Abstract:

With the rapid advancement of AI and big data, machine learning-based mineral prospectivity mapping has become a research hotspot. However, models with deeply nested, nonlinear structures and abstract representations often exhibit opaque “black-box” characteristics. This lack of interpretability between predictions and metallogenic processes limits model generalization and reliability. To address this, we propose a knowledge-data driven interpretable prediction method. First, the Best-Worst Method(BWM) is used to derive geological feature weights for constructing an ensemble model, enhancing its performance. A multi-scale, multi-dimensional interpretability framework spanning from global to local levels and from feature-level to sample-level interpretations is then applied to deconstruct results and evaluate feature importance. Expert-guided corrections, informed by field validation, further refine the predictions, thereby forming a closed loop of knowledge embedding and discovery. This process improves workflow transparency and result reliability. Applied to the Keeryin area in Sichuan, the method identified eight high-potential Class A targets, occupying only 6.58% of the study area yet containing 84% of the known deposits. Key predictive features included remote sensing spectral response of albite/albite spectral signature, total alkali (Na₂O+K₂O) content, ring structures, Li/La ratio, and two-mica granite—as validated by their strong spatial correlation with known pegmatite-type lithium deposits in the field.

Key words: mineral resource quantitative prediction, Keeryin Mining Area, knowledge embedding, integrated learning, interpretability, field validation