Quantitative Characterization of Single Particle Combustion using X-ray Phase Contrast Imaging and Machine Learning
Lehansa Marambage
Jose Marti STEM Academy, Union City, United States
Publication date: October 23, 2024
Jose Marti STEM Academy, Union City, United States
Publication date: October 23, 2024
DOI: http://doi.org/10.34614/JIYRC202416
ABSTRACT
Metal powders are potential fuel additives in explosives, propellants, and the degradation of chemical warfare agents. Current studies’ high-speed imaging of metal combustion produces challenging datasets for existing machine learning (ML) algorithms to accurately process. To address this, ML and computer-vision were employed to detect combusting metal particles, (using High-speed X-ray Phase Contrast Imaging) for yielding quantitative data on reaction mechanisms, informing optimization of metal combustion performance. MATLAB was used to develop specialized object-detection models including a region-based convolutional neural network (RCNN), Cascade detector (Viola-Jones algorithm/ histogram of oriented gradients), and Foreground detector (Gaussian mixture model). Custom scripts were written to compile an open-source training dataset and validate object-detection results. Results showed while the Cascade detector has high detection-accuracy, it had a bounding-box precision <50%. The RCNN requires more computational resources but offers higher detection-accuracy as a stand-alone ML model. These results lay the groundwork for future studies integrating these models with classification/segmentation algorithms for scenarios like particle dispersion in nuclear blasts.
Metal powders are potential fuel additives in explosives, propellants, and the degradation of chemical warfare agents. Current studies’ high-speed imaging of metal combustion produces challenging datasets for existing machine learning (ML) algorithms to accurately process. To address this, ML and computer-vision were employed to detect combusting metal particles, (using High-speed X-ray Phase Contrast Imaging) for yielding quantitative data on reaction mechanisms, informing optimization of metal combustion performance. MATLAB was used to develop specialized object-detection models including a region-based convolutional neural network (RCNN), Cascade detector (Viola-Jones algorithm/ histogram of oriented gradients), and Foreground detector (Gaussian mixture model). Custom scripts were written to compile an open-source training dataset and validate object-detection results. Results showed while the Cascade detector has high detection-accuracy, it had a bounding-box precision <50%. The RCNN requires more computational resources but offers higher detection-accuracy as a stand-alone ML model. These results lay the groundwork for future studies integrating these models with classification/segmentation algorithms for scenarios like particle dispersion in nuclear blasts.