Organic solar cells (OSCs) are a promising green energy technology due to their lightweight, low cost, and flexibility. OSCs typically utilize bulk heterojunctions (BHJs) composed of electron donor and acceptor materials. Non-fullerene acceptors (NFAs), particularly A-D-A type NFAs, have shown remarkable progress, achieving PCEs of 18-19%. However, discovering new high-performance acceptor materials remains challenging due to its time-consuming and expensive nature. This necessitates the development of high-throughput screening methods. Machine learning (ML) offers a solution, enabling the prediction of material properties and the screening of potential candidates directly from molecular structures. While computational datasets, such as those generated by density functional theory (DFT), have been utilized, models trained solely on these datasets often lack accuracy due to discrepancies with experimental data. High-quality experimental data is crucial for improving the reliability and accuracy of ML models. Deep learning, with its ability to learn high-level representations from raw data, presents a significant advantage over traditional machine learning methods for molecular property prediction. This work aims to leverage advanced deep learning techniques and a comprehensive experimental dataset to enhance predictive performance and accelerate the discovery of efficient OSC materials. The authors specifically investigate the application of BERT, a powerful natural language processing model, combined with graph neural networks (GNNs), to predict PCE and discover new NFAs.

The literature review highlights the challenges associated with traditional methods for discovering high-performance organic photovoltaic materials, emphasizing the need for efficient high-throughput screening techniques. Several studies have explored the use of machine learning for predicting material properties and screening potential candidates. These studies have utilized both computational datasets (e.g., DFT calculations) and experimental datasets. However, the limitations of computational-only datasets are noted, as these often underperform compared to models trained using a combination of computational and experimental data. Existing studies successfully used various machine learning algorithms like random forest, convolutional neural networks, and graph neural networks, highlighting the potential of machine learning in this area. The application of BERT-based models for predicting molecular properties is reviewed, along with discussions on different molecular representation strategies, including SMILES strings and molecular graphs. The benefits of incorporating GNNs to extract information from molecular graphs are discussed, particularly their ability to capture structural information for enhanced prediction accuracy. The importance of high-quality experimental data is highlighted along with the challenges of data acquisition, dataset size and composition, and the impact of these on the model's performance. The limitations of existing approaches motivate the need for improved methods that seamlessly combine deep learning algorithms with experimental data for accurate PCE prediction.

The researchers developed DeepAcceptor, a deep learning-based framework comprising data collection, PCE prediction using abcBERT, and material design and discovery. A dataset of 1027 NFAs and their experimental PCEs was curated from the literature. A computational dataset of 51,256 NFAs was obtained from reference [14]. The abcBERT model, a novel architecture integrating graph representation learning of GNNs into the BERT model, was designed. The model takes as input atom types, bond types, bond lengths, and adjacency matrices, encoding molecules as graphs. A masked molecular graph task, inspired by the masked language model in BERT, was employed for self-supervised pre-training on the computational dataset. This pre-training step aims to learn fundamental representations of molecular structures. The pre-trained abcBERT model was subsequently fine-tuned on the experimental dataset using a prediction head for PCE. To evaluate the model's performance, it was compared against several state-of-the-art models using metrics like MAE, MSE, R², and Pearson's correlation coefficient. For the molecular generation and screening process, the BRICS algorithm and a variational autoencoder (VAE) were utilized to generate a large database of candidate molecules. Subsequent screening steps involved filtering based on various molecular properties (molecular weight, LogP, number of H-bond acceptors/donors, rotatable bonds, rings, nitrogens, and oxygens), HOMO/LUMO levels predicted by a trained molecularGNN, and energy offsets (ΔHOMO, ΔLUMO) between the donor (PM6) and acceptor. Finally, the abcBERT model predicted the PCE of the screened candidates, and three promising candidates were selected for experimental validation. Hyperparameter optimization was performed to determine the optimal architecture of abcBERT. An ablation study was conducted to assess the contribution of pre-training and various input features (hydrogen atoms, bond lengths, connection information) to the model's performance. The performance of the abcBERT was compared with the performance of Random Forest (RF), dilated Convolutional Neural Network (dilated CNN), MolCLR, molecularGNN, MG-BERT, ATMOL, Graph Attention Network (GAT) and Graph Convolutional Network (GCN). A user-friendly interface was developed to facilitate the use of DeepAcceptor. The interface incorporates the experimental database, a molecular editor, and the PCE predictor, enabling users to design and evaluate new NFA molecules.

The abcBERT model demonstrated superior performance compared to other state-of-the-art models in predicting PCE, achieving a MAE of 1.78 and an R² of 0.67 on the test set. The ablation study revealed that pre-training significantly improved the model's accuracy. The inclusion of hydrogen atoms, bond lengths, and connection information in the molecular representation further enhanced predictive performance. The molecular generation and screening process successfully identified three high-performance NFA candidates. These candidates were experimentally synthesized and tested, with the best-performing candidate exhibiting a PCE of 14.61%, demonstrating the effectiveness of DeepAcceptor in accelerating the discovery of new materials. The average MAE between experimental and predicted PCE values for the three candidates was 1.96%. The user-friendly interface of DeepAcceptor enhances the accessibility and efficiency of designing and discovering high-performance acceptors. The VAE used for molecular generation achieved 100% validity, 87.1% uniqueness, and 100% novelty, indicating its ability to generate diverse and novel molecular structures. The HOMO and LUMO predictors showed superior performance compared to the Tartarus benchmarking platform. The DeepAcceptor online interface is publicly available at https://huggingface.co/spaces/jinysun/DeepAcceptor.

The results demonstrate the effectiveness of DeepAcceptor in accelerating the discovery of high-performance NFA materials for OSCs. The superior performance of the abcBERT model, coupled with the efficient molecular generation and screening process, significantly reduces the time and cost associated with traditional material discovery methods. The high PCE achieved by the experimentally validated candidates (up to 14.61%) validates the accuracy and reliability of the DeepAcceptor framework. While the model shows strong predictive capabilities, the inherent limitations of any prediction model are acknowledged, as discrepancies between predicted and experimental PCEs may arise from various factors including experimental conditions, synthesis procedures, and purification methods. The work highlights the potential of integrating deep learning with experimental data to improve the accuracy and reliability of material property predictions. The successful integration of the GNN and BERT model architecture, including bond length and connection information, showcases the importance of leveraging advanced machine learning techniques and sophisticated molecular representations for accurate predictions in the field of material science.

DeepAcceptor, a deep learning framework incorporating abcBERT, has been successfully developed and validated for accelerating the discovery of high-performance NFA materials for OSCs. The model achieves state-of-the-art PCE prediction accuracy and effectively identifies promising candidate molecules. The user-friendly interface enhances accessibility, while the integration of molecular generation and screening processes streamlines the discovery workflow. Future work could focus on expanding the experimental dataset to further enhance model accuracy, incorporating additional material properties into the prediction model, exploring other deep learning architectures, and integrating additional high-throughput experimental characterization techniques to accelerate the iterative design-synthesis-characterization cycle.

While DeepAcceptor shows promising results, several limitations should be considered. The accuracy of the PCE prediction is partially limited by the size and quality of the experimental dataset. The model's performance may vary depending on the specific types of NFAs included in the dataset. Discrepancies between predicted and experimental PCE values can be attributed to factors such as experimental conditions, synthesis variations, and purification processes. Although the model incorporates structural information, other factors influencing PCE (e.g., morphology, device fabrication techniques) are not directly accounted for in the model. Future efforts should focus on addressing these limitations by expanding the dataset, improving the model's architecture, and incorporating additional factors into the prediction model.