Abstract
This study provides statistical support for X-ray Fluorescence (XRF) spectral comparisons using quantitative similarity measures. A set of electrical tapes originating from different rolls (94 rolls, 24 brands, 54 product types, four countries of manufacture) and an additional subset originating from the same source (20 samples from the same roll) are characterized via XRF. Noise in spectra is filtered using Fast Fourier Transform, and baselines are corrected using a second derivative-constrained weighted regression. Then, spectral contrast angle ratios (SCAR) are calculated for each pairwise comparison (n = 4561). The SCAR metric can capture information on the variability between the compared samples and the variability within same-source samples. Based on that measure, a threshold minimizing erroneous associations or exclusions is proposed. In addition, SCAR is used to classify samples using cluster analysis. An automated approach to sample comparison utilizing a random forest algorithm assists in identifying the basis for similarities or differences between compared spectra. This study describes a more objective approach to reporting opinions and probabilistic determinations of spectral data that can be used as a model for other fields and materials. The use of the SCAR metric can support the forensic examiner's decision-making process and add transparency in various ways.
