Soil Geochemistry Toward Lithium Pegmatite Exploration: Building a Machine-Learning Predictive Algorithm via Portable X-Ray Fluorescence
Uložené v:
| Názov: | Soil Geochemistry Toward Lithium Pegmatite Exploration: Building a Machine-Learning Predictive Algorithm via Portable X-Ray Fluorescence |
|---|---|
| Autori: | Luiza Maria Pereira Pierangeli, Mona-Liza C. Sirbescu, Sérgio Henrique Godinho Silva, David C. Weindorf, Thomas R. Benson, Nilton Curi |
| Zdroj: | Economic Geology. 120:1311-1330 |
| Informácie o vydavateľovi: | Society of Economic Geologists, Inc., 2025. |
| Rok vydania: | 2025 |
| Popis: | As demand for lithium (Li) increases, cheaper, more sustainable, and faster methods are needed for the identification and characterization of new Li deposits. Lithium-rich pegmatites are major sources of Li, but their exploration is often hindered by soil cover. Portable X-ray fluorescence (pXRF) can rapidly and accurately quantify soil chemistry to determine the bedrock economic potential, but unfortunately, Li is undetectable via pXRF. Herein, pXRF data and random forest models were used to predict both Li contents in soil samples and Li-rich soil parent material based on abundances of 15 predictors (K, Rb, Al, Ba, Ca, etc.). For comparison, support vector regression and neural network deep learning were also conducted. The data set consisted of 112 soil samples collected over spodumene-rich pegmatites, barren granitic pegmatites, peraluminous granite, and metamorphic host rocks from forested, glaciated northern Wisconsin and Michigan, United States. Lithium abundances were independently measured using inductively coupled plasma-optical emission spectroscopy (ICP-OES). The best Li prediction was achieved using neural networks, yielding a coefficient of determination (R2) of 0.90, a root mean square error (RMSE) of ~40 mg × kg–1, and residual prediction deviation of 3.2. The best parent material prediction model was achieved using random forest, with an overall accuracy of 0.88. Portable XRF analysis discriminates among soil samples formed on bedrock with distinct mineralogy. Using pXRF combined with appropriate machine learning models to predict the Li contents in the soil and the type of underlying bedrock could become an alternative, more efficient, and less invasive exploration method compared to traditional trenching. |
| Druh dokumentu: | Article |
| Jazyk: | English |
| ISSN: | 1554-0774 0361-0128 |
| DOI: | 10.5382/econgeo.5166 |
| Rights: | CC BY |
| Prístupové číslo: | edsair.doi...........d7157de36260d6f3c0e5fc7c1fa0c4b3 |
| Databáza: | OpenAIRE |
Nájsť tento článok vo Web of Science | Abstrakt: | As demand for lithium (Li) increases, cheaper, more sustainable, and faster methods are needed for the identification and characterization of new Li deposits. Lithium-rich pegmatites are major sources of Li, but their exploration is often hindered by soil cover. Portable X-ray fluorescence (pXRF) can rapidly and accurately quantify soil chemistry to determine the bedrock economic potential, but unfortunately, Li is undetectable via pXRF. Herein, pXRF data and random forest models were used to predict both Li contents in soil samples and Li-rich soil parent material based on abundances of 15 predictors (K, Rb, Al, Ba, Ca, etc.). For comparison, support vector regression and neural network deep learning were also conducted. The data set consisted of 112 soil samples collected over spodumene-rich pegmatites, barren granitic pegmatites, peraluminous granite, and metamorphic host rocks from forested, glaciated northern Wisconsin and Michigan, United States. Lithium abundances were independently measured using inductively coupled plasma-optical emission spectroscopy (ICP-OES). The best Li prediction was achieved using neural networks, yielding a coefficient of determination (R2) of 0.90, a root mean square error (RMSE) of ~40 mg × kg–1, and residual prediction deviation of 3.2. The best parent material prediction model was achieved using random forest, with an overall accuracy of 0.88. Portable XRF analysis discriminates among soil samples formed on bedrock with distinct mineralogy. Using pXRF combined with appropriate machine learning models to predict the Li contents in the soil and the type of underlying bedrock could become an alternative, more efficient, and less invasive exploration method compared to traditional trenching. |
|---|---|
| ISSN: | 15540774 03610128 |
| DOI: | 10.5382/econgeo.5166 |