High-resolution compound-specific mapping in works of art via data fusion of MA-XRPD with hyperspectral data (part 2:...

This study is the second part of an investigation into enhancing high-resolution compound-specific mapping for artworks using multi-modal chemical imaging and machine learning. It builds upon the previously introduced framework, part one: method evaluation, by the authors that integrates macroscopic X-ray powder diffraction (MA-XRPD) macroscopic X-ray fluorescence (MA-XRF) or reflectance imaging spectroscopy (RIS). While MA-XRPD is time-consuming, the unique advantage of the technique lies in its ability to visualize the distributions of crystalline phases (e.g., pigments, degradation products) at the surface of artworks and estimate in a quantitative manner their relative concentrations. By leveraging data fusion and predictive modeling, this study aims to enhance the spatial resolution and efficiency of compound-specific mapping of artworks.

This paper applies the previously established machine learning approach to three case studies, demonstrating its feasibility in analyzing heterogenous, complex artworks. The first case study validates the method on a 17th-century Dutch still life painting, comparing the predicted high-resolution compound-specific maps against experimental data. The second case study examines the potential to reduce MA-XRPD scanning time through lower resolution scans and predictive modeling while maintaining accuracy, which is assessed using root mean square error (RMSE) and structural similarity index measure (SSIM). The third case study explores model transferability, evaluating whether predictive models trained on one artwork can accurately predict compound distributions in another. Results confirm the efficacy of the machine learning approach in improving spatial resolution and reducing data acquisition constraints.

This study addresses the time-intensive nature of MA-XRPD by demonstrating a scalable approach for generating high-resolution, compound-specific maps through data fusion and predictive modeling. By integrating the high specificity of MA-XRPD with the spatial efficiency of faster hyperspectral imaging techniques (MA-XRF and RIS), and applying a random forest regression framework, the method significantly enhances the efficiency and interpretability of chemical imaging in cultural heritage research, enabling more accessible and large-scale non-invasive analysis of complex artworks.