Machine learning quantification of grain characteristics for perovskite solar cells

Crystalline grains are the fundamental building blocks of metal halide perovskite films, and their characteristics can significantly influence the charge transport and stability in films and thus the device performance of resulting solar cells. But statistical interpenetration of perovskite grain characteristics is challenging. Here, we developed a machine-learning-based methodology for analyzing top-view micrographs, enabling a reliable quantification of individual grain surface area in perovskite films for statistical analysis. A convolutional neural network with U-Net structure was trained for grain area extraction, and further, a Voronoi-inspired post-processing method was developed to enhance the quantification accuracy. Based on this grain extractor tool, we then expanded the study from localized grain surface areas to their statistical distribution over the whole film. A more reliable numerical descriptor for grain characteristics than the popularly used average-grain size parameter was established to interpret the relationship between the microscopic grain characteristics and macroscopic device performance.