Particle-shape classification using light scattering: An exercise in deep learning. (July 2019)

• Record Type: Journal Article
• Authors: Piedra, Patricio; Kalume, Aimable; Zubko, Evgenij; Mackowski, Daniel; Pan, Yong-Le; Videen, Gorden
• Is Part Of: Journal of quantitative spectroscopy & radiative transfer. Volume 231(2019)
• Page Start: 140
• Page End: 156
• Publication Date: 2019-07
• Subjects: Light scattering -- Deep learning; Spectrum analysis -- Periodicals; Radiation -- Periodicals; Analyse spectrale -- Périodiques; Rayonnement -- Périodiques; Radiation; Spectrum analysis; Periodicals; 543.0858
• DOI: 10.1016/j.jqsrt.2019.04.013
• Languages: English
• ISSNs: 0022-4073

Abstract

Highlights: Seven common naturally occurring shapes of aerosol particles were modeled: cuboids, ellipsoids, cylinders, agglomerated debris, sphere clusters, spore clusters, and fractal aggregates. All particles were of volume-equivalent size parameter 5, refractive index 1.5 + i 0, stochastically and isotropically rotated, and had Monte-Carlo shaped-dependent variations. A fully connected neural network was used for 1D, and a convolutional neural network was used for 2D scattering patterns. Classification accuracy was approximately 70% for regular particles and greater than 90% accuracy for irregular particles. Particles with pattern-reoccurring irregularities were classified with higher accuracy than regular particles. Abstract: We apply machine-learning algorithms to the calculated light-scattering patterns from particles having seven different common and naturally occurring shapes to assess the accuracy of shape classification based on light scattering. We consider different input data sets including one- and two-dimensional scattering functions of both intensity and polarization. Our scattering data set is produced from particles of volume-equivalent size parameter 5, and refractive index m = 1.5 + 0i. As expected, classification capabilities were much greater when the two-dimensional scattering data were used than when only one-dimensional data were considered. When the two-dimensional intensity patterns are considered, classification accuracies were approximately 70% for the regularly shaped particles and above 90% for the highly irregularly shaped particles. These capabilities increased slightly when linear polarization was used as input. Although all our results are specific to our particular data set, machine-learning techniques are easily generalizable. This exercise suggests that particle discrimination can be achieved in practical experiments using light-scattering patterns through deep learning.