Artificial neural network analysis of PIXE of historical silver coins

A large set of historical silver coins from the 15th to the 17th centuries has been previously analysed with PIXE and complementary techniques, establishing an extensive data base of coin composition. The impurity content of the coins was then used to differentiate the origins of the silver across various periods and production contexts. In this work, we used artificial neural networks (ANN) to analyse the PIXE data, using the raw energy spectra as input, without any pre-processing besides compression from 2048 to 512 channels. A significant challenge in training ANNs is the requirement for a sufficiently large dataset, typically tens of thousands of examples. To overcome this, we generated synthetic spectra constructed through non-linear combinations of a reduced set of experimental spectra, an approach we have previously applied successfully in the analysis of gamma ray spectra for nuclear security applications. This paper presents the results obtained, which show that the ANN developed is fit for purpose.