Inversion of electron microscopy images using atomistic simulations and machine learning

Atomic structure of materials can be characterized by transmission electron microscopy (TEM) and scanning TEM (STEM). However, determining the positions of atoms in three dimensions from two-dimensional images is non-trivial. In this work, we use atomistic modeling with first principles density functional theory (DFT) or empirical potentials, in conjunction with machine learning, to tackle the S/TEM image inversion problem. We discuss the use of single and multi-objective evolutionary and basin-hopping approaches for S/TEM-guided atomistic structure determination, incorporating comparison of simulated and experimental S/TEM images using computer vision approaches. We show that the combined use of energetic and experimental information is effective in arriving at physical solutions.