Determination of concentrations at hydrolytic potentiometric titrations with models made by artificial neural networks

Abstract

The determination of concentrations of sulphate in different samples of river and drinking waters and of concentrations of calcium in different wine samples using Kohonen and counterpropagation artificial neural network (ANN) is described. Kohonen ANN has been used to define the training and the test sets. All the samples are represented as sets of points (pH values) of titration curves. For the process of learning of counterpropagation ANN, the concentration of each sample is needed besides the pH values of its titration curve. Altogether 31 experimental titration curves obtained by the hydrolytic potentiometric titrations of sulphate in different water samples at different sulphate concentrations and 26 titration curves of different calcium concentrations in wine samples were chosen for building the two models. The models were validated by the objects from the test set and by leave-one-out procedure. The same procedure (leave-one-out) was also employed for the study of effect of training time on the prediction ability of the network. Predictions from the models were additionally tested by the experimental titration curves recorded for this purpose. The 6×6× (30+1) ANN structure was optimal for the model built for water samples, and 6×6× (36+1) for the model built for wine samples. The cross-validated squared correlation coefficient was 0.884 in the case of water samples and 0.846 in the case of wine samples. The corresponding standard errors of prediction (SDEP) were ±2.5 and ±9.5 mg/l in the case of water and wine samples, respectively. The results indicate that ANN can successfully predict the concentration of compounds from the titration curves within 10% of error which is good enough for fast screening of waters and wines.