Unlocking 3D nanoparticle shapes from 2D high-resolution transmission electron microscopy images: a deep learning approach

Abstract The analysis of high-resolution transmission electron microscopy (HRTEM) images with atomic resolution remains limited due to the lack of automated methods capable of processing a statistically significant number of nano-objects. In particular, the accurate characterization of nanoparticle (NP) morphology at the nanoscale remains a major challenge. To address this issue, we have developed a deep learning (DL) framework trained on a dataset of simulated HRTEM images of gold NPs nanoparticles adopting a face-centred cubic structure and ranging in size from 4 to 8 nm, labeled according to their 3D shape. The dataset is generated by constructing atomistic models of NPs deposited on amorphous carbon and subjected to random rotations in order to account for all possible orientations observed in experimental samples. HRTEM images are then simulated using the multislice method, to mimic experimental acquisition of an aberration-corrected transmission electron microscope. A systematic study is carried out to evaluate the impact of different key physical and numerical quantities (amorphous carbon structure, image resolution, focusing conditions, NP size and orientation) on the predictive performance of the DL model. Finally, a robust and accurate framework is developped and proposed for inferring 3D gold NP morphologies from 2D HRTEM images, validated on both simulated and experimental datasets.