Perovskite solar cells (PSCs) have demonstrated impressive power conversion efficiencies (PCEs), reaching 26.1%, making them a promising technology for large-scale solar energy applications. However, their long-term stability remains significantly lower than that of silicon solar cells, hindering commercialization. While several studies have utilized machine learning to understand PSC stability, challenges remain, particularly concerning data quality and homogeneity. Existing databases, while large, often lack sufficient high-quality degradation data for reliable supervised machine learning. This study addresses this gap by analyzing a large, homogeneous dataset of maximum power point tracking (MPPT) operational ageing data collected under controlled conditions over three years. The dataset includes 2,245 MPPT ageing curves from devices with various perovskite absorbers, charge selective layers, contact layers, and architectures. The overarching goal is to explore the relationship between a cell's initial efficiency and its subsequent stability during operation.

Previous research has explored PSC stability using machine learning techniques. The Perovskite Database Project, although substantial (>42,400 data points), suffers from limited and inhomogeneous degradation data (less than 20% with degradation information). Graniero et al. highlighted the low quality of degradation data within this database, emphasizing the need for higher data quality over quantity for effective machine learning. Zhang et al. attempted to address this by introducing a new stability metric (*T_sso_m*) and projecting measured stability under varying conditions to reference conditions. However, this approach involves uncertainties stemming from co-dependencies between stressors and a range of parameters. Critically, these studies did not consider the variations in degradation curve shapes, which can significantly impact stability assessment. This study aims to address these limitations by focusing on a highly homogeneous dataset and incorporating degradation curve shape analysis.

The study utilized a custom-built High-throughput Ageing System to collect 2245 MPPT ageing curves of PSCs under controlled conditions (continuous 1-sun illumination, controlled temperature and atmosphere) from August 2019 to August 2022. The devices comprised various architectures (n-i-p and p-i-n), perovskite absorbers (triple cation, CsPbI3, FAPbI3), charge-selective layers (small molecules, polymers, and inorganics), and contact layers (silver, copper, gold). Ageing conditions varied in temperature (25, 45, 65, 85 °C) and the use of a UV filter. Data pre-processing involved resampling to 10-minute intervals, interpolation using the Akima method, normalization using MaxAbsScaler, and noise reduction with a Savitzky-Golay filter. The first 150 hours of ageing data were analyzed to ensure maximum comparability. Devices were grouped based on their maximum PCE reached within the first 150 hours into five groups with equal numbers of cells. The relative change in PCE (ΔPCE,_rel) after 150 hours was calculated relative to the maximum PCE for each cell. To analyze the diverse degradation curve shapes, an unsupervised machine learning method, the self-organizing map (SOM), was employed to cluster the normalized PCE data. Linear regression was used to analyze the relationship between maximum PCE and ΔPCE,_rel, and k-means clustering was used for comparison.

The statistical analysis revealed a strong correlation between the maximum PCE reached during the first 150 hours of ageing and the relative PCE loss after 150 hours (ΔPCE,_rel). Higher maximum PCEs were statistically associated with lower ΔPCE,_rel, suggesting that more efficient cells are also more stable. This trend is supported by linear regression analysis, showing that for every 1% increase in maximum PCE, ΔPCE,_rel decreased by approximately 1.5%. Two potential explanations were proposed: 1) An energy conservation model suggests that less efficient cells retain more energy not converted into electricity, potentially causing increased degradation. 2) The presence of pinholes, defects, or poor device quality could negatively impact both efficiency and stability. The SOM analysis identified four distinct degradation curve shapes: initial gain, slow exponential decay, medium exponential decay, and fast exponential decay. The frequency of these shapes correlated with the maximum PCE reached. Higher maximum PCE groups exhibited a greater proportion of initial gain curves and a lower proportion of fast exponential decay curves. This suggests that the degradation curve shape can serve as an early indicator of stability, with initial gain curves indicating higher likelihood of stability within the first 150 hours.

The findings suggest a significant link between PSC efficiency and long-term stability. The observed correlation, while statistical, provides valuable insights into PSC ageing behavior. The energy conservation model and the defect density hypothesis offer plausible explanations for the observed relationship. However, the wide range of devices and degradation mechanisms in the dataset makes it difficult to definitively attribute the findings to a single physical cause. The ability to predict stability based on initial efficiency and degradation curve shape holds promise for accelerating the development of highly stable PSCs. Further research is needed to elucidate the underlying physical mechanisms responsible for this correlation and to explore the use of degradation curve shape as a predictive tool.

This study demonstrates a statistically significant correlation between initial PCE and long-term stability in PSCs, providing valuable insights for device optimization. The use of a large, homogeneous dataset and the incorporation of degradation curve shape analysis significantly enhanced the study's power and reliability. Future research could focus on identifying specific material and device characteristics responsible for the observed correlation and refining the use of degradation curve shape as an early indicator of stability. Further investigation into the underlying physical mechanisms will be crucial for developing high-efficiency and highly stable PSCs.

The study's conclusions are based on a dataset collected under specific controlled conditions, and the generalizability to real-world conditions may be limited. The proposed explanations for the observed correlation between efficiency and stability are hypothetical and require further investigation. Although the SOM clustering provided valuable insights, it is important to note that other clustering methods could yield different results.