LAUSR

Abstract The inference of deep alterations plays an important role in the analysis of deep geological processes and the targeting of buried mineral resources. However, the construction of three-dimensional (3D) models of alteration zones based on geochemical data is challenging due to the diverse geological processes and scarcity of information related to deep alterations. In this paper, we propose an interpretable machine-learning method to infer 3D geological models of alteration zones based on geochemical data from the Dayingezhuang area, China. To expose the information hidden in geochemical data that is indirectly and weakly related to alteration zones, a generalized additive model (GAM) is used. It learns a nonlinear, intelligible model associating geochemical anomalies (extracted by PCA) with alteration features. The model learned via GAM and is then combined with a Bayesian framework to infer the thickness of alterations in a no-drilling area, thus enabling 3D modelling of deep alteration zones. With such an intelligible model, geologists are capable of inspecting the reliability of the model and better understanding the association between the distribution of geochemical elements and deep alterations. The inference results show that the relationship between geochemical anomalies and alteration depth is significant, while the relationship with alteration thickness is complex. The results of this study provide the underlying insights needed to guide the analysis of deep geological mineralization processes.