Partial least squares and artificial neural networks for multicomponent analysis from derivative UV-Vis spectra
| Authors | |
|---|---|
| Year of publication | 2002 |
| Type | Article in Proceedings |
| Conference | International Chemometric Conference - CHEMOMETRICS VI |
| MU Faculty or unit | |
| Citation | |
| Field | Analytic chemistry |
| Keywords | Partial least squares (PLS);artificial neural networks(ANN);multicomponent analysis;derivative UV-Vis spectra;adenine;cytosine |
| Description | This contribution presents a comparative study of the use of PLS and ANNs to analyze A and C mixtures using UV-Vis derivative spectra. The optimum ANN architecture enabling to model the system was established by means of TRAJAN 6.0 program. Several algorithms (Back propagation, Conjugate gradients, Quick propagation, and Delta-Bar Delta algorithm) were used for the training of the ANN to obtain a reliable model. With help of a suitable experimental design in combination with soft ANN modelling, the concentration of both A and C in mixtures can be quantified with an excellent accuracy (about 1 %). The quality of the testing set was evaluated on the basis of the average root mean square error for prediction (RMSEP) calculated from true and found values of A and C concentrations (RMSEP = 0.07 for A and 0.09 for C). It was found that ANN gives better results for the first and second derivative spectra than for original spectra. Furthermore, in comparison with PLS the ANN provides a more reliable and precise approach in the multicomponent analysis of A and C mixtures, where a number of different interactions take place. |
| Related projects: |
|