Machine learning quantification of grain boundary defects for high efficiency perovskite solar cells

The power conversion efficiency of perovskite solar cells has been significantly improved in recent years. One of the key factors leading to this change is that the microstructure of the perovskite layer and its neighboring layers can be controlled. Grain size and grain boundaries, as basic components of perovskite film, have a significant impact on the device performance of perovskite solar cells. Therefore, the statistics of perovskite microstructure properties are important for improving power conversion efficiency. However, determining the statistical properties of perovskite grain boundaries is challenging. Here, a machine learning approach was employed for analyzing top and cross-sectional views of micrographs, which allowed more reliable and easily quantified lengths of unit grain boundaries in perovskite thin films, and from there, statistics generated. The grains in the perovskite films were extracted using a convolutional neural network, and the accuracy of the model improved using post-processing of the Voronoi diagrams. With this model, a more reliable feature descriptor than usual was able to be built, relating grain boundary lengths in cross-section plots to efficiency. These results were compared with literature knowledge to establish the model reliability. At the same time, this was more conducive to a deeper understanding of the relationship between grain boundary length and device performance.