Evaluation of a generalized regression artificial neural network for extending cadmium's working calibration range in graphite furnace atomic absorption spectrometry

A generalized regression artificial neural network (GRANN) was developed and evaluated for modeling cadmium's nonlinear calibration curve in order to extend its upper concentration limit from 4.0 μg L^-1 up to 22.0 μg L^-1. This type of neural network presents important advantages over the more popular backpropagation counterpart which are worth exploiting in analytical applications, namely, (1) a smaller number of variables have to be optimized, with the subsequent reduction in "development hassle"; and, (2) shorter development times, thanks to the fact that the adjustment of the weights (the artificial synapses) is a non-iterative, one-pass process. A backpropagation artificial neural network (BPANN), a second-order polynomial, and some less frequently employed polynomial and exponential functions (e.g., Gaussian, Lorentzian, and Boltzmann), were also evaluated for comparison purposes. The quality of the fit of the various models, assessed by calculating the root mean square of the percentage deviations, was as follows: GRANN > Boltzmann > second-order polynomial > BPANN > Gauss > Lorentz. The accuracy and precision of the models were further estimated through the determination of cadmium in the certified reference material "Trace Metals in Drinking Water" (High Purity Standards, Lot No. 490915), which has a cadmium certified concentration (12.00±0.06 μg L^-1) that lies in the nonlinear regime of the calibration curve. Only the models generated by the GRANN and BPANN accurately predicted the concentrations of a series of solutions, prepared by serial dilution of the CRM, with cadmium concentrations below and above the maximum linear calibration limit (4.0 μg L^-1). Extension of the working range by using the proposed methodology represents an attractive alternative from the analytical point of view, since it results in less specimen manipulation and consequently reduced contamination risks without compromising either the accuracy or the precision of the analyses. The implementation of artificial neural networks also helps to reduce the trial-and-error task of looking for the right mathematical model from among the many possibilities currently available in the various scientific and statistic software packages.