Abstract. Single particle soot photometers (SP2) use laser-induced incandescence to detect aerosols on a single particle basis. SP2s that have been modified to provide greater spectral contrast between their narrow and broad-band incandescent detectors have previously been used to characterize both refractory black carbon (rBC) and light-absorbing metallic aerosols, including iron oxides (FeOx). However, single particles cannot be unambiguously identified from their incandescent peak height (a function of particle mass) and color ratio (a measure of blackbody temperature) alone. Machine learning offers a promising approach for improving the classification of these aerosols. Here we explore the advantages and limitations of classifying single particle signals obtained with a modified SP2 using a supervised machine learning algorithm. Laboratory samples of different aerosols that incandesce in the SP2 (fullerene soot, mineral dust, volcanic ash, coal fly ash, Fe2O3, and Fe3O4) were used to train a random forest algorithm. The trained algorithm was then applied to test data sets of laboratory samples and atmospheric aerosols. This method provides a systematic approach for classifying incandescent aerosols by providing a score, or conditional probability, that a particle is likely to belong to a particular aerosol class (rBC, FeOx, etc.) given its observed single particle features. We consider two alternative approaches for identifying aerosols in mixed populations based on their single particle SP2 response: one with specific class labels for each species sampled, and one with three broader classes (rBC, anthropogenic FeOx, and dust-like) for particles with similar SP2 responses. Predictions of the most likely particle class (the one with the highest mean probability) based on applying the trained random forest algorithm to the single particle features for test data sets comprising examples of each class are compared with the true class for those particles to estimate generalization performance. While the specific class approach performed well for rBC and Fe3O4 (≥99 % of these aerosols are correctly identified), its classification of other aerosol types is significantly worse (only 47 %–66 % of other particles are correctly identified). Using the broader class approach, we find a classification accuracy of 99 % for FeOx samples measured in the laboratory. The method allows for classification of FeOx as anthropogenic or dust-like for aerosols with effective spherical diameters from 170 to >1200 nm. The misidentification of both dust-like aerosols and rBC as anthropogenic FeOx is small, with <3 % of the dust-like aerosols and <0.1 % of rBC misidentified as FeOx for the broader class case. When applying this method to atmospheric observations taken in Boulder, CO, a clear mode consistent with FeOx was observed, distinct from dust-like aerosols.
Computing single-particle flotation kinetics using automated mineralogy data and machine learning
