Residual geochemical gold grade prediction using extreme gradient boosting

Abstract

The increasing cost of exploration and decreasing recovery in mineral exploration necessitates the use of reliable and robust models for spatial gold grade prediction. In comparison, ordinary kriging (OK) is widely applied in geospatial data analysis for target generation however, its assumption of spatial variability renders it unsatisfactorily for problems, where geological processes such as faulting have altered the ore deposits. In this study, the accuracy and robustness of the Extreme Gradient Boosting (XGBoost) model for regolith geochemical grade prediction was investigated within the Tarkwaian paleo-placer formation of Ghana, West Africa. The geochemical spatial data (total=891) was partitioned into training and testing for model development and evaluation, respectively. The optimal parameters of the model were selected by grid search with 10-fold cross-validation. The predictive performance of the model was evaluated and compared with Random Forest (RF), Generalised regression neural network (GRNN) and OK. The results show that the XGBoost model used in this study produced the lowest error score (MSE = 0.7107, MAPE = 9.3948, RMSE = 0.8430) and the highest predictive efficiency (R2 = 0.9189). In terms of accuracy and generalization, this investigation has proven that XGBoost is a very efficient and robust model for geochemical gold grade prediction. Importantly, all the machine learning models used in this study proved superiority over the OK.

Keywords

• Extreme gradient boosting: Machine learning
• Ordinary kriging
• Gold grade prediction
• Cross-validation