The challenge of studying perovskite solar cells' stability with machine learning

Perovskite solar cells are the most dynamic emerging photovoltaic technology and attracts the attention of thousands of researchers worldwide. Recently, many of them are targeting device stability issues-the key challenge for this technology-which has resulted in the accumulation of a significant amount of data. The best example is the "Perovskite Database Project," which also includes stability-related metrics. From this database, we use data on 1,800 perovskite solar cells where device stability is reported and use Random Forest to identify and study the most important factors for cell stability. By applying the concept of learning curves, we find that the potential for improving the models' performance by adding more data of the same quality is limited. However, a significant improvement can be made by increasing data quality by reporting more complete information on the performed experiments. Furthermore, we study an in-house database with data on more than 1,000 solar cells, where the entire aging curve for each cell is available as opposed to stability metrics based on a single number. We show that the interpretation of aging experiments can strongly depend on the chosen stability metric, unnaturally favoring some cells over others. Therefore, choosing universal stability metrics is a critical question for future databases targeting this promising technology.