Jonas Bals, Kateryna Loza, Matthias Epple

Inorganic Chemistry, University of Duisburg-Essen, 45117 Essen, Germany

The classification of (nano-)particles with respect to size and shape is an important task in particle synthesis and characterization that is often tediously performed in a manual way. An automated image analysis was applied to scanning transmission electron micrographs (STEM) of metallic nanoparticles. The particles were separated from the background and classified according to shape and size, using concepts of machine learning (ML). Six distinct two-dimensional projections of particles were used for classification and training, i.e. spheres, rods, triangles, squares, pentagons, and hexagon. For comparison, particle classifications were also done by chemically experienced persons (reviewers) concerning occurring shapes in STEM images. The results indicated a diverging classification, especially for particles with distorted geometric shapes, between human reviewers.

Training data sets used for machine learning contained real and artificially created (ideal) particles. The automated analysis of the particle size was well possible if the STEM images had a sufficient contrast. However, overlapping particles could not be safely assigned. Furthermore, a correct assignment of particle shapes to the different classes by automated analysis was difficult. This was especially pronounced for particles with non-ideal shapes like triangles with cut-off corners, ellipsoidally-distorted spheres, tetragonally-distorted squares. This leads to fundamental questions and caused arguments between reviewers as "real" particles never have an "ideal" (i.e. mathematically well-defined) shape. Such ambiguities in shape assignment occurred with human image evaluators and by machine learning and constitute a considerable obstacle in the training process for automated image analysis.