Concentrations in GC–MS using
electron-ionization mass spectrometry
can be determined without pure calibration standards through prediction
of relative total-ionization cross sections. An atom- and group-based
artificial neural network (FF-NN-AG) model is created to generate
EI cross sections and calibrations for organic compounds. This model
is easy to implement and is more accurate than the widely used atom-additivity-based
correlation of Fitch and Sauter (Anal. Chem. 1983).
Ninety-two new measurements of experimental EI cross sections (70–75
eV) are joined with different interlaboratory datasets, creating a
396-compound cross-section database, the largest to date. The FF-NN-AG
model uses 16 atom-type descriptors, 79 structural-group descriptors,
and one hidden layer of 10 nodes, trained 500 times. In each cycle,
96% of the compounds in this database are freshly chosen at random,
and then the model is tested with the remaining 4%. The resulting r
2 is 0.992 versus 0.904 for the Fitch and Sauter
correlation, root mean square deviation is 2.8 versus 9.2, and maximum
relative error is 0.30 versus 0.73. As an example of the model’s
use, a list of cross sections is generated for various sugars and
anhydrosugars.