A stacking methodology of machine learning for 3D geological modeling with geological-geophysical datasets, Laochang Sn camp, Gejiu (China)

Abstract

A stacking method was followed for the classification and prediction of rock types and for 3D geological modeling in the Laochang Sn camp, Geijiu (China). The sequence of the methodology is as follows. Initially, multi-source datasets were compiled GOCAD software, including a borehole dataset with rock types and 1:50,000 scale gravity (residual density) and magnetic susceptibility inversion interpretation using 3D-grid model. Then, the datasets were subjected to first-level classifiers using support vector machine, gradient boosting decision tree, neural network, random forest, and XGBoost. This was followed by subjecting the datasets to second-level classifier using XGBoost. Next, features (XYZ coordinates, residual density and magnetic susceptibility) were input for training and testing (70% of data for training, 30% for testing). Then, upon completion of training the model, the machine learning metrics were used to evaluate comprehensively the performance of the model, and the 3D geological model was visualized. Finally, the 3D geological modeling was validated by comparison with a robust geological dataset with mining channel dataset. The research results show the stacking method of machine learning can be used to construct effectively and quickly a 3D multi-type rocks model using geological and geophysical datasets. The obtained areas under the ROC curves were 0.93 for the granites, 0.98 for the basalts, and 0.94 for "other" rocks. The methodology can be extended to multi-source datasets of geoscience big data and even to different metallogenic models or geological settings.