Evaluation of equilibria with use of artificial neural networks (ANN). II.ANN and experimental design as a tool in electrochemical data evaluation for fully dynamic (labile) metal complexes

A use of artificial neural networks (ANN) and various experimental designs (ED) for refinement of experimental data obtained in a polarographic metal- ligand equilibrium study of fully dynamic (labile) metal complexes was thor oughly examined. ANN were tested on evenly and randomly distributed experim ental error-free and error-corrupted data. It was found that randomly distr ibuted experimental data did not influence the prediction power of ANN. Num erous tests demonstrated that ANN with appropriate ED can provide accurate pre diction in the stability constants with the absolute errors in the rang e of +/- 0.05 log unit or smaller. ANNs were found exceptionally robust. Ra ndom experimental errors have not influenced estimates in stability constan ts much even when errors in pH up to the value of +/- 0.1 pH unit were intr oduced. A special procedure has been worked out that allows to minimize the influence of error-corrupted data even further; no significant difference was observed between results obtained on error-free and error-corrupted dat a. This procedure makes it also possible to obtain a standard deviation in the calculated stability constants that is usually a difficult task when AN Ns are used. The results obtained from ANN were compared with those obtaine d from a hard model based nonlinear regression techniques. No significant d ifference in evaluated data from these two, soft and hard model based appro aches, was found. The use of ANN described here for polarographic data is o f general nature and, in principal, can be adopted to other analytical tech niques commonly used in metal-ligand equilibrium studies.