[Zaf16] Statistical learning approaches applied to the calculation of scaling factors for radioactive waste characterization

Atelier, Poster ou Démonstration dans une Conférence Nationale : IRPA 2016, January 2016, pp.xx, Capetown, South Africa,

Auteurs: B. Zaffora

Mots clés: scaling factor; radioactive waste; decision tree; random forest; bagging; boosting; linear models

Résumé: In the frame of radiological characterization of radioactive waste, scaling factors are typically used to quantify the activity of difficult-to-measure radionuclides from the direct measurement of dominant gamma emitters. The application of this characterization method relies on statistical considerations, such as the probability distribution of the scaling factors themselves, along with a deep understanding of the activation mechanisms in place. The present study assesses the use of statistical learning techniques for the prediction of scaling factors, for the characterization of very-low-level radioactive waste produced at CERN (European Organization for Nuclear Research, Geneva, CH). In particular, decision trees are used to identify sorting criteria of future radioactive waste based on the minimization of the Residual Sum of Squared errors coupled with binary splitting. Other methods (e.g. pruning, bagging, random forest, boosting and linear models) are also used for the optimization of the calculation of the scaling factors together with the prediction intervals. The estimation of scaling factors for the present study is based on over 1 million analytical calculations. These calculations cover the most typical activation scenarios and chemical composition of metals used at CERN. The calculations were performed with the code ActiWiz, which was developed at CERN based on FLUKA Monte Carlo simulations of reference particle spectra. The analytical predictions were benchmarked with extensive radiochemical and gamma-spectrometry measurements of samples taken from activated metals. The results demonstrate the applicability of statistical learning techniques on waste characterization, especially for the prediction of the activity of difficult-to-measure radionuclides. Average scaling factors and prediction intervals are systematically calculated and can be easily applied on routine characterization processes at CERN. A first analysis of the errors introduced by the use of scaling factors is also presented.


@inproceedings {
title="{Statistical learning approaches applied to the calculation of scaling factors for radioactive waste characterization}",
author=" B. Zaffora ",
booktitle="{IRPA 2016}",
address="Capetown, South Africa",