Numerical neural network approach to simultaneous material classification and sizing of aerosolized particles

Abstract We employ a neural network approach to classify aerosolised spherical particles according to both constituent material and size. The classification is based on simulated light scattered intensities in order to emulate the data obtained from an optical particle counter. Detection configurations of varying complexity, involving up to three wavelengths,forward- and backward-scattering detection and two polarizations, are simulated.Rather than attempting to identify specific constituent materials, distinction of particles into general material categories was attempted. The neural network approach generally showed good accuracy in both sizing and classification of particles into the four material categories non-metal, metal, soot, and water when testing on the same particle types the network was trained on. Significant improvements over conventional optical particle counter geometries were predicted when introducing additional wavelengths and an additional detector to the simulated setup, and overall size and classification accuracies of > 95% were achieved, in compliance with current standards. When testing the predictive power of the neural network on particle materials outside of the training ensemble, classification accuracies as high as 96% were observed for materials with properties within the region of the training data, while materials with optical properties far removed from the training particles, or near the boundaries of the chosen material categories, were more easily misclassified.