Estimation of uranium concentration in ore samples with machine learning methods on HPGe gamma-ray spectra

Within the scope of determining concentration uranium in ore samples by gamma-ray spectrometry, we tested a series machine-learning (ML) algorithms with database including 1288 HPGe gamma spectra measured Orano Mining. Instead detecting and identifying peaks, global interpretation is carried out. Two different approaches were used. First, reduced size dimension dataset selecting 728 acquired same experimental setup resampling their 8192 channels into 168 energy bands according to important peaks due natural uranium, thorium potassium activity. Classical ML have been trained on this best predictions show Symmetric Mean Absolute Percentage Error lower than 6%. In second step, complete six measurement setups was used train deep neural network re-sampling 2048 channels. Despite small dataset, Convolutional Neural Network (CNN) model provides satisfactory results mean errors 15% broader more complex terms concentrations setups. These outcomes shows that methods can predict similar uncertainties as classical spectroscopy (10% 20%), but without requiring an expert knowledge interpret spectra.