Machine Learning Prediction of Deep Potential Ores and its Explanation Based on Integration of 3D Geological Model and Numerical Dynamics Simulation: An Example from Dongguashan Orefield, Tongling Copper District, China

Abstract

The complex geological architecture, complicated dynamics processes and nonlinear association in mineral systems are the major intrinsic hindrances to predictive mineral exploration. For effectively overcoming such difficulties to achieve credible prediction, 3D geological modeling, numerical dynamics simulation (NDS) and machine learning (ML) were applied to characterize the complex geological architecture, to replay the complicated dynamics processes and to predict mineralization-favor spaces by extracting nonlinear association of multi-features with mineralization in the Dongguashan orefield. The method of SHapley Additive exPlanations (SHAP) was used to explain the correlations between different features and mineralization in the predictive model. The results of the 3D geological modeling revealed that the orebodies are unevenly distributed around the intrusion and closely related to the features of the intrusion’s contact zone and wall rocks. The 3D distribution of resistivity can provide some evidence to infer underground geological architecture rather than a threshold to separate orebodies from wall rocks. The NDS results showed that dilation zones developed around the intrusion and within some beds, being closely associated with the known orebodies. By applying the most popular ML algorithm, random forest, and combining different geological, geophysical and dynamics features as evidence variables, eight ML models were run to predict potential orebodies. The predictive model performance on the test samples indicates that the integration of dynamics evidence with geological evidence significantly improves the predictive capacity of the ML model. The SHAP values demonstrate that volumetric strain is the most important feature, while the inclination of the contact zone has the greatest positive contribution to the predictions. The SHAP values of variable interactions indicate that complex intrusion contact zones and low-pressure, high-dilation areas are closely related to mineralization. The 3D ML prediction evidenced synthetically by geological, geophysical and geodynamical features demonstrates that there are substantial potential ores at depth of the northern east and southern east parts of the orefield.