Genetic algorithms and neural networks for the quantitative analysis of ternary mixtures using surface plasmon resonance

Recently, time-resolved measurements have been proposed for sensor research to reduce the number of sensors needed for a multicomponent analysis. These measurements usually generate many variables, which are unfortunately highly correlated, creating several problems in data analysis. In this study, a variable selection algorithm is presented, which is optimized for limited data sets with correlated variables. The algorithm is based on many parallel runs of genetic algorithms (GA). The calibration is performed using neural networks. The algorithm is successfully applied to the selection of time points of time-resolved measurements performed by a single transducer surface plasmon resonance (SPR) setup. The selection of the time points enables an improved calibration of the vapor concentrations of three analytes in ternary mixtures. The relative root mean square errors of prediction of an external validation data set by the optimized models were 3.6% for methanol, 5.9% for ethanol and 7.6% for 1-propanol. The variable selection is reproducible and not affected by chance correlation of variables. The selected time points give insight into characteristic sensor responses of the pure analytes, and it is shown that the analysis time can be significantly reduced.