Thanks to the brilliant progress in machine learning, many research works have conducted data-driven mineral prospectivity mapping. However, it is challenging to integrate highly multidisciplinary geoscientific data with machine learning algorithms. Especially, geological data are heterogeneous and non-numerical even though they are crucial for mineral exploration. In this work, we introduce how to preprocess the geoscientific data and design a machine learning model based on knowledge to make the best use of both geoscientific information and the advantages of machine learning. We focus on the region-scale prospectivity mapping for the komatiite-hosted nickel in Yilgarn craton, Western Australia. We extract second and thirdorder features from geophysical data to enable machine learning models to capture various patterns of mineral deposits. In terms of geology, faults, interpreted geology, and isotopic mapping data are converted into numerical features that could be related to the komatiite-hosted nickel deposits. Based on domain knowledge, we design a deep learning model that systemically combines geophysical and geological features. First, our model generates a feature map and initial prospectivity map using geological data and geophysical worms which could reveal the crustal structures. Next, the model produces a final prospectivity map that delineates potential komatiite-hosted nickel deposits using whole data including geophysics. The model is trained with the locations of the known nickel deposits. We divide the Yilgarn craton area into a train and test region to validate our model. We adopt the AUC score and prospectivity score percentile of known deposits to evaluate our model in various aspects. Our model achieved a high AUC and percentile score and it can be efficiently used for early-stage nickel exploration. The suggested workflow could be applied to the exploration of the other mineral types with a slight modification reflecting the characteristics of the mineralizations.

Open-Access Online Publication: March 01, 2023