Machine Learning Toolkit for Automatic Quantification of Perovskite Microstructural Characteristics

Metal halide perovskites (MHPs), especially perovskite solar cells (PSCs), have emerged as a focal point of due to their favorable optoelectronic properties and photovoltaic (PV) performance. While the commercialization of PSCs still requires the field to overcome a few key challenges including the relatively low stability and processing reproducibility. Artificial intelligence (AI) and machine learning (ML) methods have been examined as transformative tools in the frontier of material science. And in recent years, many AI-assisted studies including automatic synthesis, automatic characterization, and close-loop experiment optimization have attracted lots of attention in the field of perovskite, for their potential to accelerate the development of MHPs and PSCs. While for the large number of experiment results generated, they still lack efficient methods to analyze and extract useful information. Here, we developed a ML-based toolkit for extracting and quantifying microstructural characteristics of MHP from atomic force microscope, enabling a high-throughput and reliable statistical analysis. A convolutional neural network with U-Net structure was trained for grain region extraction, and multiple kinds of algorithms were designed for quantifying the microstructural characteristics including grain surface area, grain boundary groove angles, groove width, grain surface depression and bulge. Based on this toolkit, we then expanded the study from localized measurement to their statistical distribution over the whole film, and reveal their correlation with the perovskite solar cells performance. This work presents a methodology on automatic analysis of microscope images, especially the information transfer from 2D images to 1D quantified variables. This methodology puts one step closer to fully self-driven laboratory and could be further applied to other kinds of materials.