How machine learning can help select capping layers to suppress perovskite degradation

Perovskite solar cells (PSCs), despite achieving efficiencies exceeding 25%, still fall short of the ~25-year stability needed for widespread adoption in the photovoltaic (PV) market. Improving environmental stability is crucial. While incorporating low-dimensional (LD) perovskites has shown promise, it often compromises carrier transport and device performance. The capping-layer method offers a solution by using conductive organic materials to intercalate the 2D perovskite, improving carrier transport and surface passivation, leading to increased short-circuit current (Jsc), open-circuit voltage (Voc), and environmental stability. The choice of organic halide significantly impacts stability, but the structure-stability relationship remains largely unexplored. Various organic halides have been explored, including those with benzene rings/phenyl with amine, long carbon chains with amine, fluorous amine, branched amines, and complex structures. While these materials show improved stability and efficiency, the underlying chemical properties governing this improvement are not well-understood. This research uses a machine-learning framework to analyze LD organic-inorganic perovskites as capping layers for MAPbI3, aiming to identify key properties for enhanced stability and to establish materials-design guidelines.

The literature extensively documents the use of various organic halides as capping layers to enhance the stability of perovskite solar cells. Studies have shown that incorporating molecules with specific functional groups, such as phenyl groups or branched alkyl chains, can improve stability by passivating defects and preventing moisture ingress. However, a systematic study correlating the chemical properties of these capping layers with the resulting stability is lacking. Existing research often focuses on individual materials or limited compound classes, without developing a comprehensive understanding of the underlying chemical principles driving stability enhancement. This study aims to bridge this gap by employing a data-driven approach to uncover the key chemical properties that contribute to the effectiveness of capping layers in suppressing perovskite degradation.

This study employed a machine-learning framework to investigate the relationship between the chemical properties of organic capping layers and the stability of MAPbI3 perovskite films. Twenty-one organic halide salts with varying sizes, branches, and chemical properties were selected as capping-layer materials, encompassing both iodine and bromine-based salts. These capping layers were deposited onto 300 nm thick MAPbI3 films using spin coating, and twelve different processing conditions (varying annealing temperature and precursor solution concentration) were explored for each material. The films were aged under rigorous accelerated conditions (85% RH, 85 °C, and 0.16 Sun illumination), with in situ imaging every 3 minutes. A thin-plate spline color warping method was used to calibrate the color images and extract numerical values for degradation onset and rate. Supervised learning algorithms were applied, using structural and chemical features of the organic molecules (from the PubChem 2019 database) and processing conditions as inputs, and degradation onset and rate as outputs. Shapley values were used to rank feature importance and infer design rules. Finally, detailed materials characterization (XRD, SEM, GIWAXS, XPS, FTIR) was performed on the top-performing material to elucidate the underlying mechanism of stability enhancement.

The machine learning model, specifically random forest regression, revealed that the number of hydrogen-bond donors and the topological polar surface area (TPSA) of the organic molecules are the two most important factors determining the stability of the capped MAPbI3 films. A low number of hydrogen-bond donors and small TPSA strongly correlate with increased stability. Phenyltriethylammonium iodide (PTEAI), the top-performing material, demonstrated significantly extended stability (4 ± 2 times longer than bare MAPbI3 and 1.3 ± 0.3 times longer than the state-of-the-art OABr). Characterization revealed that PTEAI forms a Ruddlesden-Popper perovskite, (PTEA)2(MA)3Pb4I13, as a capping layer. XPS and FTIR analysis indicated that the PTEAI capping layer modifies the surface chemistry, suppressing methylammonium loss and the formation of oxygen-containing compounds, thereby enhancing the stability of the MAPbI3 perovskite. The model also indicated the importance of molecular weight and precursor solution concentration in determining stability. Other factors, such as the partition coefficient (x log P), showed less influence.

The findings directly address the research question by identifying key molecular properties that contribute to enhanced perovskite stability. The identification of low hydrogen-bond donor count and small TPSA as critical factors provides valuable insights for the rational design of effective capping layers. The superior performance of PTEAI, backed by detailed characterization, confirms the predictive power of the machine-learning model and provides a concrete example of a highly stable capping-layer material. This research has significant implications for the field, as it offers a data-driven approach for screening and optimizing capping layer materials, thereby accelerating the development of highly stable and efficient perovskite solar cells. The ability to predict the stability of novel materials using readily available chemical descriptors could significantly reduce the experimental burden associated with trial-and-error approaches.

This study successfully demonstrates the application of machine learning to guide the design of more stable perovskite solar cells. The key finding is the identification of low hydrogen-bond donors and small TPSA as crucial features for effective capping layers. The high performance of PTEAI validates this approach. Future research could explore the synthesis and testing of novel quaternary ammonium compounds with optimized properties as suggested by the model, further enhancing the long-term stability of perovskite solar cells and bringing them closer to commercial viability. Furthermore, expanding the dataset and incorporating additional material properties into the model could lead to more accurate and comprehensive predictions.

The study's accelerated aging tests, while rigorous, may not perfectly replicate real-world conditions. The focus on degradation onset might not fully capture the long-term degradation behavior. The machine learning model's accuracy is limited by the size and variability of the dataset. The prediction of stability for materials not included in the training set relies on the model's ability to generalize, and may not be completely accurate for all materials. Additionally, the study focused on MAPbI3; further research is needed to validate these findings for other perovskite compositions.