Characterization of mixing in nanoparticle hetero‐aggregates by convolutional neural networks

Formation of hetero‐contacts between particles of different materials in nanoparticle hetero‐aggregates can lead to new functional properties. Improvement of the functional behavior requires a detailed characterization of mixing between the two types of particles, in order to correlate different mixing with the performance of the material. Scanning transmission electron microscopy (STEM) is an option for this task. To obtain statistically relevant results, STEM‐images of many hetero‐aggregates have to be acquired and evaluated. This can be time‐consuming if it is done manually. In the present work, the applicability of convolutional neural networks for the automated analysis of STEM‐images acquired from TiO2_{2}$–WO3_{3}$ nanoparticle hetero‐aggregates is investigated. Hetero‐aggregates are obtained in a double flame spray pyrolysis (DFSP) setup, in which a variation of setup parameters is expected to affect the mixing of TiO2_{2}$ and WO3_{3}$. Mixing is investigated by a measurement of cluster sizes (the number of connected particles of the same material within an aggregate) and coordination numbers (the number of particle contacts with particles of the same or the different material). Results show that the distribution of measured values is wide for both quantities, rendering it challenging to correlate mixing with parameters varied in the DFSP setup. Characterization of mixing in nanoparticle hetero‐aggregates. A convolutional neural network is trained with simulated artificial scanning transmission electron microscopy images to detect particle positions in the images and classify their material. The output is used to measure cluster sizes and coordination numbers and finally to compare the mixing in a series of experimental samples.