Application of neural networks to quantitative spectrometry analysis

Accurate quantitative analysis of complex spectra (fission and activation products), relies upon experts' knowledge. In some cases several hours, even days of tedious calculations are needed. ?his is because current software is unable to solve deconvolution problems when several rays overlap. We have shown that such analysis can be correctly handled by a neural network, and the procedure can be automated with minimum laboratory measurements for networks training, as long as all the elements of the analysed solution figure in the training set and provided that adequate scaling of input data is performed. Once the network has been trained, analysis is carried out in a few seconds. On submitting to a test between several well-known laboratories, where unknown quantities of (CO)-C-57. Co-58 Sr-85, Y-88, I-131, Ce-139, Ce-141 present in a sample had to be determined, the results yielded by our network classed it amongst the best. The method is described, including experimental device and measures, training set designing, relevant input parameters definition, input data scaling and networks training. Main results are presented together with a statistical model allowing networks error prediction. (C) 1999 Elsevier Science B.V. All rights reserved.