Neural network classification of mobility spectra for volatile organic compounds using tandem differential mobility spectrometry with field induced fragmentation

Abstract

A spectral library of field induced fragmentation (FIF) spectra for 45 oxygen-containing volatile organic compounds from 5 chemical classes was obtained using tandem differential mobility spectrometry (DMS). Protonated monomers were mobility isolated in a first DMS stage, fragmented with electric fields >10,000 V/cm in a middle (or reactive) stage, and mobility characterized in a second DMS stage. Other spectral libraries were obtained for protonated monomers and for complete mobility spectra from a single DMS stage. Neural networks from Python/Tensorflow software, prepared in-house, and from commercial NeuralWorks Professional II/PLUS were trained to assign spectra into a chemical class. The success at classification was determined for familiar and unfamiliar spectra from these three libraries. Classification test scores were best with FIF spectra with >0.99 for familiar compounds and 0.52 for unfamiliar compounds and were consistent with neural network learning of structural information from fragment ions when compared to other spectral libraries. Radar charts are introduced as measures of classification and as a tool to explore mis-classification. This work shows that ion fragmentation with multi-stage tandem DMS portends molecular identification with the portability and robustness of ambient pressure ion mobility analyzers.