Improved mineral prospectivity mapping using graph neural networks

• Mineral prospectivity mapping (MPM) that utilize GNN shows better performance when trained using lithology information. • GNN model can implicitly learn boundaries that highly prospective without a priori information. • GNN also shows better performance when the geochemistry data was added as input. • GNN can maximize the use of geological relationship and can be valuable for exploration information system (EIS) Prospectivity analysis is an important part of the exploration information system (EIS) and mineral prospectivity mapping (MPM) is a widely used technique for prospectivity analysis. Despite MPM studies employing various machine learning (ML) algorithms for various purposes, MPM is still considered insufficient to capture the complexity of many important ore-forming geological processes. One potential issue concerns the difficulty with which conventional application of ML algorithms can learn from relational information. For instance, the conventional application of ML algorithms to tabular data for mineral exploration does not effectively consider spatial relations across or between geological terranes, because point data are treated as independent entities that do not influence their neighbor’s characteristics. Here, we demonstrate how re-designing exploration data into a graph format that focuses on spatial relationships can increase the effectiveness of mineral prospectivity mapping (MPM). We demonstrate that the use of graph deep learning can be beneficial when utilizing categorical and numerical exploration data. This approach was applied to three different commodities (copper, iron, and tin) in the southern United Kingdom in order to compare the effectiveness of the graph neural network (GNN) method with conventional ML techniques. We show that graph-based ML that utilizes immediate neighbor relationships in the training process significantly improves performance in three key metrics when compared to tabular data, particularly so when trained with a dataset where there are more barren (non-occurrence) data points than mineralized (occurrence) ones. To produce the most useful prospectivity maps, we recommend training a GNN algorithm by using an imbalanced training dataset comprising more barren data than mineralized data. We expect that further testing of the GNN method will lead to optimization of the supervised ML techniques used in MPM and EIS to prospect for key metal commodities in regions with incomplete geological data.