Machine Learning-Enabled Fast-Scan STEM for the Precise Identification of Labile Single-Atom Sites

The catalytic performance of single-atom catalysts (SACs) is predominantly governed by their local atomic environment. However, directly resolving the transient configurations of individual atoms remains challenges for inherent thermal fluctuations and electron-beam-induced dynamics. Conventional scanning transmission electron microscopy (STEM) relies on long dwell times, covering such dynamic structural fluctuations and constrained by serious noise. Here, a machine-learning-enhanced fast-scan STEM methodology is introduced, permitting accurate identification and localization of individual Pt atoms on MoS₂ supports, circumventing the limitations of high-noise imaging. Analysis reveals random displacements of ∼3.2% relative to ideal lattice sites, offering a direct glimpse into intrinsic structural disorder. Furthermore, we observe persistent heterogeneity in the local electric field across Pt atomic columns in differential phase contrast (DPC) imaging. These results suggest that single-atom sites locate in a highly dynamic, nonideal equilibrium state. This study also establishes a rigorous computational framework for resolving atomic-scale structures under noisy imaging conditions.