Advancing automatic classification of single fine particles via multimodal deep learning framework

The accurate identification and classification of single fine particles are essential for understanding emission sources, atmospheric processes, and their subsequent impacts on climate and health. However, this task remains challenging due to the highly diverse and overlapping physicochemical characteristics of various particle types. Conventional single particle analysis relies heavily on manual interpretation, which is inherently slow, subjective, and limited in scalability. To overcome these limitations, this study implemented a multimodal deep learning fusion framework that classifies fine particles into seven distinct types: dust (206), fly ash (64), metals (164), organic-rich (551), sea spray (91), soot (299), and sulfur-rich (247), by simultaneously fusing three modalities (morphology, size, and elemental composition). A comprehensive dataset of 1,622 single particles, collected from diverse emission sources and ambient environments, was utilized for model development and evaluation. Three multimodal fusion strategies were systematically evaluated: early fusion (concatenating all modalities at an early stage), traditional joint fusion (integrating independently extracted features), and attention-based fusion (dynamically focusing on the most informative features within and across modalities). The attention-based fusion architecture achieved the highest predictive performance, yielding an overall accuracy of 89.6%, a sensitivity of 0.90, and a specificity of 0.98. Furthermore, an interpretability analysis revealed that the model’s classification decisions were predominantly governed by distinctive elemental composition and morphological patterns. By rapidly classifying thousands of individual particles within minutes, this automated, high-throughput framework provides a highly scalable solution for single-particle analysis, thereby facilitating a more profound understanding of atmospheric aerosol sources, transformation mechanisms, and their broader climate and health impacts.