The development of sustainable energy production technologies, such as new types of solar cells, has traditionally taken decades. These research cycles need to be accelerated in order to respond to the urgent climate change crisis. Perovskite solar cells (PSCs) are a recent technology that boast high electricity production efficiencies, however they suffer from insufficient lifetimes. Since there are thousands of material candidates even for a single layer of a PSC, it is challenging to search for stable materials and infer device degradation mechanisms.
This project aimed to implement machine learning (ML) as a part of the solution. The objectives were to create an accelerated, ML-assisted research cycle for improving the lifetime of PSCs; with focus on optimizing the composition of the light absorbing layer – perovskite – for stability. This included showing the resulting stability improvements in full PSCs and develop faster inference of the remaining PSC degradation mechanisms. The acceleration in the cycle arises from i) using ML to save experimental bandwidth, ii) designing experiments that are directly compatible with ML (opposed to first collecting data and then figuring out how to use it with ML), and iii) maximizing effectiveness by teaching physics to the ML algorithm. For example, at the materials level, the algorithm needs to know that certain perovskite compositions do not exist as mixed materials according to density functional theory. The physics integration aligns ML algorithms with physical reality and alleviates data scarcity that has traditionally hindered the use of ML in experimental science.
