Machine learning and multi-surface modeling improve handheld XRF for rapid cadmium quantification in soils

Cadmium poses severe threats to ecosystems and human health due to its high toxicity and bioaccumulative nature. The global regulatory limits for Cd are often low at 0.1-0.6 mg kg−1, which is 1 to 2 orders of magnitude lower than other heavy metals, making rapid and reliable trace-level detection challenging. To overcome this limitation, this study developed an integrated framework based on handheld energy-dispersive X-ray fluorescence (HXRF) for efficient trace-level detection and fraction analysis. System sensitivity was first elevated through hardware optimization, including optimization of the X-ray source and selection of an optimized filter set (0.2 mm Cu + 0.2 mm Mo). The refined spectral data were then used to train machine learning models (ML-HXRF) to predict total Cd concentration and key soil properties (pH, SOC, and clay content). On the test set, the models achieved mean absolute errors of 0.66 mg kg−1 for total Cd, 0.20 for pH, 2.93 g kg−1 for SOC, and 0.9% for clay, respectively. Furthermore, ML-augmented multi-surface models (ML-MSMs) were applied to predict soil Cd fraction and available fractions. A subsequent sensitivity analysis demonstrated that the propagated errors from the predicted soil properties resulted in less than 1.0% deviation in the final dissolved Cd estimates, confirming the framework's robustness. Interpretability analysis revealed statistical associations consistent with expected geochemical mechanisms underlying these predictions, elucidating elemental controls on processes including pH-dependent solubility, soil organic matter complexation, and surface binding. The resulting ML-HXRF-MSMs framework was validated in a mining-affected paddy field, revealing a clear spatial dissociation between total and dissolved Cd, governed primarily by site-specific pH variation. This proof-of-concept study demonstrates the framework's potential to support field monitoring and inform remediation priority across broader geographical and soil contexts.