Manuel Kober-Czerny

Abstract:

Monolithic all-perovskite tandem solar cells (TSCs) offer a route beyond single-junction efficiency limits through band-gap engineering. However, stability is hampered by hygroscopic degradation and phase segregation of poly(3,4-ethylenedioxythiophene) polystyrene sulfonate (PEDOT:PSS), the most common hole-transport material for narrow band-gap subcells. Here, we investigate the interface-mediated crystallization dynamics in mixed tin-lead (Sn-Pb) perovskites through in situ studies. We find that solvent-underlayer synergetic interactions with PEDOT:PSS induce metastable phase segregation during crystallization. Replacing PEDOT:PSS with a phenothiazine-functionalized interface facilitates direct phase transition and achieves preferential (100) orientation, yielding high-quality perovskite films. This enables a single-junction narrow band-gap subcell with 23.2% efficiency. Furthermore, we apply a hybrid interlayer integrating thiol and phosphonic acid anchoring groups on SnO2/Au, achieving a dense interconnecting layer for monolithic all-perovskite TSCs with 29.1% efficiency. The device retains 90% of the initial efficiency over 800 h of maximum power point tracking under simulated 1-sun illumination at 40°C, demonstrating robust operational stability.

Abstract:

Metal halide perovskites (MHPs) are emerging semiconducting materials with the potential to replace established silicon-based technologies in sustainable energy applications. These materials show exceptional optoelectronic properties, including high absorption coefficients, long carrier diffusion lengths, and tunable bandgaps. Despite these advantages, their long-term operational reliability remains a key challenge. Therefore their development requires costly and time-consuming accelerated aging experiments to probe the evolution of important performance metrics under application relevant environments. To address these limitations, time series forecasting approaches have emerged as a solution to accelerate these studies. In this work, we benchmark five different forecasting methods, including conventional machine learning models, such as Light Gradient Boosting Machine (LightGBM) and Random Forest, and specialised time series models, including Temporal Fusion Transformers (TFT), Neural Hierarchical Interpolation for Time Series Forecasting (N-HiTS), and a hybrid architecture combining convolutional neural networks and long short-term memory neural networks (CNNLSTM). The results show that tree-based models consistently achieve strong performance across all materials, including those with high variability and extreme values, whereas neural network models fail when data are highly variable or autocorrelation is low. These findings highlight that model robustness and data characteristics are more important than model complexity for MHP degradation forecasting, suggesting that simpler models are often more reliable than complex state-of-the-art architectures. Future work should focus on increasing the dataset, improving data quality, and optimising model hyperparameters.