Source Identification of Atmospheric Particles via Low-Voltage Electron Microscopy Image Recognition: A Case Study of Submicrometer Particles

Atmospheric particulate matter from different sources exhibits distinct morphological characteristics, which can be automatically extracted through deep learning for source identification. Computer-controlled scanning electron microscopy (CCSEM) offers exceptionally high analytical throughput (>1000 particles/hour), enabling the broad application of morphology-based source apportionment. However, conventional CCSEM analysis excels with micrometer-sized particles but struggles to capture the fine morphological characteristics of submicrometer particles, limiting its effectiveness in submicrometer particle source apportionment. Herein, we proposed a low-voltage (5 kV) CCSEM analysis method to precisely capture particle morphology and developed a deep learning model, ImageTracerNet, to achieve accurate submicrometer particle (0.2–1.0 μm range) source apportionment. In test set evaluations, the 5 kV-image-driven ImageTracerNet model improved accuracy by over 40% compared to the same model driven by 20 kV images (from conventional CCSEM analysis) and outperformed the multidimensional-data-driven RX model (using 20 kV images, elements, size) by over 20%. Additionally, we further demonstrated the model’s effectiveness in real atmospheric environments. Overall, this study provides an efficient method for source apportionment of atmospheric submicrometer particles, supporting the targeted control of submicrometer particle pollution sources, particularly primary emissions.