LAUSR

Abstract The device performance of thin film electronics is known to be dependent on morphological properties such as the size, shape, and orientation of aggregates and crystalline domains. So far, descriptions of morphology have been semi-quantitative and loosely linked to device performance. However, by using our recently reported quantitative analysis tool m2py, we create morphology labels and, for the first time, extract quantitative morphology information from scanning probe microscopy measurements. In this work, we use the labels and extracted morphology measurements to validate a novel approach of generating quantitative structure–property relationships through the use of machine learning and regression models; we present this generalizable approach by demonstrating it on organic photovoltaics. In our approach, the open-source toolkit, m2py, is used to label and describe a set of organic photovoltaic devices that have received different thermal annealing treatments, thereby altering their morphologies and, subsequently, their performance. Different regressors and types of morphology descriptions are used to examine the efficacy of quantitative morphology descriptors in predicting device performance. Despite the fact that scanning probe measurements only image the thin film surface, we find that the information-rich nature of morphology data enables accurate device performance predictions, even with datasets that would traditionally be considered too small to generate high-quality predictions. The implications of introducing quantitative morphology information for predictive analysis of other devices and materials is also discussed; given the general and material-agnostic nature of m2py, it is anticipated that the predictive capability of these labels will be ubiquitous to all thin film applications.