Predicting column heights and elemental composition in experimental transmission electron microscopy images of high-entropy oxides using deep learning

A novel approach is presented by integrating images-driven deep learning (DL) with high entropy oxides (HEOs) analysis. A fully convolutional neural network (FCN) is used to interpret experimental scanning transmission electron microscopy (STEM) images of HEO of various sizes. The FCN model is designed to predict column heights (CHs) and elemental distributions from single, experimentally acquired STEM images of complex (Mn, Fe, Ni, Cu, Zn)3O4 HEO nanoparticles (NPs) at atomic resolution. The model's ability to predict elemental distributions was tested across various crystallographic zones. It was found that the model could effectively adapt to different atomic configurations and operational conditions. One of the significant outcomes was the identification of substantial elemental inhomogeneities in all experimental NPs, which highlighted the random and complex nature of element distribution within HEOs. The developed FCN DL method can be applied to assist experimental HEO and beyond NP analysis in various operating conditions.