Predicting Gold Grades from Magnetic Susceptibility: A Machine Learning Approach

Traditional mineral exploration faces escalating costs and limited spatial coverage, creating demand for cost-effective targeting methods. This study demonstrates how machine learning can predict gold grades from rapid magnetic susceptibility measurements, using 2,277 borehole samples from the Ashanti Gold Belt, Ghana. Random Forest, Support Vector Machine, K-Nearest Neighbors, and ensemble methods are systematically compared using comprehensive validation. Exploratory analysis reveals substantial variability in gold grades (mean = 0.587 g/t, range = 0.005 to 33.900 g/t) and magnetic susceptibility (mean = 0.000274 SI). Random Forest achieved superior performance (RMSE = 0.5019 g/t, MAE = 0.2777 g/t, R 2 = 0.9362), substantially outperforming Support Vector Machine (R 2 = 0.8215), K-Nearest Neighbors (R 2 = 0.8605), and ensemble methods (R 2 = 0.9104). Feature importance analysis reveals that magnetic susceptibility transformations, particularly logarithmic transformation, account for 31.4% of predictive power, while lithological indicators contribute 16.6%. Cross-validation confirms model stability with narrow confidence intervals, and residual diagnostics demonstrate unbiased predictions across grade ranges. This research establishes a practical methodological framework for integrating machine learning with geophysical measurements, enabling cost reductions of 30-50% through selective assay targeting while maintaining exploration effectiveness. The approach is applicable across diverse mineral-rich geological settings.