In the last few years scanning tunneling microscopy (STM) has made it possible and easily accessible to visualize surfaces of conducting specimens at the atomic scale. Such performance allows the detailed characterization of surface morphology in an increasing spectrum of applications in a wide variety of fields. Because the basic imaging process in STM differs fundamentally from its equivalent in other well-established microscopies, good understanding of the imaging mechanism in STM enables one to grasp the correct information content in STM images. It thus appears appropriate to explore by STM the structure of amorphous carbon films because they are used in many applications, in particular in the investigation of delicate biological specimens that may be altered through the preparation procedures.All STM images in the present study were obtained with the commercial instrument Nanoscope II (Digital Instruments, Inc., Santa Barbara, California). Since the importance of the scanning tip for image optimization and artifact reduction cannot be sufficiently emphasized, as stressed by early analyses of STM image formation, great attention has been directed toward adopting the most satisfactory tip geometry. The tips used here consisted either of mechanically sheared Pt/Ir wire (90:10, 0.010" diameter) or of etched W wire (0.030" diameter). The latter were eventually preferred after a two-step procedure for etching in NaOH was found to produce routinely tips with one or more short whiskers that are essentially rigid, uniform and sharp (Fig. 1) . Under these circumstances, atomic-resolution images of cleaved highly-ordered pyro-lytic graphite (HOPG) were reproducibly and readily attained as a standard criterion for easily recognizable and satisfactory performance (Fig. 2).
Machine learning identification of impurities in the STM images
