Non-negative matrix factorization and principal component analysis to investigate similarities in macro X-ray fluorescence scanning

This study presents a methodology for investigating recurring elemental and spectral patterns in macro X-ray fluorescence (MA-XRF) datasets acquired from paintings by the same artist. The approach combines Non-negative Matrix Factorization (NMF), applied to the MA-XRF data cube, with Principal Component Analysis (PCA), used to identify similarities and differences among the NMF bases. The methodology was applied to MA-XRF datasets collected from 15 paintings by the Brazilian artist Djanira da Motta e Silva, including 14 authenticated works spanning the 1940s to the 1970s, and one painting attributed to the 1960s. The elemental maps exhibited marked visual correspondences, including the presence of zinc (Zn) across large areas in all compositions. The NMF bases showed a predominance of peaks associated with Ca, Ti, Ba, Zn, and Fe, suggesting the recurrent use of specific pigments and fillers across the investigated paintings. PCA showed broad overlap between the NMF bases of the attributed painting and those of the authenticated works, whereas the point estimates of the centroid distances suggested that the attributed painting was most consistent with works produced between 1940 and 1960. However, the large propagated uncertainties did not allow for a reliable chronological discrimination among the decade-based groups.. Infrared reflectography (IRR) analyses of the authenticated works indicated that Djanira frequently executed underdrawings and pentimenti in her paintings, a pattern also observed in the IRR image of the attributed work. The results demonstrate the potential of the proposed methodology as a complementary tool for the technical characterization of paintings, based on MA-XRF datasets in addition to the visual comparison of elemental maps.