milliseconds, which significantly reduces the consumption of computing resources, making it possible to perform high-throughput screening on large-scale molecular libraries, and has the potential to expand to a wider range of organic optoelectronic material fields such as organic light-emitting diode materials and organic field-effect transistor materials, and provides a general technical platform for data-driven design of the whole chain of organic optoelectronic materials.

Description

Technical Field

[0001] This invention relates to the field of data mining technology for organic solar cell materials, specifically to a method and system for predicting the dihedral torsional potential energy of organic solar cell materials. Background Technology

[0002] Organic solar cells (OSCs) have become a research hotspot in next-generation photovoltaic technology due to their advantages such as light weight, flexibility, and solution-processability. The photoelectric conversion efficiency of OSCs is closely related to the microscopic aggregated structure of donor / acceptor molecules in the active layer. Among them, molecular conformation is the key bridge connecting molecular structure and macroscopic morphology. Specifically, the torsional barrier of key dihedral angles within the molecule directly determines the planarity and conformational freedom of the molecule, thereby affecting intermolecular π-π packing, phase separation scale, and charge transport efficiency. Therefore, accurate analysis and prediction of dihedral angle torsional potential energy is a prerequisite for the rational design and high-throughput screening of OSC materials.

[0003] Currently, the mainstream method for obtaining the dihedral torsional potential energy (dihedral surface) of molecules is based on density functional theory (DFT) calculations in quantum chemistry. A flexible scanning strategy is typically employed, where the target dihedral angle in the molecule is fixed, and the energies of each conformation are calculated progressively, ultimately fitting a complete potential energy curve. While the DFT method can provide highly accurate energy information, it suffers from the following significant drawbacks in practical applications: (1) The computational cost is extremely high and the efficiency is low. For a typical OSC molecule containing hundreds of atoms, a complete DFT flexible scan requires huge computational resources and time, which severely restricts its application in high-throughput screening and large-scale virtual computing of materials. This core bottleneck has become the main obstacle to improving the efficiency of OSC material research and development.

[0004] (2) Lack of interpretability, making it difficult to guide molecular design. Traditional DFT calculations can only output numerical results (such as energy values ​​and energy barrier heights), which is a "black box" operation. It cannot reveal the intrinsic molecular structural factors that determine the height of the dihedral torsion energy barrier. Although researchers know the results, they cannot know "which functional group or electronic effect caused the result?", thus failing to provide clear and physicochemically significant guidance for subsequent rational molecular modification and optimization.

[0005] In recent years, machine learning (ML), as a powerful data-driven tool, has been gradually applied to the field of materials science. For example, existing studies have used ML models to establish the correlation between molecular structure and the optoelectronic properties of devices, verifying the feasibility of this paradigm. However, these studies mainly focus on the direct correlation between "molecular structure and macroscopic properties," while research on the more fundamental and underlying physicochemical link of "molecular structure and conformational energy" is relatively scarce.

[0006] More importantly, existing approaches to applying machine learning (ML) to predict molecular properties often lack effective interpretation of the model's predictions. While models such as decision trees, support vector machines, and even deep neural networks can make predictions, their internal mechanisms are complex, making it difficult for researchers to extract clear physicochemical laws. This limits a deeper understanding of the underlying physical mechanisms and further optimization of molecular design strategies. Summary of the Invention

[0007] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0008] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution: A method for predicting the dihedral torsional potential energy of an organic solar cell material includes the following steps: S1: Initial Dataset Establishment: Organic solar cell donors and non-fullerene acceptor molecules with high photoelectric conversion efficiency reported in the literature were obtained. After fragmentation and recombination, a diverse molecular fragment library was obtained as the first-stage dataset. The molecular fragments in the molecular fragment library were subjected to dihedral flexible scanning using density functional theory to obtain torsional potential energy curves as training labels. With molecular structure information as the input independent variable and torsional potential energy value as the output dependent variable, the second-stage dataset was established. The first-stage dataset and the second-stage dataset together are referred to as the initial dataset. S2: Data preprocessing: Preprocess the initial dataset to obtain the preprocessed first-stage and second-stage datasets; S3: Machine Learning Algorithm Model Building and Evaluation: Build different machine learning algorithm models, use the preprocessed first-stage dataset to train and test different machine learning algorithm models, obtain the optimal results and save the model weights, then transfer the model weights, and train and test the preprocessed second-stage dataset based on the first-stage model weights, evaluate the performance of each machine learning algorithm model, and select the optimized machine learning algorithm model. S4: Validation: Validate organic photovoltaic materials and materials molecules in other fields using optimized machine learning models; S5: Interpretability Analysis: Utilize optimized machine learning algorithm models to perform interpretability analysis on the preprocessed dataset; S6: Molecular Design and Dihedral Torsional Potential Energy Prediction: Molecular design is performed based on interpretability results, and the dihedral torsional potential energy parameters of the designed molecules are predicted using an optimized machine learning algorithm model.

[0009] As a preferred embodiment of the method for predicting the dihedral torsional potential energy of an organic solar cell material according to the present invention, the molecular fragment library in S1 is constructed by fragment cutting and recombination of known high-efficiency OPV molecules and then screening for deduplication; the dihedral flexible scanning is performed in the range of 0° to 180° at the theoretical level of B97XD / 6-31G(d,p), with a step size of 10°.

[0010] As a preferred embodiment of the method for predicting the dihedral torsional potential energy of an organic solar cell material according to the present invention, the molecular structure information in step S1 is a molecular descriptor generated by cheminformatics tools, including two-dimensional structural fingerprints, physicochemical descriptors, and conformational features.

[0011] As a preferred embodiment of the method for predicting the dihedral torsional potential energy of an organic solar cell material according to the present invention, step S2, the preprocessing of the initial dataset includes: removing meaningless label columns, filling missing values, truncating extreme outliers in the label columns using the quantile method, and randomly dividing the dataset into training and test sets by stratified sampling according to a preset ratio.

[0012] As a preferred embodiment of the method for predicting the dihedral torsional potential energy of an organic solar cell material according to the present invention, the step S3, the step of establishing and evaluating the machine learning algorithm model, specifically includes: establishing different machine learning algorithm models using XGBoost, random forest, extreme random tree, decision tree, neural network, K-nearest neighbor and support vector machine algorithms; The preprocessed datasets for the first and second stages are divided into training, validation, and test sets. On the first-stage dataset, different machine learning algorithm models are trained and tested using the training, validation, and test sets to optimize each model and save the optimal weights. Then, on the second-stage dataset, the weights saved from the first stage are loaded and fine-tuned using the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). 2 Evaluate the performance of each machine learning algorithm model, and perform minimum MAE, RMSE, and maximum R-value tests on the test set. 2 The value of the machine learning algorithm model is used as an optimized machine learning algorithm model for subsequent predictions; The formula for calculating MAE is as follows:

[0013] The formula for calculating RMSE is as follows:

[0014] Calculate the coefficient of determination R 2 The formula is as follows:

[0015] in yi This represents the experimental dataset, specifically the i-th value of the true dataset. xi This represents the prediction dataset, specifically the i-th predicted value. This represents the average value of the experimental dataset. n This indicates the number of data points in the experimental dataset.

[0016] As a preferred embodiment of the method for predicting the dihedral torsional potential energy of organic solar cell materials according to the present invention, the specific method of step S4 is as follows: collecting molecular structure information of organic photovoltaic materials and other fields, including organic light-emitting diodes, organic photodetectors, and organic field-effect transistors, from the literature; using an optimized machine learning algorithm model to predict the dihedral torsional potential energy; and comparing it with the true value obtained by quantum chemical calculation to verify the reliability of the machine learning model.

[0017] As a preferred embodiment of the method for predicting the dihedral torsional potential energy of an organic solar cell material according to the present invention, the specific method in step S5 is as follows: using the SHapley Additive ex Planations method to perform interpretability analysis on the selected optimized machine learning algorithm model, quantitatively calculating the contribution of each molecular descriptor, and identifying key molecular structural features that have a positive or negative correlation with the dihedral torsional potential energy.

[0018] As a preferred embodiment of the method for predicting the dihedral torsional potential energy of an organic solar cell material according to the present invention, the specific method in step S6 is as follows: molecular design is performed by screening key molecular structural features to construct an organic solar cell material; structural descriptors of the designed material molecules are obtained using cheminformatics tools; based on the structural descriptors, an optimized machine learning algorithm is used to predict them to obtain the corresponding output values ​​of the material molecules of the organic solar cell. The output values ​​include dihedral torsional potential energy values ​​or complete torsional potential energy curves.

[0019] As a preferred embodiment of the method for predicting the dihedral torsional potential energy of an organic solar cell material according to the present invention, after completing the molecular design and dihedral torsional potential energy prediction steps, corresponding experimental verification is carried out on the organic solar cell material molecules selected based on the prediction results of the optimized machine learning algorithm model. Corresponding theoretical verification is carried out using device simulation, first-principles calculation and related theories to evaluate the accuracy of the optimized machine learning algorithm model.

[0020] A system for predicting the dihedral torsional potential energy of an organic solar cell material, comprising: The data construction module is used to obtain the initial dataset and perform preprocessing. The model training and evaluation module is used to build and train various machine learning algorithm models, and select the optimal model through transfer learning and performance evaluation. The validation module is used to validate the generalization ability of the optimized model. The interpretability analysis module is used to identify key molecular structural features using the SHAP method; The prediction and design module is used to design molecules based on interpretability results and predict their dihedral torsional potential.

[0021] Compared with the prior art, the beneficial effects of the present invention are: 1. Significantly reduce computational costs and improve prediction efficiency: This invention uses a machine learning model to replace the traditional DFT flexible scanning, reducing the prediction time for the dihedral torsional potential energy of a single molecule from hours or days to milliseconds. This significantly reduces the consumption of computational resources, making high-throughput screening of large-scale molecular libraries possible and breaking through the efficiency bottleneck of traditional methods.

[0022] 2. High prediction accuracy and strong reliability: By constructing a high-quality, highly consistent dataset, selecting multi-dimensional molecular descriptors with physicochemical significance, and systematically comparing and optimizing the hyperparameters of multiple regression models, the prediction results of this invention have good consistency with the DFT calculation values, which can meet the accuracy requirements of conformational energy analysis and material screening.

[0023] 3. Exhibiting interpretability and enabling rational design: By introducing SHAP interpretability analysis, this invention overcomes the shortcomings of traditional machine learning "black box" models, enabling qualitative and quantitative identification of key molecular structural features affecting dihedral torsional potential energy and revealing their regulatory mechanisms. Researchers can then target molecular structure modifications based on these key features (such as optimizing steric hindrance, adjusting electron distribution, and controlling hydrogen bonding) to design high-performance OPV materials with ideal dihedral conformations, achieving a paradigm shift from "experience-based trial and error" to "precise control."

[0024] 4. The method has strong generalizability and broad application prospects: The dataset and model framework constructed in this invention are not only applicable to existing OPV donor / acceptor molecular systems, but its methodology is also applicable to conformational energy prediction of other organic conjugated materials. By expanding the dataset and optimizing the descriptor system, this method has the potential to be extended to a wider range of organic optoelectronic materials such as organic light-emitting diode materials and organic field-effect transistor materials, providing a general technical platform for data-driven design of the entire organic optoelectronic materials chain. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart illustrating the technical process of predicting the dihedral torsional potential energy of an organic solar cell material based on machine learning and high-throughput first-principles calculations, as described in this invention. (a) involves collecting and labeling 584 organic ligands and performing first-principles calculations; (b) involves constructing seven regression models and completing model training and performance evaluation; (c) involves parsing key feature descriptors using SHAP feature importance values; and (d) involves conducting validation tests using the trained models. Figure 2 Here is a table of dihedral torsional potential energy descriptors for a method of predicting dihedral torsional potential energy in an organic solar cell material according to the present invention: Figure 3 This is a descriptor correlation diagram for the prediction method of dihedral torsional potential energy of an organic solar cell material according to the present invention. (a) is the Pearson correlation matrix of 14 feature descriptors, where the values ​​in the squares represent the Pearson correlation coefficient (PCC), used to indicate the positive or negative correlation between the features; (b) is the principal component analysis (PCA) of the training sample distribution, where red dots represent high energy values, pink dots represent medium energy values, and blue dots represent low energy values. Figure 4 The diagram shows the model performance evaluation of the prediction method for the dihedral torsional potential energy of an organic solar cell material according to the present invention. (a) is a scatter plot of the predicted and actual values ​​and a consistency / correlation analysis diagram; (b) is a residual distribution and a normal fitting curve diagram. Detailed Implementation

[0026] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0027] Secondly, the present invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of the present invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not according to the usual scale. Furthermore, the schematic diagrams are merely examples and should not limit the scope of protection of the present invention. In addition, actual fabrication should include three-dimensional spatial dimensions of length, width, and depth.

[0028] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0029] A machine learning-based method for predicting the dihedral torsional potential energy of organic photovoltaic materials; the overall workflow of which can be found in [reference needed]. Figure 1 The method includes the following specific steps: Step 1: Construction and Preprocessing of a High-Quality Dataset: Organic photovoltaic donor molecules and non-fullerene acceptor molecules with high photoelectric conversion efficiency were collected from the literature. Density functional theory (DFT) calculations were used to fragment and recombine the molecules, followed by rigorous deduplication screening to construct a basic molecular library. For the key dihedral angles of all molecular fragments, a flexible scan from 0° to 180° (step size 10°) was performed using DFT. High-precision torsional potential energy curves were obtained by calculating the energy under each torsional configuration, which were used as the true label values ​​(HF values) for subsequent machine learning model training. Next, the original data was preprocessed, including cleaning outliers and erroneous data from the initial dataset, resulting in a preprocessed dataset.

[0030] Step 2: Generation and Optimization of Multi-Dimensional Molecular Descriptors: Descriptor Calculation: Using the RDKit cheminformatics toolkit and the Mordred descriptor calculation tool, a comprehensive structural characterization of each molecular fragment is performed to generate an initial set of molecular descriptors. Then, post-processing operations are performed on the descriptors.

[0031] See Figure 2The model uses 14 descriptors. The descriptor matrix X constructed in this application is not a simple list, but a 14-dimensional vector formed after deep feature engineering. Its design follows the multi-dimensional fusion principle of "structure-physicochemical-conformation", aiming to capture the intrinsic molecular properties that affect the dihedral torsional potential energy in all aspects. According to its regulation mechanism on dihedral conformation, these 14 descriptors can be systematically divided into three categories: compositional feature descriptors, structural property descriptors, and topological property descriptors. Compositional feature descriptors: NsNH2, SsOH; Structural property descriptors: Cos Angle, MATS5m, MDEC-23; Topological property descriptors: ATS8s, Xp-4dv, AATSC6Z, MATS5are, AATSC2p, SMR VSA2 file1, AATSC4v, MATS3pe, ATSC7pe. These three types of descriptors, from different scales and physicochemical levels, synergistically constitute a high-precision characterization of molecular conformation-energy relationships.

[0032] To explore the correlation of the model's descriptors, see [reference needed]. Figure 3 (a) A Pearson correlation heatmap of the descriptors used in the model to validate the model. This visualization helps identify strong linear correlations between descriptors, which may lead to redundancy in the dataset. By calculating the Pearson correlation coefficient (PCC) for each pair of descriptors, we can intuitively understand their linear relationship. Descriptors with a PCC greater than 0.90 are considered highly correlated.

[0033] To further verify the rationality and representativeness of the distribution of samples corresponding to the optimized descriptor set in the feature space, refer to... Figure 3 (b) To verify the principal component analysis (PCA) plot of the descriptors used in the model, a two-dimensional scatter plot was used to show the distribution of all data samples. Light blue represents low HF values, pink represents medium HF values, and red represents high HF values. From the distribution characteristics, the samples of different energy value categories did not show obvious category clustering or spatial isolation, but covered the PCA space in an interleaved and dispersed manner. On the one hand, there is no bias of local concentration of samples in the space, indicating that the dataset has sufficient coverage of organic donor-acceptor molecules of different energy levels, which fully proves that the dataset is representative and very suitable for subsequent analysis.

[0034] Step 3: Construction and Training of Machine Learning Models: Seven types of machine learning algorithms are employed: Extreme Gradient Boosting Tree (XGBoost), Extreme Random Trees (ExtraTrees), Random Forest (RandomForest), Decision Tree (DecisionTree), Neural Network (NeuralNetwork), K-Nearest Neighbors (KNN), and Support Vector Machine (SVM). The preprocessed datasets for the first and second stages are divided into training, validation, and test sets. On the first stage dataset, different machine learning algorithm models are trained and tested using the training, validation, and test sets to optimize each model and save the optimal weights. Then, on the second stage dataset, the weights saved from the first stage are transferred for fine-tuning. The performance of each machine learning algorithm model is evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Coefficient of Determination (R²). On the test set, the machine learning algorithm model with the lowest MAE, RMSE, and highest R² value is selected as the optimized model for subsequent predictions.

[0035] Step 4: Model Interpretability Analysis - SHAP Analysis: This step aims to overcome the "black box" nature of machine learning models, quantify the correlation mechanism between descriptors and prediction targets, and apply the SHAP (SHapley Additive Explanations) method to interpret the optimal model after training. Step 5: Rapidly predict and verify the dihedral torsional potential energy of the target molecule.

[0036] See Figure 4 ,in, Figure 4-a shows a scatter plot of predicted and actual values ​​and the fitted regression line. The black dashed line in the figure is the ideal fitted line (y=x), representing the case where the predicted value is exactly equal to the actual value; the red solid line is the linear regression fitted line based on the data points, used to visually show the overall bias trend of the model's predictions. The key performance indicators in the figure are as follows: the consistency correlation coefficient (CCC) is 0.77, indicating that the predicted value and the actual value have high consistency; the coefficient of determination (R2) is 0.59, indicating that the model can explain about 59% of the data variability; the root mean square error (RMSE) is 2.51, reflecting the overall level of deviation between the predicted value and the actual value; the mean absolute error (MAE) is 2.36, characterizing the average magnitude of the prediction error. The scatter points are mainly around the y=x line, indicating that the model has good prediction accuracy. Figure Xb shows the frequency distribution histogram of the prediction residuals and the normal fitted curve. The residuals are calculated by subtracting the actual value from the predicted value. The blue bars in the figure represent the sample frequency for each residual interval, and the red curve is the normal distribution curve (mean u = -0.58, standard deviation 0 = 1.00) based on the sample residual fit. As shown in the figure, the residual distribution approximates a symmetrical bell-shaped (normal distribution) curve centered at zero, indicating that the model's prediction error has no obvious systematic bias and the error distribution is relatively concentrated, further confirming that the regression model has good fit and stability. This visual analysis and quantitative indicators jointly verify the reliability and robustness of the machine learning prediction model constructed in this application.

[0037] Specifically, step 1 above includes: collecting organic photovoltaic donor molecules and non-fullerene acceptor molecules with high photoelectric conversion efficiency from the literature; fragmenting and recombination of the collected molecules using density functional theory (DFT) calculations; and rigorous deduplication screening to construct a high-quality basic molecular library. For the key dihedral angles of all molecular fragments, DFT is used for flexible scanning from 0° to 180° (step size 10°). By calculating the energy under each torsional configuration, a high-precision torsional potential energy curve is obtained, which is used as the true label value (HF value) for subsequent machine learning model training.

[0038] Specifically, step 2 above includes: using a cheminformatics toolkit to calculate the physicochemical descriptors of molecular fragments, and screening and optimizing the initial set of molecular descriptors. Highly relevant redundant descriptors are removed.

[0039] Specifically, step 3 above includes: using Python as the machine learning platform to build various machine learning computation models, such as extreme gradient boosting trees, extreme random trees, random forests, decision trees, artificial neural networks, K-nearest neighbors, and support vector machines. The preprocessed dataset is divided into training, validation, and test sets according to a preset ratio. Each model is optimized, and then, on the second-stage dataset, the weights saved from the first stage are transferred for fine-tuning. The mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) are used to evaluate the performance. 2 Evaluate the performance of each model. The formula for calculating the MAE index is as follows:

[0040] The formula for calculating RMSE is as follows:

[0041] Calculate the coefficient of determination R 2 The formula is as follows:

[0042] Where yi represents the experimental dataset, i.e., the i-th true value, and xi represents the prediction dataset, i.e., the i-th predicted value. represents the average value of the experimental dataset, and n represents the number of points in the experimental dataset.

[0043] Specifically, step 4 above includes: using the SHAP (SHapley Additive exPlanations) method to perform interpretability analysis on the trained optimized machine learning algorithm model. Specifically, step 5 above includes: using the saved optimized machine learning algorithm model to predict the structure of a new target molecule. The model can quickly output the predicted HF values ​​of the dihedral angle of the molecule at various torsion angles within milliseconds, thereby rapidly generating the complete dihedral torsion potential energy curve of the target molecule.

[0044] This invention also provides a system for predicting the dihedral torsional potential energy of organic solar cell materials, comprising: The data construction module is used to obtain the initial dataset and perform preprocessing. The model training and evaluation module is used to build and train various machine learning algorithm models, and select the optimal model through transfer learning and performance evaluation. The validation module is used to validate the generalization ability of the optimized model. The interpretability analysis module is used to identify key molecular structural features using the SHAP method; The prediction and design module is used to design molecules based on interpretability results and predict their dihedral torsional potential.

[0045] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for predicting the dihedral torsional potential energy of an organic solar cell material, characterized in that, Includes the following steps: S1: Initial Dataset Establishment: Organic solar cell donors and non-fullerene acceptor molecules with high photoelectric conversion efficiency reported in the literature were obtained. After fragmentation and recombination, a diverse molecular fragment library was obtained as the first-stage dataset. The molecular fragments in the molecular fragment library were subjected to dihedral flexible scanning using density functional theory to obtain torsional potential energy curves as training labels. With molecular structure information as the input independent variable and torsional potential energy value as the output dependent variable, the second-stage dataset was established. The first-stage dataset and the second-stage dataset together are referred to as the initial dataset. S2: Data preprocessing: Preprocess the initial dataset to obtain the preprocessed first-stage and second-stage datasets; S3: Machine Learning Algorithm Model Building and Evaluation: Build different machine learning algorithm models, use the preprocessed first-stage dataset to train and test different machine learning algorithm models, obtain the optimal results and save the model weights, then transfer the model weights, and train and test the preprocessed second-stage dataset based on the first-stage model weights, evaluate the performance of each machine learning algorithm model, and select the optimized machine learning algorithm model. S4: Validation: Validate organic photovoltaic materials and materials molecules in other fields using optimized machine learning models; S5: Interpretability Analysis: Utilize optimized machine learning algorithm models to perform interpretability analysis on the preprocessed dataset; S6: Molecular Design and Dihedral Torsional Potential Energy Prediction: Molecular design is performed based on interpretability results, and the dihedral torsional potential energy parameters of the designed molecules are predicted using an optimized machine learning algorithm model.

2. The method for predicting the dihedral torsional potential energy of an organic solar cell material according to claim 1, characterized in that, The molecular fragment library in S1 is constructed by fragment cutting and recombination of known high-efficiency OPV molecules and then deduplication screening; the dihedral angle flexible scanning is performed in the range of 0° to 180° at the theoretical level of B97XD / 6-31G(d,p), with a step size of 10°.

3. The method for predicting the dihedral torsional potential energy of an organic solar cell material according to claim 1, characterized in that, In step S1, the molecular structure information is a molecular descriptor generated by cheminformatics tools, including two-dimensional structural fingerprints, physicochemical descriptors, and conformation-related features.

4. The method for predicting the dihedral torsional potential energy of an organic solar cell material according to claim 1, characterized in that, In step S2, the preprocessing of the initial dataset includes: removing meaningless label columns, filling in missing values, truncating extreme outliers in the label columns using the quantile method, and randomly dividing the dataset into training and test sets by stratified sampling according to a preset ratio.

5. The method for predicting the dihedral torsional potential energy of an organic solar cell material according to claim 1, characterized in that, In step S3, the machine learning algorithm model establishment and evaluation steps specifically include: establishing different machine learning algorithm models using XGBoost, random forest, extreme random tree, decision tree, neural network, K-nearest neighbor and support vector machine algorithms; The preprocessed datasets for the first and second stages are divided into training, validation, and test sets. On the first-stage dataset, different machine learning algorithm models are trained and tested using the training, validation, and test sets to optimize each model and save the optimal weights. Then, on the second-stage dataset, the weights saved from the first stage are loaded and fine-tuned using the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). 2 Evaluate the performance of each machine learning algorithm model, and perform minimum MAE, RMSE, and maximum R-value tests on the test set. 2 The value of the machine learning algorithm model is used as an optimized machine learning algorithm model for subsequent predictions; The formula for calculating MAE is as follows: ； The formula for calculating RMSE is as follows: ； Calculate the coefficient of determination R 2 The formula is as follows: ； in yi This represents the experimental dataset, specifically the i-th value of the true dataset. xi This represents the prediction dataset, specifically the i-th predicted value. This represents the average value of the experimental dataset. n This indicates the number of data points in the experimental dataset.

6. The method for predicting the dihedral torsional potential energy of an organic solar cell material according to claim 1, characterized in that, The specific method of step S4 is as follows: collect material molecular structure information of organic photovoltaic materials and other fields, including organic light-emitting diodes, organic photodetectors, and organic field-effect transistors, from the literature; use an optimized machine learning algorithm model to predict the dihedral torsional potential energy; and compare it with the real value obtained by quantum chemical calculation to verify the reliability of the machine learning model.

7. The method for predicting the dihedral torsional potential energy of an organic solar cell material according to claim 1, characterized in that, The specific method in step S5 is to use the SHapley Additive ex Planations method to perform interpretability analysis on the selected optimized machine learning algorithm model, quantitatively calculate the contribution of each molecular descriptor, and identify key molecular structural features that are positively or negatively correlated with the dihedral torsional potential energy.

8. The method for predicting the dihedral torsional potential energy of an organic solar cell material according to claim 1, characterized in that, The specific method in step S6 is as follows: molecular design is performed through the selected key molecular structural features to construct organic solar cell materials; structural descriptors of the designed material molecules are obtained using cheminformatics tools; based on the structural descriptors, an optimized machine learning algorithm is used to predict them to obtain the corresponding output values ​​of the material molecules of the organic solar cell. The output values ​​include dihedral torsional potential energy values ​​or complete torsional potential energy curves.

9. The method for predicting the dihedral torsional potential energy of an organic solar cell material according to claim 8, characterized in that, After completing the molecular design and dihedral torsional potential energy prediction steps, corresponding experimental verification was carried out on the organic solar cell material molecules selected based on the prediction results of the optimized machine learning algorithm model. Theoretical verification was carried out using device simulation, first-principles calculations and related theories to evaluate the accuracy of the optimized machine learning algorithm model.

10. A system for predicting the dihedral torsional potential energy of an organic solar cell material, used to implement the method for predicting the dihedral torsional potential energy of an organic solar cell material according to any one of claims 1-9, characterized in that, include: The data construction module is used to obtain the initial dataset and perform preprocessing. The model training and evaluation module is used to build and train various machine learning algorithm models, and select the optimal model through transfer learning and performance evaluation. The validation module is used to validate the generalization ability of the optimized model. The interpretability analysis module is used to identify key molecular structural features using the SHAP method; The prediction and design module is used to design molecules based on interpretability results and predict their dihedral torsional potential.

Images