Shape-independent particle classification for discrimination of single crystals and agglomerates

• Published in: Powder technology 2019-03, Vol.345, p.425-437
• Main Authors: Heisel, Stefan, Ernst, Janine, Emshoff, Arne, Schembecker, Gerhard, Wohlgemuth, Kerstin
• Format: Article
• Language: English
• Subjects: Agglomerates; Agglomeration; Artificial neural networks; Aspect ratio; Classification; Classifiers; Crystal structure; Crystallinity; Crystallization; Crystals; Discrimination; Glutamic acid; Image analysis; Multivariate analysis; Neural networks; Particle classification; Particulates; Polymorphism; Process parameters; Similarity; Single crystals
• ISSN: 0032-5910; 1873-328X
• DOI: 10.1016/j.powtec.2019.01.018

While agglomeration has significant effects on particulate products, quantification is still a time-consuming process. Particle classification using multivariate analysis can help gain an understanding of these agglomeration processes, but the necessary classifiers are often applicable to one type of particles only. This study focuses on the generation of a particle classifier for the discrimination of single particles/agglomerates which is applicable to a variety of particulate systems of a different shape. This might be of importance for solids that change their shape, e.g., crystalline systems that may change their aspect ratio or habit according to different process parameters, impurity concentrations, or polymorphic form. It was found that artificial neural networks can perform the discrimination task of single crystal/agglomerate for several crystalline systems when the training set with whose help the classifier is generated contains a selection of crystals that cover a wide range of possible crystal shapes. Variable selection using proportional similarity generated a highly accurate classifier while only a little time needed to be invested. Proportional similarity not only proved helpful for the discrimination task of single crystal/agglomerate but differentiation of the α and β polymorphs of l-glutamic acid as well. Using the information from particle classification, a more in-depth characterization of how single particles, agglomerates, or different particle shapes are distributed can be given. [Display omitted] •Artificial neural network for particle discrimination (single particles/agglomerates)•ANN training process with the 3 different crystal habits•Variable selection of lesser importance than training set design•Visualization of PSD subpopulations made possible•Procedure extensible to discrimination of crystal shapes (needles and platelets)