Pengju Ren1,2, Xueqian Pan2,3, Xiao-Dong Wen1,2
1Synfuels China Co. Ltd.
2Institute of coal chemistry, Chinese Academy of Sciences
3University of Chinese Academy of Sciences
A prerequisite for understanding the physical and chemical properties of the surface is detailed knowledge of the atomic structure of surfaces. Scanning tunneling microscopy (STM) is a high-resolution imaging technique for surfaces, which can obtain atomic level images of surfaces and assembly structures of molecules on substrates. Using complex theoretical models combined with electronic-structure calculations, increasingly efforts attempt to determine the surface atomic structure in experimental STM images. However, STM image analysis heavily rely on scientists' expertise and experience. Furthermore, high computational cost on solving electronic structures limits the extensive exploration of surface compositional and conformational space. Therefore, accurate and efficient analysis of experimental STM images remains an urgent challenge. Our work proposes a machine-driven approach for deciphering surface atomic structures from STM images. This approach can efficiently analyze experimental STM spectra and automatically recommend the most probable atomic structure models. Through the machine learning method driven by chemical knowledge and theoretical data, the extensive exploration of surface conformational space is accelerated. By effectively integrating multiple machine-driven filtering processes, the reliability and efficiency of analysis results are greatly improved. Taking complex iron carbides system as model system, this method deciphers the surface atomic structures which are most consistent with the experimental observation. This approach will facilitate the studies of the materials field and catalysis field related to surface structure.