An adaptive KAN with Toptheta attention and dynamic thresholding for interpretable mineral prospectivity prediction

Mineral Prospectivity Mapping (MPM) can provide key decision-making information for mineral exploration task. Deep learning provides a powerful tool for MPM. However, when dealing with high-dimensional and large-scale geological spatial data, many deep learning models are unable to deeply explore the interaction information between geological characteristics and mineralization, thereby reducing the reliability of the model results. Thus, this study proposes an adaptive Kolmogorov-Arnold network (A-KAN) model, by introducing the Toptheta attention mechanism and dynamic threshold training strategy into the Kolmogorov-Arnold network (KAN) model. Through learnable multivariate continuous functions, the A-KAN model can accurately capture the nonlinear relationships between key controlling mineralization characteristic in high-dimensional geospatial data. Specifically, this study took the MPM of the tungsten polymetallic deposits in Nanling Metallogenic Belt, China, as an example, and constructed a 41-dimensional geospatial dataset as the input of the model. The results show that compared with KAN, CNN, and SVM models, the A-KAN model has significant advantages in identifying known deposits, demonstrating stronger classification and prediction performance. More importantly, this study conducted interpretability and visualization analysis of model training and results from four aspects. This improves the rationality and reliability of the model results, providing more scientific, and reliable technical support for MPM.