This paper presents a machine learning workflow that integrates (scanning) transmission electron microscopy ((S)TEM) data with Python-based molecular dynamics (MD) and density functional theory (DFT) simulations. Pre-trained neural networks convert imaging data into physical descriptors, bridging the gap between experimental data and simulation environments. The workflow addresses challenges like timescale disparities and observational biases, enabling the reconstruction of optimized geometries and simulation of temperature-dependent dynamics. The approach is demonstrated using a graphene system with Cr ad-atom adsorption and graphene healing effects, but is applicable to other material systems.

Ayana Ghosh, Maxim Ziatdinov, Ondrej Dyck, Bobby G. Sumpter, Sergei V. Kalinin