COMBINING PORTABLE X-RAY FLUORESCENCE AND MACHINE LEARNING: ADVANCEMENTS FOR SOIL AND ROCK EXPLORATION
Soil geochemistry; Parent material; Pegmatites; Green-energy; sustainability; proximal sensors.
Soil parent material (PM) emerges as a critical factor in understanding soil variability, with challenges in detailed assessments due to the complexity and inaccessibility of PM in deep soils. Proximal sensor approaches, including portable X-ray fluorescence (pXRF) and magnetic susceptibility (MS), offer practical solutions for predicting soil PM. The global quest for sustainable energy alternatives has surged in tandem, with lithium (Li) emerging as a pivotal element integral to rechargeable Li-ion batteries. The demand for Li necessitates the development of cost-effective and expeditious exploration methods to enhance the identification of new deposits, with lithium-cesium-tantalum (LCT) pegmatites standing as primary sources of Li. This dissertation is divided into four chapters: (I) a pilot study focuses on creating spatial PM predictive models for three distinct rock types (charnockite, mudstone, and alluvial sediments) at the Palmital Experimental Farm (Brazil), using Random Forest (RF) algorithm combined with pXRF and MS data from A and B horizons; (II) the use of microwaved-assisted (MW) total digestion for soil and rocks; (III) evaluation of the efficacy of pXRF data and RF algorithm in predicting Li content in soil samples, utilizing Li pathfinder elements and Li-rich soil PM prediction; (IV) comparing pXRF vs. total digestion Inductively Coupled Plasma Optical Emission Spectroscopy (ICP-OES) for soils and rocks analysis. The findings in this dissertation propose alternative, cost-effective methods for assessing soil PM spatial variability. Paper (I) had a strong validation for PM prediction results (Kappa coefficient = 0.85 and overall accuracy = 0.93). Meanwhile, paper (III) 's PM prediction model achieved reasonable results with a Kappa coefficient of 0.77 and an overall accuracy of 0.85. The Li prediction model tested in paper (III) achieved a coefficient of determination (R2) of 0.86, root mean square error (RMSE) of 68.5 mgkg-1, and residual prediction deviation (RPD) of 1.78. Paper (II) tested different wet digestion recipes to optimize the decomposition of different matrices. The best recipe was a mixture of HNO3, HF, and H2O2, which presented better results for the digestion of geochemical samples. Paper (IV) assessed the performance and comparability of two pXRF systems, Alpha and Beta. Both pXRFs achieved similar results compared to reported concentrations of certified reference materials between systems and methods, showing tendencies of overestimating or underestimating elements. The findings in this dissertation are intended to facilitate informed agronomic and environmental decision-making. They could also represent an alternative, pXRF-based method for more sustainable prospecting methods for PM and Li content determination and exploration. Also, helping with complex and careful decisions in choosing total digestion methods and clarifies some doubts regarding the use and comparability of results obtained from the same sample with different pXRF equipment.