Characterizing Ensembles of Platelike Particles via Machine Learning

Abstract

The size and shape characterization of platelike particles using a dual projection imaging device is presented. Based on the published algorithm to estimate the three lengths of the particles from the image pairs (method 1), the average error on the shortest length is around 140% on a test set of particles with different sizes, shapes, and alignment. Therefore, two potential improvements are tested in a simulation setting. The first approach uses an additional projection coupled with an oriented bounding box to estimate the three lengths (method 2). The second approach uses a machine learning model to estimate the three lengths from two projections (method 3). It is shown that method 2 marginally increases the accuracy, while method 3 leads to a significant improvement with an average error of 33% for the shortest characteristic length of particles in the test set and even less for the other lengths. It is shown that this remaining error is caused by the particle alignment with respect to the cameras. Since platelets cannot always be automatically distinguished from quasi-equant particles due to alignment issues, the model was also trained to estimate the three characteristic lengths of quasi-equant particles. The training set and the test set, used for model training and validation, respectively, comprise particles that exhibit quasi-equant, needlelike, and platelike shapes. Several machine learning models are identified and optimized to predict the three particle lengths based on the two projections from the imaging device. Artificial neural networks were chosen because of their superior predictive performance.

A novel image analysis technique for 2D characterization of overlapping needle-like crystals

“…While the presented algorithm is very effective for needle-like particles, it is indeed specialized and limited to separating and sizing such objects and would perform poorly on objects that are not defined by two parallel edges, e.g., round overlapping objects or objects with entirely different characteristics. One avenue to resolve this would be to develop further branches of methodology that could be used to deal with non needle-like morphologies along with a pre-classification step to decide which algorithm to use on the shapes at hand, as other authors have shown previously [25,26].…”

“…Data from one-dimensional techniques have been used for the derivation of two-dimensional PSSDs by assuming constant aspect ratios or narrow size distributions of the particles' width [12,14,19,20]. For a highly accurate description of anisotropic morphologies, however, imaging provides the best solution since it allows for direct measurement of the length and width of particles [21,22] and even for a reconstruction of three-dimensional shapes from dual projections [23][24][25][26].…”

Online 3D Characterization of Micrometer‐Sized Cuboidal Particles in Suspension

“…This is attributed to the effect of particle orientation, which is further elucidated through simulations. [40] Visually, the different outputs of the two methods can be appreciated by means of violin plots, shown on the right-handside of panels (a), (b), and (c) of Figure 3. Violin plots are useful graphical tools to display the density distributions in a compact fashion, thus allowing a rapid comparison among distributions.…”

“…Accordingly, the goals of the current study are twofold: (a)to show how monodisperse populations of arbitrarily shaped micrometer‐sized cuboidal particles can be fabricated by photolithography and employed as novel analytical standards; (b)to experimentally demonstrate how a multiprojection imaging device [ 39 ] coupled with a machine learning (ML) algorithm [ 40 ] is able to yield online and accurate estimates of the three characteristic lengths of such cuboidal particles in suspension. …”

In Situ Imaging Combined with Deep Learning for Crystallization Process Monitoring: Application to Cephalexin Production

“…However, they either still deal with dilute slurries with a low solid density and limited object overlap in the captured images or rely on external loops to generate analyzable images. Neural networks as regression models are also used for shape characterization of plate-like crystals based on dual-projection images 28 or to extract information about solution and solid-phase concentrations from Raman and IR spectra in different crystallization processes. 29,30 In the present work, we develop a computer vision model, based on CNNs, that can be combined with high-throughput in situ imaging using a PVM probe for live monitoring of crystallization processes.…”