Development of a neural network approach to characterise (226)Ra contamination at legacy sites using gamma-ray spectra taken from boreholes.

There are a large number of sites across the UK and the rest of the world that are known to be contaminated with (226)Ra owing to historical industrial and military activities. At some sites, where there is a realistic risk of contact with the general public there is a demand for proficient risk assessments to be undertaken. One of the governing factors that influence such assessments is the geometric nature of contamination particularly if hazardous high activity point sources are present. Often this type of radioactive particle is encountered at depths beyond the capabilities of surface gamma-ray techniques and so intrusive borehole methods provide a more suitable approach. However, reliable spectral processing methods to investigate the properties of the waste for this type of measurement have yet to be developed since a number of issues must first be confronted including: representative calibration spectra, variations in background activity and counting uncertainty. Here a novel method is proposed to tackle this issue based upon the interrogation of characteristic Monte Carlo calibration spectra using a combination of Principal Component Analysis and Artificial Neural Networks. The technique demonstrated that it could reliably distinguish spectra that contained contributions from point sources from those of background or dissociated contamination (homogenously distributed). The potential of the method was demonstrated by interpretation of borehole spectra collected at the Dalgety Bay headland, Fife, Scotland. Predictions concurred with intrusive surveys despite the realisation of relatively large uncertainties on activity and depth estimates. To reduce this uncertainty, a larger background sample and better spatial coverage of cores were required, alongside a higher volume better resolution detector.