Recognition of Nanoparticle and Microdroplet Parameters in Levitating Microdroplets of Suspension by Speckle Image Analysis Using Convolutional Neural Networks

Abstract

Retrieving suspension parameters from light scattered by a suspension microdroplet is a difficult inverse problem because the recorded optical pattern is shaped by droplet morphology and by perturbations of the internal optical field introduced by the dispersed phase. Here we demonstrate experimental, data-driven recognition of microdroplet and nanoparticle-suspension parameters from laser speckle images recorded for individual levitated suspension microdroplets. We confined slowly evaporating TiO${}_2$ nanoparticle suspensions in diethylene glycol in a linear electrodynamic quadrupole trap and classified their speckle patterns using a convolutional neural network (CNN). The CNN was trained and tested to recognize droplet diameter, nanoparticle mass concentration, and nanoparticle diameter, first separately and then in simultaneous two- and three-parameter tasks. The CNN recognized droplet diameter most robustly, with an estimated accuracy of approximately 6% on independent test data. Nanoparticle concentration required sufficiently separated classes. Nanoparticle-diameter classes were also recognized, although particle-size effects were coupled to changes in particle number at fixed mass concentration. We demonstrated simultaneous three-parameter classification across 27 classes. As an exploratory out-of-distribution test, we applied a CNN trained on single-component suspensions to dual-component suspensions containing two nanoparticle diameters. The CNN retained meaningful predictive capability, especially at higher concentrations, although robust decomposition of mixed nanoparticle populations was not achieved. These results show that speckle images of single levitated suspension microdroplets contain sufficient information for CNN-based multi-parameter recognition and support further development of speckle-based optical diagnostics for complex microdroplet and aerosol-like systems.

17 pages, 11 figures, 3 tables; associated with LIPLII2026 : Light Interaction with Particles / Laser Induced Incandescence conference, Rouen, France