Fast and accurate screening framework for organic solar cells based on molecular structure and deep learning

doi:10.21203/rs.3.rs-2730424/v1

Download PDF

Article

Fast and accurate screening framework for organic solar cells based on molecular structure and deep learning

https://doi.org/10.21203/rs.3.rs-2730424/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 23 Oct, 2023

Read the published version in npj Computational Materials →

Version 1

posted

You are reading this latest preprint version

Organic photovoltaics have attracted worldwide interest due to their unique advantages in developing low-cost, light-weight, and flexible power sources. Functional molecular design and synthesis have been put forward to accelerate the discovery of ideal organic semiconductors, however, it is extremely expensive to conduct experimental screening of the wide organic compound space. Here we develop a framework by combing deep learning model (graph neural network) and ensemble learning model (light Gradient Boosting Machine), which enables rapid and accurate screening of OPV molecules. This framework establishes the relationship between molecular structure, molecular properties, and device efficiency. Our framework evaluates from the chemical structure of the OPV molecules directly and accurately. Since it does not involve DFT calculations, it makes fast predictions. The reliability of our framework is verified with data from previous reports and our newly synthesized organic molecules. Our work provides an efficient method for developing new organic optoelectronic materials.

Physical sciences/Materials science/Materials for energy and catalysis/Solar cells

Physical sciences/Materials science/Theory and computation

Organic semiconducting materials exhibit great synthetic flexibility, which allows for excellent tunability over the bandgap, energy level, and carrier mobility, offering great potential in the design of efficient optoelectronic devices like organic solar cells (OSCs). In comparison with inorganic counterparts, OSCs show unique advantages like light-weight, good flexibility, semi-transparency etc.^1–3. Advances in the last decade in functional materials design, morphology optimization and device architecture engineering have led to certified power conversion efficiencies (PCEs) of over 19%, demonstrating great potential for emerging photovoltaic technology. However, exploring suitable organic molecules in the vast organic compound space is extremely difficult, and efficiency breakthrough in the lab needs the constant input of intensive labor and time.

Although DFT calculations allow us to acquire many electronic structural properties of organic molecules without complex organic synthesis, we still lack an effective mathematical model to calculate the PCEs directly from the physical properties of the molecules^4–5. In addition, although DFT calculations save economic costs, the huge time cost still limits their application in the high-throughput screening of molecules. Therefore, it’s an urgent problem to establish a Quantitative Structure-Property Relationship (QSPR) model that can conduct the high-throughput screening of the organic compound space to find more suitable molecules.

As a powerful technology for mining relationships hidden in big data, artificial intelligence has brought great development and prosperity to the field of machine learning^6–8. With the development of material informatics, a new generation of material research and development paradigms is gradually formed: 1) Use the material database to train machine learning models. 2) Use the model to predict new materials. 3) Verify the results with experiments or calculations^9–11. Machine learning shows an excellent performance in accelerating the discovery of new materials, guiding the design of new materials, and exploring the QSPR of materials^12–14.

The recent application of machine learning in the field of OSC shows its potential in performing the high-throughput screening of organic molecules effectively^15–17. Scharber et al. built a model that can calculate the PCEs with a function of the bandgap and the energy levels of the conjugated polymer¹⁸. Harvard Clean Energy Project (CEP) collected the calculations and experimental data of thousands of organic photoelectric molecules and predicted their PCEs using Scharber’s model^19–20. The same team later used the Gaussian Process Regression (GPR) method in machine learning to correct Scharber’s model, which increased its Pearson correlation coefficient (r) from 0.3 to 0.43 and make a rough estimate of the PCEs²¹. However, due to the low accuracy and the time-consuming quantum mechanics calculations of the energy levels, Scharber’s model is not competent for the fast and accurate high-throughput screening²².

By removing the expensive input of quantitative microscopic properties, Sun et al. established a model with deep learning that can quickly classify photoelectric molecules ^23–24. This model can use molecular graphs or fingerprint information as input to predict the PCEs interval (0–3% or 3-14.6%). Since the acquisition of molecular fingerprints does not require additional quantum mechanics calculations, this model can achieve a fast classification of molecules, but cannot predict the value of PCEs. Moreover, the accuracy of this model is not satisfactory (69.41%) since the existing data cannot meet the high demand for deep learning. The input of the molecular microscopic properties is very helpful to improve the accuracy of the model. For example, Alessandro et al. trained a KRR model that can predict the PCEs better(r = 0.68) by combining both structural and electronic descriptors, and such accuracy has met the requirements of the high-throughput screening²². With the data collection of published literature and quantitative calculations, Ma et al. trained a model that can directly predict the PCEs using the GBRT method^25–26. This model takes the 13 microscopic properties of molecules as input and shows a high accuracy (r = 0.79). Obviously, the addition of molecular microscopic properties improves the accuracy of existing models, which meets the requirements of high-throughput screening.

However, many expensive calculations of microscopic properties (especially excited states) greatly limited the high-throughput screening of the large-scale organic compound space for suitable molecules. Therefore, we have to train an accurate machine learning model for the high-throughput screening of organic optoelectronic molecules with input that can be easily obtained.

In this work, we established an automated framework that can quickly predict the PCEs of OSCs. First, a small dataset containing high-quality experimental data were used to train an ensemble learning model which can predict the PCEs based on the physical and chemical properties of molecules. Then we trained a deep learning model that can predict the molecular properties accurately by using a graph neural network (GNN) architecture and a dataset containing a large number of molecular structures and properties. Specifically, we used SLI-GNN (Self Learning Input GNN), which was recently developed by ourselves²⁷. Based on these two models, we designed this framework that can directly predict the PCEs based on the molecular structure. Finally, our framework shows excellent performance and universality in high throughput screening and is verified by our experimental results. By combination of deep learning and ensemble learning, we achieve direct, fast and accurate prediction of PCEs based on molecular structure.

Figure 1 shows our efficient prediction workflow for OPV, which includes two parts: 1) Predict molecular properties based on molecular structures with the Graph Neural Networks model (Model 1). 2) Train an ensemble learning model to realize the prediction of the PCEs based on the molecular physical properties (Model 2). Here we call the graph neural network model as Model 1, and the ensemble learning model as Model 2. The details of model construction are given in Section Quantitative structure-property relationship modeling.

We collected data and performed DFT calculations to build a database including the experimental device efficiencies and the physicochemical properties of molecules. After that, Model 2 was trained using this database, and the relationship between the molecular microscopic properties and the macroscopic properties of devices was established owing to the excellent ability of machine learning to extract complex relationships. In order to avoid the time-consuming DFT calculations for the microscopic properties of molecules, we established Model 1 with a database containing a large number of molecular structures and properties, to predict molecular microscopic properties based on molecular structures. With the input of molecular structures, Model 1 can be used to quickly predict desired molecular microscopic properties, and Model 2 can be used to predict the PCEs of the device based on microscopic properties predicted by Model 1. In our framework, Model 2 is critical for accuracy and Model 1 is critical for efficiency. The combination of these 2 models leads to a direct, fast and accurate prediction of OPV from molecular structures.

Our database contains 440 small molecule/fullerene pairs and their corresponding PCEs. These data come from published literature. Since one molecule was synthesized by multiple experiments and may correspond to multiple PCEs, the highest PCEs was chosen as the criterion. In order to improve Model 1 by transfer learning, 200,000 pieces of data from the Clean Energy Project Database (CEPDB) were used for pre-training²⁸. These data only include structures and microscopic properties of the molecule without the experimental PCEs. The details of data selection are given in the section 1 of Supporting Information.

The 440 pieces of data are divided into the training set and the test set according to the ratio of 10:1, and the 2 datasets are used to train and test model respectively. The PCEs distributions of the training set and the test set are shown in Fig. 2a and Fig. 2b, respectively. The ratio of OPVs with the low (0 ~ 3%), the medium (3 ~ 6%) and the high (6%~) PCEs is about 2:5:3. From Fig. 2a and Fig. 2b, we can see there is little difference between the training set and the test set for the PCEs distributions, as the average (avg), the median (Me), and the dispersion coefficient (σ) is close, which ensures that the test set is suitable for examining the prediction accuracy of the models.

Feature selection is an important step in the process of building machine learning models. Generally, the stronger the correlation between the features and the targets, the less difficult the learning task would be. The feature selection process is to find suitable physical and chemical properties of molecules showing strong correlation with the PCEs for Model 2. Earlier studies proved that microscopic properties of molecule are helpful to improve the accuracy of the model^29–30. Since our ultimate aim is to achieve the high-throughput screening, we need to consider the availability of data when selecting microscopic properties. Although the final DFT calculation will be skipped, these properties still need to be calculated to generate the training set for training Model 1. Considering some of the descriptors previously proposed and the availability of data²⁶, we finally selected nine properties shown in Table 1 as the learning features for Model 2. The nine features include the ground-state properties, the excited-state properties and the characters of molecular structure. Figure S1 shows the Pearson’s correlation coefficient for all features that help us obtain an initial insight on the data. Most of these features show weak correlation with each other. Since the addition of redundant features can reduce the difficulty of learning, these 9 selected features are not independent of each other, and some features, such as Nd, can be obtained easily without DFT calculation. These features of 440 data are calculated by DFT, and constitute the dataset for training the model. Considering the consistency of the data distribution, the same calculation method as that of CEPDB²⁸ was used. The DFT calculation details are given in section Methods.

Model 2: From molecular property to device performance

Support Vector Machine (SVM), Random Forest (RF), Gradient Boosting Decision Tree (GBDT) and LightGBM are common machine learning models that have been widely used in the field of materials science. Based on the training set of 400 small molecule donors, 4 different models (SVM RF GBDT and LightGBM) were used to capture the correlation between the physicochemical properties of molecules and the PCEs of devices. Leave-One-Out validation was used to assess how well each model generalizes to another dataset, and ensure they are not over-fitting³¹. Hyperparameters was optimized with grid searches in the training and the Leave-One-Out validation. The implementation details and hyperparameters of 4 models are given in the Methods and the Table S1. The PCEs distributions corresponding to the prediction of the models and the experimental determinations, as well as the error distributions on the dataset are shown in Fig. 2 and Figure S2. We employed the violet plot to visualize the error distributions of four models in Fig. 2d and in the violet plot, the fatter part means the more centralized distribution of the data. For the SVM and RF model, there are several abnormal points with absolute errors exceeds 4%, which deteriorate the performance of the model. For GBDT and LightGBM model, the errors are basically distributed ranging from − 2%~2%, and most of the points are concentrated from − 1–1%. The comparison of these four models shows that the 2 integrated models using LightGBM and GBDT have better performance than SVM and RF. From Fig. 2e and Fig. 2f, it can be seen that the performances of LightGBM on training set and test set are similar (r = 0.82, r = 0.87), which indicates that model is not overfitting. Figure 2e, Fig. 2f, and Figure S2 shows that the RMSE and the r of LightGBM are also better than SVM and RF models. Since the principles of GBDT and LightGBM are similar, the performance of the two models are equivalent. LightGBM has better generalization ability than traditional boosting models like GBDT, it was selected for Model 2 to predict the PCEs in the end.

Table 1

The selected features and their physical interpretations
Feature	Physical interpretations
HOMO	The highest occupied molecular orbital energy level of donors
LUMO	The lowest unoccupied molecular orbital energy level of donors
E_DAHL	The difference between HOMO of donor and LUMO of acceptor
E_DALL	The difference between LUMO of donor and LUMO of acceptor
E_T1	Energy of the electronic transition to the lowest-lying triplet state
N_d	Number of unsaturated atoms in the donor molecules
\({\varDelta }_{LA}\)	The difference between LUMO and LUMO + 1 of acceptor
\({\varDelta }_{HD}\)	The difference between HOMO and HOMO − 1 of donor
\({\varDelta }_{LD}\)	The difference between LUMO and LUMO + 1 of donor

The influence of each feature on PCEs is compared by analyzing the weight of each feature. Figure 2c shows the weight ranking of each feature used in trained LightGBM model. It can be seen that the molecular orbital energy level has a significant effect on the PCEs, especially the difference between LUMO and LUMO + 1 and the difference between HOMO and HOMO-1. The physical meaning of these differences is the degeneracy of the HOMO and the LUMO of a molecule, and their important influences on the photoelectric properties of the molecule have been proven^32–33. In addition, the energy of the electronic transition to the lowest-lying triplet state is quite informative among the features^34–35. It is clear that the excited-state properties have an influence on the photo-physical processes, thus having a relation with the short-circuit current density^36–38. The distinction between the LUMO and LUMO + 1 energy levels of the acceptor appears to be less significant for the model. This is likely because, in our dataset, the acceptors are fullerenes, and there are only two distinct molecules present (PC₆₁BM and PC₇₁BM).

Model 1: From molecular structure to molecular property

Model 2 realizes the prediction of the PCEs based on the molecular properties. However, molecular microscopic properties that require time-consuming DFT calculations are inapplicable to high-throughput calculations. Therefore, Model 1 is trained to establish the relationship between the structures and the molecular properties. Classical machine learning methods cannot capture the molecular structure information well, and it is difficult to achieve molecular structure-to-property prediction. As a new deep learning architecture, graph neural network has a wide range of applications in chemistry due to their natural adaptability to molecular structures. We constructed Model 1 using a graph neural network to capture the relationship between molecular structure and properties. Figure 3 shows the framework of Model 1.

First, we convert the input molecules structure into a graph, with atoms as nodes and chemical bonds as edges. Three matrices are used to record the properties of the molecular graph. Matrix 1 describes the node properties of atoms, matrix 2 describes the properties of edges, and matrix 3 is an adjacency matrix that describes the adjacency properties of nodes. The three matrices are passed to the graph convolution layer after the embedding operation. The message passing mechanism of graph convolution can well describe the influence of connections between nodes. The output of the graph convolutional layer is pooled as input to the fully connected layer. The fully connected layers (FC layers) give the final prediction of molecular properties. More details on model building are given in Methods.

The complexity of the graphical model enables it to handle more complex and variable molecular structures, but at the same time increases the data volume requirements. In order to improve the accuracy of the model, the strategy of transfer learning was adopted^39–41. Based on the original 400 training set, 200,000 CEP data containing molecular structures and physicochemical properties are used to train Model 1 which can predict properties based on molecular structures. The CEP data were used to pre-train the model. Figure S3 and Fig. 4 show the error distributions corresponding to the direct training with 400 molecular training data and the transfer learning respectively. It can be seen that the transfer learning improves training accuracy effectively. The results of the 5 features to be predicted are given in Fig. 4. It can be clearly seen that the prediction errors of the 5 features are almost distributed in the range of -0.2eV and − 0.2 eV, which are close to the errors of DFT calculations, therefore Model 1 can accurately predict the microscopic properties of the molecules we need. The prediction accuracy of Model 1 for the molecular ground state properties (HOMO, LUMO, \({\varDelta }_{HD}\), \({\varDelta }_{LD}\)) is better than E_T1, which is consistent with the noise distribution of our dataset. The reason is the accuracy of the DFT calculations of the ground state energy level is better than that of the excited state energy.

Model 1 and Model 2 utilize a combination of Graph Neural Networks and LightGBM to implement the framework from molecular structure to device efficiency prediction. To verify the performance of this framework, we validate the framework by the following three parts.

Accuracy and speed validation

40 test data from previous reports were used to verify the entire framework. The 40 test data have never been used in the training progress of the above 2 models, so they are 40 "unseen" data for this framework. First, molecular structures were used as the input onto Model 2 to obtain the corresponding physical and chemical properties. After that, we used the predicted physical and chemical properties as the input onto Model 1 to obtain the PCEs. The predicted PCEs and the error distributions are shown in Fig. 5a and Fig. 5b. From Fig. 5a and Fig. 5b, it can be seen that the prediction results are still accurate (r = 0.79, RMSE = 1.20). Due to the bias of Model 1, the RMSE increases and r decreases versus the PCEs that directly predicted based on the properties calculated by DFT. For these 40 data, we only fed their molecular structures into the framework and skipped the DFT calculations. Therefore, we have a full progress in approaching the accurate prediction of the efficiency from the molecular structures directly. We contrast the time consumption of DFT calculation of molecular properties with the time consumption of ML prediction in Figure. 5e. It can be clearly seen that the ML method is significantly more time-saving than DFT, and the time-consuming of DFT is exponentially related to the number of atoms. The prediction progress takes only a few minutes, and the whole framework is compiled to an easy-to-use python package, which was described with details in Supporting Information.

Generalization ability verification

The construction of Model 1 and Model 2 is based on the fullerene acceptor OSC dataset. Recently, non-fullerene acceptors received more attention due to various advantages⁴². Model 2 considered the characteristics of the molecular acceptors during construction and thus can be easily generalized to non-fullerene acceptors. We reconstructed Model 2 by adding 11 non-fullerene acceptor device efficiencies and their molecular structure of donor and acceptor to the original dataset as new datasets. 3 molecules that did not participate in training and the original test set were used as a new test set to test new model. Details of the non-fullerene acceptor device data are given in Table S4. The tested PCEs and errors are given in Fig. 5c and Fig. 5d. RMSE decrease and r increase in Fig. 5c compared to Fig. 5a. For non-fullerene acceptors device, the model shows excellent generalization ability.

Novel molecular screening and experimental verification

Novel molecules were designed to test accuracy and rapid screening capabilities of our framework. Conjugated molecules used in OPVs are composed of different types of building blocks such as donor (D), p-spacer (S), acceptor (A), and end-capping (C) units shown in Fig. 6a. Through the arrangement and combination of the 20 fragments in Fig. 6a, 375 molecules of 3 configurations (CADAC, CSASC, CSDSC) were designed as input for our framework. Their PCEs were predicted by our model in few minutes with a personal computer. This demonstrates the capability of the model to rapidly screen molecules. The predicted results for all designed molecules are given in Table S5. The distribution of PCEs is shown in Fig. 6b. As can be seen from Fig. 6b, the PCEs of the molecule in the CADAC configuration are higher than that of the other two configurations.

Table 2

The device performances of SM1 ~ SM4 and their corresponding PCEs predicted from our model.
Donor	Acceptor	V_oc (V)	J_sc (mA/cm²)	FF (%)	PCEs (%)	Predicted PCEs (%)
SM1	PC₇₁BM	0.85	9.40	0.54	4.29	5.94
SM2	PC₇₁BM	0.97	9.06	0.59	5.19	6.83
SM3	PC₇₁BM	0.92	7.42	0.63	4.27	5.74
SM4	PC₇₁BM	0.90	9.71	0.62	5.44	5.34
SM4	Y6	0.81	23.05	0.61	11.35	12.01

Table 3

The calculated and predicted features of SM1 ~ SM4. The subscript ‘cal’ means the value calculated from DFT calculations while the subscript ‘pre’ means the value predicted from Model 1.
	SM1/eV	SM2/eV	SM3/eV	SM4/eV
HOMO_cal	-4.771	-4.987	-4.908	-4.826
HOMO_pre	-4.780	-4.917	-4.882	-4.847
LUMO_cal	-2.979	-2.967	-2.929	-2.923
LUMO_pre	-3.007	-2.954	-2.919	-2.967
E_T1_cal	1.224	1.410	1.403	1.449
E_T1_pre	1.277	1.554	1.558	1.469
\({\varDelta }_{HD}\)_cal	0.464	0.247	0.159	0.261
\({\varDelta }_{HD}\)_pre	0.323	0.404	0.355	0.316
\({\varDelta }_{LD}\)_cal	0.158	0.134	0.088	0.075
\({\varDelta }_{LD}\)_pre	0.197	0.113	0.105	0.101

In order to validate our new framework, 4 small molecules (SM1-SM4) with CADAC configuration that have not been reported were synthesized in this work. The formulas of 4 molecules were shown in Fig. 6c, with detailed synthetic routine displayed in Figure S4 according previous report. They are the representatives of the first generation of organic photovoltaics molecules (SM1)⁴³, the second generation of organic photovoltaics molecules (SM2 and SM3)⁴⁴, and the third generation of organic photovoltaics molecules (SM4)⁴⁵. These molecules were manufactured into OSC devices with PC₇₁BM (SM1-SM4) and Y6(SM-4) acceptors. We further investigated the photovoltaic performance of these functional molecules using conventional device structure described in the experimental part. In addition, we adopted both conventional fullerene acceptor PC₇₁BM as well as recently emerged non-fullerene molecule Y6 as the electron acceptor. The I-V curve of the optimal device for each molecule is shown in Fig. 6d and Fig. 6e. The PCEs determined by experiments and predicted by our framework are shown in Fig. 6f and Table 2. It can be seen that the predicted PCEs of PC₇₁BM based devices are very close to the actual experimental PCEs with an error below 2%. The predicted PCEs of Y6 based device (12.01%) is also very close to the actual experimental PCEs (11.35%). The trend of the predicted value is also consistent with that of the experimental value, while the predicted PCEs are higher than the experimental PCEs. In order to analyze the predicted results, the features predicted by Model 1 for the 4 molecules are given in Table 3. As shown in Table 3, the HOMO and the LUMO of the molecules predicted by Model 2 are very accurate, as well as the predicted E_T1 and \({\varDelta }_{LD}\). The slight deviation from the calculated value of \({\varDelta }_{HD}\) should be the major factor that influences the predicted PCEs.

By using machine learning, a general approach to predict the PCEs of organic solar cells was developed, which shows the best performance (r = 0.79) without any DFT calculation. Model 2 built a quantitative model from molecular properties to device performance by using the ability of machine learning to derive relationships from large amounts of data. This model solves the problem that traditional calculation cannot directly obtain accurate device efficiency theoretically. Model 1 established the corresponding relationship between molecular structure and properties utilizing the natural adaptability between graph neural network and molecular structure. The problem in Model 2 that large-scale screening cannot be achieved due to the high computational cost of molecular properties is solved by Model 1. An OPV database which has nearly 500 molecules with the calculated properties (HOMO, LUMO, E_T1 et al.) was established. The workflow was developed into an easy-to-use Python software package, which is open access. Our work is expected to significantly assist the development of OPVs and the design of new efficient materials like OLEDs or organic catalysts with high-throughput screening.

The combination of two models makes it possible to predict device performance from molecular structure. The reliability of the model has been verified by experiments. However, since the feature selection of Model 2 only considers some properties of the electronic structure of a single molecule, it ignores some factors that can affect the device efficiency, such as molecular side chains, which are highly correlated with solubility. Therefore, the accuracy of this working framework can be further improved if more features at larger scales are considered. But it also increases the difficulty of building the model.

In addition, since the space of organic molecules is almost infinite, how to construct the molecular structure that needs to be screened is another problem that needs to be explored. This approach involves generating completely new molecular structures from scratch using computational algorithms, often guided by desired properties or specific design criteria. More active learning strategies such as Generative Adversarial Network and Reinforcement learning combined with our model could enable the transition from high throughput screening to automatic design. This will help researchers efficiently explore the vast chemical landscape and identify promising candidates for organic solar cells.

Quantum chemical calculations

The ground state structures of all molecules were optimized by DFT/BP86 with the def2svp basis set^46–48. The energy levels of the molecules were calculated with B3LYP/def2svp based on the optimized geometries^{46, 48–49}. The energy of the electronic transition to the lowest-lying triplet state was calculated by TD-DFT/M062X with the 6-311g(d) basis set^50–51. To avoid a very high computational cost, side alkyl chains were not considered in the calculations though they have negligible influences on electronic properties. All the calculations above were performed using Gaussian 09 software package⁵².

Machine learning models

Scikit-Learn was used to obtain the SVM, GBDT, and RF algorithms which were chosen to build models for control study⁵³. The details of the model evaluation metrics are given in Supporting Information. SVM is an algorithm that learns the importance of each training data point for representing the decision boundary between the 2 categories^54–55. RF and GBDT are 2 sub-models of Decision Tree (DT) models⁵⁶. To address the problem of overfitting that limits the application of DT, RF utilizes randomness injection into the tree building while GBDT tries to correct the mistakes of the previous tree continuously^57–59. The Light Gradient Boosting Machine (LightGBM) model is a tree based learning algorithm improved from GBDT, hence this algorithm has the advantages in regularization and multiple loss functions⁶⁰. Instead of using all the sample points to calculate the gradient, GOSS (Gradient-based One-Side Sampling) was used to calculate the gradient. EFB (Exclusive Feature Bundling) combine certain features together to reduce the dimensionality of the features and find the best segmentation point to reduce the consumption. Hyperparameters of these models are given in Table S1.

Model 1 is implemented in PyTorch Geometric, a library built upon PyTorch to easily write and train Graph Neural Networks. The atom, bond, and neighbor feature embedding layers produce 64–256 dimensional inputs to the graph convolution layers. The main body of the network consists 3–9 graph convolution (GCN) layers, each with hidden dimension 128. The final atom representations are reduced by global mean pooling and mapped to regression outputs by full connection layer. Hyperparameters of this model are given in Table S3.

Devices fabrication

First, PEDOT:PSS was spin-coated onto the pre-cleaned ITO substrate at 4500 rpm for 40 s after the filtration through a 0.45 µm filter and then the substrates were baked at 150 ℃ for 10 min under ambient conditions. Subsequently, for PC₇₁BM based device, a blend of the small molecule and PC₇₁BM was dissolved in chloroform (the total concentration: 14 mg mL^− 1, D/A weight ratio: 1/0.8, 0.5% DIO as additive), and the blend solutions were spin-cast onto the PEDOT:PSS layer at a spin-coating rate of 2500 rpm, after the active layers were treated with thermal annealing at 100°C for 10 mins; for Y6 based device, a blend of the small molecule and Y6 was dissolved in chloroform (the total concentration: 16 mg/mL, D/A weight ratio: 2/1). The blend solutions were spin-cast onto the PEDOT:PSS layer at a spin-coating rate of 3000 rpm, and the active layers were treated with thermal annealing at 150°C for 5 mins. Then, 8 nm PFN-Br was spin-coated onto the active layer. Finally, Al at a speed of 2 Å/s (100 nm) was thermally evaporated to accomplish the device fabrication. The J-V characterization was performed by Keithley 2400 digital source meter under simulated AM 1.5G solar irradiation at 100 mW cm^− 2. The device area is 0.0725 cm², solar cell devices were measured in forward scan (-1.0 V → 1.0 V, step 0.0125 V, scan rate: 0.1 V s^− 1) in the glovebox.

Acknowledgments

This work was supported by the National Key Research Program of China (Grant No. 2022YFA1503101), Science and Technology Project of Jiangsu Province (Grant No. BZ2020011), National Natural Science Foundation of China (Grant No. 22173067), Science and Technology Development Fund, Macau SAR (FDCT No. 0052/2021/A), Collaborative Innovation Center of Suzhou Nano Science & Technology, Priority Academic Program Development of Jiangsu Higher Education Institutions (PAPD), 111 Project, and Joint International Research Laboratory of Carbon-Based Functional Materials and Devices.

Author contributions

Y.L. and H.W. conceived the idea and initiated this project. H.W. and J.F. collected and built the datasets, H.W., J.F., Z.D., and L.J. constructed and optimized the deep learning, SVM, RF, and LightGBM models, respectively. J.Y. and M.L. designed and synthesized the OPV materials and fabricated OPV devices. H.S. wrote the manuscript. Y.L., L.J., J.F., and J.Y. contributed to the fruitful discussions and supervision of the project. All authors discussed the results and commented on the manuscript.

Cheng, P.,Li, G.,Zhan, X.&Yang, Y., Next-generation organic photovoltaics based on non-fullerene acceptors. Nat. Photonics 12, 131-142 (2018).
Wan, X.,Li, C.,Zhang, M.&Chen, Y., Acceptor–donor–acceptor type molecules for high performance organic photovoltaics–chemistry and mechanism. Chem. Soc. Rev. 49, 2828-2842 (2020).
Kini, G. P.,Jeon, S. J.&Moon, D. K., Design principles and synergistic effects of chlorination on a conjugated backbone for efficient organic photovoltaics: A critical review. Adv. Mater. 32, 1906175 (2020).
Jain, A.,Shin, Y.&Persson, K. A., Computational predictions of energy materials using density functional theory. Nat. Rev. Mater. 1, 1-13 (2016).
Cui, Y.,Zhu, P.,Liao, X.&Chen, Y., Recent advances of computational chemistry in organic solar cells. J. Chem. Mater. C (2020).
Russell, S.&Norvig, P., Artificial intelligence: A modern approach. (2002).
Nilsson, N. J., Principles of artificial intelligence; Morgan Kaufmann, 2014.
Minsky, M., Steps toward artificial intelligence. Proceedings of the IRE 49, 8-30 (1961).
Wang, H.,Ji, Y.&Li, Y., Simulation and design of energy materials accelerated by machine learning. WIRES Comput. Mol. Sci.. 10, e1421 (2020).
Chen, C., et al., A critical review of machine learning of energy materials. Adv. Energy Mater. 10, 1903242 (2020).
Schmidt, J.,Marques, M. R.,Botti, S.&Marques, M. A., Recent advances and applications of machine learning in solid-state materials science. npj Comput. Mater. 5, 1-36 (2019).
Carrasquilla, J.&Melko, R. G., Machine learning phases of matter. Nat. Phys. 13, 431-434 (2017).
Ma, S.&Liu, Z.-P., Machine learning for atomic simulation and activity prediction in heterogeneous catalysis: Current status and future. ACS Catal. 10, 13213-13226 (2020).
Butler, K. T., et al., Machine learning for molecular and materials science. Nature 559, 547-555 (2018).
Mahmood, A.&Wang, J.-L., Machine learning for high performance organic solar cells: Current scenario and future prospects. Energy Environ. Sci. 14, 90-105 (2021).
Feng, J.,Wang, H.,Ji, Y.&Li, Y., Molecular design and performance improvement in organic solar cells guided by high‐throughput screening and machine learning. Nano Select (2021).
Saeki, A.&Kranthiraja, K., A high throughput molecular screening for organic electronics via machine learning: Present status and perspective. Jpn. J. Appl. Phys. 59, SD0801 (2019).
Scharber, M. C., et al., Design rules for donors in bulk‐heterojunction solar cells—towards 10% energy‐conversion efficiency. Adv. Mater. 18, 789-794 (2006).
Hachmann, J., et al., Lead candidates for high-performance organic photovoltaics from high-throughput quantum chemistry–the harvard clean energy project. Energy Environ. Sci. 7, 698-704 (2014).
Hachmann, J., et al., The harvard clean energy project: Large-scale computational screening and design of organic photovoltaics on the world community grid. J. Phys. Chem. Lett. 2, 2241-2251 (2011).
Lopez, S. A.,Sanchez-Lengeling, B.,de Goes Soares, J.&Aspuru-Guzik, A., Design principles and top non-fullerene acceptor candidates for organic photovoltaics. Joule 1, 857-870 (2017).
Padula, D.,Simpson, J. D.&Troisi, A., Combining electronic and structural features in machine learning models to predict organic solar cells properties. Mater. Horiz. 6, 343-349 (2019).
Sun, W., et al., Machine learning–assisted molecular design and efficiency prediction for high-performance organic photovoltaic materials. Sci. Adv. 5, eaay4275 (2019).
Sun, W., et al., The use of deep learning to fast evaluate organic photovoltaic materials. Adv. Theor. simul. 2, 1800116 (2019).
Sahu, H.&Ma, H., Unraveling correlations between molecular properties and device parameters of organic solar cells using machine learning. J. Phys. Chem. Lett. 10, 7277-7284 (2019).
Sahu, H.,Rao, W.,Troisi, A.&Ma, H., Toward predicting efficiency of organic solar cells via machine learning and improved descriptors. Adv. Energy Mater. 8, 1801032 (2018).
Dong, Z.,Feng, J.,Ji, Y.&Li, Y., Sli-gnn: A self-learning-input graph neural network for predicting crystal and molecular properties. ChemRxiv. Cambridge: Cambridge Open Engage (2023).
Lopez, S. A., et al., The harvard organic photovoltaic dataset. Sci. Data. 3, 1-7 (2016).
Zhao, Z.-W.,del Cueto, M.,Geng, Y.&Troisi, A., Effect of increasing the descriptor set on machine learning prediction of small molecule-based organic solar cells. Chem. Mater. 32, 7777-7787 (2020).
Padula, D.&Troisi, A., Concurrent optimization of organic donor–acceptor pairs through machine learning. Adv. Energy Mater. 9, 1902463 (2019).
Vehtari, A.,Gelman, A.&Gabry, J., Practical bayesian model evaluation using leave-one-out cross-validation and waic. Stat. Comput. 27, 1413-1432 (2017).
Li, H.,Brédas, J.-L.&Lennartz, C., First-principles theoretical investigation of the electronic couplings in single crystals of phenanthroline-based organic semiconductors. J. Chem. Phys. 126, 164704 (2007).
Kuzmich, A.,Padula, D.,Ma, H.&Troisi, A., Trends in the electronic and geometric structure of non-fullerene based acceptors for organic solar cells. Energy Environ. Sci. 10, 395-401 (2017).
Schwarz, K. N., et al., Suppressing subnanosecond bimolecular charge recombination in a high-performance organic photovoltaic material. J. Phys. Chem. C 120, 24002-24010 (2016).
Zhang, J.,Zhu, L.&Wei, Z., Toward over 15% power conversion efficiency for organic solar cells: Current status and perspectives. Small Methods 1, 1700258 (2017).
Mühlbacher, D., et al., High photovoltaic performance of a low‐bandgap polymer. Adv. Mater. 18, 2884-2889 (2006).
Murphy, A. R.&Frechet, J. M., Organic semiconducting oligomers for use in thin film transistors. Chem. Rev. 107, 1066-1096 (2007).
Oberhofer, H.,Reuter, K.&Blumberger, J., Charge transport in molecular materials: An assessment of computational methods. Chem. Rev. 117, 10319-10357 (2017).
Paul, A., et al. In Transfer learning using ensemble neural networks for organic solar cell screening, International Joint Conference on Neural Networks, 2019.
Pan, S. J.&Yang, Q., A survey on transfer learning. IEEE T. Knowl. Data. En. 22, 1345-1359 (2009).
Wang, X., et al., Electric dipole descriptor for machine learning prediction of catalyst surface–molecular adsorbate interactions. J. Am. Chem. Soc. 142, 7737-7743 (2020).
Hou, J.,Inganäs, O.,Friend, R. H.&Gao, F., Organic solar cells based on non-fullerene acceptors. Nat Mater. 17, 119-128 (2018).
Sun, Y., et al., Solution-processed small-molecule solar cells with 6.7% efficiency. Nat Mater. 11, 44-48 (2012).
Zhang, Q., et al., Small-molecule solar cells with efficiency over 9%. Nat. Photonics 9, 35-41 (2015).
Guo, J., et al., 15.71% efficiency all-small-molecule organic solar cells based on low-cost synthesized donor molecules. Adv. Funct. Mater. 32, 2110159 (2022).
Becke, A. D., Density-functional exchange-energy approximation with correct asymptotic behavior. Phys. Rev. A 38, 3098 (1988).
Perdew, J. P., Density-functional approximation for the correlation energy of the inhomogeneous electron gas. Phys. Rev. B 33, 8822 (1986).
Weigend, F.&Ahlrichs, R., Balanced basis sets of split valence, triple zeta valence and quadruple zeta valence quality for h to rn: Design and assessment of accuracy. Phys. Chem. Chem. Phys. 7, 3297-3305 (2005).
Beck, A. D., Density-functional thermochemistry. Iii. The role of exact exchange. J. Chem. Phys 98, 5648-5646 (1993).
Zhao, Y.&Truhlar, D. G., The m06 suite of density functionals for main group thermochemistry, thermochemical kinetics, noncovalent interactions, excited states, and transition elements: Two new functionals and systematic testing of four m06-class functionals and 12 other functionals. Theor. Chem. Soc. 120, 215-241 (2008).
Zhao, Y.&Truhlar, D. G., Density functionals for noncovalent interaction energies of biological importance. J. Chem. Theory Comput. 3, 289-300 (2007).
Frisch, M., et al., Gaussian 03, revision c. 02; gaussian, inc.: Wallingford, ct, 2004. There is no corresponding record for this reference.[Google Scholar] (2013).
Pedregosa, F., et al., Scikit-learn: Machine learning in python. J. Mach. Learn. Res. 12, 2825-2830 (2011).
Cherkassky, V.&Ma, Y., Practical selection of svm parameters and noise estimation for svm regression. Neural Networks 17, 113-126 (2004).
Hearst, M. A., et al., Support vector machines. IEEE Intelligent Systems and their applications 13, 18-28 (1998).
Myles, A. J., et al., An introduction to decision tree modeling. J. Chemometr 18, 275-285 (2004).
Svetnik, V., et al., Random forest: A classification and regression tool for compound classification and qsar modeling. J. Chem. Inf. Model. 43, 1947-1958 (2003).
Breiman, L., Random forests. Mach. Learn. 45, 5-32 (2001).
Friedman, J. H., Stochastic gradient boosting. Comput. Stat. Data An. 38, 367-378 (2002).
Ke, G., et al., Lightgbm: A highly efficient gradient boosting decision tree. Advances in neural information processing systems 30, 3146-3154 (2017).

There is NO Competing Interest.

20230324SI.docx
Supplementary Information

Download PDF

Journal Publication

published 23 Oct, 2023

Read the published version in npj Computational Materials →

Version 1

posted

You are reading this latest preprint version

Fast and accurate screening framework for organic solar cells based on molecular structure and deep learning

Status:

Journal Publication

Version 1

Abstract

Figures

Introduction

Workflow

Dataset & Feature selection

Quantitative structure-property relationship modelling

Model 2: From molecular property to device performance

Model 1: From molecular structure to molecular property

Framework verification

Accuracy and speed validation

Generalization ability verification

Novel molecular screening and experimental verification

Conclusion

Method

Quantum chemical calculations

Machine learning models

Devices fabrication

Declarations

References

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1