Machine learning approaches for automatic classification of single-particle mass spectrometry data

Abstract. The chemical composition of aerosol particles is a key parameter for human health and climate effects. Single-particle mass spectrometry (SPMS) has evolved to a mature technology with unique chemical coverage and the capability to analyze the distribution of aerosol components in the particle ensemble in real-time. With the fully automated characterization of the chemical profile of the aerosol particles, selective real-time monitoring of air quality could be performed e.g. for urgent risk assessments due to particularly harmful pollutants. For aerosol particle classification, mostly unsupervised clustering algorithms (ART-2a, K-means and their derivatives) are used, which require manual post-processing. In this work, we focus on supervised algorithms to tackle the problem of automatic classification of large amounts of aerosol particle data. Supervised learning requires data with labels to train a predictive model. Therefore, we created a labeled benchmark dataset containing ~24,000 particles with eight different coarse categories that were highly abundant at a measurement in summer in Central Europe: Elemental Carbon (EC), Organic Carbon and Elemental Carbon (OC-EC), Potassium-rich (K-rich), Calcium-rich (Ca-rich), Iron-rich (Fe-rich), Vanadium-rich (V-rich), Magnesium-rich (Mg-rich) and Sodium-rich (Na-rich). Using the chemical features of particles the performance of the following classical supervised algorithms was tested: K-nearest neighbors, support vector machine, decision tree, random forest and multi-layer perceptron. This work shows that despite the entrenched position of unsupervised clustering algorithms in the field, the use of supervised algorithms has the potential to replace the manual step of clustering algorithms in many applications, where real-time data analysis is essential. For the classification of the eight classes, the prediction accuracy of several supervised algorithms exceeded 97 %. The trained model was used to classify ~49,000 particles from a blind dataset in 0.2 seconds, taking into account also a class of “unclassified” particles. The predictions are highly consistent with the results obtained in a previous study using ART-2a.