Interpretable Machine Learning for Identifying Key Variables Influencing Gold Recovery and Grade.

Area of Science:

• Mineral Processing
• Applied Machine Learning
• Geochemistry

Background:

• Gold flotation performance relies on complex, interacting variables.
• Existing predictive models often prioritize accuracy over interpretability, hindering practical application for process engineers.
• Limited experimental data presents a challenge for traditional modeling approaches.

Purpose of the Study:

• To apply explainable machine learning (XAI) techniques for identifying and interpreting key variables influencing gold flotation recovery and grade.
• To address the limitations of accuracy-focused models by emphasizing interpretability for process optimization.
• To demonstrate the utility of XAI in data-constrained environments for mineral processing.

Main Methods:

• Utilized a Gradient Boosting Regressor model.
• Employed SHAP (Shapley Additive Explanations), permutation importance, and feature importance analyses.
• Applied these methods to a small, experimentally derived dataset (n=11) from Ballarat gold ore flotation.

Main Results:

• Consistently identified power, head grade, and processing time as dominant predictors of gold recovery and grade.
• Uncovered significant linear and non-linear relationships between variables.
• Revealed important interaction effects, such as head grade × collector and size × head grade.

Conclusions:

• Interpretable machine learning provides actionable insights for process engineers, bridging the gap between modeling and optimization.
• Findings highlight trade-offs between energy input and flotation efficiency.
• Operational conditions for improved gold recovery and grade can be identified through transparent, domain-specific insights.