Developing and Testing Machine Learning Algorithms to Process SEM-EDX Images

BORIS GALVIS, Natali Zambrano, David Restrepo, Olga Lucia Quintero Montoya, Nestor Rojas, Elena Montilla Rosero, Jose Duque, Universidad de La Salle

     Abstract Number: 594
     Working Group: Source Apportionment

Abstract
In this work, we developed machine learning models to analyze the results SEM (Scanning Electron Microscope)-EDX (Energy Dispersive X-Ray Spectometry) technique of individual particles to later relate their morphological and chemical characteristics. Five models were developed with a database of 466 NASA Cosmic Dust Catalog Vol 15 (1997) images, classified into four groups: Cosmic (C), Artificial Terrestrial Contamination (TCA), Natural Terrestrial Contamination (TCN) and Aluminum or Aluminum Oxide Sphere (AOS) in order to identify which model best fit the cosmic dust images. We are also performing individual particle analysis with the digital image processing program ImageJ (developed at the National Institutes of Health) in order to obtain morphological characteristics sucha as area, perimeter, aspecto ratio, circularity and solidity. The results show that the optimal model was the one developed by transfer-learnig with the pre-trained model "efficientnetv2-b1". In the future, we intend to use the developed machine learning code to characterize SEM/EDS results from coarse particulate matter sampling in the cities of Bogota, Cali and Medellin, Colombia.