Empirical portable x-ray fluorescence quantification limits and accuracies for 28 elements within 1942 diverse geological standards using open-source machine learning

Portable X-Ray Fluorescence (pXRF) instruments are commonly used to quantify elements within materials due to their portability and quick results using a fundamental parameters approach and a data library. Results can also be customized by assembling alternative data libraries instead of the included calibrations that are typically provided with such systems. In contrast, this paper presents an alternate approach to the standard fundamental parameters (FP) protocol. We instead empirically quantify elemental abundances directly from pXRF spectra using multivariate (MVA) methods trained on 1942 geochemical standards. Prediction accuracies (RMSE-P) and limits of quantification (LOQ) are presented for six major (Al₂O₃, CaO, Fe₂O₃, MgO, SiO₂, TiO₂), two minor (MnO, P₂O₅), and 20 trace elements (As, Bi, Cr, Cu, Mo, Nb, Ni, Pb, Rb, SO₃, Sn, Sr, Ta, Th, U, V, W, Y, Zn, Zr). Partial least squares (PLS) regression was chosen as the optimal MVA method from five other linear methods (ElasticNet, Ridge, OMP, PCR, LASSO). The resultant MVA models have accuracies comparable to instrument-based FP calibrations, and are more sensitive to low concentrations. This work demonstrates a powerful and adaptable open-source method for quantifying elements using XRF spectra that should be considered as a viable alternative to traditional FP approaches. The large and diverse geochemical dataset used here is made publicly available to encourage further study and combination of datasets with calibration transfer.