Machine Learning Classification of Single-Particle Mass Spectra

Compositional analysis of atmospheric and laboratory aerosols is essential to understanding their impact on climate and the environment. Single-particle mass spectrometry (SPMS) is an in situ and real-time analytical technique that provides mass spectra for single particles. However, the large amount of data generated by SPMS can be challenging to analyze. In this study, we leverage the power of machine learning to develop classifiers using a comprehensive dataset of SPMS spectra. These classifiers enable automatic differentiation of aerosol particles based on their chemistry and size, facilitating more accurate and efficient aerosol classification. We trained supervised support vector machine (SVM) algorithms and a stacked autoencoder classifier. We trained the models on both labeled and unlabeled mass spectra. The unsupervised behavior was enabled by pseudo label propagation for the SVM algorithm and the Mean-teacher framework for the stacked autoencoder model. Our preliminary results show increased accuracy when including unlabeled data through both frameworks for SVM and stacked autoencoders. Here we present a comparison of both models' architectures, their increased performance after unlabeled data is included, and finally a comparison on performance across models. Our work suggests that machine learning has the potential to be a powerful tool for classifying SPMS data. Our work provides a foundation for further research in this area.