Multi-Parameter Dataset Enables Machine Learning Analysis of Organic Acid Dissolution of Mine Tailings for Sustainable Critical Mineral Recovery

Abstract

Mine tailings are residual materials generated after extracting valuable minerals from ores and represent the largest global waste stream. These materials contain residual metals that serve as secondary resources for modern technologies, and their recovery can reduce the environmental burden of tailings management and dependence on conventional mining. Traditional recovery methods often rely on corrosive inorganic acids that generate acidic effluents, motivating interest in organic acids as a more sustainable alternative due to their lower toxicity and biodegradability. However, optimal elemental dissolution remains difficult to predict due to complex interactions among tailings composition, acid properties, and leaching conditions. Here, we compile a 44-parameter machine learning dataset containing 114 valid experimental cases extracted from 16 published organic acid dissolution studies of mine tailings. 6 regression models (Elastic Net, Bayesian Ridge, Support Vector Regression, Random Forest Regressor, Extra Trees Regressor, Gradient Boosting Regressor) are trained on these feature matrices to identify experimental parameters that maximize target element dissolution. Comparison between Group Out-of-Fold R² values for the 6 models identified Support Vector Regressor as the top-performing model architecture with a Group Out-of-Fold R² value of 0.156768, demonstrating the model’s moderate ability to generalize and predict across unseen literature. Factor importance analysis determined dosage as the most critical primary factor driving leaching rate predictions. Nonlinear pairwise interaction strength analysis highlighted stirring and time as the strongest interacting pair, with optimal synergy occurring at 500 RPM and ~457 minutes, suggesting potential to guide experimental design. By predicting optimal organic acid molecules and leaching conditions, this machine learning framework may reduce trial-and-error experimentation and identify promising combinations for recovering valuable metals, supporting sustainable critical mineral recovery and reducing dependence on conventional extraction methods.