Machine learning: our future spotlight into single-particle ICP-ToF-MS analysis

Abstract

Using the multi-element capabilities of single-particle inductively coupled plasma time-of-flight mass spectrometry (spICP-ToF-MS) in combination with a laser ablation introduction system, environmentally relevant road runoff samples from three different sampling points were measured. Pearson correlations were used to find trends of element correlations, and t-distributed stochastic neighbour embedding (TSNE) to reduce data set dimensions for more effective visualization. Finally, classes of particles for multi-elemental particles (MEPs) were proposed. The particle elemental trends and correlations were compared with literature-reported elemental particle fingerprints. Ultimately, three major classes of particles were identified namely based on the literature, rare earth elements (REEs) which include both potential anthropogenic and geogenic sources, brake and tire wear, and platinum group elements (PGEs). All samples were compared based on these arbitrary classes, which showed discernible differences between the sample locations. The information gained from correlation and TSNE analysis in combination with the reported literature elemental markers was used to manually label a dataset of particles for testing of a supervised classification algorithm. Using a LightGBM multiclass classifier, an effective data processing pipeline was created. The machine learning model ultimately automates the work of dataset labeling and classification, allowing for a quick and efficient method for inter/intra sample comparison in terms of MEP elemental correlations.