The dating of French gilt bronzes with ED-XRF analysis and machine learning
Arlen Heginbotham, Robert G. Erdmann, Lee‐Ann C. Hayek · Journal of the American Institute for Conservation · 2018
This study presents a method for estimating the date of production of Parisian cast gilt bronze artifacts from the late seventeenth to the early twenty-first centuries based on their elemental composition as measured by energy dispersive X-ray florescence spectroscopy (ED-XRF). Gilt bronzes are notoriously difficult to date; however, during this time period in France, both metallurgy and trade in metals were in a state of continual transformation. These technological and commercial transformations drove changes in the elemental composition of gilt bronzes produced by Parisian craftsmen over time. These changes in turn allow estimation of the manufacture date of gilt bronzes of unknown provenance based on their elemental composition. The composition of undated French gilt bronzes can be compared to a database of compositions derived from well-dated examples. In this paper, a database of 466 such observations collected using ED-XRF is presented as a basis for comparison. The use of an ED-XRF calibration protocol designed to deliver accurate and reproducible quantitative results (CHARMed PyMca) allows for data collected on multiple instruments to be aggregated in the database and for data collected by similarly calibrated instruments to be evaluated. A method for estimating the date of manufacture for unknown samples is presented using two machine learning techniques, support vector regression (SVR) and random forest regression (RFR). By using a simple combination of both models, an estimate of the time of manufacture may be made within ±37 years with a confidence of approximately 90%. This level of precision can be useful for distinguishing eighteenth century castings from nineteenth century copies as well as from castings made in the twentieth century.