Machine Learning Models Forecasting Optoelectronic Service Lifetimes of Low-Bandgap Perovskite Solar Cells

Abstract

Accurately predicting the lifetime of perovskite solar cells (PSCs) is essential for their successful commercialization. Mixed Sn-Pb halide perovskites are promising materials because of their relatively low bandgap (~1.2-1.3 eV) making them a strong candidate for use as the absorber layer in single-junction solar cells or as the low-bandgap subcell absorber in all perovskite tandem solar cells. However, this composition suffers from stability issues when exposed to environmental conditions. By collecting experimental data from accelerated degradation, we learn that efficiency losses are driven by short-circuit current and indicate that the loss in perovskite material is a major contributor to degradation. Additionally, we have identified significant correlations between device lifetime and the perovskite degradation rate (derived from previous studies), partial pressure of water, and early time derivative of specific device measurements. To predict the lifetime of PSCs, machine learning models were trained using a menu of features that were split into 3 distinct categories: (1) a priori known features, (2) initial device measurements, and (3) measurements of the initial rates of change of device parameters. With a dataset consisting of 39 degradation experiments, trained models reveal that the derived degradation rate of the FA0.75Cs0.25Pb0.5Sn0.5I3 perovskite material is a valuable feature for forecasting PSC lifetime and was selected as the most dominant feature in the best performing models which demonstrate a median test error of 26%. These findings underscore the importance of quantitative measurements of perovskite degradation in accurately forecasting the lifetime of PSCs and highlight the utility of integrating such metrics into predictive models.