226Ra activity concentration determined directly from the 186 keV photopeak using gamma spectrometry and a neural network

Abstract

Determining 226Ra activity concentration directly from the 186 keV photopeak using gamma spectrometry avoids the 21-day waiting time required for indirect calculation from 214Pb or 214Bi. Calculation methods using the 186 keV photopeak, suppressing the 235U interference, are satisfactory but do not allow the analyst to correct for possible biases due to the calculation. This paper therefore examines the use of an alternative method based on a neural network that allows enrichment and monitoring of the solution to the problem, taking into account the concentrations of reference activities or those obtained from 214Pb in the samples analysed. The 226Ra counts at the 186 keV peak were determined by fitting two Gaussian functions whose figure of merit was minimised using the Levenberg–Marquardt algorithm. The results obtained were validated and fed into the neural network, using 70% for training and 30% for testing. Network learning was estimated using samples of comparative exercises, which were analysed and then used to train the network. The possibility of increasing the number of values used to train the network was also investigated using interpolation of linear and quadratic splines. The accuracy and precision of the 226Ra activity concentrations obtained by the neural network, as determined by Student's pair-wise t-test and Fisher's F-test, in samples from different intercomparison exercises were statistically comparable (p > 0.05) at 60% and 73%, respectively. Finally, the percentage of acceptability of the 226Ra activity concentrations with respect to the usual criteria of intercomparison exercises (En ≤ ±1 and D% ≤ ±20%) was 55%, which is comparable to the 60% obtained by the numerical algorithm developed in a previous study.