Molecule graph reconstruction from atomic force microscope images with machine learning

Abstract: Despite the success of noncontact atomic force microscopy (AFM) in providing atomic-scale insight into the structure and properties of matter on surfaces, the wider applicability of the technique faces challenges in the difficulty of interpreting the measurement data. We tackle this problem by proposing a machine learning model for extracting molecule graphs of samples from AFM images. The predicted graphs contain not only atoms and their bond connections but also their coordinates within the image and elemental identification. The model is shown to be effective on simulated AFM images, but we also highlight some issues with robustness that need to be addressed before generalization to real AFM images. Impact statement: Developing better techniques for imaging matter at the atomic scale is important for advancing our fundamental understanding of physics and chemistry as well as providing better tools for materials R&D of nanotechnologies. State-of-the-art high-resolution atomic force microscopy experiments are providing such atomic-resolution imaging for many systems of interest. However, greater automation of processing the measurement data is required in order to eliminate the need for subjective evaluation by human operators, which is unreliable and requires specialized expertise. The ability to convert microscope images into graphs would provide an easily understandable and precise view into the structure of the system under study. Furthermore, a graph consisting of a discrete set of objects, rather than an image that describes a continuous domain, is much more amenable to further processing and analysis using symbolic reasoning based on physically motivated rules. This type of image-to-graph conversion is also relevant to other machine learning tasks such as scene understanding.