Machine Learning Techniques for Rapid Analysis of Particle Morphology and Elemental Composition

BORIS GALVIS, Fabiana Franceschi, Natali Zambrano, David Restrepo, Olga Lucia Quintero Montoya, Jose Duque, Elena Montilla Rosero, Daniela Bustos, Nestor Rojas, Universidad del Valle

     Abstract Number: 130
     Working Group: Source Apportionment

Abstract
Morphological characterization and elemental analysis of single particles using scanning electron microscopy (SEM) and energy-dispersive X-ray spectroscopy (EDX) have been effective techniques for identifying the sources of ambient particles. However, analyzing a large number of particles using these methods can be time-consuming. To address this issue, we developed an image classifier based on machine learning algorithms that utilizes computer vision and pattern recognition to automate a significant portion of the SEM-EDX particle analysis process.

We first tested five different models using SEM-EDX images from NASA's cosmic dust catalog VOL XV and evaluated their performance on the same image set. The models based on EfficientNet and ResNet exhibited the highest accuracy. We then further trained and tested our algorithms using thousands of SEM-EDX images of particles collected in three cities in Colombia (Bogotá, Cali, and Medellín). The algorithm was able to classify particles into four morpho-chemical groups (organic biogenic, metallic, mineral, and tire wear) within hours.

Our image classifier based on EfficientNet and ResNet achieved accuracies of 80% and 85%, respectively. Our ultimate goal is to develop a source apportionment tool based on this technology and compare it to traditional methods.