MPM-KAN: An Interpretable Data-Driven Framework for Mineral Prospectivity Mapping Based on Kolmogorov–Arnold Networks

Transport in Porous Media · Springer Nature · 2026-10-07 · DOI: 10.1007/s11004-026-10351-4

Аннотация

A major limitation of data-driven deep learning models for mineral prospectivity mapping (MPM) is their lack of intrinsic interpretability, which often obscures the underlying geological regularities and hinders acceptance among geologists. To address this problem, a highly interpretable data-driven network (MPM-KAN) for MPM is proposed by integrating multi-source geoscience data, including lithology, faults, geophysical, geochemical, and digital elevation model (DEM) data. In MPM-KAN, multi-stream feature extraction encoders are utilized to extract features from different kinds of geoscience data, thereby decoupling the contributions of different geoscience features to mineralization probability. Interpretable Kolmogorov–Arnold Networks (KANs) are used to extract the complex nonlinear relationship between key geological features and mineralization. The Norseman–Wiluna greenstone belt in Western Australia is known for its abundant gold resources and is used to evaluate MPM-KAN. Comparative experiments with the existing models demonstrate the necessity and superiority of MPM-KAN’s structure. Furthermore, MPM-KAN achieves the highest prediction efficiency among all compared methods. Based on the results of MPM-KAN, highly promising prospective targets are delineated, and an interpretable formula between key geological features and mineralization is derived. MPM-KAN can solve the limitations of traditional data-driven models, providing a new idea for interpretable MPM.

Читать у издателя →

Похожие статьи