Qualitative Analysis of Chlorinated Solvents using Raman Spectroscopy and One-Sided Classification Machine Learning Techniques.

Frank G. Glavin & Michael G. Madden
National University of Ireland, Galway & Analyze IQ Limited.

Raman spectroscopy is an established technique for non-destructive analysis of complex mixtures in solid or liquid state, which can be used as the basis for both quantitative and qualitative analysis of materials. In this paper, we explore th application of Machine Learning (ML) to qualitative analysis of Raman spectra.

In particular, we focus on ML classification algorithms, which are concerned with predicting the class of unknown samples after being trained with a set of known samples. Our research considers a practical application in separating hazardous solvents from other materials, more specifically, the task of identifying whether a substance contains a chlorinated solvent, based on its Raman spectrum.  While the training data set contains many negatives, they do not represent everything that is not a solvent. Accordingly, propose the use of one-class classification algorithms, which are distinguished form conventional (multi-class) classification algorithms in that they assume that only one class is well-characterized in the training set. Objects of this class are distinguished from all others, referred to as outliers that consist of all the other objects that are not targets.

Using a new one-sided classification toolkit that we have developed, we compare a one-sided classification algorithm with two well-known conventional algorithms and conclude that the one-sided approach is more robust in subsequent use, because the outlier concept (materials that are not chlorinated solvents) is not exhaustively characterized in the training set.