An integrated machine learning framework with uncertainty quantification for three-dimensional lithological modeling from multi-source geophysical data and drilling data

Abstract

Nowadays, it is commonplace for geological surveys to integrate multi-source geophysical data and drilling data in order to construct three-dimensional (3D) lithological models. In this context, manual translation of complex geophysical data into parameters used for 3D lithological modeling is challenging. Machine learning has recently shown great potential in 3D lithological modeling. However, the performance of machine learning algorithm is influenced by the imbalance in number of categories of lithological samples. In addition, the uncertainty associated with 3D lithological modeling by machine learning has rarely been quantified. This study presents a novel integrated machine learning framework to address the imbalance issue and to quantify uncertainty in 3D lithological modeling. As its novelty, our integrated machine learning framework can subdivide total uncertainty into aleatoric and epistemic uncertainties in the 3D lithological modeling procedure by stochastic gradient Langevin boosting. Another innovation of this study is the use of Bayesian hyperparameter optimization for automatic tuning of hyperparameters of the integrated machine learning framework. The 3D lithological and uncertainty modeling case study in the Jiaojia–Sanshandao gold district of China demonstrated the superiority of our proposed integrated machine learning framework. The proposed framework has great potential in integrating multi-source geophysical and drilling data for 3D lithological and uncertainty modeling in engineering geology . • Construct 3D lithological and uncertainty modeling using machine learning from geophysical and drilling data. • Total uncertainty was subdivided into aleatoric and epistemic uncertainties in the 3D lithological modeling. • SMOTE-ENN was utilized to address the lithological imbalance issue in training for 3D lithological modeling. • The Bayesian hyperparameter optimization was applied to tune the hyperparameters of machine learning model automatically.