Unsupervised Deep Learning-Based Mineral Prospectivity Mapping for Rare Earth Elements in the Rairakhol Alkaline Complex Using Integrated Geochemical and Geophysical Data

The Rairakhol Alkaline Complex in Eastern India represents a favourable geological setting for Rare Earth Element (REE) mineralisation. This study presents an advanced mineral prospectivity mapping approach by the integration of stream sediment geochemical data with gravity and magnetic geophysical datasets using a deep learning framework. Conventional mineral targeting approaches often face difficulties due to the complex, high dimensional nature of geoscientific data, particularly when clear labels or known deposits are scattered and nonuniform. To address these challenges, we implemented a deep autoencoder network for unsupervised anomaly detection, enabling the identification of prospective zones without depending on predefined training labels. The geochemical data, consisting concentrations of 17 REEs and associated trace elements, were processed using centred log-ratio transformation to address compositional effects. Synchronously, geophysical datasets which include Bouguer gravity and total magnetic intensity (TMI) data, were optimised through the calculation of derivative layers such as first vertical derivatives and tilt angle maps. All datasets were spatially coordinated, scaled, normalised, and integrated into a consolidated feature matrix for model input. The deep autoencoder was trained to reconstruct the multidimensional geoscientific data, with reconstruction error serving as an indicator of anomalous zones. Areas showing high reconstruction errors are considered as geologically distinct regions potentially associated with REE enrichment. The resulting Prospectivity map, generated from the normalised reconstruction error values, reveals several spatially coherent high anomalous zones that correlate with known alkaline intrusives in the region. This approach demonstrates the possibility of deep learning-based unsupervised techniques for mineral exploration in data-limited regions. By integrating heterogeneous datasets within a nonlinear learning framework, the approach significantly enhances the objectivity, resolution, and interpretability of REE prospectivity mapping. The results provide a scientific foundation for focusing on future exploration strategies in the Rairakhol Alkaline Complex and equivalent terrains globally.