A causal-aware artificial intelligence framework for mineral prospectivity mapping

Mineral prospectivity mapping (MPM) plays a vital role in locating potential mineral resources and informing exploration planning. Although data-driven machine learning (ML) techniques have demonstrated effectiveness in handling complex geological patterns, they often depend purely on statistical dependencies, ignoring the underlying causal mechanisms responsible for mineralization. This shortcoming can introduce spurious links and reduce the robustness of the model. In this study, we introduce a new causal-aware artificial intelligence (CausalAI) framework tailored for MPM, which incorporates both traditional metallogenic features and their causal interdependencies as input data. Firstly, we propose the Spatial Granger Causality (SGC) model to evaluate the causal relations between geochemical elements, drawing inspiration from the principles of geostatistics and the Granger causality. Secondly, we design a causal-aware feature extractor (Causactor) to mitigate spurious correlations between predictor variables and predicted responses, facilitating causal feature extraction. Finally, causal-guided ML methods are developed to predict mineralization and classify mineral resource types. A case study in the northern Cu-polymetallic metallogenic belt of the Sanjiang region, southwestern China, demonstrates the high performance of our framework for MPM. Overall, the development of the proposed CausalAI framework effectively integrates causal relations throughout the modeling process, promoting the synergy between causal inference and Earth sciences.