A machine learning approach to aerosol classification for single-particle mass spectrometry

Costa D Christopoulos,Daniel J Cziczo,Ottmar Möhler,Sarvesh Garimella,Maria A Zawadowicz

doi:10.5194/amt-11-5687-2018

Abstract

Abstract. Compositional analysis of atmospheric and laboratory aerosols is often conducted via single-particle mass spectrometry (SPMS), an in situ and real-time analytical technique that produces mass spectra on a single-particle basis. In this study, classifiers are created using a data set of SPMS spectra to automatically differentiate particles on the basis of chemistry and size. Machine learning algorithms build a predictive model from a training set for which the aerosol type associated with each mass spectrum is known a priori. Our primary focus surrounds the growing of random forests using feature selection to reduce dimensionality and the evaluation of trained models with confusion matrices. In addition to classifying ∼20 unique, but chemically similar, aerosol types, models were also created to differentiate aerosol within four broader categories: fertile soils, mineral/metallic particles, biological particles, and all other aerosols. Differentiation was accomplished using ∼40 positive and negative spectral features. For the broad categorization, machine learning resulted in a classification accuracy of ∼93 %. Classification of aerosols by specific type resulted in a classification accuracy of ∼87 %. The “trained” model was then applied to a “blind” mixture of aerosols which was known to be a subset of the training set. Model agreement was found on the presence of secondary organic aerosol, coated and uncoated mineral dust, and fertile soil.

Highlights

Following the introduction of random forests in the 1990s, recent developments in deep learning and neural networks have helped to trigger a renewed interest in machine learning
We show that supervised training with random forests can differentiate aerosols in single-particle mass spectrometry (SPMS) data more accurately than simpler approaches
Confusion matrices represent model predictions as columns i and true aerosol type or category as rows j,where class names are mapped to integers i, j ∈ {1, 2, . . ., y}

Summary

Introduction

Following the introduction of random forests in the 1990s, recent developments in deep learning and neural networks have helped to trigger a renewed interest in machine learning. While random forests have been used for complex classification and regression analysis in various fields, studies that employ random forests in aerosol mass spectrometry remain sparse Utilizing these tools, the primary purpose of our study is to introduce a framework for growing random forests, reducing dimensionality, ranking chemical features, and evaluating performance using confusion matrices. The primary purpose of our study is to introduce a framework for growing random forests, reducing dimensionality, ranking chemical features, and evaluating performance using confusion matrices Such properties are desirable for SPMS studies, where input variables can become redundant and interpretability is more limited with more advanced methods such as neural networks. Analysis techniques such as those coming out of recent artificial intelligence research can prove useful for helping to tease out the subtle yet significant impact that aerosol chemistry has on the climate system

Objectives

Results

Conclusion