Studies by Environment Canada and Health Canada have shown that levels of airborne particles in major Canadian urban centers are high enough to pose human health risks. [1]   This particulate matter (PM) is also involved in a number of atmospheric issues including acid precipitation, visibility degradation, and smog formation.  Physical and chemical characterization of PM, depending on the sophistication of the PM monitoring technique applied, can provide important insights into the origin and evolution of aerosol particles during their lifetime and any potential inhalation hazards.  A recent advancement in PM characterization has been the initiation of techniques to permit on-line, single particle laser ablation mass spectrometry (LAMS).  As distinct from conventional ‘off-line’ filter collections of 24h bulk PM samples, LAMS can provide the simultaneous size and inorganic/organic analysis of individual particles on a continuous, ‘on-line’ basis.  Some aerosol LAMS instruments can analyze up to 10 particles/second. [2] However, the vast ambient PM dataset from on-line operation of an aerosol LAMS requires classification of mass spectra into an interpretable summarized format. 

A first approach to identifying groups within a dataset without a priori knowledge of the classes expected has been to apply a pattern recognition program such as a hierarchical cluster analysis (HCA), or an unsupervised learning neural network such as the adaptive resonance theory-2a (ART-2a). [3] ^{, [4]}   Typically these methods identify similarities between spectra and apply a linkage criterion to decide which spectra to group together. Once the chemical classes expected in a dataset are known, HCA and the ART-2a programs can refine such operating criteria, but inherently they do not train their analyses to include known information about the spectral data.  For instance, it has been shown that the spectral response in aerosol LAMS can vary significantly for some standard chemical particles, and be subtly different but fairly reproducible for others; however HCA and ART-2a only apply one linkage criterion causing either inherently varying spectra to be inaccurately separated, or reproducible spectra with slight distinctions to be erroneously grouped.  As for the analysis rates of HCA and ART-2a, they are run in batch or semi-batch mode where groups from one batch cannot always be matched to groups from another, thus requiring slower user interpretation of the classes that are similar.  A method that incorporates more chemical information, and runs continuously without any manual interaction can hence improve spectral classification. 

Discriminant analysis (DA) is a multivariate method that groups data into assigned classes, but has not been explored for classification of mass spectra.  One reason is because early aerosol LAMS researchers did not know the chemical classes to expect in ambient PM and hence exploratory HCA and ART-2a were applied so that even small ambient particle variability was completely investigated. But these minor classes have been mainly ignored in the literature. Another reason against DA of mass spectra is because DA alone cannot group spectra into new classes should they arise in the PM population, and it would be unreasonable to assume that every chemical class could be assigned beforehand. 

In this paper, a newly-developed, Algorithm for Discriminant Analysis of Mass Spectra – ADAMS - is presented which successfully classified on-line aerosol LAMS mass spectral data into chemically-assigned groups.  ADAMS incorporated a Remainder feature to allow DA the flexibility to classify new, unassigned classes and overcome the limitations in conventional DA.  The design objective of ADAMS was to create a reproducible, chemically interpretable set of classes for dominant ambient spectra by utilizing the advantages of DA, while allowing minor classes to still be categorized, but not emphasized, in a Remainder.  ADAMS validation and application to ambient PM data are also described.  This work was part of the development of Canada’s first aerosol LAMS system.

---