Quantitative analysis of heavy metals in ores by X-ray fluorescence with improved CARS and support vector regression using ACOR optimization algorithm

X-ray fluorescence (XRF) spectrometry is widely applied for the quantitative determination of elemental concentrations in ores due to its rapid, non-destructive, and multi-element analytical capabilities. However, the high dimensionality, spectral redundancy, and strong matrix-dependent nonlinear effects inherent in XRF spectra often limit the predictive accuracy and robustness of conventional quantitative models. Therefore, the effective selection of informative spectral variables and the construction of reliable regression models remain critical challenges for accurate ore analysis. In this study, an integrated quantitative framework combining adaptive heuristic optimization and machine-learning regression is proposed to improve elemental concentration prediction from ore XRF spectra. First, a competitive adaptive reweighted sampling (CARS) strategy optimized by the ant colony optimization for continuous domains (ACOR) algorithm is employed to perform robust spectral variable selection. This process removes redundant noisy spectral information while preserving chemically relevant features. Subsequently, an ACOR-optimized support vector regression (ACOR-SVR) model is developed for quantitative analysis. In this model, key SVR hyperparameters are automatically tuned to enhance nonlinear modelling capability and improve generalization performance. The proposed ACSVR framework is evaluated on multi-element ore datasets involving Fe, Cu, Zn, and Pb. The experimental results demonstrate that ACOR-CARS provides more stable and compact feature subsets than conventional CARS, whereas ACOR-SVR consistently outperforms baseline SVR and other classical machine-learning models. High prediction accuracy is achieved for Fe, Cu, Zn, and Pb, with coefficients of determination (R²) reaching 0.9940, 0.9984, 0.9968, and 0.9995, respectively. These results confirm the effectiveness of integrating optimized variable selection with nonlinear regression modelling for XRF-based quantitative analyses. Overall, this study provides a practical and efficient strategy for high-precision elemental quantification in ores, offering strong potential for mineral exploration, resource evaluation and intelligent geochemical analysis.