THE APPLICATION OF MACHINE LEARNING TO HHXRF-DERIVED GEOCHEMICAL DATA OF FLOOR SAMPLES FROM THE NESS OF BRODGAR, ORKNEY, SCOTLAND: ASSIGNING CONTEXTS WITHOUT THE USE OF A STANDARD

A perceived limitation of handheld X-ray fluoresce spectroscopy (hhXRF) analysis in archaeological sediment studies is the difficulty of accurately determining geochemical abundances due to the matrix effect and insufficient standards. This is a very real concern if the archaeological question being addressed requires accurate concentrations to be known. This paper explores the feasibility of using machine learning to develop statistical models using semi-quantitative hhXRF data to compare and differentiate archaeological contexts in meaningful and informative ways without the use of a standard. The hhXRF data used for this study was derived from hundreds of analyses of floor sediments from multiple monumental structures at the Neolithic site of the Ness of Brodgar in Orkney, Scotland. Assays were run on floor samples using a 50 cm x 50 cm grid across multiple structures and all phases of occupation, structure remodeling, and floor resurfacing. Data were initially collected using a Bruker Tracer III-SD and later a Bruker Tracer 5g using similar parameters to target both major and trace elements. Element counts were calculated using the deconvolution tool in Bruker’s Artex software, and the resultant values were normalized using the Compton backscatter values of Rh. Using supervised regression methods, a model was generated to identify and determine coefficient weights of those elements contributing most to group identification. The element coefficients were then used to build unsupervised clustering models. This paper reports on the success and limitations of such an approach to archaeological site investigations and provides a workflow for future hhXRF studies when the use of standards is not feasible.