Applying benefits and avoiding pitfalls of 3D computational modeling-based machine learning prediction for exploration targeting: Lessons from two mines in the Tongling-Anqing district, eastern China

Abstract

The 3D computational modeling-based machine learning (ML) prediction is an innovative methodology for exploration-targeting. This paper presents our studies on its application in the Fenghuangshan (FHS) and Anqing (AQ) mines, its benefits and pitfalls and the strategies for applying benefits and avoiding pitfalls. The 3D geological models of the FHS and AQ orefields show that the topography and attitudes of the intrusions’ contact zones have spatial constraints on orebodies. The 3D variation of resistivity can provide some ambiguous evidences for inferring contact zone and orebodies. The dynamic simulations of the intrusions’ cooling processes suggest that the dilatant deformation produced by the coupled mechano-thermo-hydrological (MTH) processes is favor for mineralization. Based on the results of 3D geometric and geodynamic modeling, we conducted the random forest (RF) ML model in the FHS mine and the ML models respectively of artificial neural network (ANN), support vector machine (SVM) and RF in the AQ mine to predict mineral potentials. The RF prediction indicates that there is no significant potential in the FHS mine down to depth of 1400 m. The drills of the FHS Deep Drilling Program (FHSDDP) are all at the locations with low RF prediction probability (PP) of mineralization, which can explain why the FHSDDP had not discovered any orebodies and verifies the reliability of the RF prediction. The ANN, SVM and RF models of AQ mine gave different predictions, although the same data were imputed. The RF prediction indicates that there is no significant potential there down to −1520 m, whilst the SVM prediction shows a few small high potentials there, and ANN prediction shows prominent high potentials there. The drills of the AQ Mine Outer Drilling Program (AQMODP) that failed in ore discovery are all at the locations with low RF PP, but with high ANN PP, suggesting that the RF prediction is more reliable than that of ANN. Such inconsistent predictions of different algorithms and the unsuccessful exploration stories in FHS and AQ mines demonstrate that the 3D computational modeling-based ML technologies are a two edged sword. On the one hand, they can facilitate mineral exploration by characterizing the mineralization system in 3D visual and quantitative prediction of ore potentials. But on the other hand, they have some hidden pitfalls to impede or misguide mineral exploration, such as over extrapolation producing false model, over interpolation creating false big data, overfitting generating false knowledge and over simplification causing false prediction. The right strategies for applying benefits and avoiding pitfalls of 3D computational modeling-based ML prediction are pursuing high quality data, applying a capable ML algorithm, reinforcing field geology-based investigation and ratiocination, and keeping full feedback coupling with exploration processes to ensure correct decision-making and timely updating of data.