113-6 Machine Learning Approaches to Estimating Lithium in Spodumene from Color Metrics, Hardness, and Portable XRF Trace Geochemistry

Session: Mineralogical Characterization of Economic Resources: From Critical Minerals to Gemstones (Posters)

Presenting Author:

Authors:

Abstract:

Lithium (Li) is becoming increasingly important in recent years because of its use in green energy technologies. One of the most common Li ore minerals is spodumene, found in lithium-cesium-tantalum (LCT) pegmatites. As spodumene undergoes hydrothermal alteration, its color changes, its hardness and Li content decrease, foreign elements are introduced, and its atomic structure is destroyed while external crystal shape is maintained. Because of the increasing demand for Li, new methods are needed to quickly and reliably evaluate the Li content and ore quality of spodumene. In this study, we analyze 173 distinct spots from 36 spodumene samples (one to eight spots each) from 12 LCT pegmatite localities across the US and Canada. By correlating the alteration of spodumene with its color and Li content, we aim to develop a novel method of estimating Li ore quality. 

Sample photographs were taken using a smartphone camera and L*, a*, and b* color values were quantified in Adobe Photoshop. Hardness ranged between 1 and 7 on the Mohs scale. Trace-element geochemistry was measured using a portable X-ray fluorescence (pXRF) analyzer. To predict Li contents using these data, Random Forest, Cubist, and Support Vector Regression machine learning models were constructed in RStudio, with samples randomly separated into training (70%) and validation (30%) data. Each model was run on datasets comprising color and hardness data, pXRF data, and all data together. Predicted Li values were validated with portable laser-induced breakdown spectroscopy (LIBS) measurements using a custom calibration. Four models achieved an R² value above 0.8: Random Forest based on all data (R²=0.845, RMSE=5186 ppm), Cubist based on all data (R²=0.843, RMSE=5227 ppm), Random Forest based on color and hardness data (R²=0.814, RMSE=5693 ppm) and Random Forest based on pXRF data (R²=0.806, RMSE=5809 ppm). Scanning electron microscopy – energy dispersive spectroscopy was used on polished thin-sections of selected samples to identify the alteration minerals and/or primary inclusions altering macroscopic spodumene color. For example, nearly black coloration appears to be related to abundant As and Fe mineral inclusions such as löllingite (FeAs₂). These promising lithium predictions integrating rapid physical and geochemical readouts of spodumene may further contribute to automation protocols to assess ore quality during exploration or ore processing.

Geological Society of America Abstracts with Program. Vol. 57, No. 6, 2025

doi: 10.1130/abs/2025AM-6982

© Copyright 2025 The Geological Society of America (GSA), all rights reserved.