Advancing environmental and food system monitoring with machine-learning-enhanced X-ray fluorescence analytics

Deficiencies of essential and excesses of toxic elements contribute to substantial, yet frequently under‑recognized, global disease burden. Energy Dispersive (ED) X-ray fluorescence (XRF) spectroscopy is widely used for rapid, multi-elemental analysis. However, different instruments and XRF-system configurations vary in performance, constrained by sensitivity, matrix effects and spectral interferences. Presented here is a novel analytics solution. Integrating across common XRF platforms, sequences of systematic model building, and validation were performed with diverse machine-learning algorithms. A total of 85 certified reference materials, including 20 soils and 65 food and biological samples, were used to provide a comparable validation framework. Matrix-specific limits of detection were calculated, and platform performance was assessed. Hybrid calibration strategies using linear, partial least squares, and random forest enhanced XRF models, were developed to correct systematic biases. Unlike approaches based on direct prediction from high-dimensional spectra, this framework uses XRF-derived concentrations as inputs and corrects residual deviations. Machine learning-assisted calibration improved measurement accuracy, particularly for handheld XRF. However, no single correction model was universally applicable. Performance varied by element, matrix, and instrument. These results show that hybrid calibration improves the reliability of multi-elemental XRF analysis and supports its application in environmental and food system monitoring.