The classification and discrimination of glass fragments using non destructive energy dispersive X-ray microfluorescence

Frequency of analytical characteristics is best estimated on glass recovered at random. However, as such data were not available to us, we decided to use control windows for this estimation. In order to use such a database, one has to establish that the recovered fragment comes from a window. Therefore, elemental analysis was used both for classification and discrimination of glass fragments. Several articles have been published on the subject, but most methods alter the glass sample. The use of non destructive energy dispersive X-ray microfluorescence (microXRF) for the analysis of small glass fragments has been evaluated in this context. The refractive index (RI) has also been measured in order to evaluate the complementarity of techniques. Classification of fragments has been achieved using Fisher's linear discriminant analysis (LDA) and neural networks (NN). Discrimination was based on Hotelling's T2 test. Only pairs that were not differentiated by RI followed by the Welch test were studied. The results show that neural network and linear discriminant analysis using qualitative and semi-quantitative data establishes a classification of glass specimens with a high degree of reliability. For discrimination, 119 windows collected from crime scene were compared: using RI it was possible to distinguish 6892 pairs. Out of 129 remaining pairs, 112 were distinguished by microXRF.