The atomic-scale structure of interfacial water plays a central role in electrochemistry, catalysis, friction, and biological engineering. Although atomic force microscopy (AFM) provides high spatial resolution, direct determination of atomic water structures remains challenging due to weak hydrogen contrast and the complex relationship between AFM images and underlying atomic configurations. Here, we develop a closed-loop, physics-informed structural inversion framework for interfacial water from multi-height AFM images. This framework combines conditional generative adversarial learning with an explicit and interpretable structural descriptor that explicitly encodes atomic positions and hydrogen orientations, establishing a direct link between AFM contrast and atomic configuration. Trained on simulated AFM data, the method achieves high accuracy in localizing atomic positions and determining hydrogen orientation. For experimental AFM images, automated preprocessing and structure-aware postprocessing procedures yield physically plausible atomic structures that reproduce the observed AFM contrast after relaxation, despite experimental noise and limited height sampling. Rather than targeting a unique solution, this approach provides a robust initialization for AFM inverse problems, substantially reducing the configurational search space and offering a general strategy applicable to other hydrogen-rich and weakly bonded interfacial systems.
Luo et al. (Fri,) studied this question.