Logo of the Physikalisch-Technische Bundesanstalt
Production sequence of Si-spheres and interferometrical determination of the sphere volume

Machine Learning in the context of electron microscopy

22.12.2020

The use of machine learning is intended to make better use of the diverse information in electron microscopic images. The work is embedded in the two European EMPIR projects "nPSize" and "Black Carbon", in which both non-spherical nanoparticles and soot particles are investigated. In addition to other measurement techniques, electron microscopy is used for this purpose. Before the size and shape of the particles can be determined, they must be reliably detected in the electron micrograph and distinguished from artifacts (e.g. particle agglomerations, contamination). Using machine learning, this can be done more reliably and conveniently than by manual selection.

Soot particles are produced by combustion processes and not only pose a health hazard but also have a harmful effect on the climate due to their optical properties. In particular the fractal dimension of the particles influences their behavior. The fractal dimension describes the three-dimensional shape of an object with values between 1 (long chain) and 3 (sphere). There are several conventional approaches to determine the fractal dimension from electron micrographs (which are only two-dimensional projections of the soot particles). In initial experiments, simulated images were used to show that machine learning yields comparable results for the fractal dimension as conventional algorithms. In the future, these approaches will be further developed and transferred to real images.

Machine learning is based on learning and generalizing from training data. Therefore, the success of the method largely depends on the size and quality of the available training data sets. By Monte Carlo simulation of the image formation in the electron microscope, the working group is able to generate extensive and high-quality training data sets. This will contribute significantly to the further development of the described approaches and enable new applications.


Schematic representation of an intermediate step in determining the fractal dimension: a machine learning network is used to determine the number and mean diameter of the primary particles of the soot aggregate.

Contact

Address

Physikalisch-Technische Bundesanstalt
Bundesallee 100
38116 Braunschweig