Evaluating Lithium Concentration Predictions in the U.S. Gulf Coast: Machine Learning Approaches and Outlier Impact Analysis

Lithium-ion battery technology is a key component of energy storage and the demand for lithium (Li) will continue to increase in the next decades. Dissolved Li in subsurface brines represents a growing segment of the overall Li resource, and accurate estimates of Li concentrations in geological formations is crucial for its efficient management. This study employs advanced machine learning models to predict Li concentration in the Gulf Coast Basin using a comprehensive geochemical data set developed by the USGS. Six machine learning models-Gradient Boosting, Random Forest, K-Nearest Neighbor, Support Vector Regression, Neural Networks, and Extreme Gradient Boosting-were implemented and their accuracy compared and contrasted. In order to improve prediction accuracy, we also introduced a Multi-Target Sequential Chaining approach that imputes missing data from the data set sequentially. We tested three different data scenarios: (1) using the original data set (with outliers), (2) using the data set with outliers removed, and (3) introducing altered outliers to analyze model sensitivity. Results indicate that outliers impact model performance, with altered outliers helping assess robustness. The results of this work suggest spatial extensions of already known Li-rich areas (e.g., Smackover Formation) and confirm the low Li potential of most of the Gulf Coast.