Applying Machine Learning to Determine the Biogenic Source of Atmospheric Aerosols

The composition of aerosol in the atmosphere is inherently linked to effects on climate. However, apportioning aerosol sources is often difficult and time consuming due in part to the diversity of origins. This project therefore developed applied machine learning algorithms to classify aerosols by source according to their measured organic composition. Samples were taken from the Southern Great Plains Atmospheric Radiation Measurement site and analyzed with a high-resolution mass spectrometer. These samples were used to train and test multiple models, including random forest, XGBoost, gradient boost, and neural network, using Python libraries. These models predicted whether the aerosol originated from isoprene, terpene, or a different primary aerosol source. F1 scores for the first three models were generally greater than 0.90, an indicator of a good model. These models predicted the percent compositions of all samples with high accuracy, within 0.5% of the true values for all classes. Hyperparameter tuning of the XGBoost and gradient boosting classifier using a grid search algorithm provided improved F1 scores, increasing by 0.01-0.06. After predicting aerosol class, any previously unidentified aerosols predicted to be isoprene or terpene-related were compared to the aerosols used to label the training data through linear regression. Many formulas had R² values greater than 0.7, indicating good correlation and suggesting the possibility of relation to the sources. Future work will divide the “other” class into anthropogenic aerosol types and introduce a biomass burning component. Other classes can be created as the supporting science, such as chamber studies, advances. Such chamber studies can also be used to support or repudiate the identified correlations.