(24c) Determination of Primary Particle Size Distributions of Agglomerates on Transmission Electron Microscopy Images By Artificial Neural Networks

The properties of nanomaterials (e.g. mechanical, optical, catalytic, biological, etc.) are determined by their characteristic length as well as their shape [1]. In case of nanoparticles, the characteristic length is usually some kind of equivalent diameter (e.g. hydrodynamic, volume-based, surface-based, etc.) [2]. There are various techniques available for the determination of the different types of equivalent diameters, one of them being transmission electron microscopy (TEM) [3].

The determination of particle size distributions (PSDs) of non-agglomerated particles on TEM images can already be performed partially or fully automated since several decades [4]. By contrast, the determination of PSDs of agglomerated particles still depends on the recognition of primary particles by a human operator. This practice is not only laborious, expensive and repetitive but also error-prone due to the subjectivity and exhaustion of the operator [5].

To solve this problem, a new approach to imaging particle analysis, with help of artificial neural networks (ANNs), was proposed, implemented and validated. ANNs were trained on the determination of the projected areas of primary particles of agglomerates via regression and the number of primary particles per agglomerate via classification.

A major challenge of the proposed method is the fact that the training of ANNs requires up to several hundreds of thousands of samples with known properties to avoid overfitting [6], i.e. that the ANNs are not able to generalize when being confronted with previously unknown data. Unfortunately, there is no publicly available source of already evaluated samples and a manual evaluation of such a large number of images is hardly feasible, due to the reasons given before. Therefore, TEM images with defined characteristics were synthesized and used for the training of the ANNs.

To speed up the image synthesis as well as the training of the ANNs the necessary calculations were performed on graphics processing units (GPUs), whenever possible.

[1] G. Guisbiers, S. Mejía-Rosales and F. Leonard Deepak. ‘Nanomaterial Properties: Size and Shape Dependencies’. In: Journal of Nanomaterials 2012 (2012), pp. 1–2.

[2] J. G. Calvert. ‘Glossary of atmospheric chemistry terms (Recommendations 1990)’. In: Pure and Applied Chemistry 62.11 (1990).

[3] E. J. Cho et al. ‘Nanoparticle characterization: state of the art, challenges, and emerging technologies’. In: Molecular pharmaceutics 10.6 (2013), pp. 2093–2110.

[4] R. P. King. ‘Determination of the distribution of size of irregularly shaped particles from measurements on sections or projected areas’. In: Powder Technology 32.1 (1982), pp. 87–100.

[5] T. Allen. Powder sampling and particle size determination. 1st ed. Amsterdam and Boston: Elsevier, 2003.

[6] M. A. Babyak. ‘What you see may not be what you get: a brief, nontechnical introduction to overfitting in regression-type models’. In: Psychosomatic medicine 66.3 (2004), pp. 411–421.