Neural networks for classification and segmentation of thermally-induced droplet breakup in spray-flame synthesis

Abstract

Spray-flame synthesis (SFS) enables one-step production of functional metal-oxide nanoparticles from inexpensive precursors such as metal salts. Precursor-laden droplets show thermally-induced breakup, i.e., puffing and micro-explosion, considered a key step in SFS. In this work, shadowgraphy images of droplets are investigated with neural networks with the aim of extracting quantitative data on this breakup process. Faster and Mask region-based convolutional neural networks (R-CNNs) were applied to segment in-focus droplet shadowgraphs (Mask R–CNN) and to distinguish regular from disrupting droplets and detect sequences of consecutive droplet shadowgraphs (Faster R–CNN). The networks were trained with about 400 manually annotated images containing a total of about 1200 objects. Precision and recall of the trained networks reach 80% and 90%, respectively. The results from the two networks were combined and compared with those from conventional particle/droplet image analysis (PDIA). Mask R–CNN and PDIA identify similar droplets as being in-focus and segment similar regions of the droplets. Puffing or micro-explosion sometimes leave behind secondary droplets that are too small or defocused for reliable segmentation. Such shadowgraphs are classified erroneously by PDIA. In contrast to this, Faster R–CNN is more sensitive as it does not rely on the segmentation and thus classifies such low-contrast shadowgraphs correctly. Therefore, the CNNs find on average 30% higher droplet-disruption probabilities than PDIA.